diff -Nru gmp-ecm-7.0.3+ds/acinclude.m4 gmp-ecm-7.0.4+ds/acinclude.m4
--- gmp-ecm-7.0.3+ds/acinclude.m4	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/acinclude.m4	2016-08-23 12:25:18.000000000 +0000
@@ -605,4 +605,31 @@
 
 ])
 
-
+dnl Checks whether the stack can be marked nonexecutable by passing an option
+dnl to the C-compiler when acting on .s files. Appends that option to ASMFLAGS.
+dnl This macro is adapted from one found in GMP 6.1.1.
+dnl FIXME: This test looks broken. It tests that a file with .note.GNU-stack...
+dnl can be compiled/assembled with -Wa,--noexecstack.  It does not determine
+dnl if that command-line option has any effect on general asm code.
+AC_DEFUN([CL_AS_NOEXECSTACK],[
+dnl AC_REQUIRE([AC_PROG_CC]) GMP uses something else
+AC_CACHE_CHECK([whether assembler supports --noexecstack option],
+cl_cv_as_noexecstack, [dnl
+  cat > conftest.c <<EOF
+void foo() {}
+EOF
+  if AC_TRY_COMMAND([${CC} $CFLAGS $CPPFLAGS
+                     -S -o conftest.s conftest.c >/dev/null]) \
+     && grep .note.GNU-stack conftest.s >/dev/null \
+     && AC_TRY_COMMAND([${CC} $CFLAGS $CPPFLAGS -Wa,--noexecstack
+                       -c -o conftest.o conftest.s >/dev/null])
+  then
+    cl_cv_as_noexecstack=yes
+  else
+    cl_cv_as_noexecstack=no
+  fi
+  rm -f conftest*])
+  if test "$cl_cv_as_noexecstack" = yes; then
+    LIBECM_LDFLAGS="$LIBECM_LDFLAGS -Wl,-znoexecstack"
+  fi
+])
diff -Nru gmp-ecm-7.0.3+ds/build.vc12/gen_ecm_h.bat gmp-ecm-7.0.4+ds/build.vc12/gen_ecm_h.bat
--- gmp-ecm-7.0.3+ds/build.vc12/gen_ecm_h.bat	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/build.vc12/gen_ecm_h.bat	2016-10-11 09:22:42.000000000 +0000
@@ -5,7 +5,7 @@
 for /f "tokens=1,2*" %%a in (..\ecm.h.in) do (
   if "%%a" EQU "#undef" (
     if "%%b" EQU "ECM_VERSION" (
-      echo #define ECM_VERSION "7.0.3">>tmp.h
+      echo #define ECM_VERSION "7.0.4">>tmp.h
     )
   ) else echo %%a %%b %%c>>tmp.h
 )
diff -Nru gmp-ecm-7.0.3+ds/build.vc12/tests.py gmp-ecm-7.0.4+ds/build.vc12/tests.py
--- gmp-ecm-7.0.3+ds/build.vc12/tests.py	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/build.vc12/tests.py	2016-10-11 09:14:19.000000000 +0000
@@ -7,7 +7,12 @@
 from re import match
 from subprocess import Popen, PIPE, STDOUT
 from tempfile import *
-from time import clock
+
+try:
+  from time import perf_counter as timer
+except ImportError:
+  from time import clock as timer
+
 
 x64 = True
 debug = False
@@ -16,8 +21,8 @@
 run_gpu_tests = True
 
 class Timer() :
-  def __enter__(self): self.start = clock()
-  def __exit__(self, *args): print(' time {:.3f} milliseconds'.format(1000 * (clock() - self.start)))
+  def __enter__(self): self.start = timer()
+  def __exit__(self, *args): print(' time {:.3f} milliseconds'.format(1000 * (timer() - self.start)))
 
 cpath = os.path.dirname(__file__)
 config = 'x64' if x64 else 'Win32'
@@ -46,8 +51,8 @@
         elif start:
           sp = line.split('="')
           if len(sp) == 2:
-            if sp[1].startswith('$1'):
-              sub[sp[0]] = sp[1][2:-1]
+            if sp[1].startswith('${1:-./ecm}'):
+              sub[sp[0]] = sp[1][12:-1]
             else:
               sub[sp[0]] = sp[1][:-1]
           continue
@@ -122,9 +127,9 @@
     output_complex_tests(ctests)
   if not err_cnt:
     if ctests:
-      print('  all other tests passed')
+      print('all other tests passed')
     else:
-      print('  all tests passed')
+      print('all tests passed')
 
 with Timer():
   if os.path.exists('test.pm1.save'):
diff -Nru gmp-ecm-7.0.3+ds/champions.h gmp-ecm-7.0.4+ds/champions.h
--- gmp-ecm-7.0.3+ds/champions.h	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/champions.h	2016-10-11 09:26:27.000000000 +0000
@@ -11,4 +11,4 @@
  "http://www.loria.fr/~zimmerma/records/Pminus1.html",
  "http://www.loria.fr/~zimmerma/records/Pplus1.html"};
 /* minimal number of digits to enter the champions table for ECM, P-1, P+1 */
-static unsigned int champion_digits[3] = { 72, 54, 48 };
+static unsigned int champion_digits[3] = { 73, 54, 48 };
diff -Nru gmp-ecm-7.0.3+ds/ChangeLog gmp-ecm-7.0.4+ds/ChangeLog
--- gmp-ecm-7.0.3+ds/ChangeLog	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/ChangeLog	2016-10-11 09:25:46.000000000 +0000
@@ -1,4 +1,266 @@
 ------------------------------------------------------------------------
+r2990 | brian_gladman | 2016-09-29 13:13:04 +0200 (Thu, 29 Sep 2016) | 1 line
+Changed paths:
+   M /trunk/build.vc14/ecm_gpu/ecm_gpu.vcxproj
+   M /trunk/build.vc14/libecm_gpu/libecm_gpu.vcxproj
+   M /trunk/build.vc14/readme.txt
+
+Update default CUDA compute capabilities and the readme.txt file 
+------------------------------------------------------------------------
+r2989 | brian_gladman | 2016-09-29 12:52:25 +0200 (Thu, 29 Sep 2016) | 1 line
+Changed paths:
+   M /trunk/build.vc12/tests.py
+   A /trunk/build.vc14
+   A /trunk/build.vc14/Makefile.am
+   A /trunk/build.vc14/assembler
+   A /trunk/build.vc14/assembler/Makefile.am
+   A /trunk/build.vc14/assembler/a_win32a_redc.asm
+   A /trunk/build.vc14/assembler/a_win32p_mulredc.asm
+   A /trunk/build.vc14/assembler/a_win32p_redc.asm
+   A /trunk/build.vc14/assembler/a_x64_mulredc.asm
+   A /trunk/build.vc14/assembler/a_x64_redc.asm
+   A /trunk/build.vc14/assembler/mulredc.asm
+   A /trunk/build.vc14/assembler/mulredc.h
+   A /trunk/build.vc14/assembler/redc.asm
+   A /trunk/build.vc14/assembler/test_mulredc.c
+   A /trunk/build.vc14/bench_mulredc
+   A /trunk/build.vc14/bench_mulredc/Makefile.am
+   A /trunk/build.vc14/bench_mulredc/bench_mulredc.vcxproj
+   A /trunk/build.vc14/bench_mulredc/bench_mulredc.vcxproj.filters
+   A /trunk/build.vc14/config.h
+   A /trunk/build.vc14/ecm
+   A /trunk/build.vc14/ecm/Makefile.am
+   A /trunk/build.vc14/ecm/ecm.vcxproj
+   A /trunk/build.vc14/ecm/ecm.vcxproj.filters
+   A /trunk/build.vc14/ecm.sln
+   A /trunk/build.vc14/ecm_gpu
+   A /trunk/build.vc14/ecm_gpu/Makefile.am
+   A /trunk/build.vc14/ecm_gpu/ecm.vcxproj.filters
+   A /trunk/build.vc14/ecm_gpu/ecm_gpu.vcxproj
+   A /trunk/build.vc14/ecm_gpu/ecm_gpu.vcxproj.filters
+   A /trunk/build.vc14/ecm_gpu.sln
+   A /trunk/build.vc14/file_copy.bat
+   A /trunk/build.vc14/gen_ecm_h.bat
+   A /trunk/build.vc14/getopt.c
+   A /trunk/build.vc14/getopt.h
+   A /trunk/build.vc14/getrusage.c
+   A /trunk/build.vc14/getrusage.h
+   A /trunk/build.vc14/gettimeofday.c
+   A /trunk/build.vc14/gettimeofday.h
+   A /trunk/build.vc14/libecm
+   A /trunk/build.vc14/libecm/Makefile.am
+   A /trunk/build.vc14/libecm/libecm.vcxproj
+   A /trunk/build.vc14/libecm/libecm.vcxproj.filters
+   A /trunk/build.vc14/libecm_gpu
+   A /trunk/build.vc14/libecm_gpu/Makefile.am
+   A /trunk/build.vc14/libecm_gpu/libecm_gpu.vcxproj
+   A /trunk/build.vc14/libecm_gpu/libecm_gpu.vcxproj.filters
+   A /trunk/build.vc14/mp_lib.props
+   A /trunk/build.vc14/multiecm
+   A /trunk/build.vc14/multiecm/Makefile.am
+   A /trunk/build.vc14/multiecm/multiecm.vcxproj
+   A /trunk/build.vc14/multiecm/multiecm.vcxproj.filters
+   A /trunk/build.vc14/out_copy_rename.bat
+   A /trunk/build.vc14/prebuild.bat
+   A /trunk/build.vc14/python.bat
+   A /trunk/build.vc14/readme.txt
+   A /trunk/build.vc14/tests.py
+   A /trunk/build.vc14/tune
+   A /trunk/build.vc14/tune/Makefile.am
+   A /trunk/build.vc14/tune/tune.vcxproj
+   A /trunk/build.vc14/tune/tune.vcxproj.filters
+   A /trunk/build.vc14/vacopy.c
+   A /trunk/build.vc14/vsyasm.props
+   A /trunk/build.vc14/vsyasm.targets
+   A /trunk/build.vc14/vsyasm.xml
+
+Add build files for Visual Studio 2015
+------------------------------------------------------------------------
+r2988 | zimmerma | 2016-09-06 09:41:17 +0200 (Tue, 06 Sep 2016) | 2 lines
+Changed paths:
+   M /trunk/mpmod.c
+
+[mpmod.c] define mpres_equal when HAVE_ADDLAWS=1
+
+------------------------------------------------------------------------
+r2987 | zimmerma | 2016-08-30 15:46:13 +0200 (Tue, 30 Aug 2016) | 2 lines
+Changed paths:
+   M /trunk/ecm-gmp.h
+
+[ecm-gmp.h] fixed typo
+
+------------------------------------------------------------------------
+r2986 | zimmerma | 2016-08-29 08:28:41 +0200 (Mon, 29 Aug 2016) | 2 lines
+Changed paths:
+   M /trunk/test.pm1
+   M /trunk/test.pp1
+
+added more test cases
+
+------------------------------------------------------------------------
+r2985 | zimmerma | 2016-08-26 15:17:56 +0200 (Fri, 26 Aug 2016) | 2 lines
+Changed paths:
+   M /trunk/NEWS
+
+[NEWS] updated
+
+------------------------------------------------------------------------
+r2984 | zimmerma | 2016-08-26 15:03:39 +0200 (Fri, 26 Aug 2016) | 3 lines
+Changed paths:
+   M /trunk/mpmod.c
+   M /trunk/pm1fs2.c
+   M /trunk/test.pm1
+   M /trunk/test.pp1
+
+fixed Segmentation fault with -pm1 and -pp1 (presumably only happens
+with small parameters)
+
+------------------------------------------------------------------------
+r2983 | zimmerma | 2016-08-26 13:15:40 +0200 (Fri, 26 Aug 2016) | 2 lines
+Changed paths:
+   M /trunk/test.pm1
+
+[test.pm1] added more test cases
+
+------------------------------------------------------------------------
+r2982 | zimmerma | 2016-08-26 09:18:26 +0200 (Fri, 26 Aug 2016) | 2 lines
+Changed paths:
+   M /trunk/configure.ac
+
+patch from François Bissey to address issue on Mac OS X
+
+------------------------------------------------------------------------
+r2981 | zimmerma | 2016-08-25 15:44:58 +0200 (Thu, 25 Aug 2016) | 2 lines
+Changed paths:
+   M /trunk/x86_64/mulredc.m4
+   M /trunk/x86_64/mulredc1.asm
+   M /trunk/x86_64/mulredc1.m4
+
+removed (non-portable) WANT_ASSERT stuff in assembly files
+
+------------------------------------------------------------------------
+r2980 | zimmerma | 2016-08-25 13:58:10 +0200 (Thu, 25 Aug 2016) | 2 lines
+Changed paths:
+   M /trunk/NEWS
+
+updated NEWS
+
+------------------------------------------------------------------------
+r2979 | zimmerma | 2016-08-25 13:56:54 +0200 (Thu, 25 Aug 2016) | 3 lines
+Changed paths:
+   M /trunk/mpmod.c
+
+[mpmod.c] fixed mpres_pow() for negative EXP
+(https://gforge.inria.fr/tracker/?func=detail&atid=623&aid=20712&group_id=135)
+
+------------------------------------------------------------------------
+r2978 | zimmerma | 2016-08-23 14:30:23 +0200 (Tue, 23 Aug 2016) | 2 lines
+Changed paths:
+   M /trunk/Makefile.am
+   M /trunk/acinclude.m4
+   M /trunk/configure.ac
+
+added configure test to check whether ld accepts -znoexecstack
+
+------------------------------------------------------------------------
+r2977 | zimmerma | 2016-08-23 11:22:28 +0200 (Tue, 23 Aug 2016) | 2 lines
+Changed paths:
+   M /trunk/README.gpu
+
+[README.gpu] changed Cyril's email (preferred one)
+
+------------------------------------------------------------------------
+r2976 | zimmerma | 2016-08-23 08:51:49 +0200 (Tue, 23 Aug 2016) | 7 lines
+Changed paths:
+   M /trunk/configure.ac
+
+[configure.ac] applied patch from Erik Bray to set the WINDOWS64 flag under
+Cygwin.
+
+cf https://gforge.inria.fr/tracker/?func=detail&atid=623&aid=20697&group_id=135
+and https://trac.sagemath.org/ticket/21223
+	       
+
+------------------------------------------------------------------------
+r2975 | zimmerma | 2016-08-22 13:45:25 +0200 (Mon, 22 Aug 2016) | 2 lines
+Changed paths:
+   M /trunk/Makefile.am
+
+fixed make longcheck with out-of-source build
+
+------------------------------------------------------------------------
+r2974 | zimmerma | 2016-08-22 13:32:04 +0200 (Mon, 22 Aug 2016) | 2 lines
+Changed paths:
+   M /trunk/x86_64/mulredc.m4
+
+[x86_64/mulredc.m4] Opteron -> x86_64
+
+------------------------------------------------------------------------
+r2971 | zimmerma | 2016-08-22 13:11:21 +0200 (Mon, 22 Aug 2016) | 2 lines
+Changed paths:
+   M /trunk/ks-multiply.c
+
+[ks-multiply.c] fix when FFT_WRAP is not defined
+
+------------------------------------------------------------------------
+r2970 | zimmerma | 2016-08-22 09:49:08 +0200 (Mon, 22 Aug 2016) | 2 lines
+Changed paths:
+   M /trunk/README.gpu
+
+[README.gpu] updated Cyril's email
+
+------------------------------------------------------------------------
+r2969 | zimmerma | 2016-08-22 09:23:55 +0200 (Mon, 22 Aug 2016) | 3 lines
+Changed paths:
+   M /trunk/README.gpu
+
+[README.gpu] added hint about Cuda error "too many resources requested for launch"
+(see https://lists.gforge.inria.fr/pipermail/ecm-discuss/2016-July/004368.html)
+
+------------------------------------------------------------------------
+r2968 | dcleaver | 2016-07-31 20:12:55 +0200 (Sun, 31 Jul 2016) | 1 line
+Changed paths:
+   M /trunk/addlaws.c
+   M /trunk/listz_handle.c
+   M /trunk/main.c
+   M /trunk/test.ecm
+   M /trunk/tune.c
+
+Fix zu printf problems on Windows by replacing zu with PRIu64.  This also fixed a crash that could occur on Windows when using both -v and -bloads.  Include test to check for that.
+------------------------------------------------------------------------
+r2967 | zimmerma | 2016-07-25 16:12:17 +0200 (Mon, 25 Jul 2016) | 3 lines
+Changed paths:
+   M /trunk/test.gpuecm
+
+increased initial sigma so that test works with ECM_GPU_CURVES_BY_BLOCK=16
+(see https://lists.gforge.inria.fr/pipermail/ecm-discuss/2016-July/004368.html)
+
+------------------------------------------------------------------------
+r2966 | dcleaver | 2016-07-24 16:06:00 +0200 (Sun, 24 Jul 2016) | 1 line
+Changed paths:
+   M /trunk/cudawrapper.c
+   M /trunk/ecm.c
+
+Correct two verbose print problems on Windows.  1 remove extra 0 in expected curves output 2 fix zu format modifier in computing batch product output.
+------------------------------------------------------------------------
+r2965 | zimmerma | 2016-07-04 10:15:30 +0200 (Mon, 04 Jul 2016) | 2 lines
+Changed paths:
+   M /trunk/build.vc12/gen_ecm_h.bat
+   M /trunk/configure.ac
+
+increased version number
+
+------------------------------------------------------------------------
+r2964 | zimmerma | 2016-07-04 10:13:28 +0200 (Mon, 04 Jul 2016) | 2 lines
+Changed paths:
+   M /trunk/ChangeLog
+   M /trunk/README.dev
+   M /trunk/build.vc12/gen_ecm_h.bat
+   M /trunk/configure.ac
+
+this is release 7.0.3
+
+------------------------------------------------------------------------
 r2963 | zimmerma | 2016-06-29 15:55:37 +0200 (Wed, 29 Jun 2016) | 2 lines
 Changed paths:
    M /trunk/NEWS
diff -Nru gmp-ecm-7.0.3+ds/configure.ac gmp-ecm-7.0.4+ds/configure.ac
--- gmp-ecm-7.0.3+ds/configure.ac	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/configure.ac	2016-10-11 09:23:36.000000000 +0000
@@ -1,4 +1,4 @@
-m4_define([ECM_VERSION_AC], [7.0.3])
+m4_define([ECM_VERSION_AC], [7.0.4])
 
 AC_PREREQ([2.57])
 AC_INIT([ecm], ECM_VERSION_AC, [ecm-discuss@lists.gforge.inria.fr])
@@ -272,7 +272,7 @@
   GMP_ASM_TYPE 
 
   case $host in
-    *-*-mingw32) GMP_DEFINE([WINDOWS64_ABI], 1)
+    *-*-mingw32|*cygwin*) GMP_DEFINE([WINDOWS64_ABI], 1)
                  AC_DEFINE([WINDOWS64_ABI], 1,[Define to 1 if x86_64 mulredc*() functions should be called with Windows ABI]);;
     *) ;;
   esac
@@ -481,13 +481,18 @@
   GMPLDFLAGS="-L$with_gmp_lib"
 fi
 GMPLIB="-lgmp"
-if test "x$enable_shared" != xyes; then
-  if test -r "$with_gmp_lib/libgmp.a"; then
-    GMPLIB="$with_gmp_lib/libgmp.a"
-    dnl Don't need -L flag since we give full path to libgmp.a
-    GMPLDFLAGS=""
-  fi
-fi
+dnl On OS X we should not give libgmp.a
+dnl (see https://trac.sagemath.org/ticket/20385#comment:103)
+case $host_os in
+  darwin*) ;;
+  *) if test "x$enable_shared" != xyes; then
+       if test -r "$with_gmp_lib/libgmp.a"; then
+         GMPLIB="$with_gmp_lib/libgmp.a"
+         dnl Don't need -L flag since we give full path to libgmp.a
+         GMPLDFLAGS=""
+       fi
+     fi
+esac
 AC_SUBST([GMPLIB])
 LDFLAGS="$LDFLAGS $GMPLDFLAGS"
 
@@ -585,6 +590,7 @@
   fi
   ;;
 esac
+CL_AS_NOEXECSTACK
 AC_SUBST([LIBECM_LDFLAGS])
 
 
diff -Nru gmp-ecm-7.0.3+ds/cudawrapper.c gmp-ecm-7.0.4+ds/cudawrapper.c
--- gmp-ecm-7.0.3+ds/cudawrapper.c	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/cudawrapper.c	2016-07-25 11:06:54.000000000 +0000
@@ -333,8 +333,8 @@
       st = cputime ();
       /* construct the batch exponent */
       compute_s (batch_s, B1, NULL);
-      outputf (OUTPUT_VERBOSE, "Computing batch product (of %zu bits) of "
-                               "primes up to B1=%1.0f took %ldms\n", 
+      outputf (OUTPUT_VERBOSE, "Computing batch product (of %" PRIu64
+                               " bits) of primes up to B1=%1.0f took %ldms\n",
                                mpz_sizeinbase (batch_s, 2), B1, cputime () - st);
     }
 
diff -Nru gmp-ecm-7.0.3+ds/debian/changelog gmp-ecm-7.0.4+ds/debian/changelog
--- gmp-ecm-7.0.3+ds/debian/changelog	2016-07-04 14:13:01.000000000 +0000
+++ gmp-ecm-7.0.4+ds/debian/changelog	2016-10-14 00:25:48.000000000 +0000
@@ -1,3 +1,9 @@
+gmp-ecm (7.0.4+ds-1) unstable; urgency=medium
+
+  * New upstream patch release.
+
+ -- Jerome Benoit <calculus@rezozer.net>  Fri, 14 Oct 2016 00:25:48 +0000
+
 gmp-ecm (7.0.3+ds-1) unstable; urgency=medium
 
   * New upstream patch release.
diff -Nru gmp-ecm-7.0.3+ds/ecm.c gmp-ecm-7.0.4+ds/ecm.c
--- gmp-ecm-7.0.3+ds/ecm.c	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/ecm.c	2016-07-25 11:06:54.000000000 +0000
@@ -1,7 +1,7 @@
 /* Elliptic Curve Method: toplevel and stage 1 routines.
 
 Copyright 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010, 2011,
-2012 Paul Zimmermann, Alexander Kruppa, Cyril Bouvier, David Cleaver.
+2012, 2016 Paul Zimmermann, Alexander Kruppa, Cyril Bouvier, David Cleaver.
 
 This file is part of the ECM Library.
 
@@ -849,7 +849,7 @@
 {
   double prob;
   int i, j;
-  char sep, outs[128];
+  char sep, outs[128], flt[16];
   double smoothness_correction;
 
   if (param == ECM_PARAM_SUYAMA || param == ECM_PARAM_BATCH_2)
@@ -861,8 +861,8 @@
   else /* This case should never happen */
       smoothness_correction = 0.0; 
 
-  for (i = DIGITS_START, j = 0; i <= DIGITS_END; i += DIGITS_INCR, j += 3)
-    sprintf (outs + j, "%2u%c", i, (i < DIGITS_END) ? '\t' : '\n');
+  for (i = DIGITS_START, j = 0; i <= DIGITS_END; i += DIGITS_INCR)
+    j += sprintf (outs + j, "%u%c", i, (i < DIGITS_END) ? '\t' : '\n');
   outs[j] = '\0';
   outputf (OUTPUT_VERBOSE, "Expected number of curves to find a factor "
            "of n digits:\n%s", outs);
@@ -876,7 +876,14 @@
       if (prob > 1. / 10000000)
         outputf (OUTPUT_VERBOSE, "%.0f%c", floor (1. / prob + .5), sep);
       else if (prob > 0.)
-        outputf (OUTPUT_VERBOSE, "%.2g%c", floor (1. / prob + .5), sep);
+        {
+          /* on Windows: 2.6e+009   or   2.6e+025  (8 characters in string) */
+          /* on Linux  : 2.6e+09    or   2.6e+25   (7 characters in string) */
+          /* This will normalize the output so that the Windows ouptut will match the Linux output */
+          if (sprintf (flt, "%.2g", floor (1. / prob + .5)) == 8)
+            memmove (&flt[5], &flt[6], strlen(flt) - 5);
+          outputf (OUTPUT_VERBOSE, "%s%c", flt, sep);
+        }
       else
         outputf (OUTPUT_VERBOSE, "Inf%c", sep);
     }
@@ -900,8 +907,8 @@
   else /* This case should never happen */
       smoothness_correction = 0.0; 
   
-  for (i = DIGITS_START, j = 0; i <= DIGITS_END; i += DIGITS_INCR, j += 3)
-    sprintf (outs + j, "%2u%c", i, (i < DIGITS_END) ? '\t' : '\n');
+  for (i = DIGITS_START, j = 0; i <= DIGITS_END; i += DIGITS_INCR)
+    j += sprintf (outs + j, "%u%c", i, (i < DIGITS_END) ? '\t' : '\n');
   outs[j] = '\0';
   outputf (OUTPUT_VERBOSE, "Expected time to find a factor of n digits:\n%s",
            outs);
@@ -1219,9 +1226,9 @@
       st = cputime ();
       /* construct the batch exponent */
       compute_s (batch_s, B1, NULL);
-      outputf (OUTPUT_VERBOSE, "Computing batch product (of %zu bits) of "
-                               "primes up to B1=%1.0f took %ldms\n", 
-               mpz_sizeinbase (batch_s, 2), B1, cputime () - st);
+      outputf (OUTPUT_VERBOSE, "Computing batch product (of %" PRIu64
+                               " bits) of primes up to B1=%1.0f took %ldms\n",
+                               mpz_sizeinbase (batch_s, 2), B1, cputime () - st);
     }
 
   st = cputime ();
diff -Nru gmp-ecm-7.0.3+ds/ecm-gmp.h gmp-ecm-7.0.4+ds/ecm-gmp.h
--- gmp-ecm-7.0.3+ds/ecm-gmp.h	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/ecm-gmp.h	2016-08-30 13:45:14.000000000 +0000
@@ -143,7 +143,7 @@
 #if __MPIR_RELEASE == 20600
 #error MPIR 2.6 does not support GMP-ECM, please use an alternative version
 #endif
-/* WARNING - the following two defintions map the internal interface
+/* WARNING - the following two definitions map the internal interface
    of the new FFT in MPIR 2.6 (and later) to the GMP FFT interface -
    they work in this context but the parameters for mpn_fft_next_size
    and fft_adjust_limbs have different semantics, which means that
@@ -152,7 +152,7 @@
 #  define mpn_fft_best_k(n, k)             (0) 
 #  define mpn_fft_next_size(n, k)          fft_adjust_limbs(n)
 
-#else
+#else /* we are using the real GMP */
 
 #define mpn_mul_fft __gmpn_mul_fft
 mp_limb_t __gmpn_mul_fft (mp_ptr, mp_size_t, mp_srcptr, mp_size_t, mp_srcptr, 
diff -Nru gmp-ecm-7.0.3+ds/ecm.h gmp-ecm-7.0.4+ds/ecm.h
--- gmp-ecm-7.0.3+ds/ecm.h	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/ecm.h	2016-10-11 09:28:12.000000000 +0000
@@ -27,7 +27,7 @@
 #include <stdio.h> /* for FILE */
 #include <gmp.h>
 
-#define ECM_VERSION "7.0.3"
+#define ECM_VERSION "7.0.4"
 
 #ifdef __cplusplus
 extern "C" {
diff -Nru gmp-ecm-7.0.3+ds/ks-multiply.c gmp-ecm-7.0.4+ds/ks-multiply.c
--- gmp-ecm-7.0.3+ds/ks-multiply.c	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/ks-multiply.c	2016-08-22 11:10:25.000000000 +0000
@@ -578,6 +578,10 @@
             listz_t B, unsigned int l,
 	    mpz_t n)
 {
+#ifndef FFT_WRAP
+  ASSERT_ALWAYS(0); /* ks_wrapmul should not be called in that case */
+  return 0;
+#else
   unsigned long i, m, t;
   mp_size_t s, size_t0, size_t1, size_tmp;
   mp_ptr t0_ptr, t1_ptr, t2_ptr, r_ptr, tp;
@@ -621,15 +625,11 @@
     if (SIZ(B[i]))
       MPN_COPY (t1_ptr + i * s, PTR(B[i]), SIZ(B[i]));
 
-#ifdef FFT_WRAP
   i = mpn_mulmod_bnm1_next_size (m0 * s);
   /* the following loop ensures we don't cut in the middle of a
      coefficient */
   while (i % s)
     i = mpn_mulmod_bnm1_next_size (i + 1);
-#else
-  ASSERT_ALWAYS(0); /* ks_wrapmul should not be called in that case */
-#endif
   ASSERT(i % s == 0);
   m = i / s;
   ASSERT(m <= 2 * m0 - 3 + list_mul_mem (m0 - 1));
@@ -671,4 +671,5 @@
   free (t2_ptr);
   
   return m;
+#endif /* FFT_WRAP */
 }
diff -Nru gmp-ecm-7.0.3+ds/main.c gmp-ecm-7.0.4+ds/main.c
--- gmp-ecm-7.0.3+ds/main.c	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/main.c	2016-08-22 11:05:41.000000000 +0000
@@ -1349,7 +1349,7 @@
               exit (EXIT_FAILURE);
             }
           else if (verbose >= OUTPUT_VERBOSE)
-              fprintf (stdout, "Reading batch product (of %zu bits) of "
+              fprintf (stdout, "Reading batch product (of %"PRIu64" bits) of "
                                "primes up to B1=%1.0f from %s took %ldms\n", 
                                mpz_sizeinbase (params->batch_s, 2), B1,
                                loadfile_s, cputime () - st);
diff -Nru gmp-ecm-7.0.3+ds/Makefile.am gmp-ecm-7.0.4+ds/Makefile.am
--- gmp-ecm-7.0.3+ds/Makefile.am	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/Makefile.am	2016-08-23 12:25:36.000000000 +0000
@@ -50,7 +50,7 @@
 # www.gnu.org/software/libtool/manual/html_node/Updating-version-info.html
 # If any interfaces have been added, removed, or changed since the last
 # update, increment current, and set revision to 0.
-libecm_la_LDFLAGS = $(LIBECM_LDFLAGS) -version-info 1:0:0 -g -Wl,-znoexecstack
+libecm_la_LDFLAGS = $(LIBECM_LDFLAGS) -version-info 1:0:0 -g
 libecm_la_LIBADD = $(MULREDCLIBRARY)
 if WANT_GPU 
   libecm_la_SOURCES += cudakernel.cu
@@ -143,11 +143,11 @@
 endif
 
 TESTS = $(dist_check_SCRIPTS)
-TESTS_ENVIRONMENT = GMPECM_DATADIR=$(srcdir) $(VALGRIND)
+TESTS_ENVIRONMENT = $(VALGRIND)
 
 # see https://www.gnu.org/software/automake/manual/html_node/Scripts_002dbased-Testsuites.html
 # this is only needed for "make longcheck" below
-# export GMPECM_DATADIR = $(srcdir)
+export GMPECM_DATADIR = $(srcdir)
 
 longcheck: ecm$(EXEEXT)
 	$(srcdir)/test.pp1 "$(VALGRIND) ./ecm$(EXEEXT)"
diff -Nru gmp-ecm-7.0.3+ds/mpmod.c gmp-ecm-7.0.4+ds/mpmod.c
--- gmp-ecm-7.0.3+ds/mpmod.c	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/mpmod.c	2016-09-06 07:39:34.000000000 +0000
@@ -1111,7 +1111,7 @@
 }
 
 /* R <- BASE^EXP mod modulus.
-   EXP can be negative.
+   Warning: EXP can be negative (only occurs for P-1 in revision 2982).
  */
 void 
 mpres_pow (mpres_t R, const mpres_t BASE, const mpz_t EXP, mpmod_t modulus)
@@ -1129,17 +1129,23 @@
     {
       size_t expidx;
       mp_limb_t bitmask, expbits;
+      int sgn = mpz_sgn (EXP);
+      mpz_t exp;
 
       /* case EXP=0 */
-      if (mpz_sgn (EXP) == 0)
+      if (sgn == 0)
         {
           mpres_set_ui (R, 1UL, modulus); /* set result to 1 */
           ASSERT_NORMALIZED (R);
           return;
         }
 
-      expidx = mpz_size (EXP) - 1;         /* point at most significant limb */
-      expbits = mpz_getlimbn (EXP, expidx); /* get most significant limb */
+      /* exp <- |EXP| */
+      PTR(exp) = PTR(EXP);
+      SIZ(exp) = ABSIZ(EXP);
+
+      expidx = mpz_size (exp) - 1;         /* point at most significant limb */
+      expbits = mpz_getlimbn (exp, expidx); /* get most significant limb */
       ASSERT (expbits != 0);
 
       /* Scan for the MSB in expbits */
@@ -1150,7 +1156,7 @@
       /* bitmask is set to mask in the msb of expbits */
 
       k = 1; /* sliding window size */
-      expnbits = mpz_sizeinbase (EXP, 2);
+      expnbits = mpz_sizeinbase (exp, 2);
       /* the average number of multiplications is 2^(k-1) + expnbits / (k+1) */
       while ((1 << (k-1)) + expnbits / (k + 1) > (1 << k) + expnbits / (k + 2))
         k ++;
@@ -1197,9 +1203,9 @@
                     }
                   /* if the current bit is the smallest one of the window,
                      we multiply by BASE^w */
-                  if ((mpz_sgn (EXP) > 0 && mpz_scan1 (EXP, n0) == expnbits - 1)
-                    || (mpz_sgn (EXP) < 0 && mpz_scan0 (EXP, n0) == expnbits - 1))
+                  if (mpz_scan1 (exp, n0) == expnbits - 1)
                     {
+                      ASSERT(w/2 < K);
                       mpres_pow_mul (modulus->temp2, (w == 1) ? BASE : B[w/2],
                                      modulus->temp2, modulus);
                       w = lw = 0;
@@ -1209,9 +1215,10 @@
           if (expidx == 0)		/* if we just processed the least */
             break;			/* significant limb, we are done */
           expidx --;
-          expbits = mpz_getlimbn (EXP, expidx);
+          expbits = mpz_getlimbn (exp, expidx);
           bitmask = (mp_limb_t) 1 << (GMP_NUMB_BITS - 1);
         }
+      ASSERT(w/2 < K);
       if (w != 0)
         mpres_pow_mul (modulus->temp2, (w == 1) ? BASE : B[w/2],
                        modulus->temp2, modulus);
@@ -1222,10 +1229,8 @@
 
       /* mpz_getlimbn() ignores sign of argument, so we computed BASE^|EXP|.
          If EXP was negative, do a modular inverse */
-      if (mpz_sgn (EXP) < 0)
-        {
-          mpres_invert (R, R, modulus);
-        }
+      if (sgn < 0)
+        mpres_invert (R, R, modulus);
     } /* if (modulus->repr == ECM_MOD_BASE2 || ... ) */
   ASSERT_NORMALIZED (R);
 }
@@ -1693,11 +1698,11 @@
   ASSERT_NORMALIZED (R);  
 }
 
-/* Compagnon function to mpres_set_z_for_gcd(). It divides by c^n. 
+/* Companion function to mpres_set_z_for_gcd(). It divides by c^n. 
    In case of MULREDC with k b-bit words, c = 1/2^(b*k), so we multiply 
    by 2^(n*b*k). The purpose is to fix products of n terms collected by 
    using mpres_set_z_for_gcd(), so that we can still get the exact residue. */
-   
+
 void 
 mpres_set_z_for_gcd_fix (mpres_t R, const mpres_t S, const mpz_t n, mpmod_t modulus)
 {
@@ -1712,6 +1717,7 @@
           mpz_init (nb);
           mpres_init (po2, modulus);
 
+          /* Note: modulus->bits is always non-negative for MODMULN and REDC */
           mpz_mul_ui (nb, n, modulus->bits);
           mpres_set_ui (po2, 2, modulus);
           mpres_pow (po2, po2, nb, modulus);
@@ -2035,6 +2041,18 @@
   mpz_gcd (R, S, modulus->orig_modulus);
 }
 
+#ifdef HAVE_ADDLAWS
+/* Returns non-zero if the two residues are equal, 
+   and zero if they are not */
+int 
+mpres_equal (const mpres_t S1, const mpres_t S2, mpmod_t modulus)
+{
+  mpz_mod (modulus->temp1, S1, modulus->orig_modulus);
+  mpz_mod (modulus->temp2, S2, modulus->orig_modulus);
+  return (mpz_cmp (modulus->temp1, modulus->temp2) == 0);
+}
+#endif
+
 #if 0 /* those routines are not called in normal operation */
 
 void 
@@ -2117,7 +2135,7 @@
 
 /* Returns 1 if successful, 0 if not */
 static int
-test_mpres_set_z_for_gcd_fix(const int maxk, mpmod_t modulus)
+test_mpres_set_z_for_gcd_fix (const int maxk, mpmod_t modulus)
 {
   mpres_t m, prod;
   mpz_t n;
@@ -2158,16 +2176,6 @@
   return 1;
 }
 
-/* Returns non-zero if the two residues are equal, 
-   and zero if they are not */
-int 
-mpres_equal (const mpres_t S1, const mpres_t S2, mpmod_t modulus)
-{
-  mpz_mod (modulus->temp1, S1, modulus->orig_modulus);
-  mpz_mod (modulus->temp2, S2, modulus->orig_modulus);
-  return (mpz_cmp (modulus->temp1, modulus->temp2) == 0);
-}
-
 int
 mpmod_selftest (const mpz_t n)
 {
diff -Nru gmp-ecm-7.0.3+ds/NEWS gmp-ecm-7.0.4+ds/NEWS
--- gmp-ecm-7.0.3+ds/NEWS	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/NEWS	2016-08-26 13:16:29.000000000 +0000
@@ -1,3 +1,10 @@
+Changes between GMP-ECM 7.0.3 and 7.0.4:
+* fixed a bug in mpres_pow:
+  https://gforge.inria.fr/tracker/?func=detail&atid=623&aid=20712&group_id=135
+* fixed a bug with P-1 or P+1 and small B2
+* improved portability (OS X, Windows, Cygwin, non GNU ld)
+* fixed make longcheck with out-of-source build
+
 Changes between GMP-ECM 7.0.2 and GMP-ECM 7.0.3:
 * fixed compilation issue on powerpc64
 
diff -Nru gmp-ecm-7.0.3+ds/pm1fs2.c gmp-ecm-7.0.4+ds/pm1fs2.c
--- gmp-ecm-7.0.3+ds/pm1fs2.c	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/pm1fs2.c	2016-08-26 11:59:15.000000000 +0000
@@ -1752,6 +1752,38 @@
 }
 #endif
 
+/* return the memory needed in the F[] array for poly_from_sets_V */
+static unsigned long
+mem_poly_from_sets_V (sets_long_t *sets)
+{
+  unsigned long c, deg, i, nr;
+  set_long_t *set = sets->sets;
+  unsigned long mem, maxmem = 0;
+  
+  deg = 1UL;
+  for (nr = sets->nr - 1UL; nr > 0UL; nr--)
+    {
+      set = sets_nextset (sets->sets); /* Skip first set */
+      for (i = 1UL; i < nr; i++) /* Skip over remaining sets but one */
+        set = sets_nextset (set);
+      c = set->card;
+      if (c == 2UL)
+        deg *= 2UL;
+      else
+        {
+          i = (c - 1UL) / 2UL - 1; /* maximal value of i */
+          /* list_scale_V needs 2*deg+1 entries starting at
+             F + (2UL * i + 1UL) * (deg + 1UL) */
+          mem = (2UL * i + 1UL) * (deg + 1UL) + (2 * deg + 1UL);
+          if (mem > maxmem)
+            maxmem = mem;
+	  deg *= c;
+	}
+    }
+
+  return maxmem;
+}
+
 /* Build a polynomial with roots r^2i, i in the sumset of the sets in "sets".
    The parameter Q = r + 1/r. This code uses the fact that the polynomials 
    are symmetric. Requires that the first set in "sets" has cardinality 2,
@@ -1885,7 +1917,7 @@
 }
 
 static int
-build_F_ntt (listz_t F, const mpres_t P_1, sets_long_t *S_1, 
+build_F_ntt (listz_t F, const mpres_t P_1, sets_long_t *S_1,
 	     const faststage2_param_t *params, mpmod_t modulus)
 {
   mpzspm_t F_ntt_context;
@@ -1919,8 +1951,8 @@
   tmplen = params->s_1 + 100;
   tmp = init_list2 (tmplen, (unsigned int) abs (modulus->bits));
   F_ntt = mpzspv_init (1UL << ceil_log2 (params->s_1 / 2 + 1), F_ntt_context);
-  
-  i = poly_from_sets_V (F, P_1, S_1, tmp, tmplen, modulus, F_ntt, 
+
+  i = poly_from_sets_V (F, P_1, S_1, tmp, tmplen, modulus, F_ntt,
                         F_ntt_context);
   ASSERT_ALWAYS(2 * i == params->s_1);
   ASSERT_ALWAYS(mpz_cmp_ui (F[i], 1UL) == 0);
@@ -2039,6 +2071,7 @@
   mpz_neg (t, t);
   mpz_mul_2exp (t, t, 1UL);
   mpz_add_ui (t, t, 1UL);
+  /* Warning: t might be negative here. */
   mpres_pow (r[1], r[0], t, modulus);    /* r[1] = r^{2(-M+i)+1}, i = 0 */
   mpz_set_si (t, M);
   mpz_mul (t, t, t);                     /* t = M^2 */
@@ -2054,6 +2087,7 @@
     outputf (OUTPUT_TRACE, "/* pm1_sequence_g */ 2*%ld + (2*%Zd + 1)*P == "
 	     "%Zd /* PARI C */\n", k_2, m_1, t);
 
+  /* Warning: t might be negative here. */
   mpres_pow (x_0, b_1, t, modulus);  /* x_0 = b_1^{2*k_2 + (2*m_1 + 1)*P} */
   if (want_output && test_verbose (OUTPUT_TRACE))
     {
@@ -2670,7 +2704,7 @@
   /* First compute X + 1/X */
   mpres_invert (mr, X, modulus);
   mpres_add (mr, mr, X, modulus);
-  
+
   i = poly_from_sets_V (F, mr, S_1, tmp, tmplen, modulus, NULL, NULL);
   ASSERT_ALWAYS(2 * i == params->s_1);
   ASSERT(mpz_cmp_ui (F[i], 1UL) == 0);
@@ -2849,7 +2883,7 @@
   unsigned long nr;
   unsigned long l, lenF;
   sets_long_t *S_1; /* This is stored as a set of sets (arithmetic 
-                       progressions of prime length */
+                       progressions of prime length) */
   set_long_t *S_2; /* This is stored as a regular set */
   listz_t F;   /* Polynomial F has roots X^{k_1} for k_1 \in S_1, so has 
 		  degree s_1. It is symmetric, so has only s_1 / 2 + 1 
@@ -2897,6 +2931,14 @@
   mpres_init (tmpres, modulus);
   lenF = params->s_1 / 2 + 1 + 1; /* Another +1 because poly_from_sets_V stores
 				     the leading 1 monomial for each factor */
+  {
+    /* in some cases the above value of lenF is not enough, for example with
+       s_1 = 10, which gives lenF = 7, but 9 entries are needed */
+    unsigned long mem = mem_poly_from_sets_V (S_1);
+    if (mem > lenF)
+      lenF = mem;
+  }
+
   F = init_list2 (lenF, (unsigned int) abs (modulus->bits));
 
   mpres_get_z (mt, X, modulus); /* mpz_t copy of X for printing */
@@ -4269,6 +4311,14 @@
   /* Allocate memory for F with correct amount of space for each mpz_t */
   lenF = params->s_1 / 2 + 1 + 1; /* Another +1 because poly_from_sets_V stores
 				     the leading 1 monomial for each factor */
+  {
+    /* in some cases the above value of lenF is not enough, for example with
+       s_1 = 10, which gives lenF = 7, but 9 entries are needed */
+    unsigned long mem = mem_poly_from_sets_V (S_1);
+    if (mem > lenF)
+      lenF = mem;
+  }
+
   F = init_list2 (lenF, (unsigned int) abs (modulus->bits) + GMP_NUMB_BITS);
   
   /* Build F */
diff -Nru gmp-ecm-7.0.3+ds/README.gpu gmp-ecm-7.0.4+ds/README.gpu
--- gmp-ecm-7.0.3+ds/README.gpu	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/README.gpu	2016-08-23 09:22:08.000000000 +0000
@@ -7,7 +7,7 @@
 1. How to enable GPU code in GMP-ECM
 2. Basic Usage
 3. Advanced Usage
-
+4. Known issues
 
 ##############################################################################
 
@@ -99,5 +99,13 @@
 
 ##############################################################################
 
-Please report to Cyril.Bouvier@inria.fr any problems, bugs, ...
+4. Known issues
+
+On some configurations (GTX 570 with compute capability 2.0 for example)
+one gets the Cuda error "too many resources requested for launch". This
+can be solved by decreasing ECM_GPU_CURVES_BY_BLOCK from 32 to 16 in ecm-gpu.h.
+
+##############################################################################
+
+Please report to mail@cyrilbouvier.fr any problems, bugs, ...
 
diff -Nru gmp-ecm-7.0.3+ds/test.ecm gmp-ecm-7.0.4+ds/test.ecm
--- gmp-ecm-7.0.3+ds/test.ecm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/test.ecm	2016-08-22 11:05:41.000000000 +0000
@@ -441,10 +441,10 @@
 
 # test -bsaves/-bloads
 echo 18446744073709551557 | $ECM -param 1 -A 1 -bsaves test.ecm.s 11000
-echo 18446744073709551557 | $ECM -param 1 -A 312656731337392125 -bloads test.ecm.s 11000
-C=$?
+echo 18446744073709551557 | $ECM -param 1 -A 312656731337392125 -bloads test.ecm.s 11000; checkcode $? 8
+# The following test (with -v and -bloads) would crash on Windows prior to svn 2968.
+echo 18446744073709551557 | $ECM -param 1 -A 312656731337392125 -bloads test.ecm.s -v 11000; checkcode $? 8
 /bin/rm -f test.ecm.s
-checkcode $C 8
 
 # non-regression test for bug fixed by changeset r1819 on 32-bit
 echo 4294967291 | $ECM -param 1 -A 17 1000
diff -Nru gmp-ecm-7.0.3+ds/test.gpuecm gmp-ecm-7.0.4+ds/test.gpuecm
--- gmp-ecm-7.0.3+ds/test.gpuecm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/test.gpuecm	2016-07-25 14:11:06.000000000 +0000
@@ -113,7 +113,7 @@
 # bug reported by David Cleaver (was also reported by Paul Leyland)
 # should report p6 factor and c210 cofactor instead of "Input number found"
 # fixed with svn 2867
-echo "(2^718+1)/5" | $ECM -sigma 3:2000 40 48
+echo "(2^718+1)/5" | $ECM -sigma 3:2016 40 48
 checkcode $? 6
 
 
diff -Nru gmp-ecm-7.0.3+ds/test.pm1 gmp-ecm-7.0.4+ds/test.pm1
--- gmp-ecm-7.0.3+ds/test.pm1	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/test.pm1	2016-08-29 06:27:01.000000000 +0000
@@ -168,4 +168,19 @@
 # exercise bug fixed in revision 2949 on 64-bit processor (occurred with -redc)
 echo "(2^1049-1)/33569/459463" | $PM1 1049 21686243; checkcode $? 14
 
+# exercise bug (Segmentation fault) in ecm 7.0.x for x <= 3
+echo 1000033 | $PM1 32 947; checkcode $? 8
+echo 10001801 | $PM1 43 1163; checkcode $? 8
+echo 100002943 | $PM1 23 1093; checkcode $? 8
+echo 1000020313 | $PM1 37 859; checkcode $? 8
+echo 10000026037 | $PM1 49 1367; checkcode $? 8
+
+# Segmentation fault in 6.4.4
+echo 100000003897 | $PM1 787 1223; checkcode $? 8
+echo 1000000671451 | $PM1 757 1249; checkcode $? 8
+echo 10000000097557 | $PM1 761 1279; checkcode $? 8
+echo 100000002915569 | $PM1 1361 1823; checkcode $? 8
+echo 1000000001405311 | $PM1 1373 1759; checkcode $? 8
+echo 10000000003317199 | $PM1 2269 2803; checkcode $? 8
+
 echo "All P-1 tests are ok."
diff -Nru gmp-ecm-7.0.3+ds/test.pp1 gmp-ecm-7.0.4+ds/test.pp1
--- gmp-ecm-7.0.3+ds/test.pp1	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/test.pp1	2016-08-29 06:28:03.000000000 +0000
@@ -180,5 +180,14 @@
 # exercise -x0 0 option
 echo 18446744073709551653 | $PP1 -x0 0 257687 141775993; checkcode $? 8
 
+# exercise bug (Segmentation fault) in ecm 7.0.x for x <= 3
+echo 1000039 | $PP1 -x0 4 23 1087; checkcode $? 8
+echo 10001779 | $PP1 -x0 4 23 1279; checkcode $? 8
+echo 100002941 | $PP1 -x0 4 23 1093; checkcode $? 8
+echo 1000024339 | $PP1 -x0 4 49 1249; checkcode $? 8
+echo 10000098767 | $PP1 -x0 3 41 937; checkcode $? 8
+echo 100000049993 | $PP1 -x0 3 43 1061; checkcode $? 8
+echo 1000008849503 | $PP1 -x0 3 47 859; checkcode $? 8
+
 echo "All P+1 tests are ok."
 
diff -Nru gmp-ecm-7.0.3+ds/tune.c gmp-ecm-7.0.4+ds/tune.c
--- gmp-ecm-7.0.3+ds/tune.c	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/tune.c	2016-08-22 11:05:41.000000000 +0000
@@ -429,7 +429,7 @@
   for (n = 1; n < TUNE_LIST_MUL_N_MAX_SIZE; n++)
     {
       if (tune_verbose)
-        printf ("%zu:", n);
+        printf ("%"PRIu64":", n);
       __k = 1;
       TUNE_FUNC_LOOP(list_mul_n_basecase(z, x, y, n));
       st[0] = (double) __st / (double) __k;
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc10.asm gmp-ecm-7.0.4+ds/x86_64/mulredc10.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc10.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc10.asm	2016-08-25 13:42:10.000000000 +0000
@@ -23,7 +23,7 @@
 include(`config.m4')
 
 	TEXT
-.p2align 6 # Opteron L1 code cache line is 64 bytes long
+.p2align 6 # x86_64 L1 code cache line is 64 bytes long
 	GLOBL GSYM_PREFIX`'mulredc10
 	TYPE(GSYM_PREFIX`'mulredc`'10,`function')
 
@@ -143,15 +143,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	2f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(2)
-	popf')
 define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
 define(`T0', defn(`T1'))dnl
@@ -179,12 +170,6 @@
 	movq	8(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -223,12 +208,6 @@
 	movq	16(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -267,12 +246,6 @@
 	movq	24(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -311,12 +284,6 @@
 	movq	32(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -355,12 +322,6 @@
 	movq	40(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -399,12 +360,6 @@
 	movq	48(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -443,12 +398,6 @@
 	movq	56(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -487,12 +436,6 @@
 	movq	64(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -567,15 +510,6 @@
 
 	movq	8(%YP), %rax		# Fetch y[1]
 
-ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	4f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(4)
-	popf')
 define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
 define(`T0', defn(`T1'))dnl
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc1_10.asm gmp-ecm-7.0.4+ds/x86_64/mulredc1_10.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc1_10.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc1_10.asm	2016-08-25 13:42:13.000000000 +0000
@@ -93,16 +93,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-ifdef(`WANT_ASSERT', `
-        pushf
-	testq	T0, T0
-	jz	assert1
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-assert1:
-	popf
-')
 define(`TT', defn(`T0'))dnl
 define(`T0', defn(`T1'))dnl
 define(`T1', defn(`TT'))dnl
@@ -126,13 +116,6 @@
 	movq	8(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -167,13 +150,6 @@
 	movq	16(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -208,13 +184,6 @@
 	movq	24(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -249,13 +218,6 @@
 	movq	32(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -290,13 +252,6 @@
 	movq	40(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -331,13 +286,6 @@
 	movq	48(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -372,13 +320,6 @@
 	movq	56(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -413,13 +354,6 @@
 	movq	64(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc1_11.asm gmp-ecm-7.0.4+ds/x86_64/mulredc1_11.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc1_11.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc1_11.asm	2016-08-25 13:42:13.000000000 +0000
@@ -93,16 +93,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-ifdef(`WANT_ASSERT', `
-        pushf
-	testq	T0, T0
-	jz	assert1
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-assert1:
-	popf
-')
 define(`TT', defn(`T0'))dnl
 define(`T0', defn(`T1'))dnl
 define(`T1', defn(`TT'))dnl
@@ -126,13 +116,6 @@
 	movq	8(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -167,13 +150,6 @@
 	movq	16(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -208,13 +184,6 @@
 	movq	24(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -249,13 +218,6 @@
 	movq	32(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -290,13 +252,6 @@
 	movq	40(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -331,13 +286,6 @@
 	movq	48(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -372,13 +320,6 @@
 	movq	56(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -413,13 +354,6 @@
 	movq	64(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -454,13 +388,6 @@
 	movq	72(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc1_12.asm gmp-ecm-7.0.4+ds/x86_64/mulredc1_12.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc1_12.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc1_12.asm	2016-08-25 13:42:13.000000000 +0000
@@ -93,16 +93,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-ifdef(`WANT_ASSERT', `
-        pushf
-	testq	T0, T0
-	jz	assert1
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-assert1:
-	popf
-')
 define(`TT', defn(`T0'))dnl
 define(`T0', defn(`T1'))dnl
 define(`T1', defn(`TT'))dnl
@@ -126,13 +116,6 @@
 	movq	8(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -167,13 +150,6 @@
 	movq	16(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -208,13 +184,6 @@
 	movq	24(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -249,13 +218,6 @@
 	movq	32(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -290,13 +252,6 @@
 	movq	40(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -331,13 +286,6 @@
 	movq	48(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -372,13 +320,6 @@
 	movq	56(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -413,13 +354,6 @@
 	movq	64(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -454,13 +388,6 @@
 	movq	72(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -495,13 +422,6 @@
 	movq	80(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc1_13.asm gmp-ecm-7.0.4+ds/x86_64/mulredc1_13.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc1_13.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc1_13.asm	2016-08-25 13:42:13.000000000 +0000
@@ -93,16 +93,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-ifdef(`WANT_ASSERT', `
-        pushf
-	testq	T0, T0
-	jz	assert1
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-assert1:
-	popf
-')
 define(`TT', defn(`T0'))dnl
 define(`T0', defn(`T1'))dnl
 define(`T1', defn(`TT'))dnl
@@ -126,13 +116,6 @@
 	movq	8(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -167,13 +150,6 @@
 	movq	16(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -208,13 +184,6 @@
 	movq	24(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -249,13 +218,6 @@
 	movq	32(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -290,13 +252,6 @@
 	movq	40(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -331,13 +286,6 @@
 	movq	48(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -372,13 +320,6 @@
 	movq	56(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -413,13 +354,6 @@
 	movq	64(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -454,13 +388,6 @@
 	movq	72(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -495,13 +422,6 @@
 	movq	80(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -536,13 +456,6 @@
 	movq	88(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc1_14.asm gmp-ecm-7.0.4+ds/x86_64/mulredc1_14.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc1_14.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc1_14.asm	2016-08-25 13:42:13.000000000 +0000
@@ -93,16 +93,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-ifdef(`WANT_ASSERT', `
-        pushf
-	testq	T0, T0
-	jz	assert1
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-assert1:
-	popf
-')
 define(`TT', defn(`T0'))dnl
 define(`T0', defn(`T1'))dnl
 define(`T1', defn(`TT'))dnl
@@ -126,13 +116,6 @@
 	movq	8(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -167,13 +150,6 @@
 	movq	16(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -208,13 +184,6 @@
 	movq	24(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -249,13 +218,6 @@
 	movq	32(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -290,13 +252,6 @@
 	movq	40(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -331,13 +286,6 @@
 	movq	48(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -372,13 +320,6 @@
 	movq	56(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -413,13 +354,6 @@
 	movq	64(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -454,13 +388,6 @@
 	movq	72(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -495,13 +422,6 @@
 	movq	80(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -536,13 +456,6 @@
 	movq	88(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -577,13 +490,6 @@
 	movq	96(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc1_15.asm gmp-ecm-7.0.4+ds/x86_64/mulredc1_15.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc1_15.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc1_15.asm	2016-08-25 13:42:14.000000000 +0000
@@ -93,16 +93,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-ifdef(`WANT_ASSERT', `
-        pushf
-	testq	T0, T0
-	jz	assert1
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-assert1:
-	popf
-')
 define(`TT', defn(`T0'))dnl
 define(`T0', defn(`T1'))dnl
 define(`T1', defn(`TT'))dnl
@@ -126,13 +116,6 @@
 	movq	8(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -167,13 +150,6 @@
 	movq	16(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -208,13 +184,6 @@
 	movq	24(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -249,13 +218,6 @@
 	movq	32(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -290,13 +252,6 @@
 	movq	40(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -331,13 +286,6 @@
 	movq	48(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -372,13 +320,6 @@
 	movq	56(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -413,13 +354,6 @@
 	movq	64(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -454,13 +388,6 @@
 	movq	72(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -495,13 +422,6 @@
 	movq	80(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -536,13 +456,6 @@
 	movq	88(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -577,13 +490,6 @@
 	movq	96(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -618,13 +524,6 @@
 	movq	104(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc1_16.asm gmp-ecm-7.0.4+ds/x86_64/mulredc1_16.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc1_16.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc1_16.asm	2016-08-25 13:42:14.000000000 +0000
@@ -93,16 +93,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-ifdef(`WANT_ASSERT', `
-        pushf
-	testq	T0, T0
-	jz	assert1
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-assert1:
-	popf
-')
 define(`TT', defn(`T0'))dnl
 define(`T0', defn(`T1'))dnl
 define(`T1', defn(`TT'))dnl
@@ -126,13 +116,6 @@
 	movq	8(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -167,13 +150,6 @@
 	movq	16(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -208,13 +184,6 @@
 	movq	24(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -249,13 +218,6 @@
 	movq	32(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -290,13 +252,6 @@
 	movq	40(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -331,13 +286,6 @@
 	movq	48(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -372,13 +320,6 @@
 	movq	56(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -413,13 +354,6 @@
 	movq	64(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -454,13 +388,6 @@
 	movq	72(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -495,13 +422,6 @@
 	movq	80(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -536,13 +456,6 @@
 	movq	88(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -577,13 +490,6 @@
 	movq	96(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -618,13 +524,6 @@
 	movq	104(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -659,13 +558,6 @@
 	movq	112(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc1_17.asm gmp-ecm-7.0.4+ds/x86_64/mulredc1_17.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc1_17.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc1_17.asm	2016-08-25 13:42:14.000000000 +0000
@@ -93,16 +93,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-ifdef(`WANT_ASSERT', `
-        pushf
-	testq	T0, T0
-	jz	assert1
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-assert1:
-	popf
-')
 define(`TT', defn(`T0'))dnl
 define(`T0', defn(`T1'))dnl
 define(`T1', defn(`TT'))dnl
@@ -126,13 +116,6 @@
 	movq	8(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -167,13 +150,6 @@
 	movq	16(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -208,13 +184,6 @@
 	movq	24(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -249,13 +218,6 @@
 	movq	32(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -290,13 +252,6 @@
 	movq	40(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -331,13 +286,6 @@
 	movq	48(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -372,13 +320,6 @@
 	movq	56(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -413,13 +354,6 @@
 	movq	64(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -454,13 +388,6 @@
 	movq	72(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -495,13 +422,6 @@
 	movq	80(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -536,13 +456,6 @@
 	movq	88(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -577,13 +490,6 @@
 	movq	96(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -618,13 +524,6 @@
 	movq	104(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -659,13 +558,6 @@
 	movq	112(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -700,13 +592,6 @@
 	movq	120(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc1_18.asm gmp-ecm-7.0.4+ds/x86_64/mulredc1_18.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc1_18.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc1_18.asm	2016-08-25 13:42:14.000000000 +0000
@@ -93,16 +93,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-ifdef(`WANT_ASSERT', `
-        pushf
-	testq	T0, T0
-	jz	assert1
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-assert1:
-	popf
-')
 define(`TT', defn(`T0'))dnl
 define(`T0', defn(`T1'))dnl
 define(`T1', defn(`TT'))dnl
@@ -126,13 +116,6 @@
 	movq	8(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -167,13 +150,6 @@
 	movq	16(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -208,13 +184,6 @@
 	movq	24(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -249,13 +218,6 @@
 	movq	32(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -290,13 +252,6 @@
 	movq	40(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -331,13 +286,6 @@
 	movq	48(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -372,13 +320,6 @@
 	movq	56(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -413,13 +354,6 @@
 	movq	64(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -454,13 +388,6 @@
 	movq	72(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -495,13 +422,6 @@
 	movq	80(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -536,13 +456,6 @@
 	movq	88(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -577,13 +490,6 @@
 	movq	96(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -618,13 +524,6 @@
 	movq	104(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -659,13 +558,6 @@
 	movq	112(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -700,13 +592,6 @@
 	movq	120(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -741,13 +626,6 @@
 	movq	128(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc1_19.asm gmp-ecm-7.0.4+ds/x86_64/mulredc1_19.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc1_19.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc1_19.asm	2016-08-25 13:42:14.000000000 +0000
@@ -93,16 +93,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-ifdef(`WANT_ASSERT', `
-        pushf
-	testq	T0, T0
-	jz	assert1
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-assert1:
-	popf
-')
 define(`TT', defn(`T0'))dnl
 define(`T0', defn(`T1'))dnl
 define(`T1', defn(`TT'))dnl
@@ -126,13 +116,6 @@
 	movq	8(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -167,13 +150,6 @@
 	movq	16(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -208,13 +184,6 @@
 	movq	24(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -249,13 +218,6 @@
 	movq	32(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -290,13 +252,6 @@
 	movq	40(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -331,13 +286,6 @@
 	movq	48(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -372,13 +320,6 @@
 	movq	56(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -413,13 +354,6 @@
 	movq	64(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -454,13 +388,6 @@
 	movq	72(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -495,13 +422,6 @@
 	movq	80(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -536,13 +456,6 @@
 	movq	88(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -577,13 +490,6 @@
 	movq	96(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -618,13 +524,6 @@
 	movq	104(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -659,13 +558,6 @@
 	movq	112(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -700,13 +592,6 @@
 	movq	120(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -741,13 +626,6 @@
 	movq	128(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -782,13 +660,6 @@
 	movq	136(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc11.asm gmp-ecm-7.0.4+ds/x86_64/mulredc11.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc11.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc11.asm	2016-08-25 13:42:10.000000000 +0000
@@ -23,7 +23,7 @@
 include(`config.m4')
 
 	TEXT
-.p2align 6 # Opteron L1 code cache line is 64 bytes long
+.p2align 6 # x86_64 L1 code cache line is 64 bytes long
 	GLOBL GSYM_PREFIX`'mulredc11
 	TYPE(GSYM_PREFIX`'mulredc`'11,`function')
 
@@ -143,15 +143,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	2f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(2)
-	popf')
 define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
 define(`T0', defn(`T1'))dnl
@@ -179,12 +170,6 @@
 	movq	8(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -223,12 +208,6 @@
 	movq	16(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -267,12 +246,6 @@
 	movq	24(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -311,12 +284,6 @@
 	movq	32(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -355,12 +322,6 @@
 	movq	40(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -399,12 +360,6 @@
 	movq	48(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -443,12 +398,6 @@
 	movq	56(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -487,12 +436,6 @@
 	movq	64(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -531,12 +474,6 @@
 	movq	72(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -611,15 +548,6 @@
 
 	movq	8(%YP), %rax		# Fetch y[1]
 
-ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	4f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(4)
-	popf')
 define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
 define(`T0', defn(`T1'))dnl
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc1_20.asm gmp-ecm-7.0.4+ds/x86_64/mulredc1_20.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc1_20.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc1_20.asm	2016-08-25 13:42:14.000000000 +0000
@@ -93,16 +93,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-ifdef(`WANT_ASSERT', `
-        pushf
-	testq	T0, T0
-	jz	assert1
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-assert1:
-	popf
-')
 define(`TT', defn(`T0'))dnl
 define(`T0', defn(`T1'))dnl
 define(`T1', defn(`TT'))dnl
@@ -126,13 +116,6 @@
 	movq	8(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -167,13 +150,6 @@
 	movq	16(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -208,13 +184,6 @@
 	movq	24(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -249,13 +218,6 @@
 	movq	32(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -290,13 +252,6 @@
 	movq	40(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -331,13 +286,6 @@
 	movq	48(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -372,13 +320,6 @@
 	movq	56(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -413,13 +354,6 @@
 	movq	64(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -454,13 +388,6 @@
 	movq	72(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -495,13 +422,6 @@
 	movq	80(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -536,13 +456,6 @@
 	movq	88(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -577,13 +490,6 @@
 	movq	96(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -618,13 +524,6 @@
 	movq	104(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -659,13 +558,6 @@
 	movq	112(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -700,13 +592,6 @@
 	movq	120(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -741,13 +626,6 @@
 	movq	128(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -782,13 +660,6 @@
 	movq	136(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -823,13 +694,6 @@
 	movq	144(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc1_2.asm gmp-ecm-7.0.4+ds/x86_64/mulredc1_2.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc1_2.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc1_2.asm	2016-08-25 13:42:12.000000000 +0000
@@ -93,16 +93,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-ifdef(`WANT_ASSERT', `
-        pushf
-	testq	T0, T0
-	jz	assert1
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-assert1:
-	popf
-')
 define(`TT', defn(`T0'))dnl
 define(`T0', defn(`T1'))dnl
 define(`T1', defn(`TT'))dnl
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc12.asm gmp-ecm-7.0.4+ds/x86_64/mulredc12.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc12.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc12.asm	2016-08-25 13:42:10.000000000 +0000
@@ -23,7 +23,7 @@
 include(`config.m4')
 
 	TEXT
-.p2align 6 # Opteron L1 code cache line is 64 bytes long
+.p2align 6 # x86_64 L1 code cache line is 64 bytes long
 	GLOBL GSYM_PREFIX`'mulredc12
 	TYPE(GSYM_PREFIX`'mulredc`'12,`function')
 
@@ -143,15 +143,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	2f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(2)
-	popf')
 define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
 define(`T0', defn(`T1'))dnl
@@ -179,12 +170,6 @@
 	movq	8(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -223,12 +208,6 @@
 	movq	16(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -267,12 +246,6 @@
 	movq	24(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -311,12 +284,6 @@
 	movq	32(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -355,12 +322,6 @@
 	movq	40(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -399,12 +360,6 @@
 	movq	48(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -443,12 +398,6 @@
 	movq	56(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -487,12 +436,6 @@
 	movq	64(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -531,12 +474,6 @@
 	movq	72(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -575,12 +512,6 @@
 	movq	80(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -655,15 +586,6 @@
 
 	movq	8(%YP), %rax		# Fetch y[1]
 
-ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	4f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(4)
-	popf')
 define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
 define(`T0', defn(`T1'))dnl
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc1_3.asm gmp-ecm-7.0.4+ds/x86_64/mulredc1_3.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc1_3.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc1_3.asm	2016-08-25 13:42:12.000000000 +0000
@@ -93,16 +93,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-ifdef(`WANT_ASSERT', `
-        pushf
-	testq	T0, T0
-	jz	assert1
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-assert1:
-	popf
-')
 define(`TT', defn(`T0'))dnl
 define(`T0', defn(`T1'))dnl
 define(`T1', defn(`TT'))dnl
@@ -126,13 +116,6 @@
 	movq	8(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc13.asm gmp-ecm-7.0.4+ds/x86_64/mulredc13.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc13.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc13.asm	2016-08-25 13:42:10.000000000 +0000
@@ -23,7 +23,7 @@
 include(`config.m4')
 
 	TEXT
-.p2align 6 # Opteron L1 code cache line is 64 bytes long
+.p2align 6 # x86_64 L1 code cache line is 64 bytes long
 	GLOBL GSYM_PREFIX`'mulredc13
 	TYPE(GSYM_PREFIX`'mulredc`'13,`function')
 
@@ -143,15 +143,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	2f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(2)
-	popf')
 define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
 define(`T0', defn(`T1'))dnl
@@ -179,12 +170,6 @@
 	movq	8(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -223,12 +208,6 @@
 	movq	16(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -267,12 +246,6 @@
 	movq	24(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -311,12 +284,6 @@
 	movq	32(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -355,12 +322,6 @@
 	movq	40(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -399,12 +360,6 @@
 	movq	48(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -443,12 +398,6 @@
 	movq	56(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -487,12 +436,6 @@
 	movq	64(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -531,12 +474,6 @@
 	movq	72(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -575,12 +512,6 @@
 	movq	80(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -619,12 +550,6 @@
 	movq	88(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -699,15 +624,6 @@
 
 	movq	8(%YP), %rax		# Fetch y[1]
 
-ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	4f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(4)
-	popf')
 define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
 define(`T0', defn(`T1'))dnl
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc1_4.asm gmp-ecm-7.0.4+ds/x86_64/mulredc1_4.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc1_4.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc1_4.asm	2016-08-25 13:42:12.000000000 +0000
@@ -93,16 +93,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-ifdef(`WANT_ASSERT', `
-        pushf
-	testq	T0, T0
-	jz	assert1
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-assert1:
-	popf
-')
 define(`TT', defn(`T0'))dnl
 define(`T0', defn(`T1'))dnl
 define(`T1', defn(`TT'))dnl
@@ -126,13 +116,6 @@
 	movq	8(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -167,13 +150,6 @@
 	movq	16(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc14.asm gmp-ecm-7.0.4+ds/x86_64/mulredc14.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc14.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc14.asm	2016-08-25 13:42:10.000000000 +0000
@@ -23,7 +23,7 @@
 include(`config.m4')
 
 	TEXT
-.p2align 6 # Opteron L1 code cache line is 64 bytes long
+.p2align 6 # x86_64 L1 code cache line is 64 bytes long
 	GLOBL GSYM_PREFIX`'mulredc14
 	TYPE(GSYM_PREFIX`'mulredc`'14,`function')
 
@@ -143,15 +143,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	2f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(2)
-	popf')
 define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
 define(`T0', defn(`T1'))dnl
@@ -179,12 +170,6 @@
 	movq	8(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -223,12 +208,6 @@
 	movq	16(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -267,12 +246,6 @@
 	movq	24(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -311,12 +284,6 @@
 	movq	32(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -355,12 +322,6 @@
 	movq	40(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -399,12 +360,6 @@
 	movq	48(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -443,12 +398,6 @@
 	movq	56(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -487,12 +436,6 @@
 	movq	64(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -531,12 +474,6 @@
 	movq	72(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -575,12 +512,6 @@
 	movq	80(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -619,12 +550,6 @@
 	movq	88(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -663,12 +588,6 @@
 	movq	96(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -743,15 +662,6 @@
 
 	movq	8(%YP), %rax		# Fetch y[1]
 
-ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	4f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(4)
-	popf')
 define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
 define(`T0', defn(`T1'))dnl
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc1_5.asm gmp-ecm-7.0.4+ds/x86_64/mulredc1_5.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc1_5.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc1_5.asm	2016-08-25 13:42:12.000000000 +0000
@@ -93,16 +93,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-ifdef(`WANT_ASSERT', `
-        pushf
-	testq	T0, T0
-	jz	assert1
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-assert1:
-	popf
-')
 define(`TT', defn(`T0'))dnl
 define(`T0', defn(`T1'))dnl
 define(`T1', defn(`TT'))dnl
@@ -126,13 +116,6 @@
 	movq	8(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -167,13 +150,6 @@
 	movq	16(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -208,13 +184,6 @@
 	movq	24(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc15.asm gmp-ecm-7.0.4+ds/x86_64/mulredc15.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc15.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc15.asm	2016-08-25 13:42:11.000000000 +0000
@@ -23,7 +23,7 @@
 include(`config.m4')
 
 	TEXT
-.p2align 6 # Opteron L1 code cache line is 64 bytes long
+.p2align 6 # x86_64 L1 code cache line is 64 bytes long
 	GLOBL GSYM_PREFIX`'mulredc15
 	TYPE(GSYM_PREFIX`'mulredc`'15,`function')
 
@@ -143,15 +143,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	2f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(2)
-	popf')
 define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
 define(`T0', defn(`T1'))dnl
@@ -179,12 +170,6 @@
 	movq	8(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -223,12 +208,6 @@
 	movq	16(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -267,12 +246,6 @@
 	movq	24(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -311,12 +284,6 @@
 	movq	32(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -355,12 +322,6 @@
 	movq	40(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -399,12 +360,6 @@
 	movq	48(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -443,12 +398,6 @@
 	movq	56(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -487,12 +436,6 @@
 	movq	64(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -531,12 +474,6 @@
 	movq	72(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -575,12 +512,6 @@
 	movq	80(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -619,12 +550,6 @@
 	movq	88(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -663,12 +588,6 @@
 	movq	96(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -707,12 +626,6 @@
 	movq	104(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -787,15 +700,6 @@
 
 	movq	8(%YP), %rax		# Fetch y[1]
 
-ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	4f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(4)
-	popf')
 define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
 define(`T0', defn(`T1'))dnl
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc1_6.asm gmp-ecm-7.0.4+ds/x86_64/mulredc1_6.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc1_6.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc1_6.asm	2016-08-25 13:42:12.000000000 +0000
@@ -93,16 +93,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-ifdef(`WANT_ASSERT', `
-        pushf
-	testq	T0, T0
-	jz	assert1
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-assert1:
-	popf
-')
 define(`TT', defn(`T0'))dnl
 define(`T0', defn(`T1'))dnl
 define(`T1', defn(`TT'))dnl
@@ -126,13 +116,6 @@
 	movq	8(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -167,13 +150,6 @@
 	movq	16(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -208,13 +184,6 @@
 	movq	24(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -249,13 +218,6 @@
 	movq	32(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc16.asm gmp-ecm-7.0.4+ds/x86_64/mulredc16.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc16.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc16.asm	2016-08-25 13:42:11.000000000 +0000
@@ -23,7 +23,7 @@
 include(`config.m4')
 
 	TEXT
-.p2align 6 # Opteron L1 code cache line is 64 bytes long
+.p2align 6 # x86_64 L1 code cache line is 64 bytes long
 	GLOBL GSYM_PREFIX`'mulredc16
 	TYPE(GSYM_PREFIX`'mulredc`'16,`function')
 
@@ -143,15 +143,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	2f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(2)
-	popf')
 define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
 define(`T0', defn(`T1'))dnl
@@ -179,12 +170,6 @@
 	movq	8(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -223,12 +208,6 @@
 	movq	16(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -267,12 +246,6 @@
 	movq	24(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -311,12 +284,6 @@
 	movq	32(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -355,12 +322,6 @@
 	movq	40(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -399,12 +360,6 @@
 	movq	48(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -443,12 +398,6 @@
 	movq	56(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -487,12 +436,6 @@
 	movq	64(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -531,12 +474,6 @@
 	movq	72(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -575,12 +512,6 @@
 	movq	80(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -619,12 +550,6 @@
 	movq	88(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -663,12 +588,6 @@
 	movq	96(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -707,12 +626,6 @@
 	movq	104(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -751,12 +664,6 @@
 	movq	112(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -831,15 +738,6 @@
 
 	movq	8(%YP), %rax		# Fetch y[1]
 
-ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	4f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(4)
-	popf')
 define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
 define(`T0', defn(`T1'))dnl
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc1_7.asm gmp-ecm-7.0.4+ds/x86_64/mulredc1_7.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc1_7.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc1_7.asm	2016-08-25 13:42:12.000000000 +0000
@@ -93,16 +93,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-ifdef(`WANT_ASSERT', `
-        pushf
-	testq	T0, T0
-	jz	assert1
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-assert1:
-	popf
-')
 define(`TT', defn(`T0'))dnl
 define(`T0', defn(`T1'))dnl
 define(`T1', defn(`TT'))dnl
@@ -126,13 +116,6 @@
 	movq	8(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -167,13 +150,6 @@
 	movq	16(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -208,13 +184,6 @@
 	movq	24(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -249,13 +218,6 @@
 	movq	32(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -290,13 +252,6 @@
 	movq	40(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc17.asm gmp-ecm-7.0.4+ds/x86_64/mulredc17.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc17.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc17.asm	2016-08-25 13:42:11.000000000 +0000
@@ -23,7 +23,7 @@
 include(`config.m4')
 
 	TEXT
-.p2align 6 # Opteron L1 code cache line is 64 bytes long
+.p2align 6 # x86_64 L1 code cache line is 64 bytes long
 	GLOBL GSYM_PREFIX`'mulredc17
 	TYPE(GSYM_PREFIX`'mulredc`'17,`function')
 
@@ -143,15 +143,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	2f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(2)
-	popf')
 define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
 define(`T0', defn(`T1'))dnl
@@ -179,12 +170,6 @@
 	movq	8(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -223,12 +208,6 @@
 	movq	16(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -267,12 +246,6 @@
 	movq	24(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -311,12 +284,6 @@
 	movq	32(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -355,12 +322,6 @@
 	movq	40(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -399,12 +360,6 @@
 	movq	48(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -443,12 +398,6 @@
 	movq	56(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -487,12 +436,6 @@
 	movq	64(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -531,12 +474,6 @@
 	movq	72(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -575,12 +512,6 @@
 	movq	80(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -619,12 +550,6 @@
 	movq	88(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -663,12 +588,6 @@
 	movq	96(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -707,12 +626,6 @@
 	movq	104(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -751,12 +664,6 @@
 	movq	112(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -795,12 +702,6 @@
 	movq	120(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -875,15 +776,6 @@
 
 	movq	8(%YP), %rax		# Fetch y[1]
 
-ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	4f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(4)
-	popf')
 define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
 define(`T0', defn(`T1'))dnl
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc1_8.asm gmp-ecm-7.0.4+ds/x86_64/mulredc1_8.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc1_8.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc1_8.asm	2016-08-25 13:42:13.000000000 +0000
@@ -93,16 +93,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-ifdef(`WANT_ASSERT', `
-        pushf
-	testq	T0, T0
-	jz	assert1
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-assert1:
-	popf
-')
 define(`TT', defn(`T0'))dnl
 define(`T0', defn(`T1'))dnl
 define(`T1', defn(`TT'))dnl
@@ -126,13 +116,6 @@
 	movq	8(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -167,13 +150,6 @@
 	movq	16(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -208,13 +184,6 @@
 	movq	24(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -249,13 +218,6 @@
 	movq	32(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -290,13 +252,6 @@
 	movq	40(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -331,13 +286,6 @@
 	movq	48(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc18.asm gmp-ecm-7.0.4+ds/x86_64/mulredc18.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc18.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc18.asm	2016-08-25 13:42:11.000000000 +0000
@@ -23,7 +23,7 @@
 include(`config.m4')
 
 	TEXT
-.p2align 6 # Opteron L1 code cache line is 64 bytes long
+.p2align 6 # x86_64 L1 code cache line is 64 bytes long
 	GLOBL GSYM_PREFIX`'mulredc18
 	TYPE(GSYM_PREFIX`'mulredc`'18,`function')
 
@@ -143,15 +143,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	2f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(2)
-	popf')
 define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
 define(`T0', defn(`T1'))dnl
@@ -179,12 +170,6 @@
 	movq	8(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -223,12 +208,6 @@
 	movq	16(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -267,12 +246,6 @@
 	movq	24(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -311,12 +284,6 @@
 	movq	32(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -355,12 +322,6 @@
 	movq	40(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -399,12 +360,6 @@
 	movq	48(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -443,12 +398,6 @@
 	movq	56(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -487,12 +436,6 @@
 	movq	64(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -531,12 +474,6 @@
 	movq	72(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -575,12 +512,6 @@
 	movq	80(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -619,12 +550,6 @@
 	movq	88(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -663,12 +588,6 @@
 	movq	96(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -707,12 +626,6 @@
 	movq	104(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -751,12 +664,6 @@
 	movq	112(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -795,12 +702,6 @@
 	movq	120(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -839,12 +740,6 @@
 	movq	128(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -919,15 +814,6 @@
 
 	movq	8(%YP), %rax		# Fetch y[1]
 
-ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	4f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(4)
-	popf')
 define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
 define(`T0', defn(`T1'))dnl
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc1_9.asm gmp-ecm-7.0.4+ds/x86_64/mulredc1_9.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc1_9.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc1_9.asm	2016-08-25 13:42:13.000000000 +0000
@@ -93,16 +93,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-ifdef(`WANT_ASSERT', `
-        pushf
-	testq	T0, T0
-	jz	assert1
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-assert1:
-	popf
-')
 define(`TT', defn(`T0'))dnl
 define(`T0', defn(`T1'))dnl
 define(`T1', defn(`TT'))dnl
@@ -126,13 +116,6 @@
 	movq	8(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -167,13 +150,6 @@
 	movq	16(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -208,13 +184,6 @@
 	movq	24(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -249,13 +218,6 @@
 	movq	32(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -290,13 +252,6 @@
 	movq	40(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -331,13 +286,6 @@
 	movq	48(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -372,13 +320,6 @@
 	movq	56(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc19.asm gmp-ecm-7.0.4+ds/x86_64/mulredc19.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc19.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc19.asm	2016-08-25 13:42:11.000000000 +0000
@@ -23,7 +23,7 @@
 include(`config.m4')
 
 	TEXT
-.p2align 6 # Opteron L1 code cache line is 64 bytes long
+.p2align 6 # x86_64 L1 code cache line is 64 bytes long
 	GLOBL GSYM_PREFIX`'mulredc19
 	TYPE(GSYM_PREFIX`'mulredc`'19,`function')
 
@@ -143,15 +143,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	2f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(2)
-	popf')
 define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
 define(`T0', defn(`T1'))dnl
@@ -179,12 +170,6 @@
 	movq	8(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -223,12 +208,6 @@
 	movq	16(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -267,12 +246,6 @@
 	movq	24(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -311,12 +284,6 @@
 	movq	32(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -355,12 +322,6 @@
 	movq	40(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -399,12 +360,6 @@
 	movq	48(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -443,12 +398,6 @@
 	movq	56(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -487,12 +436,6 @@
 	movq	64(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -531,12 +474,6 @@
 	movq	72(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -575,12 +512,6 @@
 	movq	80(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -619,12 +550,6 @@
 	movq	88(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -663,12 +588,6 @@
 	movq	96(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -707,12 +626,6 @@
 	movq	104(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -751,12 +664,6 @@
 	movq	112(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -795,12 +702,6 @@
 	movq	120(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -839,12 +740,6 @@
 	movq	128(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -883,12 +778,6 @@
 	movq	136(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -963,15 +852,6 @@
 
 	movq	8(%YP), %rax		# Fetch y[1]
 
-ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	4f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(4)
-	popf')
 define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
 define(`T0', defn(`T1'))dnl
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc1.asm gmp-ecm-7.0.4+ds/x86_64/mulredc1.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc1.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc1.asm	2016-08-25 13:40:28.000000000 +0000
@@ -48,11 +48,6 @@
 	mulq	INV_M           # compute u
 	mulq	M               # compute u*m
 	addq	TMP1, %rax      # rax is 0, now (carry is important)
-ifdef(`WANT_ASSERT', 
-`	jz	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx
-	call	abort@plt
-LABEL_SUFFIX(1)')
 	adcq	TMP2, %rdx
 	movq	%rdx, (Z)
 	adcq	$0, %rax
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc1.m4 gmp-ecm-7.0.4+ds/x86_64/mulredc1.m4
--- gmp-ecm-7.0.3+ds/x86_64/mulredc1.m4	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc1.m4	2016-08-25 13:40:56.000000000 +0000
@@ -100,16 +100,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-`ifdef(`WANT_ASSERT', `
-        pushf
-	testq	T0, T0
-	jz	assert1
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-assert1:
-	popf
-')'
 dnl Cycle ring buffer. Only mappings of T0 and T1 to regs change, no MOVs!
 `define(`TT', defn(`T0'))dnl
 define(`T0', defn(`T1'))dnl
@@ -138,13 +128,6 @@
 	movq	J`'(MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-`ifdef(`WANT_ASSERT', `
-	jnc	1f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-1:
-',`')'
 	
 	mulq	U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc20.asm gmp-ecm-7.0.4+ds/x86_64/mulredc20.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc20.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc20.asm	2016-08-25 13:42:11.000000000 +0000
@@ -23,7 +23,7 @@
 include(`config.m4')
 
 	TEXT
-.p2align 6 # Opteron L1 code cache line is 64 bytes long
+.p2align 6 # x86_64 L1 code cache line is 64 bytes long
 	GLOBL GSYM_PREFIX`'mulredc20
 	TYPE(GSYM_PREFIX`'mulredc`'20,`function')
 
@@ -143,15 +143,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	2f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(2)
-	popf')
 define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
 define(`T0', defn(`T1'))dnl
@@ -179,12 +170,6 @@
 	movq	8(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -223,12 +208,6 @@
 	movq	16(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -267,12 +246,6 @@
 	movq	24(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -311,12 +284,6 @@
 	movq	32(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -355,12 +322,6 @@
 	movq	40(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -399,12 +360,6 @@
 	movq	48(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -443,12 +398,6 @@
 	movq	56(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -487,12 +436,6 @@
 	movq	64(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -531,12 +474,6 @@
 	movq	72(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -575,12 +512,6 @@
 	movq	80(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -619,12 +550,6 @@
 	movq	88(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -663,12 +588,6 @@
 	movq	96(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -707,12 +626,6 @@
 	movq	104(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -751,12 +664,6 @@
 	movq	112(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -795,12 +702,6 @@
 	movq	120(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -839,12 +740,6 @@
 	movq	128(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -883,12 +778,6 @@
 	movq	136(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -927,12 +816,6 @@
 	movq	144(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -1007,15 +890,6 @@
 
 	movq	8(%YP), %rax		# Fetch y[1]
 
-ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	4f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(4)
-	popf')
 define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
 define(`T0', defn(`T1'))dnl
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc2.asm gmp-ecm-7.0.4+ds/x86_64/mulredc2.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc2.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc2.asm	2016-08-25 13:42:08.000000000 +0000
@@ -23,7 +23,7 @@
 include(`config.m4')
 
 	TEXT
-.p2align 6 # Opteron L1 code cache line is 64 bytes long
+.p2align 6 # x86_64 L1 code cache line is 64 bytes long
 	GLOBL GSYM_PREFIX`'mulredc2
 	TYPE(GSYM_PREFIX`'mulredc`'2,`function')
 
@@ -143,15 +143,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	2f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(2)
-	popf')
 define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
 define(`T0', defn(`T1'))dnl
@@ -215,15 +206,6 @@
 
 	movq	8(%YP), %rax		# Fetch y[1]
 
-ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	4f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(4)
-	popf')
 define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
 define(`T0', defn(`T1'))dnl
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc3.asm gmp-ecm-7.0.4+ds/x86_64/mulredc3.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc3.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc3.asm	2016-08-25 13:42:09.000000000 +0000
@@ -23,7 +23,7 @@
 include(`config.m4')
 
 	TEXT
-.p2align 6 # Opteron L1 code cache line is 64 bytes long
+.p2align 6 # x86_64 L1 code cache line is 64 bytes long
 	GLOBL GSYM_PREFIX`'mulredc3
 	TYPE(GSYM_PREFIX`'mulredc`'3,`function')
 
@@ -143,15 +143,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	2f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(2)
-	popf')
 define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
 define(`T0', defn(`T1'))dnl
@@ -179,12 +170,6 @@
 	movq	8(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -259,15 +244,6 @@
 
 	movq	8(%YP), %rax		# Fetch y[1]
 
-ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	4f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(4)
-	popf')
 define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
 define(`T0', defn(`T1'))dnl
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc4.asm gmp-ecm-7.0.4+ds/x86_64/mulredc4.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc4.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc4.asm	2016-08-25 13:42:09.000000000 +0000
@@ -23,7 +23,7 @@
 include(`config.m4')
 
 	TEXT
-.p2align 6 # Opteron L1 code cache line is 64 bytes long
+.p2align 6 # x86_64 L1 code cache line is 64 bytes long
 	GLOBL GSYM_PREFIX`'mulredc4
 	TYPE(GSYM_PREFIX`'mulredc`'4,`function')
 
@@ -143,15 +143,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	2f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(2)
-	popf')
 define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
 define(`T0', defn(`T1'))dnl
@@ -179,12 +170,6 @@
 	movq	8(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -223,12 +208,6 @@
 	movq	16(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -303,15 +282,6 @@
 
 	movq	8(%YP), %rax		# Fetch y[1]
 
-ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	4f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(4)
-	popf')
 define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
 define(`T0', defn(`T1'))dnl
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc5.asm gmp-ecm-7.0.4+ds/x86_64/mulredc5.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc5.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc5.asm	2016-08-25 13:42:09.000000000 +0000
@@ -23,7 +23,7 @@
 include(`config.m4')
 
 	TEXT
-.p2align 6 # Opteron L1 code cache line is 64 bytes long
+.p2align 6 # x86_64 L1 code cache line is 64 bytes long
 	GLOBL GSYM_PREFIX`'mulredc5
 	TYPE(GSYM_PREFIX`'mulredc`'5,`function')
 
@@ -143,15 +143,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	2f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(2)
-	popf')
 define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
 define(`T0', defn(`T1'))dnl
@@ -179,12 +170,6 @@
 	movq	8(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -223,12 +208,6 @@
 	movq	16(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -267,12 +246,6 @@
 	movq	24(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -347,15 +320,6 @@
 
 	movq	8(%YP), %rax		# Fetch y[1]
 
-ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	4f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(4)
-	popf')
 define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
 define(`T0', defn(`T1'))dnl
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc6.asm gmp-ecm-7.0.4+ds/x86_64/mulredc6.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc6.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc6.asm	2016-08-25 13:42:09.000000000 +0000
@@ -23,7 +23,7 @@
 include(`config.m4')
 
 	TEXT
-.p2align 6 # Opteron L1 code cache line is 64 bytes long
+.p2align 6 # x86_64 L1 code cache line is 64 bytes long
 	GLOBL GSYM_PREFIX`'mulredc6
 	TYPE(GSYM_PREFIX`'mulredc`'6,`function')
 
@@ -143,15 +143,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	2f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(2)
-	popf')
 define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
 define(`T0', defn(`T1'))dnl
@@ -179,12 +170,6 @@
 	movq	8(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -223,12 +208,6 @@
 	movq	16(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -267,12 +246,6 @@
 	movq	24(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -311,12 +284,6 @@
 	movq	32(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -391,15 +358,6 @@
 
 	movq	8(%YP), %rax		# Fetch y[1]
 
-ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	4f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(4)
-	popf')
 define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
 define(`T0', defn(`T1'))dnl
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc7.asm gmp-ecm-7.0.4+ds/x86_64/mulredc7.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc7.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc7.asm	2016-08-25 13:42:09.000000000 +0000
@@ -23,7 +23,7 @@
 include(`config.m4')
 
 	TEXT
-.p2align 6 # Opteron L1 code cache line is 64 bytes long
+.p2align 6 # x86_64 L1 code cache line is 64 bytes long
 	GLOBL GSYM_PREFIX`'mulredc7
 	TYPE(GSYM_PREFIX`'mulredc`'7,`function')
 
@@ -143,15 +143,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	2f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(2)
-	popf')
 define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
 define(`T0', defn(`T1'))dnl
@@ -179,12 +170,6 @@
 	movq	8(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -223,12 +208,6 @@
 	movq	16(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -267,12 +246,6 @@
 	movq	24(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -311,12 +284,6 @@
 	movq	32(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -355,12 +322,6 @@
 	movq	40(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -435,15 +396,6 @@
 
 	movq	8(%YP), %rax		# Fetch y[1]
 
-ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	4f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(4)
-	popf')
 define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
 define(`T0', defn(`T1'))dnl
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc8.asm gmp-ecm-7.0.4+ds/x86_64/mulredc8.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc8.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc8.asm	2016-08-25 13:42:09.000000000 +0000
@@ -23,7 +23,7 @@
 include(`config.m4')
 
 	TEXT
-.p2align 6 # Opteron L1 code cache line is 64 bytes long
+.p2align 6 # x86_64 L1 code cache line is 64 bytes long
 	GLOBL GSYM_PREFIX`'mulredc8
 	TYPE(GSYM_PREFIX`'mulredc`'8,`function')
 
@@ -143,15 +143,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	2f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(2)
-	popf')
 define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
 define(`T0', defn(`T1'))dnl
@@ -179,12 +170,6 @@
 	movq	8(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -223,12 +208,6 @@
 	movq	16(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -267,12 +246,6 @@
 	movq	24(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -311,12 +284,6 @@
 	movq	32(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -355,12 +322,6 @@
 	movq	40(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -399,12 +360,6 @@
 	movq	48(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -479,15 +434,6 @@
 
 	movq	8(%YP), %rax		# Fetch y[1]
 
-ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	4f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(4)
-	popf')
 define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
 define(`T0', defn(`T1'))dnl
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc9.asm gmp-ecm-7.0.4+ds/x86_64/mulredc9.asm
--- gmp-ecm-7.0.3+ds/x86_64/mulredc9.asm	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc9.asm	2016-08-25 13:42:10.000000000 +0000
@@ -23,7 +23,7 @@
 include(`config.m4')
 
 	TEXT
-.p2align 6 # Opteron L1 code cache line is 64 bytes long
+.p2align 6 # x86_64 L1 code cache line is 64 bytes long
 	GLOBL GSYM_PREFIX`'mulredc9
 	TYPE(GSYM_PREFIX`'mulredc`'9,`function')
 
@@ -143,15 +143,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	2f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(2)
-	popf')
 define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
 define(`T0', defn(`T1'))dnl
@@ -179,12 +170,6 @@
 	movq	8(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -223,12 +208,6 @@
 	movq	16(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -267,12 +246,6 @@
 	movq	24(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -311,12 +284,6 @@
 	movq	32(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -355,12 +322,6 @@
 	movq	40(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -399,12 +360,6 @@
 	movq	48(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -443,12 +398,6 @@
 	movq	56(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -523,15 +472,6 @@
 
 	movq	8(%YP), %rax		# Fetch y[1]
 
-ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	4f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(4)
-	popf')
 define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
 define(`T0', defn(`T1'))dnl
diff -Nru gmp-ecm-7.0.3+ds/x86_64/mulredc.m4 gmp-ecm-7.0.4+ds/x86_64/mulredc.m4
--- gmp-ecm-7.0.3+ds/x86_64/mulredc.m4	2016-07-04 09:34:49.000000000 +0000
+++ gmp-ecm-7.0.4+ds/x86_64/mulredc.m4	2016-08-25 13:42:00.000000000 +0000
@@ -30,7 +30,7 @@
 `include(`config.m4')'
 
 	TEXT
-.p2align 6 # Opteron L1 code cache line is 64 bytes long
+.p2align 6 # x86_64 L1 code cache line is 64 bytes long
 	GLOBL GSYM_PREFIX``''mulredc`'LENGTH
 	TYPE(GSYM_PREFIX``''mulredc``''LENGTH,``function'')
 
@@ -154,15 +154,6 @@
 	# CY:T1:T0 <= 2*(2^64-1)^2 <= 2^2*128 - 4*2^64 + 2, hence
 	# CY:T1 <= 2*2^64 - 4
 
-`ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	2f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(2)
-	popf')'
 dnl Cycle ring buffer. Only mappings of T0 and T1 to regs change, no MOVs!
 `define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl
@@ -195,12 +186,6 @@
 	movq	J`'(%MP), %rax	# Fetch m[j] into %rax
 	adcq	%rdx, %T1	# Add high word with carry to T1
 	# T1:T0 <= 2^128 - 2*2^64 + 1 + 2*2^64 - 2 <= 2^128 - 1, no carry!
-`ifdef(`WANT_ASSERT', 
-`	jnc	3f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(3)')'
 	
 	mulq	%U		# m[j]*u
 	# rdx:rax <= 2^128 - 2*2^64 + 1, T1:T0 <= 2^128 - 1
@@ -280,15 +265,6 @@
 
 	movq	8(%YP), %rax		# Fetch y[1]
 
-`ifdef(`WANT_ASSERT', 
-`	pushf
-	testq	%T0, %T0
-	jz	4f
-	lea     _GLOBAL_OFFSET_TABLE_(%rip), %rbx # if we do PIC code, we 
-		# need to set rbx; if not, it doesnt hurt
-	call	GSYM_PREFIX`'abort@plt
-LABEL_SUFFIX(4)
-	popf')'
 dnl Cycle ring buffer. Only mappings of T0 and T1 to regs change, no MOVs!
 `define(`TT', defn(`T0'))dnl
 define(`TTl', defn(`T0l'))dnl