aaron/ispc
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Added disable-handle-pseudo-memory-ops option.

Browse Source
This commit is contained in:
Matt Pharr
2011-10-23 13:35:23 -07:00
parent d224252b5d
commit 6084d6aeaf
4 changed files with 18 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
ispc.cpp
View File

@@ -284,6 +284,7 @@ Opt::Opt() {
fastMaskedVload = false; fastMaskedVload = false;
unrollLoops = true; unrollLoops = true;
disableAsserts = false; disableAsserts = false;
disableHandlePseudoMemoryOps = false;
disableBlendedMaskedStores = false; disableBlendedMaskedStores = false;
disableCoherentControlFlow = false; disableCoherentControlFlow = false;
disableUniformControlFlow = false; disableUniformControlFlow = false;

6
ispc.h
View File

@@ -228,6 +228,12 @@ struct Opt {
performance in the generated code). */ performance in the generated code). */
bool disableAsserts; bool disableAsserts;
/** If enabled, the various __pseudo* memory ops (gather/scatter,
masked load/store) are left in their __pseudo* form, for better
understanding of the structure of generated code when reading
it. */
bool disableHandlePseudoMemoryOps;
/** On targets that don't have a masked store instruction but do have a /** On targets that don't have a masked store instruction but do have a
blending instruction, by default, we simulate masked stores by blending instruction, by default, we simulate masked stores by
loading the old value, blending, and storing the result. This can loading the old value, blending, and storing the result. This can

3
main.cpp
View File

@@ -88,6 +88,7 @@ static void usage(int ret) {
printf(" fast-masked-vload\t\tFaster masked vector loads on SSE (may go past end of array)\n"); printf(" fast-masked-vload\t\tFaster masked vector loads on SSE (may go past end of array)\n");
printf(" fast-math\t\t\tPerform non-IEEE-compliant optimizations of numeric expressions\n"); printf(" fast-math\t\t\tPerform non-IEEE-compliant optimizations of numeric expressions\n");
#if 0 #if 0
printf(" disable-handle-pseudo-memory-ops\n");
printf(" disable-blended-masked-stores\t\tScalarize masked stores on SSE (vs. using vblendps)\n"); printf(" disable-blended-masked-stores\t\tScalarize masked stores on SSE (vs. using vblendps)\n");
printf(" disable-coherent-control-flow\t\tDisable coherent control flow optimizations\n"); printf(" disable-coherent-control-flow\t\tDisable coherent control flow optimizations\n");
printf(" disable-uniform-control-flow\t\tDisable uniform control flow optimizations\n"); printf(" disable-uniform-control-flow\t\tDisable uniform control flow optimizations\n");
@@ -254,6 +255,8 @@ int main(int Argc, char *Argv[]) {
// These are only used for performance tests of specific // These are only used for performance tests of specific
// optimizations // optimizations
else if (!strcmp(opt, "disable-handle-pseudo-memory-ops"))
g->opt.disableHandlePseudoMemoryOps = true;
else if (!strcmp(opt, "disable-blended-masked-stores")) else if (!strcmp(opt, "disable-blended-masked-stores"))
g->opt.disableBlendedMaskedStores = true; g->opt.disableBlendedMaskedStores = true;
else if (!strcmp(opt, "disable-coherent-control-flow")) else if (!strcmp(opt, "disable-coherent-control-flow"))

4
opt.cpp
View File

@@ -201,8 +201,10 @@ Optimize(llvm::Module *module, int optLevel) {
// them into something that can actually execute. // them into something that can actually execute.
optPM.add(llvm::createPromoteMemoryToRegisterPass()); optPM.add(llvm::createPromoteMemoryToRegisterPass());
optPM.add(CreateGatherScatterFlattenPass()); optPM.add(CreateGatherScatterFlattenPass());
if (g->opt.disableHandlePseudoMemoryOps == false) {
optPM.add(CreateLowerGatherScatterPass()); optPM.add(CreateLowerGatherScatterPass());
optPM.add(CreateLowerMaskedStorePass()); optPM.add(CreateLowerMaskedStorePass());
}
optPM.add(CreateIsCompileTimeConstantPass(true)); optPM.add(CreateIsCompileTimeConstantPass(true));
optPM.add(llvm::createFunctionInliningPass()); optPM.add(llvm::createFunctionInliningPass());
optPM.add(CreateMakeInternalFuncsStaticPass()); optPM.add(CreateMakeInternalFuncsStaticPass());
@@ -282,8 +284,10 @@ Optimize(llvm::Module *module, int optLevel) {
optPM.add(CreateLowerMaskedStorePass()); optPM.add(CreateLowerMaskedStorePass());
if (!g->opt.disableGatherScatterOptimizations) if (!g->opt.disableGatherScatterOptimizations)
optPM.add(CreateGatherScatterImprovementsPass()); optPM.add(CreateGatherScatterImprovementsPass());
if (g->opt.disableHandlePseudoMemoryOps == false) {
optPM.add(CreateLowerMaskedStorePass()); optPM.add(CreateLowerMaskedStorePass());
optPM.add(CreateLowerGatherScatterPass()); optPM.add(CreateLowerGatherScatterPass());
}
optPM.add(llvm::createFunctionInliningPass()); optPM.add(llvm::createFunctionInliningPass());
optPM.add(llvm::createConstantPropagationPass()); optPM.add(llvm::createConstantPropagationPass());
optPM.add(CreateIntrinsicsOptPass()); optPM.add(CreateIntrinsicsOptPass());