mirrors/dynarmic
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

emit_x64_crc32: PCLMULQDQ implementation of EmitCRC32ISO

Browse source
This commit is contained in:
MerryMage 2020-06-03 11:16:53 +01:00
parent b47adaee1d
commit fa6aee434e
1 changed files with 67 additions and 4 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

67
src/backend/x64/emit_x64_crc32.cpp
View file

@ -24,16 +24,79 @@ static void EmitCRC32Castagnoli(BlockOfCode& code, EmitContext& ctx, IR::Inst* i
const Xbyak::Reg value = ctx.reg_alloc.UseGpr(args[1]).changeBit(data_size); const Xbyak::Reg value = ctx.reg_alloc.UseGpr(args[1]).changeBit(data_size);
code.crc32(crc, value); code.crc32(crc, value);
ctx.reg_alloc.DefineValue(inst, crc); ctx.reg_alloc.DefineValue(inst, crc);
} else { return;
}
ctx.reg_alloc.HostCall(inst, args[0], args[1], {}); ctx.reg_alloc.HostCall(inst, args[0], args[1], {});
code.mov(code.ABI_PARAM3, data_size / CHAR_BIT); code.mov(code.ABI_PARAM3, data_size / CHAR_BIT);
code.CallFunction(&CRC32::ComputeCRC32Castagnoli); code.CallFunction(&CRC32::ComputeCRC32Castagnoli);
}
} }
static void EmitCRC32ISO(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst, const int data_size) { static void EmitCRC32ISO(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst, const int data_size) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst); auto args = ctx.reg_alloc.GetArgumentInfo(inst);
if (code.DoesCpuSupport(Xbyak::util::Cpu::tPCLMULQDQ) && data_size <= 32) {
const Xbyak::Reg32 crc = ctx.reg_alloc.UseScratchGpr(args[0]).cvt32();
const Xbyak::Reg64 value = ctx.reg_alloc.UseGpr(args[1]);
const Xbyak::Xmm xmm_crc = ctx.reg_alloc.ScratchXmm();
const Xbyak::Xmm xmm_value = ctx.reg_alloc.ScratchXmm();
const Xbyak::Xmm xmm_const = ctx.reg_alloc.ScratchXmm();
code.movd(xmm_value, value.cvt32());
code.movd(xmm_crc, crc);
code.movdqa(xmm_const, code.MConst(xword, 0x00000001'F7011641, 0x00000001'DB710641));
code.pxor(xmm_value, xmm_crc);
if (data_size < 32) {
code.pslld(xmm_value, 32 - data_size);
code.psllq(xmm_crc, 32 - data_size);
}
code.pclmulqdq(xmm_value, xmm_const, 0x00);
code.pshufd(xmm_value, xmm_value, 0b11111100);
code.pclmulqdq(xmm_value, xmm_const, 0x10);
if (data_size < 32) {
code.pxor(xmm_value, xmm_crc);
}
code.pextrd(crc, xmm_value, 1);
ctx.reg_alloc.DefineValue(inst, crc);
return;
}
if (code.DoesCpuSupport(Xbyak::util::Cpu::tPCLMULQDQ) && data_size == 64) {
const Xbyak::Reg32 crc = ctx.reg_alloc.UseScratchGpr(args[0]).cvt32();
const Xbyak::Reg64 value = ctx.reg_alloc.UseScratchGpr(args[1]);
const Xbyak::Xmm xmm_crc = ctx.reg_alloc.ScratchXmm();
const Xbyak::Xmm xmm_value = ctx.reg_alloc.ScratchXmm();
const Xbyak::Xmm xmm_const = ctx.reg_alloc.ScratchXmm();
code.movd(xmm_value, value.cvt32());
code.movd(xmm_crc, crc);
code.movdqa(xmm_const, code.MConst(xword, 0x00000001'F7011641, 0x00000001'DB710641));
code.pxor(xmm_value, xmm_crc);
code.pclmulqdq(xmm_value, xmm_const, 0x00);
code.pshufd(xmm_value, xmm_value, 0b11111100);
code.pclmulqdq(xmm_value, xmm_const, 0x10);
code.movq(xmm_crc, value);
code.pxor(xmm_value, xmm_crc);
code.pshufd(xmm_value, xmm_value, 0b11111101);
code.pclmulqdq(xmm_value, xmm_const, 0x00);
code.pshufd(xmm_value, xmm_value, 0b11111100);
code.pclmulqdq(xmm_value, xmm_const, 0x10);
code.pextrd(crc, xmm_value, 1);
ctx.reg_alloc.DefineValue(inst, crc);
return;
}
ctx.reg_alloc.HostCall(inst, args[0], args[1], {}); ctx.reg_alloc.HostCall(inst, args[0], args[1], {});
code.mov(code.ABI_PARAM3, data_size / CHAR_BIT); code.mov(code.ABI_PARAM3, data_size / CHAR_BIT);
code.CallFunction(&CRC32::ComputeCRC32ISO); code.CallFunction(&CRC32::ComputeCRC32ISO);