Remove MMX register freezes from MemcpyFast.

git-svn-id: http://pcsx2.googlecode.com/svn/trunk@3380 96395faa-99c1-11dd-bbfe-3dabce05a288
This commit is contained in:
Jake.Stine 2010-07-03 13:08:41 +00:00
parent 84c5073b7c
commit 402f19413b
2 changed files with 3 additions and 323 deletions

View File

@ -368,15 +368,9 @@ $memcpy_align_done: // destination is dword aligned
shr eax, 6 // get 64-byte block count shr eax, 6 // get 64-byte block count
jz $memcpy_ic_2 // finish the last few bytes jz $memcpy_ic_2 // finish the last few bytes
mov edx, offset _mmx_backup // will probably need this to save/restore mmx
cmp eax, IN_CACHE_COPY/64 // too big 4 cache? use uncached copy cmp eax, IN_CACHE_COPY/64 // too big 4 cache? use uncached copy
jae $memcpy_uc_test jae $memcpy_uc_test
movq [edx+0x00],mm0
movq [edx+0x08],mm1
movq [edx+0x10],mm2
movq [edx+0x18],mm3
// This is small block copy that uses the MMX registers to copy 8 bytes // This is small block copy that uses the MMX registers to copy 8 bytes
// at a time. It uses the "unrolled loop" optimization, and also uses // at a time. It uses the "unrolled loop" optimization, and also uses
// the software prefetch instruction to get the data into the cache. // the software prefetch instruction to get the data into the cache.
@ -407,11 +401,6 @@ $memcpy_ic_1: // 64-byte block copies, in-cache copy
dec eax // count down dec eax // count down
jnz $memcpy_ic_1 // last 64-byte block? jnz $memcpy_ic_1 // last 64-byte block?
movq mm0,[edx+0x00]
movq mm1,[edx+0x08]
movq mm2,[edx+0x10]
movq mm3,[edx+0x18]
$memcpy_ic_2: $memcpy_ic_2:
mov eax, ecx // has valid low 6 bits of the byte count mov eax, ecx // has valid low 6 bits of the byte count
$memcpy_ic_3: $memcpy_ic_3:
@ -430,10 +419,6 @@ $memcpy_uc_test:
// bypasses the cache and writes straight to main memory. This code also // bypasses the cache and writes straight to main memory. This code also
// uses the software prefetch instruction to pre-read the data. // uses the software prefetch instruction to pre-read the data.
movq [edx+0x00],mm0
movq [edx+0x08],mm1
movq [edx+0x10],mm2
.align 16 .align 16
$memcpy_uc_1: // 64-byte blocks, uncached copy $memcpy_uc_1: // 64-byte blocks, uncached copy
@ -460,10 +445,6 @@ $memcpy_uc_1: // 64-byte blocks, uncached copy
movntq [edi-8], mm1 movntq [edi-8], mm1
jnz $memcpy_uc_1 // last 64-byte block? jnz $memcpy_uc_1 // last 64-byte block?
movq mm0,[edx+0x00]
movq mm1,[edx+0x08]
movq mm2,[edx+0x10]
jmp $memcpy_ic_2 // almost done (not needed because large copy below was removed) jmp $memcpy_ic_2 // almost done (not needed because large copy below was removed)
// For the largest size blocks, a special technique called Block Prefetch // For the largest size blocks, a special technique called Block Prefetch

View File

@ -64,296 +64,14 @@ MEMCPY_AMD.CPP
// uses the software prefetch instruction to pre-read the data. // uses the software prefetch instruction to pre-read the data.
// USE 64 * 1024 FOR THIS VALUE IF YOU'RE ALWAYS FILLING A "CLEAN CACHE" // USE 64 * 1024 FOR THIS VALUE IF YOU'RE ALWAYS FILLING A "CLEAN CACHE"
#define BLOCK_PREFETCH_COPY infinity // no limit for movq/movntq w/block prefetch
#define CACHEBLOCK 80h // number of 64-byte blocks (cache lines) for block prefetch
// For the largest size blocks, a special technique called Block Prefetch
// can be used to accelerate the read operations. Block Prefetch reads
// one address per cache line, for a series of cache lines, in a short loop.
// This is faster than using software prefetch. The technique is great for
// getting maximum read bandwidth, especially in DDR memory systems.
// Inline assembly syntax for use with Visual C++ // Inline assembly syntax for use with Visual C++
#if defined(_MSC_VER) #if defined(_MSC_VER)
#ifdef PCSX2_DEBUG // --------------------------------------------------------------------------------------
extern u8 g_globalMMXSaved; // Fast memcpy as coded by AMD, and then improved by air.
// --------------------------------------------------------------------------------------
#endif
static __aligned16 u8 _xmm_backup[16*2];
static __aligned16 u8 _mmx_backup[8*4];
static __declspec(naked) void __fastcall _memcpy_raz_usrc(void *dest, const void *src, size_t bytes)
{
// MOVSRC = opcode used to read. I use the same code for the aligned version, with a different define :)
#define MOVSRC movdqu
#define MOVDST movdqa
__asm
{
//Reads before reads, to avoid stalls
mov eax,[esp+4];
//Make sure to save xmm0, it must be preserved ...
movaps [_xmm_backup],xmm0;
//if >=128 bytes use 128 byte unrolled loop
//i use cmp ..,127 + jna because 127 is encodable using the simm8 form
cmp eax,127;
jna _loop_1;
//since this is a common branch target it could be good to align it -- no idea if it has any effect :p
align 16
//128 byte unrolled loop
_loop_8:
MOVSRC xmm0,[edx+0x00]; //read first to avoid read-after-write stalls
MOVDST [ecx+0x00],xmm0; //then write :p
MOVSRC xmm0,[edx+0x10];
MOVDST [ecx+0x10],xmm0;
sub edx,-128; //edx won't be used for a while, so update it here. sub/-128 for simm8 encoding
sub ecx,-128; //ecx won't be used for a while, so update it here. sub/-128 for simm8 encoding
MOVSRC xmm0,[edx+0x20-128];
MOVDST [ecx+0x20-128],xmm0;
MOVSRC xmm0,[edx+0x30-128];
MOVDST [ecx+0x30-128],xmm0;
add eax,-128; //eax won't be used for a while, so update it here. add/-128 for simm8 encoding
MOVSRC xmm0,[edx+0x40-128];
MOVDST [ecx+0x40-128],xmm0;
MOVSRC xmm0,[edx+0x50-128];
MOVDST [ecx+0x50-128],xmm0;
MOVSRC xmm0,[edx+0x60-128];
MOVDST [ecx+0x60-128],xmm0;
MOVSRC xmm0,[edx+0x70-128];
MOVDST [ecx+0x70-128],xmm0;
//127~ja, 127 is encodable as simm8 :)
cmp eax,127;
ja _loop_8;
//direct copy for 0~7 qwords
//in order to avoid the inc/dec of all 3 registers
//i use negative relative addressing from the top of the buffers
//[top-current index]
_loop_1:
//prepare the regs for 'negative relative addressing'
add edx,eax;
add ecx,eax;
neg eax;
jz cleanup; //exit if nothing to do
_loop_1_inner:
MOVSRC xmm0,[edx+eax];
MOVDST [ecx+eax],xmm0;
add eax,16; //while the offset is still negative we have data to copy
js _loop_1_inner;
//done !
cleanup:
//restore xmm and exit ~)
movaps xmm0,[_xmm_backup];
ret 4;
}
#undef MOVSRC
#undef MOVDST
}
static __declspec(naked) void __fastcall _memcpy_raz_udst(void *dest, const void *src, size_t bytes)
{
// MOVDST = opcode used to read. I use the same code for the aligned version, with a different define :)
#define MOVSRC movaps
#define MOVDST movups
__asm
{
//Reads before reads, to avoid stalls
mov eax,[esp+4];
//Make sure to save xmm0, it must be preserved ...
movaps [_xmm_backup],xmm0;
//if >=128 bytes use 128 byte unrolled loop
//i use cmp ..,127 + jna because 127 is encodable using the simm8 form
cmp eax,127;
jna _loop_1;
//since this is a common branch target it could be good to align it -- no idea if it has any effect :p
align 16
//128 byte unrolled loop
_loop_8:
MOVSRC xmm0,[edx+0x00]; //read first to avoid read-after-write stalls
MOVDST [ecx+0x00],xmm0; //then write :p
MOVSRC xmm0,[edx+0x10];
MOVDST [ecx+0x10],xmm0;
sub edx,-128; //edx won't be used for a while, so update it here. sub/-128 for simm8 encoding
sub ecx,-128; //ecx won't be used for a while, so update it here. sub/-128 for simm8 encoding
MOVSRC xmm0,[edx+0x20-128];
MOVDST [ecx+0x20-128],xmm0;
MOVSRC xmm0,[edx+0x30-128];
MOVDST [ecx+0x30-128],xmm0;
add eax,-128; //eax won't be used for a while, so update it here. add/-128 for simm8 encoding
MOVSRC xmm0,[edx+0x40-128];
MOVDST [ecx+0x40-128],xmm0;
MOVSRC xmm0,[edx+0x50-128];
MOVDST [ecx+0x50-128],xmm0;
MOVSRC xmm0,[edx+0x60-128];
MOVDST [ecx+0x60-128],xmm0;
MOVSRC xmm0,[edx+0x70-128];
MOVDST [ecx+0x70-128],xmm0;
//127~ja, 127 is encodable as simm8 :)
cmp eax,127;
ja _loop_8;
//direct copy for 0~7 qwords
//in order to avoid the inc/dec of all 3 registers
//i use negative relative addressing from the top of the buffers
//[top-current index]
_loop_1:
//prepare the regs for 'negative relative addressing'
add edx,eax;
add ecx,eax;
neg eax;
jz cleanup; //exit if nothing to do
_loop_1_inner:
MOVSRC xmm0,[edx+eax];
movaps [ecx+eax],xmm0;
add eax,16; //while the offset is still negative we have data to copy
js _loop_1_inner;
//done !
cleanup:
//restore xmm and exit ~)
movaps xmm0,[_xmm_backup];
ret 4;
}
#undef MOVSRC
#undef MOVDST
}
// Custom memcpy, only for 16 byte aligned stuff (used for mtgs)
// This function is optimized for medium-small transfer sizes (<2048, >=128). No prefetching is
// used since the reads are linear and the cache logic can predict em :)
// *OBSOLETE* -- memcpy_amd_ has been optimized and is now faster.
__declspec(naked) void __fastcall memcpy_raz_(void *dest, const void *src, size_t bytes)
{
// Code Implementation Notes:
// Uses a forward copy, in 128 byte blocks, and then does the remaining in 16 byte blocks :)
// MOVSRC = opcode used to read. I use the same code for the unaligned version, with a different define :)
#define MOVSRC movaps
#define MOVDST movaps
__asm
{
//Reads before reads, to avoid stalls
mov eax,[esp+4];
//Make sure to save xmm0, it must be preserved ...
movaps [_xmm_backup],xmm0;
//if >=128 bytes use 128 byte unrolled loop
//i use cmp ..,127 + jna because 127 is encodable using the simm8 form
cmp eax,127;
jna _loop_1;
//since this is a common branch target it could be good to align it -- no idea if it has any effect :p
align 16
//128 byte unrolled loop
_loop_8:
MOVSRC xmm0,[edx+0x00]; //read first to avoid read-after-write stalls
MOVDST [ecx+0x00],xmm0; //then write :p
MOVSRC xmm0,[edx+0x10];
MOVDST [ecx+0x10],xmm0;
sub edx,-128; //edx won't be used for a while, so update it here. sub/-128 for simm8 encoding
sub ecx,-128; //ecx won't be used for a while, so update it here. sub/-128 for simm8 encoding
MOVSRC xmm0,[edx+0x20-128];
MOVDST [ecx+0x20-128],xmm0;
MOVSRC xmm0,[edx+0x30-128];
MOVDST [ecx+0x30-128],xmm0;
add eax,-128; //eax won't be used for a while, so update it here. add/-128 for simm8 encoding
MOVSRC xmm0,[edx+0x40-128];
MOVDST [ecx+0x40-128],xmm0;
MOVSRC xmm0,[edx+0x50-128];
MOVDST [ecx+0x50-128],xmm0;
MOVSRC xmm0,[edx+0x60-128];
MOVDST [ecx+0x60-128],xmm0;
MOVSRC xmm0,[edx+0x70-128];
MOVDST [ecx+0x70-128],xmm0;
//127~ja, 127 is encodable as simm8 :)
cmp eax,127;
ja _loop_8;
//direct copy for 0~7 qwords
//in order to avoid the inc/dec of all 3 registers
//i use negative relative addressing from the top of the buffers
//[top-current index]
_loop_1:
//prepare the regs for 'negative relative addressing'
add edx,eax;
add ecx,eax;
neg eax;
jz cleanup; //exit if nothing to do
_loop_1_inner:
MOVSRC xmm0,[edx+eax];
MOVDST [ecx+eax],xmm0;
add eax,16; //while the offset is still negative we have data to copy
js _loop_1_inner;
//done !
cleanup:
//restore xmm and exit ~)
movaps xmm0,[_xmm_backup];
ret 4;
}
#undef MOVSRC
#undef MOVDST
}
// This memcpy routine is for use in situations where the source buffer's alignment is indeterminate.
__forceinline void __fastcall memcpy_raz_usrc(void *dest, const void *src, size_t bytes)
{
if( ((uptr)src & 0xf) == 0 )
memcpy_raz_( dest, src, bytes );
else
_memcpy_raz_usrc( dest, src, bytes );
}
// This memcpy routine is for use in situations where the destination buffer's alignment is indeterminate.
__forceinline void __fastcall memcpy_raz_udst(void *dest, const void *src, size_t bytes)
{
if( ((uptr)dest & 0xf) == 0 )
memcpy_raz_( dest, src, bytes );
else
_memcpy_raz_udst( dest, src, bytes );
}
//////////////////////////////////////////////////////////////////////////
// Fast memcpy as coded by AMD, and thn improved by air.
//
// This routine preserves mmx registers! It's the complete real deal!
__declspec(naked) void __fastcall memcpy_amd_(void *dest, const void *src, size_t n) __declspec(naked) void __fastcall memcpy_amd_(void *dest, const void *src, size_t n)
{ {
__asm __asm
@ -398,15 +116,9 @@ $memcpy_align_done: ; destination is dword aligned
shr eax, 6 ; get 64-byte block count shr eax, 6 ; get 64-byte block count
jz $memcpy_ic_2 ; finish the last few bytes jz $memcpy_ic_2 ; finish the last few bytes
mov edx, offset _mmx_backup ; will probably need this to save/restore mmx
cmp eax, IN_CACHE_COPY/64 ; too big 4 cache? use uncached copy cmp eax, IN_CACHE_COPY/64 ; too big 4 cache? use uncached copy
jae $memcpy_uc_test jae $memcpy_uc_test
movq [edx+0x00],mm0
movq [edx+0x08],mm1
movq [edx+0x10],mm2
movq [edx+0x18],mm3
// This is small block copy that uses the MMX registers to copy 8 bytes // This is small block copy that uses the MMX registers to copy 8 bytes
// at a time. It uses the "unrolled loop" optimization, and also uses // at a time. It uses the "unrolled loop" optimization, and also uses
// the software prefetch instruction to get the data into the cache. // the software prefetch instruction to get the data into the cache.
@ -437,11 +149,6 @@ $memcpy_ic_1: ; 64-byte block copies, in-cache copy
dec eax ; count down dec eax ; count down
jnz $memcpy_ic_1 ; last 64-byte block? jnz $memcpy_ic_1 ; last 64-byte block?
movq mm0,[edx+0x00]
movq mm1,[edx+0x08]
movq mm2,[edx+0x10]
movq mm3,[edx+0x18]
$memcpy_ic_2: $memcpy_ic_2:
mov eax, ecx ; has valid low 6 bits of the byte count mov eax, ecx ; has valid low 6 bits of the byte count
$memcpy_ic_3: $memcpy_ic_3:
@ -460,10 +167,6 @@ $memcpy_uc_test:
// bypasses the cache and writes straight to main memory. This code also // bypasses the cache and writes straight to main memory. This code also
// uses the software prefetch instruction to pre-read the data. // uses the software prefetch instruction to pre-read the data.
movq [edx+0x00],mm0
movq [edx+0x08],mm1
movq [edx+0x10],mm2
align 16 align 16
$memcpy_uc_1: ; 64-byte blocks, uncached copy $memcpy_uc_1: ; 64-byte blocks, uncached copy
@ -490,10 +193,6 @@ $memcpy_uc_1: ; 64-byte blocks, uncached copy
movntq [edi-8], mm1 movntq [edi-8], mm1
jnz $memcpy_uc_1 ; last 64-byte block? jnz $memcpy_uc_1 ; last 64-byte block?
movq mm0,[edx+0x00]
movq mm1,[edx+0x08]
movq mm2,[edx+0x10]
jmp $memcpy_ic_2 ; almost done (not needed because large copy below was removed) jmp $memcpy_ic_2 ; almost done (not needed because large copy below was removed)
// For the largest size blocks, a special technique called Block Prefetch // For the largest size blocks, a special technique called Block Prefetch