Menu
▾
▴
ggnfs-commits
|
From: <brg...@us...> - 2011-10-22 19:20:47
|
Revision: 418
http://ggnfs.svn.sourceforge.net/ggnfs/?rev=418&view=rev
Author: brgladman
Date: 2011-10-22 19:20:40 +0000 (Sat, 22 Oct 2011)
Log Message:
-----------
update Windows x64 assembler code
Modified Paths:
--------------
trunk/src/experimental/lasieve4_64/x64/MMX-TD1.asm
trunk/src/experimental/lasieve4_64/x64/MMX-TD2.asm
trunk/src/experimental/lasieve4_64/x64/lasched0.asm
trunk/src/experimental/lasieve4_64/x64/lasched1.asm
trunk/src/experimental/lasieve4_64/x64/lasieve-asmprepn.asm
trunk/src/experimental/lasieve4_64/x64/lasieve-asmprepn2.asm
trunk/src/experimental/lasieve4_64/x64/medsched0.asm
trunk/src/experimental/lasieve4_64/x64/modinv1000.asm
trunk/src/experimental/lasieve4_64/x64/modinv1000b.asm
trunk/src/experimental/lasieve4_64/x64/mpqs_sieveinit.asm
trunk/src/experimental/lasieve4_64/x64/mpqs_td.asm
trunk/src/experimental/lasieve4_64/x64/pt64.asm
trunk/src/experimental/lasieve4_64/x64/search0.asm
trunk/src/experimental/lasieve4_64/x64/slinie1A.asm
trunk/src/experimental/lasieve4_64/x64/slinie2A.asm
trunk/src/experimental/lasieve4_64/x64/slinie3A.asm
trunk/src/experimental/lasieve4_64/x64/tdsieve-from-sched.asm
Modified: trunk/src/experimental/lasieve4_64/x64/MMX-TD1.asm
===================================================================
--- trunk/src/experimental/lasieve4_64/x64/MMX-TD1.asm 2011-10-22 07:34:27 UTC (rev 417)
+++ trunk/src/experimental/lasieve4_64/x64/MMX-TD1.asm 2011-10-22 19:20:40 UTC (rev 418)
@@ -14,7 +14,7 @@
mov rdi, rcx
mov esi, edx
mov rdx, r8
- mov ecx, r9d
+ mov rcx, r9
mov r8d, esi
movq mm2, [rdx]
Modified: trunk/src/experimental/lasieve4_64/x64/MMX-TD2.asm
===================================================================
--- trunk/src/experimental/lasieve4_64/x64/MMX-TD2.asm 2011-10-22 07:34:27 UTC (rev 417)
+++ trunk/src/experimental/lasieve4_64/x64/MMX-TD2.asm 2011-10-22 19:20:40 UTC (rev 418)
@@ -14,7 +14,7 @@
mov rdi, rcx
mov esi, edx
mov rdx, r8
- mov ecx, r9d
+ mov rcx, r9
mov r8d, esi
movdqa xmm2, [rdx]
Modified: trunk/src/experimental/lasieve4_64/x64/lasched0.asm
===================================================================
--- trunk/src/experimental/lasieve4_64/x64/lasched0.asm 2011-10-22 07:34:27 UTC (rev 417)
+++ trunk/src/experimental/lasieve4_64/x64/lasched0.asm 2011-10-22 19:20:40 UTC (rev 418)
@@ -23,7 +23,7 @@
mov rdx, r8
mov ecx, r9d
mov r8, [rsp+stack_use+40]
- mov r9d, dword [rsp+stack_use+48]
+ mov r9d, [rsp+stack_use+48]
shl rcx, n_i_bits
mov eax, [rsi]
Modified: trunk/src/experimental/lasieve4_64/x64/lasched1.asm
===================================================================
--- trunk/src/experimental/lasieve4_64/x64/lasched1.asm 2011-10-22 07:34:27 UTC (rev 417)
+++ trunk/src/experimental/lasieve4_64/x64/lasched1.asm 2011-10-22 19:20:40 UTC (rev 418)
@@ -23,7 +23,7 @@
mov rdx, r8
mov ecx, r9d
mov r8, [rsp+stack_use+40]
- mov r9d, dword [rsp+stack_use+48]
+ mov r9d, [rsp+stack_use+48]
shl rcx, n_i_bits
shr r9, 16
Modified: trunk/src/experimental/lasieve4_64/x64/lasieve-asmprepn.asm
===================================================================
--- trunk/src/experimental/lasieve4_64/x64/lasieve-asmprepn.asm 2011-10-22 07:34:27 UTC (rev 417)
+++ trunk/src/experimental/lasieve4_64/x64/lasieve-asmprepn.asm 2011-10-22 19:20:40 UTC (rev 418)
@@ -178,13 +178,17 @@
psrldq xmm7, 8
mov esi, ebx
pmuludq xmm12, xmm2
+ add rsp, 32
call asm_modinv32b
+ sub rsp, 32
movd edi, xmm7
pmuludq xmm12, xmm0
movd xmm11, eax
mov esi, r15d
pxor xmm10, xmm10
+ add rsp, 32
call asm_modinv32b
+ sub rsp, 32
movd xmm8, eax
psubq xmm9, xmm12
mov rax, [r12]
@@ -243,13 +247,17 @@
psrldq xmm7, 8
movdqa xmm2, xmm11
pmuludq xmm11, xmm8
+ add rsp, 32
call get_recurrence_info
+ sub rsp, 32
mov esi, r15d
lea rdi, [r13+8]
movd edx, xmm7
pmuludq xmm11, xmm2
movdqa xmm0, xmm8
+ add rsp, 32
call get_recurrence_info
+ sub rsp, 32
cmp r14, r12
pslld xmm2, 1
pshufd xmm12, xmm1, 0x98
Modified: trunk/src/experimental/lasieve4_64/x64/lasieve-asmprepn2.asm
===================================================================
--- trunk/src/experimental/lasieve4_64/x64/lasieve-asmprepn2.asm 2011-10-22 07:34:27 UTC (rev 417)
+++ trunk/src/experimental/lasieve4_64/x64/lasieve-asmprepn2.asm 2011-10-22 19:20:40 UTC (rev 418)
@@ -191,7 +191,9 @@
pshufd xmm8, xmm8, 0x4e
mov esi, ebx
pmuludq xmm12, xmm2
+ add rsp, 32
call asm_modinv32b
+ sub rsp, 32
%if %1 * (%1 - 3) == 0
sub r15d, eax
%endif
@@ -256,13 +258,17 @@
psrldq xmm7, 8
movdqa xmm2, xmm11
pmuludq xmm11, xmm8
+ add rsp, 32
call get_recurrence_info
+ sub rsp, 32
mov esi, ebx
lea rdi, [r13+8]
movd edx, xmm7
pmuludq xmm11, xmm2
movdqa xmm0, xmm8
+ add rsp, 32
call get_recurrence_info
+ sub rsp, 32
cmp r14, r12
pslld xmm2, 1
pshufd xmm12, xmm1, 0x98
Modified: trunk/src/experimental/lasieve4_64/x64/medsched0.asm
===================================================================
--- trunk/src/experimental/lasieve4_64/x64/medsched0.asm 2011-10-22 07:34:27 UTC (rev 417)
+++ trunk/src/experimental/lasieve4_64/x64/medsched0.asm 2011-10-22 19:20:40 UTC (rev 418)
@@ -15,7 +15,7 @@
mov rsi, rdx
mov rdx, r8
mov rcx, r9
- mov r8d, dword [rsp+stack_use+40]
+ mov r8d, [rsp+stack_use+40]
shl r8d, 16
cmp rdx, rsi
Modified: trunk/src/experimental/lasieve4_64/x64/modinv1000.asm
===================================================================
--- trunk/src/experimental/lasieve4_64/x64/modinv1000.asm 2011-10-22 07:34:27 UTC (rev 417)
+++ trunk/src/experimental/lasieve4_64/x64/modinv1000.asm 2011-10-22 19:20:40 UTC (rev 418)
@@ -75,7 +75,9 @@
mov eax, [rel modulo32]
sub eax, r8d
ret
-.6: call abort
+.6: add rsp, 32
+ call abort
+ sub rsp, 32
end
\ No newline at end of file
Modified: trunk/src/experimental/lasieve4_64/x64/modinv1000b.asm
===================================================================
--- trunk/src/experimental/lasieve4_64/x64/modinv1000b.asm 2011-10-22 07:34:27 UTC (rev 417)
+++ trunk/src/experimental/lasieve4_64/x64/modinv1000b.asm 2011-10-22 19:20:40 UTC (rev 418)
@@ -80,7 +80,9 @@
sub eax, r8d
ret
.6: je .8
-.7: call abort
+.7: add rsp, 32
+ call abort
+ sub rsp, 32
.8: xor eax, eax
ret
Modified: trunk/src/experimental/lasieve4_64/x64/mpqs_sieveinit.asm
===================================================================
--- trunk/src/experimental/lasieve4_64/x64/mpqs_sieveinit.asm 2011-10-22 07:34:27 UTC (rev 417)
+++ trunk/src/experimental/lasieve4_64/x64/mpqs_sieveinit.asm 2011-10-22 19:20:40 UTC (rev 418)
@@ -25,8 +25,8 @@
mov esi, edx
mov rdx, r8
mov rcx, r9
- mov r8, [rsp+stack_use+40]
- mov r9d, [rsp+stack_use+48]
+ mov r8, [rsp+stack_use+40]
+ mov r9d, [rsp+stack_use+48]
mov rbx, rdx
mov rcx, [rcx]
@@ -70,8 +70,8 @@
mov esi, edx
mov rdx, r8
mov rcx, r9
- mov r8, [rsp+stack_use+40]
- mov r9d, [rsp+stack_use+48]
+ mov r8, [rsp+stack_use+40]
+ mov r9d, [rsp+stack_use+48]
mov rbx, rdx
mov rcx, [rcx]
Modified: trunk/src/experimental/lasieve4_64/x64/mpqs_td.asm
===================================================================
--- trunk/src/experimental/lasieve4_64/x64/mpqs_td.asm 2011-10-22 07:34:27 UTC (rev 417)
+++ trunk/src/experimental/lasieve4_64/x64/mpqs_td.asm 2011-10-22 19:20:40 UTC (rev 418)
@@ -278,7 +278,9 @@
.16: xor rax, rax
emms
EXIT_PROC reg_save_list
-.17: call abort
+.17: add rsp, 32
+ call abort
+ sub rsp, 32
END_PROC reg_save_list
end
Modified: trunk/src/experimental/lasieve4_64/x64/pt64.asm
===================================================================
--- trunk/src/experimental/lasieve4_64/x64/pt64.asm 2011-10-22 07:34:27 UTC (rev 417)
+++ trunk/src/experimental/lasieve4_64/x64/pt64.asm 2011-10-22 19:20:40 UTC (rev 418)
@@ -60,7 +60,9 @@
mov rcx, rax
mov r11, rsi
jmp .5
-.4: call modsq64
+.4: add rsp,32
+ call modsq64
+ sub rsp, 32
test r9, rcx
jz .6
.5: xor rax, rax
@@ -79,7 +81,9 @@
je .9
dec r10
jz .8
+ add rsp, 32
call modsq64
+ sub rsp, 32
jmp .7
align 8
Modified: trunk/src/experimental/lasieve4_64/x64/search0.asm
===================================================================
--- trunk/src/experimental/lasieve4_64/x64/search0.asm 2011-10-22 07:34:27 UTC (rev 417)
+++ trunk/src/experimental/lasieve4_64/x64/search0.asm 2011-10-22 19:20:40 UTC (rev 418)
@@ -1,6 +1,6 @@
-; lasieve_search0(uchar*, uchar*, uchar*, uchar*, uchar*, u16_t*, uchar*);
-; GAS rax rdi, rsi, rdx, rcx, r8, r9, [rsp+8]
+; lasieve_search0(uchar*, uchar*, uchar*, uchar*, uchar*, u16_t*, uchar*);
+; GAS rax rdi, rsi, rdx, rcx, r8, r9, [rsp+8]
; YASM rax rcx, rdx, r8, r9, [rsp+40], [rsp+48], [rsp+56],
%include "ls-defs.inc"
@@ -15,9 +15,9 @@
mov rsi, rdx
mov rdx, r8
mov rcx, r9
- mov r8, qword [rsp+stack_use+40]
- mov r9, qword [rsp+stack_use+48]
- mov r14, qword [rsp+stack_use+56]
+ mov r8, [rsp+stack_use+40]
+ mov r9, [rsp+stack_use+48]
+ mov r14, [rsp+stack_use+56]
mov r10, rdi
lea r15, [rdi+128]
Modified: trunk/src/experimental/lasieve4_64/x64/slinie1A.asm
===================================================================
--- trunk/src/experimental/lasieve4_64/x64/slinie1A.asm 2011-10-22 07:34:27 UTC (rev 417)
+++ trunk/src/experimental/lasieve4_64/x64/slinie1A.asm 2011-10-22 19:20:40 UTC (rev 418)
@@ -77,6 +77,10 @@
%define j_per_strip_minus1 1
FRAME_PROC slinie1, 0, reg_save_list
+ mov rdi, rcx
+ mov rsi, rdx
+ mov rdx, r8
+
cmp rsi, rdi
jbe .3
sub rsi, 8
Modified: trunk/src/experimental/lasieve4_64/x64/slinie2A.asm
===================================================================
--- trunk/src/experimental/lasieve4_64/x64/slinie2A.asm 2011-10-22 07:34:27 UTC (rev 417)
+++ trunk/src/experimental/lasieve4_64/x64/slinie2A.asm 2011-10-22 19:20:40 UTC (rev 418)
@@ -81,6 +81,9 @@
%define j_per_strip_minus1 1
FRAME_PROC slinie2, 0, reg_save_list
+ mov rdi, rcx
+ mov rsi, rdx
+ mov rdx, r8
cmp rsi, rdi
jbe .3
Modified: trunk/src/experimental/lasieve4_64/x64/slinie3A.asm
===================================================================
--- trunk/src/experimental/lasieve4_64/x64/slinie3A.asm 2011-10-22 07:34:27 UTC (rev 417)
+++ trunk/src/experimental/lasieve4_64/x64/slinie3A.asm 2011-10-22 19:20:40 UTC (rev 418)
@@ -87,6 +87,9 @@
%define j_per_strip_minus1 1
FRAME_PROC slinie3, 0, reg_save_list
+ mov rdi, rcx
+ mov rsi, rdx
+ mov rdx, r8
cmp rsi, rdi
jbe .3
Modified: trunk/src/experimental/lasieve4_64/x64/tdsieve-from-sched.asm
===================================================================
--- trunk/src/experimental/lasieve4_64/x64/tdsieve-from-sched.asm 2011-10-22 07:34:27 UTC (rev 417)
+++ trunk/src/experimental/lasieve4_64/x64/tdsieve-from-sched.asm 2011-10-22 19:20:40 UTC (rev 418)
@@ -15,7 +15,7 @@
mov rsi, rdx
mov rdx, r8
mov rcx, r9
- mov r8, qword [rsp+stack_use+40]
+ mov r8, [rsp+stack_use+40]
mov r9, [rdi]
lea rsi, [rsi-12]
This was sent by the SourceForge.net collaborative development platform, the world's largest Open Source development site.
|
|
From: <brg...@us...> - 2011-10-22 22:07:12
|
Revision: 419
http://ggnfs.svn.sourceforge.net/ggnfs/?rev=419&view=rev
Author: brgladman
Date: 2011-10-22 22:07:05 +0000 (Sat, 22 Oct 2011)
Log Message:
-----------
correct Windows assembler handling of RIP relative addressing
Modified Paths:
--------------
trunk/src/experimental/lasieve4_64/x64/asm-zeit.asm
trunk/src/experimental/lasieve4_64/x64/mpqs_gauss.asm
trunk/src/experimental/lasieve4_64/x64/mpqs_sieve.asm
trunk/src/experimental/lasieve4_64/x64/mpqs_td.asm
Modified: trunk/src/experimental/lasieve4_64/x64/asm-zeit.asm
===================================================================
--- trunk/src/experimental/lasieve4_64/x64/asm-zeit.asm 2011-10-22 19:20:40 UTC (rev 418)
+++ trunk/src/experimental/lasieve4_64/x64/asm-zeit.asm 2011-10-22 22:07:05 UTC (rev 419)
@@ -16,16 +16,16 @@
rdtsc
shl rdx, 32
or rax, rdx
- mov rdx, [rip+asmzeitcounter]
- sub [rdx+rdi*8], rax
+ mov rdx, [rel asmzeitcounter]
+ sub [rdx+rcx*8], rax
ret
LEAF_PROC zeitB
rdtsc
shl rdx, 32
or rax, rdx
- mov rdx, [rip+asmzeitcounter]
- add [rdx+rdi*8], rax
+ mov rdx, [rel asmzeitcounter]
+ add [rdx+rcx*8], rax
ret
end
Modified: trunk/src/experimental/lasieve4_64/x64/mpqs_gauss.asm
===================================================================
--- trunk/src/experimental/lasieve4_64/x64/mpqs_gauss.asm 2011-10-22 19:20:40 UTC (rev 418)
+++ trunk/src/experimental/lasieve4_64/x64/mpqs_gauss.asm 2011-10-22 22:07:05 UTC (rev 419)
@@ -28,10 +28,10 @@
FRAME_PROC asm_gauss, 0, reg_save_list
-.1: mov eax, [rip+mpqs_gauss_k]
+.1: mov eax, [rel mpqs_gauss_k]
dec eax
mov r8d, 1
- mov [rip+mpqs_gauss_k], eax
+ mov [rel mpqs_gauss_k], eax
jl .14
mov ecx, eax
and ecx, 31
@@ -39,31 +39,31 @@
mov ecx, eax
shr ecx, 5 ; %r11d
mov r11d, ecx
- mov ecx, [rip+mpqs_gauss_j]
+ mov ecx, [rel mpqs_gauss_j]
mov rdi, mpqs_gauss_c
- mov rsi, [rip+mpqs_gauss_row]
+ mov rsi, [rel mpqs_gauss_row]
mov eax, r11d
- mov edx, [rip+mpqs_gauss_n32]
+ mov edx, [rel mpqs_gauss_n32]
mov rsi, [rsi+rcx*8]
sub rax, rdx
dec ecx
lea rsi, [rsi+rax*4]
- mov eax, [rip+mpqs_gauss_n32]
+ mov eax, [rel mpqs_gauss_n32]
align 16
.2: inc ecx
-.3: cmp ecx, [rip+mpqs_gauss_m]
+.3: cmp ecx, [rel mpqs_gauss_m]
jnc .15
lea rsi, [rsi+rax*4]
mov edx, [rsi]
and edx, r8d
jz .2
-.4: cmp ecx, [rip+mpqs_gauss_j]
+.4: cmp ecx, [rel mpqs_gauss_j]
jz .7
- mov rsi, [rip+mpqs_gauss_row]
+ mov rsi, [rel mpqs_gauss_row]
mov rax, [rsi+rcx*8] ; mpqs_gauss_row[j]
- mov edx, [rip+mpqs_gauss_j]
+ mov edx, [rel mpqs_gauss_j]
mov r9, [rsi+rdx*8] ; mpqs_gauss_row[mpqs_gauss_j]
xor rdx, rdx
.5: movq mm0, [rax+rdx*4]
@@ -71,20 +71,20 @@
movq [r9+rdx*4], mm0
movq [rax+rdx*4], mm1
add rdx, 2
- cmp edx, [rip+mpqs_gauss_n32]
+ cmp edx, [rel mpqs_gauss_n32]
jc .5
-.6: mov ecx, [rip+mpqs_gauss_j]
+.6: mov ecx, [rel mpqs_gauss_j]
.7: mov rsi, mpqs_gauss_d
- mov eax, [rip+mpqs_gauss_k]
+ mov eax, [rel mpqs_gauss_k]
mov [rsi+rax*2], cx
mov [rdi+rcx*2], ax
- inc dword[rip+mpqs_gauss_j]
+ inc dword[rel mpqs_gauss_j]
mov r10d, ecx
- mov rsi, [rip+mpqs_gauss_mat]
+ mov rsi, [rel mpqs_gauss_mat]
mov rdi, mpqs_gauss_col
mov edx, r11d
test ecx, ecx
- mov eax, [rip+mpqs_gauss_n32]
+ mov eax, [rel mpqs_gauss_n32]
lea rsi, [rsi+rdx*4]
mov rdx, 2
jz .9
@@ -102,7 +102,7 @@
lea rsi, [rsi+rax*4]
jc .8
.9: inc ecx
- cmp ecx, [rip+mpqs_gauss_m]
+ cmp ecx, [rel mpqs_gauss_m]
lea rsi, [rsi+rax*4]
jnc .11
@@ -114,14 +114,14 @@
cmovnz r9, rdx
inc rcx
add rdi, r9
- cmp ecx, [rip+mpqs_gauss_m]
+ cmp ecx, [rel mpqs_gauss_m]
lea rsi, [rsi+rax*4]
jc .10
.11: mov r8, mpqs_gauss_col
cmp rdi, r8
jz .1
mov ecx, r10d
- mov rsi, [rip+mpqs_gauss_row]
+ mov rsi, [rel mpqs_gauss_row]
mov rsi, [rsi+rcx*8]
mov edx, r11d
cmp edx, 2
@@ -133,7 +133,7 @@
align 16
.12: movzx r9, word[r8]
- mov rax, [rip+mpqs_gauss_row]
+ mov rax, [rel mpqs_gauss_row]
mov rax, [rax+r9*8]
xor ecx, ecx
.13: mov edx, [rsi+rcx*4]
@@ -150,14 +150,14 @@
EXIT_PROC reg_save_list
.15: mov rdi, mpqs_gauss_d
- mov eax, [rip+mpqs_gauss_k]
+ mov eax, [rel mpqs_gauss_k]
mov r9w, -1
mov [rdi+rax*2], r9w
mov r10d, ecx
jmp .14
.16 movzx r9, word[r8]
movq mm0, [rsi]
- mov rsi, [rip+mpqs_gauss_row]
+ mov rsi, [rel mpqs_gauss_row]
mov rax, [rsi+r9*8]
.17: movq mm1, [rax]
lea r8, [r8+2]
@@ -171,7 +171,7 @@
.18: movq mm0, [rsi]
movzx r9, word[r8]
movq mm2, [rsi+8]
- mov rsi, [rip+mpqs_gauss_row]
+ mov rsi, [rel mpqs_gauss_row]
mov rax, [rsi+r9*8]
.19: movq mm1, [rax]
lea r8, [r8+2]
@@ -189,7 +189,7 @@
movzx r9, word[r8]
movq mm2, [rsi+8]
movq mm4, [rsi+16]
- mov rsi, [rip+mpqs_gauss_row]
+ mov rsi, [rel mpqs_gauss_row]
mov rax, [rsi+r9*8]
.21: movq mm1, [rax]
lea r8, [r8+2]
Modified: trunk/src/experimental/lasieve4_64/x64/mpqs_sieve.asm
===================================================================
--- trunk/src/experimental/lasieve4_64/x64/mpqs_sieve.asm 2011-10-22 19:20:40 UTC (rev 418)
+++ trunk/src/experimental/lasieve4_64/x64/mpqs_sieve.asm 2011-10-22 22:07:05 UTC (rev 419)
@@ -17,8 +17,8 @@
FRAME_PROC asm_sieve, 0, reg_save_list
- movzx rax, word[rip+mpqs_sievebegin]
- movzx r10, word[rip+mpqs_sievelen]
+ movzx rax, word[rel mpqs_sievebegin]
+ movzx r10, word[rel mpqs_sievelen]
mov rcx, mpqs_FB
mov r8, mpqs_FB_start
mov r9, mpqs_FB_log
@@ -27,7 +27,7 @@
lea r9, [r9+rax]
mov r11, r10
shr r10, 2
- mov r13, [rip+mpqs_sievearray]
+ mov r13, [rel mpqs_sievearray]
mov r15, r13
add r15, r11
Modified: trunk/src/experimental/lasieve4_64/x64/mpqs_td.asm
===================================================================
--- trunk/src/experimental/lasieve4_64/x64/mpqs_td.asm 2011-10-22 19:20:40 UTC (rev 418)
+++ trunk/src/experimental/lasieve4_64/x64/mpqs_td.asm 2011-10-22 22:07:05 UTC (rev 419)
@@ -58,10 +58,10 @@
pxor xmm5, xmm5
movzx r9, word[rdi+8] ; %r9
- movzx rcx, word[rip+mpqs_td_begin]
+ movzx rcx, word[rel mpqs_td_begin]
mov rbx, mpqs_FB_inv_info
mov r11, mpqs_FB_start
- mov r13w, [rip+mpqs_nFBk_1]
+ mov r13w, [rel mpqs_nFBk_1]
add r13w, 4
movaps xmm0, [rbx]
movaps xmm1, [rbx+16]
@@ -139,7 +139,7 @@
jmp .1
.2:
mov r10, mpqs_FB
- movzx rdx, word[rip+mpqs_nFBk_1]
+ movzx rdx, word[rel mpqs_nFBk_1]
add rdx, rdx
add rdx, rdx
sub r10, rdx
@@ -163,7 +163,7 @@
inc r9
cmp r9, 27
jnc .16
- mov cx, [rip+mpqs_nFBk_1]
+ mov cx, [rel mpqs_nFBk_1]
mov [rdi+r9*2+8], cx
shr r8, 1
jmp .5
@@ -195,7 +195,7 @@
mul r11d
mov r8d, eax
mov r11, mpqs_FB_inv
- movzx rcx, word[rip+mpqs_nFBk_1]
+ movzx rcx, word[rel mpqs_nFBk_1]
add rcx, rcx
add rcx, rcx
sub r11, rcx
@@ -223,7 +223,7 @@
jz .15
mov r11, mpqs_FBk_inv
mov r10, mpqs_FBk
- movzx rsi, word[rip+mpqs_nFBk]
+ movzx rsi, word[rel mpqs_nFBk]
inc rsi
.10: dec rsi
mov eax, r8d
@@ -247,8 +247,8 @@
jz .15
mov r11, mpqs_FB_A_inv
mov r10, mpqs_Adiv_all
- mov r13w, [rip+mpqs_nFB]
- add r13w, [rip+mpqs_nFBk]
+ mov r13w, [rel mpqs_nFB]
+ add r13w, [rel mpqs_nFBk]
movzx rsi, word[mpqs_nAdiv_total]
inc rsi
.13: dec rsi
This was sent by the SourceForge.net collaborative development platform, the world's largest Open Source development site.
|
|
From: <brg...@us...> - 2011-10-22 22:25:04
|
Revision: 420
http://ggnfs.svn.sourceforge.net/ggnfs/?rev=420&view=rev
Author: brgladman
Date: 2011-10-22 22:24:58 +0000 (Sat, 22 Oct 2011)
Log Message:
-----------
correct Windows assembler handling of RIP relative addressing
Modified Paths:
--------------
trunk/src/experimental/lasieve4_64/x64/lasieve-asmprepn.asm
trunk/src/experimental/lasieve4_64/x64/lasieve-asmprepn2.asm
trunk/src/experimental/lasieve4_64/x64/modinv1000.asm
trunk/src/experimental/lasieve4_64/x64/modinv1000b.asm
trunk/src/experimental/lasieve4_64/x64/mpqs_td.asm
trunk/src/experimental/lasieve4_64/x64/pt64.asm
Modified: trunk/src/experimental/lasieve4_64/x64/lasieve-asmprepn.asm
===================================================================
--- trunk/src/experimental/lasieve4_64/x64/lasieve-asmprepn.asm 2011-10-22 22:07:05 UTC (rev 419)
+++ trunk/src/experimental/lasieve4_64/x64/lasieve-asmprepn.asm 2011-10-22 22:24:58 UTC (rev 420)
@@ -178,17 +178,17 @@
psrldq xmm7, 8
mov esi, ebx
pmuludq xmm12, xmm2
+ sub rsp, 32
+ call asm_modinv32b
add rsp, 32
- call asm_modinv32b
- sub rsp, 32
movd edi, xmm7
pmuludq xmm12, xmm0
movd xmm11, eax
mov esi, r15d
pxor xmm10, xmm10
+ sub rsp, 32
+ call asm_modinv32b
add rsp, 32
- call asm_modinv32b
- sub rsp, 32
movd xmm8, eax
psubq xmm9, xmm12
mov rax, [r12]
@@ -247,17 +247,17 @@
psrldq xmm7, 8
movdqa xmm2, xmm11
pmuludq xmm11, xmm8
+ sub rsp, 32
+ call get_recurrence_info
add rsp, 32
- call get_recurrence_info
- sub rsp, 32
mov esi, r15d
lea rdi, [r13+8]
movd edx, xmm7
pmuludq xmm11, xmm2
movdqa xmm0, xmm8
+ sub rsp, 32
+ call get_recurrence_info
add rsp, 32
- call get_recurrence_info
- sub rsp, 32
cmp r14, r12
pslld xmm2, 1
pshufd xmm12, xmm1, 0x98
Modified: trunk/src/experimental/lasieve4_64/x64/lasieve-asmprepn2.asm
===================================================================
--- trunk/src/experimental/lasieve4_64/x64/lasieve-asmprepn2.asm 2011-10-22 22:07:05 UTC (rev 419)
+++ trunk/src/experimental/lasieve4_64/x64/lasieve-asmprepn2.asm 2011-10-22 22:24:58 UTC (rev 420)
@@ -191,9 +191,9 @@
pshufd xmm8, xmm8, 0x4e
mov esi, ebx
pmuludq xmm12, xmm2
- add rsp, 32
+ sub rsp, 32
call asm_modinv32b
- sub rsp, 32
+ add rsp, 32
%if %1 * (%1 - 3) == 0
sub r15d, eax
%endif
@@ -258,17 +258,17 @@
psrldq xmm7, 8
movdqa xmm2, xmm11
pmuludq xmm11, xmm8
- add rsp, 32
+ sub rsp, 32
call get_recurrence_info
- sub rsp, 32
+ add rsp, 32
mov esi, ebx
lea rdi, [r13+8]
movd edx, xmm7
pmuludq xmm11, xmm2
movdqa xmm0, xmm8
+ sub rsp, 32
+ call get_recurrence_info
add rsp, 32
- call get_recurrence_info
- sub rsp, 32
cmp r14, r12
pslld xmm2, 1
pshufd xmm12, xmm1, 0x98
Modified: trunk/src/experimental/lasieve4_64/x64/modinv1000.asm
===================================================================
--- trunk/src/experimental/lasieve4_64/x64/modinv1000.asm 2011-10-22 22:07:05 UTC (rev 419)
+++ trunk/src/experimental/lasieve4_64/x64/modinv1000.asm 2011-10-22 22:24:58 UTC (rev 420)
@@ -75,9 +75,9 @@
mov eax, [rel modulo32]
sub eax, r8d
ret
-.6: add rsp, 32
+.6: sub rsp, 32
call abort
- sub rsp, 32
+ add rsp, 32
end
\ No newline at end of file
Modified: trunk/src/experimental/lasieve4_64/x64/modinv1000b.asm
===================================================================
--- trunk/src/experimental/lasieve4_64/x64/modinv1000b.asm 2011-10-22 22:07:05 UTC (rev 419)
+++ trunk/src/experimental/lasieve4_64/x64/modinv1000b.asm 2011-10-22 22:24:58 UTC (rev 420)
@@ -80,9 +80,9 @@
sub eax, r8d
ret
.6: je .8
-.7: add rsp, 32
+.7: sub rsp, 32
call abort
- sub rsp, 32
+ add rsp, 32
.8: xor eax, eax
ret
Modified: trunk/src/experimental/lasieve4_64/x64/mpqs_td.asm
===================================================================
--- trunk/src/experimental/lasieve4_64/x64/mpqs_td.asm 2011-10-22 22:07:05 UTC (rev 419)
+++ trunk/src/experimental/lasieve4_64/x64/mpqs_td.asm 2011-10-22 22:24:58 UTC (rev 420)
@@ -278,9 +278,9 @@
.16: xor rax, rax
emms
EXIT_PROC reg_save_list
-.17: add rsp, 32
+.17: sub rsp, 32
call abort
- sub rsp, 32
+ add rsp, 32
END_PROC reg_save_list
end
Modified: trunk/src/experimental/lasieve4_64/x64/pt64.asm
===================================================================
--- trunk/src/experimental/lasieve4_64/x64/pt64.asm 2011-10-22 22:07:05 UTC (rev 419)
+++ trunk/src/experimental/lasieve4_64/x64/pt64.asm 2011-10-22 22:24:58 UTC (rev 420)
@@ -60,9 +60,9 @@
mov rcx, rax
mov r11, rsi
jmp .5
-.4: add rsp,32
+.4: sub rsp,32
call modsq64
- sub rsp, 32
+ add rsp, 32
test r9, rcx
jz .6
.5: xor rax, rax
@@ -81,9 +81,9 @@
je .9
dec r10
jz .8
+ sub rsp, 32
+ call modsq64
add rsp, 32
- call modsq64
- sub rsp, 32
jmp .7
align 8
This was sent by the SourceForge.net collaborative development platform, the world's largest Open Source development site.
|
|
From: <brg...@us...> - 2011-10-23 19:58:45
|
Revision: 422
http://ggnfs.svn.sourceforge.net/ggnfs/?rev=422&view=rev
Author: brgladman
Date: 2011-10-23 19:58:39 +0000 (Sun, 23 Oct 2011)
Log Message:
-----------
More Windows x64 assembler errors
Modified Paths:
--------------
trunk/src/experimental/lasieve4_64/x64/lasieve-asmprepn.asm
trunk/src/experimental/lasieve4_64/x64/lasieve-asmprepn2.asm
trunk/src/experimental/lasieve4_64/x64/siever-config.h
Modified: trunk/src/experimental/lasieve4_64/x64/lasieve-asmprepn.asm
===================================================================
--- trunk/src/experimental/lasieve4_64/x64/lasieve-asmprepn.asm 2011-10-23 16:20:46 UTC (rev 421)
+++ trunk/src/experimental/lasieve4_64/x64/lasieve-asmprepn.asm 2011-10-23 19:58:39 UTC (rev 422)
@@ -248,6 +248,9 @@
movdqa xmm2, xmm11
pmuludq xmm11, xmm8
sub rsp, 32
+ mov r8, rdx
+ mov rdx, rsi
+ mov rcx, rdi
call get_recurrence_info
add rsp, 32
mov esi, r15d
@@ -256,6 +259,9 @@
pmuludq xmm11, xmm2
movdqa xmm0, xmm8
sub rsp, 32
+ mov r8, rdx
+ mov rdx, rsi
+ mov rcx, rdi
call get_recurrence_info
add rsp, 32
cmp r14, r12
Modified: trunk/src/experimental/lasieve4_64/x64/lasieve-asmprepn2.asm
===================================================================
--- trunk/src/experimental/lasieve4_64/x64/lasieve-asmprepn2.asm 2011-10-23 16:20:46 UTC (rev 421)
+++ trunk/src/experimental/lasieve4_64/x64/lasieve-asmprepn2.asm 2011-10-23 19:58:39 UTC (rev 422)
@@ -259,6 +259,9 @@
movdqa xmm2, xmm11
pmuludq xmm11, xmm8
sub rsp, 32
+ mov r8, rdx
+ mov rdx, rsi
+ mov rcx, rdi
call get_recurrence_info
add rsp, 32
mov esi, ebx
@@ -267,6 +270,9 @@
pmuludq xmm11, xmm2
movdqa xmm0, xmm8
sub rsp, 32
+ mov r8, rdx
+ mov rdx, rsi
+ mov rcx, rdi
call get_recurrence_info
add rsp, 32
cmp r14, r12
Modified: trunk/src/experimental/lasieve4_64/x64/siever-config.h
===================================================================
--- trunk/src/experimental/lasieve4_64/x64/siever-config.h 2011-10-23 16:20:46 UTC (rev 421)
+++ trunk/src/experimental/lasieve4_64/x64/siever-config.h 2011-10-23 19:58:39 UTC (rev 422)
@@ -32,8 +32,14 @@
typedef int i32_t;
typedef short int i16_t;
typedef unsigned short u16_t;
+#ifdef _WIN32
+typedef unsigned long long u64_t;
+typedef long long i64_t;
+#else
typedef unsigned long u64_t;
typedef long int i64_t;
+#endif
+
#define U32_MAX 0xffffffff
#define I32_MAX INT_MAX
This was sent by the SourceForge.net collaborative development platform, the world's largest Open Source development site.
|
×
Want the latest updates on software, tech news, and AI?
Get latest updates about software, tech news, and AI from SourceForge directly in your inbox once a month.
Thanks for helping keep SourceForge clean.
X