Optimized loading from scratchpad

This commit is contained in:
SChernykh
2019-09-11 11:48:22 +02:00
parent f1ff014bca
commit 93fec18991
7 changed files with 63 additions and 14 deletions
+13 -7
View File
@@ -243,7 +243,7 @@ namespace randomx {
generateProgramPrologue(prog, pcfg);
memcpy(code + codePos, codeReadDataset, readDatasetSize);
codePos += readDatasetSize;
generateProgramEpilogue(prog);
generateProgramEpilogue(prog, pcfg);
}
void JitCompilerX86::generateProgramLight(Program& prog, ProgramConfiguration& pcfg, uint32_t datasetOffset) {
@@ -254,7 +254,7 @@ namespace randomx {
emitByte(CALL);
emit32(superScalarHashOffset - (codePos + 4));
emit(codeReadDatasetLightSshFin, readDatasetLightFinSize);
generateProgramEpilogue(prog);
generateProgramEpilogue(prog, pcfg);
}
template<size_t N>
@@ -298,12 +298,13 @@ namespace randomx {
for (unsigned i = 0; i < 8; ++i) {
registerUsage[i] = -1;
}
codePos = ((uint8_t*)randomx_program_prologue_first_load) - ((uint8_t*)randomx_program_prologue);
code[codePos + sizeof(REX_XOR_RAX_R64)] = 0xc0 + pcfg.readReg0;
code[codePos + sizeof(REX_XOR_RAX_R64) * 2 + 1] = 0xc0 + pcfg.readReg1;
codePos = prologueSize;
memcpy(code + codePos - 48, &pcfg.eMask, sizeof(pcfg.eMask));
emit(REX_XOR_RAX_R64);
emitByte(0xc0 + pcfg.readReg0);
emit(REX_XOR_RAX_R64);
emitByte(0xc0 + pcfg.readReg1);
memcpy(code + codePos, codeLoopLoad, loopLoadSize);
codePos += loopLoadSize;
for (unsigned i = 0; i < prog.getSize(); ++i) {
@@ -318,7 +319,12 @@ namespace randomx {
emitByte(0xc0 + pcfg.readReg3);
}
void JitCompilerX86::generateProgramEpilogue(Program& prog) {
void JitCompilerX86::generateProgramEpilogue(Program& prog, ProgramConfiguration& pcfg) {
emit(REX_MOV_RR64);
emitByte(0xc0 + pcfg.readReg0);
emit(REX_XOR_RAX_R64);
emitByte(0xc0 + pcfg.readReg1);
emit((const uint8_t*)&randomx_prefetch_scratchpad, ((uint8_t*)&randomx_prefetch_scratchpad_end) - ((uint8_t*)&randomx_prefetch_scratchpad));
memcpy(code + codePos, codeLoopStore, loopStoreSize);
codePos += loopStoreSize;
emit(SUB_EBX);