add mullow_n_basecase common assembler code to the Windows build

build.vc14/cdata/mpn/x86_64w/sandybridge/cfg.h

@@ -34,6 +34,7 @@ mpn_modexact_1c_odd
 mpn_mul_1
 mpn_mul_2
 mpn_mul_basecase
+mpn_mullow_n_basecase
 mpn_mulmid_basecase
 mpn_nand_n
 mpn_nior_n

build.vc14/dll_mpir_sandybridge/dll_mpir_sandybridge.vcxproj

@@ -354,6 +354,10 @@ postbuild "$(TargetPath)" 14
     <ClCompile Include="..\..\mpz\lcm.c" />
     <ClCompile Include="..\..\mpz\lcm_ui.c" />
     <ClCompile Include="..\..\mpz\likely_prime_p.c" />
+    <ClCompile Include="..\..\mpz\limbs_finish.c" />
+    <ClCompile Include="..\..\mpz\limbs_modify.c" />
+    <ClCompile Include="..\..\mpz\limbs_read.c" />
+    <ClCompile Include="..\..\mpz\limbs_write.c" />
     <ClCompile Include="..\..\mpz\lucnum2_ui.c" />
     <ClCompile Include="..\..\mpz\lucnum_ui.c" />
     <ClCompile Include="..\..\mpz\mfac_uiui.c" />
@@ -385,6 +389,7 @@ postbuild "$(TargetPath)" 14
     <ClCompile Include="..\..\mpz\realloc.c" />
     <ClCompile Include="..\..\mpz\realloc2.c" />
     <ClCompile Include="..\..\mpz\remove.c" />
+    <ClCompile Include="..\..\mpz\roinit_n.c" />
     <ClCompile Include="..\..\mpz\root.c" />
     <ClCompile Include="..\..\mpz\rootrem.c" />
     <ClCompile Include="..\..\mpz\rrandomb.c" />
@@ -469,6 +474,7 @@ postbuild "$(TargetPath)" 14
     <ClCompile Include="..\..\mpn\generic\addadd_n.c" />
     <ClCompile Include="..\..\mpn\generic\addsub_n.c" />
     <ClCompile Include="..\..\mpn\generic\bdivmod.c" />
+    <ClCompile Include="..\..\mpn\generic\binvert.c" />
     <ClCompile Include="..\..\mpn\generic\cmp.c" />
     <ClCompile Include="..\..\mpn\generic\comb_tables.c" />
     <ClCompile Include="..\..\mpn\generic\dc_bdiv_q.c" />
@@ -527,7 +533,6 @@ postbuild "$(TargetPath)" 14
     <ClCompile Include="..\..\mpn\generic\mulhigh_n.c" />
     <ClCompile Include="..\..\mpn\generic\mullow_basecase.c" />
     <ClCompile Include="..\..\mpn\generic\mullow_n.c" />
-    <ClCompile Include="..\..\mpn\generic\mullow_n_basecase.c" />
     <ClCompile Include="..\..\mpn\generic\mulmid.c" />
     <ClCompile Include="..\..\mpn\generic\mulmid_n.c" />
     <ClCompile Include="..\..\mpn\generic\mulmod_2expm1.c" />
@@ -537,11 +542,14 @@ postbuild "$(TargetPath)" 14
     <ClCompile Include="..\..\mpn\generic\nsumdiff_n.c" />
     <ClCompile Include="..\..\mpn\generic\perfect_square_p.c" />
     <ClCompile Include="..\..\mpn\generic\pow_1.c" />
+    <ClCompile Include="..\..\mpn\generic\powlo.c" />
+    <ClCompile Include="..\..\mpn\generic\powm.c" />
     <ClCompile Include="..\..\mpn\generic\preinv_mod_1.c" />
     <ClCompile Include="..\..\mpn\generic\random.c" />
     <ClCompile Include="..\..\mpn\generic\random2.c" />
     <ClCompile Include="..\..\mpn\generic\randomb.c" />
     <ClCompile Include="..\..\mpn\generic\redc_2.c" />
+    <ClCompile Include="..\..\mpn\generic\redc_n.c" />
     <ClCompile Include="..\..\mpn\generic\rootrem.c" />
     <ClCompile Include="..\..\mpn\generic\rootrem_basecase.c" />
     <ClCompile Include="..\..\mpn\generic\rrandom.c" />
@@ -615,6 +623,7 @@ postbuild "$(TargetPath)" 14
     <YASM Include="..\..\mpn\x86_64w\sandybridge\mul_1.asm" />
     <YASM Include="..\..\mpn\x86_64w\sandybridge\mul_2.asm" />
     <YASM Include="..\..\mpn\x86_64w\sandybridge\mul_basecase.asm" />
+    <YASM Include="..\..\mpn\x86_64w\sandybridge\mullow_n_basecase.asm" />
     <YASM Include="..\..\mpn\x86_64w\mulmid_basecase.asm" />
     <YASM Include="..\..\mpn\x86_64w\sandybridge\nand_n.asm" />
     <YASM Include="..\..\mpn\x86_64w\sandybridge\nior_n.asm" />

build.vc14/dll_mpir_sandybridge/dll_mpir_sandybridge.vcxproj.filters

@@ -778,6 +778,18 @@
     <ClCompile Include="..\..\mpz\likely_prime_p.c">
       <Filter>Source Files\mpz</Filter>
     </ClCompile>
+    <ClCompile Include="..\..\mpz\limbs_finish.c">
+      <Filter>Source Files\mpz</Filter>
+    </ClCompile>
+    <ClCompile Include="..\..\mpz\limbs_modify.c">
+      <Filter>Source Files\mpz</Filter>
+    </ClCompile>
+    <ClCompile Include="..\..\mpz\limbs_read.c">
+      <Filter>Source Files\mpz</Filter>
+    </ClCompile>
+    <ClCompile Include="..\..\mpz\limbs_write.c">
+      <Filter>Source Files\mpz</Filter>
+    </ClCompile>
     <ClCompile Include="..\..\mpz\lucnum2_ui.c">
       <Filter>Source Files\mpz</Filter>
     </ClCompile>
@@ -871,6 +883,9 @@
     <ClCompile Include="..\..\mpz\remove.c">
       <Filter>Source Files\mpz</Filter>
     </ClCompile>
+    <ClCompile Include="..\..\mpz\roinit_n.c">
+      <Filter>Source Files\mpz</Filter>
+    </ClCompile>
     <ClCompile Include="..\..\mpz\root.c">
       <Filter>Source Files\mpz</Filter>
     </ClCompile>
@@ -1123,6 +1138,9 @@
     <ClCompile Include="..\..\mpn\generic\bdivmod.c">
       <Filter>Source Files\mpn</Filter>
     </ClCompile>
+    <ClCompile Include="..\..\mpn\generic\binvert.c">
+      <Filter>Source Files\mpn</Filter>
+    </ClCompile>
     <ClCompile Include="..\..\mpn\generic\cmp.c">
       <Filter>Source Files\mpn</Filter>
     </ClCompile>
@@ -1297,9 +1315,6 @@
     <ClCompile Include="..\..\mpn\generic\mullow_n.c">
       <Filter>Source Files\mpn</Filter>
     </ClCompile>
-    <ClCompile Include="..\..\mpn\generic\mullow_n_basecase.c">
-      <Filter>Source Files\mpn</Filter>
-    </ClCompile>
     <ClCompile Include="..\..\mpn\generic\mulmid.c">
       <Filter>Source Files\mpn</Filter>
     </ClCompile>
@@ -1327,6 +1342,12 @@
     <ClCompile Include="..\..\mpn\generic\pow_1.c">
       <Filter>Source Files\mpn</Filter>
     </ClCompile>
+    <ClCompile Include="..\..\mpn\generic\powlo.c">
+      <Filter>Source Files\mpn</Filter>
+    </ClCompile>
+    <ClCompile Include="..\..\mpn\generic\powm.c">
+      <Filter>Source Files\mpn</Filter>
+    </ClCompile>
     <ClCompile Include="..\..\mpn\generic\preinv_mod_1.c">
       <Filter>Source Files\mpn</Filter>
     </ClCompile>
@@ -1342,6 +1363,9 @@
     <ClCompile Include="..\..\mpn\generic\redc_2.c">
       <Filter>Source Files\mpn</Filter>
     </ClCompile>
+    <ClCompile Include="..\..\mpn\generic\redc_n.c">
+      <Filter>Source Files\mpn</Filter>
+    </ClCompile>
     <ClCompile Include="..\..\mpn\generic\rootrem.c">
       <Filter>Source Files\mpn</Filter>
     </ClCompile>
@@ -1557,6 +1581,9 @@
     <YASM Include="..\..\mpn\x86_64w\sandybridge\mul_basecase.asm">
       <Filter>Source Files\mpn\yasm</Filter>
     </YASM>
+    <YASM Include="..\..\mpn\x86_64w\sandybridge\mullow_n_basecase.asm">
+      <Filter>Source Files\mpn\yasm</Filter>
+    </YASM>
     <YASM Include="..\..\mpn\x86_64w\mulmid_basecase.asm">
       <Filter>Source Files\mpn\yasm</Filter>
     </YASM>

build.vc14/lib_mpir_sandybridge/lib_mpir_sandybridge.vcxproj

@@ -350,6 +350,10 @@ postbuild "$(TargetPath)" 14
     <ClCompile Include="..\..\mpz\lcm.c" />
     <ClCompile Include="..\..\mpz\lcm_ui.c" />
     <ClCompile Include="..\..\mpz\likely_prime_p.c" />
+    <ClCompile Include="..\..\mpz\limbs_finish.c" />
+    <ClCompile Include="..\..\mpz\limbs_modify.c" />
+    <ClCompile Include="..\..\mpz\limbs_read.c" />
+    <ClCompile Include="..\..\mpz\limbs_write.c" />
     <ClCompile Include="..\..\mpz\lucnum2_ui.c" />
     <ClCompile Include="..\..\mpz\lucnum_ui.c" />
     <ClCompile Include="..\..\mpz\mfac_uiui.c" />
@@ -381,6 +385,7 @@ postbuild "$(TargetPath)" 14
     <ClCompile Include="..\..\mpz\realloc.c" />
     <ClCompile Include="..\..\mpz\realloc2.c" />
     <ClCompile Include="..\..\mpz\remove.c" />
+    <ClCompile Include="..\..\mpz\roinit_n.c" />
     <ClCompile Include="..\..\mpz\root.c" />
     <ClCompile Include="..\..\mpz\rootrem.c" />
     <ClCompile Include="..\..\mpz\rrandomb.c" />
@@ -454,6 +459,7 @@ postbuild "$(TargetPath)" 14
     <ClCompile Include="..\..\mpn\generic\addadd_n.c" />
     <ClCompile Include="..\..\mpn\generic\addsub_n.c" />
     <ClCompile Include="..\..\mpn\generic\bdivmod.c" />
+    <ClCompile Include="..\..\mpn\generic\binvert.c" />
     <ClCompile Include="..\..\mpn\generic\cmp.c" />
     <ClCompile Include="..\..\mpn\generic\comb_tables.c" />
     <ClCompile Include="..\..\mpn\generic\dc_bdiv_q.c" />
@@ -512,7 +518,6 @@ postbuild "$(TargetPath)" 14
     <ClCompile Include="..\..\mpn\generic\mulhigh_n.c" />
     <ClCompile Include="..\..\mpn\generic\mullow_basecase.c" />
     <ClCompile Include="..\..\mpn\generic\mullow_n.c" />
-    <ClCompile Include="..\..\mpn\generic\mullow_n_basecase.c" />
     <ClCompile Include="..\..\mpn\generic\mulmid.c" />
     <ClCompile Include="..\..\mpn\generic\mulmid_n.c" />
     <ClCompile Include="..\..\mpn\generic\mulmod_2expm1.c" />
@@ -522,11 +527,14 @@ postbuild "$(TargetPath)" 14
     <ClCompile Include="..\..\mpn\generic\nsumdiff_n.c" />
     <ClCompile Include="..\..\mpn\generic\perfect_square_p.c" />
     <ClCompile Include="..\..\mpn\generic\pow_1.c" />
+    <ClCompile Include="..\..\mpn\generic\powlo.c" />
+    <ClCompile Include="..\..\mpn\generic\powm.c" />
     <ClCompile Include="..\..\mpn\generic\preinv_mod_1.c" />
     <ClCompile Include="..\..\mpn\generic\random.c" />
     <ClCompile Include="..\..\mpn\generic\random2.c" />
     <ClCompile Include="..\..\mpn\generic\randomb.c" />
     <ClCompile Include="..\..\mpn\generic\redc_2.c" />
+    <ClCompile Include="..\..\mpn\generic\redc_n.c" />
     <ClCompile Include="..\..\mpn\generic\rootrem.c" />
     <ClCompile Include="..\..\mpn\generic\rootrem_basecase.c" />
     <ClCompile Include="..\..\mpn\generic\rrandom.c" />
@@ -600,6 +608,7 @@ postbuild "$(TargetPath)" 14
     <YASM Include="..\..\mpn\x86_64w\sandybridge\mul_1.asm" />
     <YASM Include="..\..\mpn\x86_64w\sandybridge\mul_2.asm" />
     <YASM Include="..\..\mpn\x86_64w\sandybridge\mul_basecase.asm" />
+    <YASM Include="..\..\mpn\x86_64w\sandybridge\mullow_n_basecase.asm" />
     <YASM Include="..\..\mpn\x86_64w\mulmid_basecase.asm" />
     <YASM Include="..\..\mpn\x86_64w\sandybridge\nand_n.asm" />
     <YASM Include="..\..\mpn\x86_64w\sandybridge\nior_n.asm" />

build.vc14/lib_mpir_sandybridge/lib_mpir_sandybridge.vcxproj.filters

@@ -777,6 +777,18 @@
     <ClCompile Include="..\..\mpz\likely_prime_p.c">
       <Filter>Source Files\mpz</Filter>
     </ClCompile>
+    <ClCompile Include="..\..\mpz\limbs_finish.c">
+      <Filter>Source Files\mpz</Filter>
+    </ClCompile>
+    <ClCompile Include="..\..\mpz\limbs_modify.c">
+      <Filter>Source Files\mpz</Filter>
+    </ClCompile>
+    <ClCompile Include="..\..\mpz\limbs_read.c">
+      <Filter>Source Files\mpz</Filter>
+    </ClCompile>
+    <ClCompile Include="..\..\mpz\limbs_write.c">
+      <Filter>Source Files\mpz</Filter>
+    </ClCompile>
     <ClCompile Include="..\..\mpz\lucnum2_ui.c">
       <Filter>Source Files\mpz</Filter>
     </ClCompile>
@@ -870,6 +882,9 @@
     <ClCompile Include="..\..\mpz\remove.c">
       <Filter>Source Files\mpz</Filter>
     </ClCompile>
+    <ClCompile Include="..\..\mpz\roinit_n.c">
+      <Filter>Source Files\mpz</Filter>
+    </ClCompile>
     <ClCompile Include="..\..\mpz\root.c">
       <Filter>Source Files\mpz</Filter>
     </ClCompile>
@@ -1089,6 +1104,9 @@
     <ClCompile Include="..\..\mpn\generic\bdivmod.c">
       <Filter>Source Files\mpn</Filter>
     </ClCompile>
+    <ClCompile Include="..\..\mpn\generic\binvert.c">
+      <Filter>Source Files\mpn</Filter>
+    </ClCompile>
     <ClCompile Include="..\..\mpn\generic\cmp.c">
       <Filter>Source Files\mpn</Filter>
     </ClCompile>
@@ -1263,9 +1281,6 @@
     <ClCompile Include="..\..\mpn\generic\mullow_n.c">
       <Filter>Source Files\mpn</Filter>
     </ClCompile>
-    <ClCompile Include="..\..\mpn\generic\mullow_n_basecase.c">
-      <Filter>Source Files\mpn</Filter>
-    </ClCompile>
     <ClCompile Include="..\..\mpn\generic\mulmid.c">
       <Filter>Source Files\mpn</Filter>
     </ClCompile>
@@ -1293,6 +1308,12 @@
     <ClCompile Include="..\..\mpn\generic\pow_1.c">
       <Filter>Source Files\mpn</Filter>
     </ClCompile>
+    <ClCompile Include="..\..\mpn\generic\powlo.c">
+      <Filter>Source Files\mpn</Filter>
+    </ClCompile>
+    <ClCompile Include="..\..\mpn\generic\powm.c">
+      <Filter>Source Files\mpn</Filter>
+    </ClCompile>
     <ClCompile Include="..\..\mpn\generic\preinv_mod_1.c">
       <Filter>Source Files\mpn</Filter>
     </ClCompile>
@@ -1308,6 +1329,9 @@
     <ClCompile Include="..\..\mpn\generic\redc_2.c">
       <Filter>Source Files\mpn</Filter>
     </ClCompile>
+    <ClCompile Include="..\..\mpn\generic\redc_n.c">
+      <Filter>Source Files\mpn</Filter>
+    </ClCompile>
     <ClCompile Include="..\..\mpn\generic\rootrem.c">
       <Filter>Source Files\mpn</Filter>
     </ClCompile>
@@ -1523,6 +1547,9 @@
     <YASM Include="..\..\mpn\x86_64w\sandybridge\mul_basecase.asm">
       <Filter>Source Files\mpn\yasm</Filter>
     </YASM>
+    <YASM Include="..\..\mpn\x86_64w\sandybridge\mullow_n_basecase.asm">
+      <Filter>Source Files\mpn\yasm</Filter>
+    </YASM>
     <YASM Include="..\..\mpn\x86_64w\mulmid_basecase.asm">
       <Filter>Source Files\mpn\yasm</Filter>
     </YASM>

mpn/x86_64w/mullow_n_basecase.asm

@@ -0,0 +1,336 @@
+;  AMD64 mpn_mullow_n_basecase
+;
+;  Copyright 2015 Free Software Foundation, Inc.
+;
+;  This file is part of the GNU MP Library.
+;
+;  The GNU MP Library is free software; you can redistribute it and/or modify
+;  it under the terms of either:
+;
+;    * the GNU Lesser General Public License as published by the Free
+;      Software Foundation; either version 3 of the License, or (at your
+;      option) any later version.
+;
+;  or
+;
+;    * the GNU General Public License as published by the Free Software
+;      Foundation; either version 2 of the License, or (at your option) any
+;      later version.
+;
+;  or both in parallel, as here.
+;
+;  The GNU MP Library is distributed in the hope that it will be useful, but
+;  WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
+;  or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
+;  for more details.
+;
+;  You should have received copies of the GNU General Public License and the
+;  GNU Lesser General Public License along with the GNU MP Library.  If not,
+;  see https://www.gnu.org/licenses/.
+;
+;void mpn_mullow_n_basecase (mp_ptr rp, mp_srcptr up, mp_srcptr vp, mp_size_t n)
+;                                  rdi           rsi           rdx          rcx
+;                                  rcx           rdx            r8           r9
+
+%include 'yasm_mac.inc'
+
+	TEXT
+
+	LEAF_PROC mpn_mullow_n_basecase
+    cmp     r9, 3
+    je      asm_sym(?mpn_mullow1)
+    jg      asm_sym(?mpn_mullow2)
+	mov     rax, [rdx]
+    mov     r11, [r8+8]
+	mov     r8, [r8]
+    cmp     r9, 1
+    jg      .3
+	
+.2:	imul    rax, r8
+	mov     [rcx], rax
+	ret
+	
+.3:	mov     r9, [rdx+8]
+    imul    r11, rax
+	mul     r8
+	mov     [rcx], rax
+	imul    r8, r9
+	lea     rax, [r11+rdx]
+	add     rax, r8
+	mov     [rcx+8], rax
+    ret
+
+%define reg_save_list rsi, rdi
+	
+	FRAME_PROC ?mpn_mullow1, 0, reg_save_list
+    mov     rdi, rcx
+    mov     rsi, rdx
+    mov     rdx, r8
+    mov     rcx, r9
+    mov	    rax, [rsi]
+	mov	    r8, [rdx]
+    mov     r9, [rdx+8]
+	mov     r11, [rdx+16]
+	mul     r8
+	mov     [rdi], rax
+	mov     rax, [rsi]
+	mov     rcx, rdx
+	mul     r9
+	imul    r9, [rsi+8]
+	mov     r10, [rsi+16]
+	imul    r10, r8
+	add     rcx, rax
+	adc     r9, rdx
+	add     r9, r10
+	mov     rax, [rsi+8]
+	mul     r8
+	add     rcx, rax
+	adc     r9, rdx
+	mov     rax, r11
+	imul    rax, [rsi]
+	add     r9, rax
+	mov     [rdi+8], rcx
+	mov     [rdi+16], r9
+	END_PROC reg_save_list
+
+%define reg_save_list rsi, rdi, rbx, rbp, r13, r14, r15
+
+	align  16
+	FRAME_PROC ?mpn_mullow2, 0, reg_save_list
+    mov     rdi, rcx
+    mov     rsi, rdx
+    mov     rdx, r8
+    mov     rcx, r9
+.5:	mov     rax, [rsi]
+	mov     r13, [rdx]
+	mov     r11, rdx
+	lea     rdi, [rdi+rcx*8]
+	lea     rsi, [rsi+rcx*8]
+	neg     rcx
+	mul     r13
+	test    cl, 1
+	jz      .12
+.6: lea     rdi, [rdi-8]
+	lea     rsi, [rsi-8]
+	test    cl, 2
+	jnz     .8
+.7: lea     r9, [rcx-1]
+	lea     r8, [rel .18]
+	mov     rbx, rax
+	mov     r15, rdx
+	xor     ebp, ebp
+	xor     r10d, r10d
+	mov     rax, [rsi+rcx*8+16]
+	jmp     .11
+
+.8: lea     r9, [rcx+1]
+	lea     r8, [rel .19]
+	mov     rbp, rax
+	mov     r10, rdx
+	xor     ebx, ebx
+	jmp     .10
+
+	align  16
+.9: mov     [rdi+r9*8-16], rbx
+	add     r15, rax
+	mov     rax, [rsi+r9*8]
+	adc     rbp, rdx
+	xor     ebx, ebx
+	mul     r13
+	mov     [rdi+r9*8-8], r15
+	add     rbp, rax
+	adc     r10, rdx
+.10:mov     rax, [rsi+r9*8+8]
+	mul     r13
+	mov     [rdi+r9*8], rbp
+	add     r10, rax
+	adc     rbx, rdx
+	mov     rax, [rsi+r9*8+16]
+	mul     r13
+	mov     [rdi+r9*8+8], r10
+	xor     ebp, ebp
+	mov     r10, rbp
+	add     rbx, rax
+	mov     rax, [rsi+r9*8+24]
+	mov     r15, rbp
+	adc     r15, rdx
+.11:mul     r13
+	add     r9, 4
+	js      .9
+	mov     [rdi-16], rbx
+	add     r15, rax
+	mov     [rdi-8], r15
+	adc     rbp, rdx
+	imul    r13, [rsi]
+	add     rbp, r13
+	mov     [rdi], rbp
+	add     rcx, 1
+	jz      .23
+	mov     r13, [r11+8]
+	mov     r14, [r11+16]
+	lea     rsi, [rsi+16]
+	lea     r11, [r11+8]
+	lea     rdi, [rdi+24]
+	jmp     r8
+
+.12:mov     r14, [r11+8]
+	test    cl, 2
+	jz      .14
+
+	align  16
+.13:lea     r9, [rcx+0]
+	mov     r10, rax
+	mov     rbx, rdx
+	xor     r15d, r15d
+	mov     rax, [rsi+rcx*8]
+	lea     r8, [rel .19]
+	jmp     .17
+	
+	align  16
+.14:lea     r9, [rcx+2]
+	mov     r10d, 0
+	mov     r15, rax
+	mov     rax, [rsi+rcx*8]
+	mov     rbp, rdx
+	lea     r8, [rel .18]
+	jmp     .16
+	
+	align  16
+.15:mov     rax, [rsi+r9*8-32]
+	mul     r14
+	add     rbx, rax
+	adc     r15, rdx
+	mov     rax, [rsi+r9*8-24]
+	xor     ebp, ebp
+	mul     r13
+	add     rbx, rax
+	mov     rax, [rsi+r9*8-24]
+	adc     r15, rdx
+	adc     ebp, 0
+	mul     r14
+	add     r15, rax
+	mov     [rdi+r9*8-24], rbx
+	adc     rbp, rdx
+	mov     rax, [rsi+r9*8-16]
+	mul     r13
+	mov     r10d, 0
+	add     r15, rax
+	adc     rbp, rdx
+	mov     rax, [rsi+r9*8-16]
+	adc     r10d, 0
+.16:mov     ebx, 0
+	mov     [rdi+r9*8-16], r15
+	mul     r14
+	add     rbp, rax
+	mov     rax, [rsi+r9*8-8]
+	adc     r10, rdx
+	mov     r15d, 0
+	mul     r13
+	add     rbp, rax
+	mov     rax, [rsi+r9*8-8]
+	adc     r10, rdx
+	adc     ebx, r15d
+	mul     r14
+	add     r10, rax
+	mov     [rdi+r9*8-8], rbp
+	adc     rbx, rdx
+	mov     rax, [rsi+r9*8]
+	mul     r13
+	add     r10, rax
+	adc     rbx, rdx
+	adc     r15d, 0
+.17:add     r9, 4
+	mov     [rdi+r9*8-32], r10
+	js      .15
+	imul    r14, [rsi-16]
+	add     rbx, r14
+	imul    r13, [rsi-8]
+	add     rbx, r13
+	mov     [rdi-8], rbx
+	add     rcx, 2
+	jz      .23
+	mov     r13, [r11+16]
+	mov     r14, [r11+24]
+	lea     r11, [r11+16]
+	lea     rdi, [rdi+16]
+	jmp     r8
+.18:lea     r9, [rcx-2]
+	mov     rax, [rsi+rcx*8-16]
+	mul     r13
+	mov     r10, rax
+	mov     rax, [rsi+rcx*8-16]
+	mov     rbx, rdx
+	xor     r15d, r15d
+	lea     r8, [rel .19]
+	jmp     .22
+.19:lea     r9, [rcx+0]
+	mov     rax, [rsi+rcx*8-16]
+	xor     r10d, r10d
+	mul     r13
+	mov     r15, rax
+	mov     rax, [rsi+rcx*8-16]
+	mov     rbp, rdx
+	lea     r8, [rel .18]
+	jmp     .21
+	
+	align  16
+.20:add     [rdi+r9*8-32], r10
+	adc     rbx, rax
+	mov     rax, [rsi+r9*8-24]
+	adc     r15, rdx
+	xor     ebp, ebp
+	mul     r13
+	add     rbx, rax
+	mov     rax, [rsi+r9*8-24]
+	adc     r15, rdx
+	adc     ebp, ebp
+	mul     r14
+	xor     r10d, r10d
+	add     [rdi+r9*8-24], rbx
+	adc     r15, rax
+	mov     rax, [rsi+r9*8-16]
+	adc     rbp, rdx
+	mul     r13
+	add     r15, rax
+	mov     rax, [rsi+r9*8-16]
+	adc     rbp, rdx
+	adc     r10d, 0
+.21:mul     r14
+	add     [rdi+r9*8-16], r15
+	adc     rbp, rax
+	mov     rax, [rsi+r9*8-8]
+	adc     r10, rdx
+	mul     r13
+	xor     ebx, ebx
+	add     rbp, rax
+	adc     r10, rdx
+	mov     r15d, 0
+	mov     rax, [rsi+r9*8-8]
+	adc     ebx, r15d
+	mul     r14
+	add     [rdi+r9*8-8], rbp
+	adc     r10, rax
+	adc     rbx, rdx
+	mov     rax, [rsi+r9*8]
+	mul     r13
+	add     r10, rax
+	mov     rax, [rsi+r9*8]
+	adc     rbx, rdx
+	adc     r15d, 0
+.22:mul     r14
+	add     r9, 4
+	js      .20
+	add     [rdi-32], r10
+	adc     rbx, rax
+	imul    r13, [rsi-24]
+	add     rbx, r13
+	add     [rdi-24], rbx
+	add     rcx, 2
+	jns     .23
+	lea     r11, [r11+16]
+	mov     r13, [r11]
+	mov     r14, [r11+8]
+	lea     rsi, [rsi-16]
+	jmp     r8
+.23:
+    END_PROC reg_save_list

mpn/x86_64w/yasm_mac.inc

@@ -63,6 +63,8 @@
 %define r6b bpl
 %define r7b spl
 
+%define asm_sym(x) __g %+ x
+
 ; Standard macro for alignment (used to allow easy subsititution of
 ; alternative padding schemes)
 
@@ -105,13 +107,13 @@
 
 %macro FRAME_PROC 2-*
 
-    global  __g%1
+    global  asm_sym(%1)
     
 %ifdef DLL
-    export  __g%1
+    export  asm_sym(%1)
 %endif
 
-    PROC_FRAME __g%1
+    PROC_FRAME asm_sym(%1)
     %rotate 1
       
     %if %1 < 0
@@ -216,13 +218,13 @@
 
 %macro LEAF_PROC 1
     
-    global  __g%1
+    global  asm_sym(%1)
     
 %ifdef DLL
-    export  __g%1
+    export  asm_sym(%1)
 %endif
 
-__g%1:
+asm_sym(%1):
 
 %endmacro