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

added correction for local pointer computations

Browse Source
This commit is contained in:
Evghenii
2014-01-07 18:45:22 +01:00
parent 7d37f7b634
commit 1303b07b72
2 changed files with 11 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files

28
ctx.cpp
View File

@@ -1836,7 +1836,7 @@ static llvm::Value* lCorrectLocalPtr(FunctionEmitContext *ctx, llvm::Value* valu
{ {
assert(value->getType()->isPointerTy()); assert(value->getType()->isPointerTy());
llvm::PointerType *pt = llvm::dyn_cast<llvm::PointerType>(value->getType()); llvm::PointerType *pt = llvm::dyn_cast<llvm::PointerType>(value->getType());
assert (pt->getAddressSpace() == 3); if (pt->getAddressSpace() != 3) return value;
llvm::Function *func_tid_x = m->module->getFunction("__tid_x"); llvm::Function *func_tid_x = m->module->getFunction("__tid_x");
llvm::Function *func_warpsz = m->module->getFunction("__warpsize"); llvm::Function *func_warpsz = m->module->getFunction("__warpsize");
@@ -1852,13 +1852,13 @@ static llvm::Value* lConvertLocalToGenericPtr(FunctionEmitContext *ctx, llvm::Va
llvm::PointerType *pt = llvm::dyn_cast<llvm::PointerType>(value->getType()); llvm::PointerType *pt = llvm::dyn_cast<llvm::PointerType>(value->getType());
if (pt->getAddressSpace() != 3) return value; if (pt->getAddressSpace() != 3) return value;
value = lCorrectLocalPtr(ctx, value);
llvm::PointerType *PointerTy = llvm::PointerType::get(LLVMTypes::Int64Type, 3); llvm::PointerType *PointerTy = llvm::PointerType::get(LLVMTypes::Int64Type, 3);
llvm::Value *cast = ctx->BitCastInst(value, PointerTy, "__cvt_log2gen_i64ptr1_"); value = ctx->BitCastInst(value, PointerTy, "cvtLog2Gen_i64ptr");
value = lCorrectLocalPtr(ctx, value);
llvm::Function *__cvt_loc2gen = m->module->getFunction("__cvt_loc2gen"); llvm::Function *__cvt_loc2gen = m->module->getFunction("__cvt_loc2gen");
std::vector<llvm::Value *> __cvt_loc2gen_args; std::vector<llvm::Value *> __cvt_loc2gen_args;
__cvt_loc2gen_args.push_back(cast); __cvt_loc2gen_args.push_back(value);
return ctx->CallInst(__cvt_loc2gen, NULL, __cvt_loc2gen_args, "__cvt_loc2gen1_"); return ctx->CallInst(__cvt_loc2gen, NULL, __cvt_loc2gen_args, "cvtLoc2Gen");
} }
llvm::Value * llvm::Value *
@@ -1875,7 +1875,7 @@ FunctionEmitContext::PtrToIntInst(llvm::Value *value, const char *name) {
if (name == NULL) if (name == NULL)
name = LLVMGetName(value, "_ptr2int"); name = LLVMGetName(value, "_ptr2int");
value = lConvertLocalToGenericPtr(this, value); /* NVPTX */
llvm::Type *type = LLVMTypes::PointerIntType; llvm::Type *type = LLVMTypes::PointerIntType;
llvm::Instruction *inst = new llvm::PtrToIntInst(value, type, name, bblock); llvm::Instruction *inst = new llvm::PtrToIntInst(value, type, name, bblock);
AddDebugPos(inst); AddDebugPos(inst);
@@ -1909,19 +1909,7 @@ FunctionEmitContext::PtrToIntInst(llvm::Value *value, llvm::Type *toType,
} }
} }
if (value->getType()->isPointerTy() && g->target->getISA() == Target::NVPTX) value = lConvertLocalToGenericPtr(this, value); /* NVPTX */
{
llvm::PointerType *pt = llvm::dyn_cast<llvm::PointerType>(value->getType());
if (pt->getAddressSpace() == 3)
{
llvm::PointerType *PointerTy3 = llvm::PointerType::get(LLVMTypes::Int64Type, 3);
llvm::Value *cast = BitCastInst(value, PointerTy3, "__cvt_log2gen_i64ptr2_");
llvm::Function *__cvt_loc2gen = m->module->getFunction("__cvt_loc2gen");
std::vector<llvm::Value *> __cvt_loc2gen_args;
__cvt_loc2gen_args.push_back(cast);
value = CallInst(__cvt_loc2gen, NULL, __cvt_loc2gen_args, "__cvt_loc2gen2_");
}
}
llvm::Instruction *inst = new llvm::PtrToIntInst(value, toType, name, bblock); llvm::Instruction *inst = new llvm::PtrToIntInst(value, toType, name, bblock);
AddDebugPos(inst); AddDebugPos(inst);
return inst; return inst;
@@ -2500,6 +2488,7 @@ FunctionEmitContext::LoadInst(llvm::Value *ptr, const char *name) {
if (name == NULL) if (name == NULL)
name = LLVMGetName(ptr, "_load"); name = LLVMGetName(ptr, "_load");
ptr = lCorrectLocalPtr(this, ptr); /* NVPTX */
llvm::LoadInst *inst = new llvm::LoadInst(ptr, name, bblock); llvm::LoadInst *inst = new llvm::LoadInst(ptr, name, bblock);
if (g->opt.forceAlignedMemory && if (g->opt.forceAlignedMemory &&
@@ -2632,6 +2621,7 @@ FunctionEmitContext::LoadInst(llvm::Value *ptr, llvm::Value *mask,
// it's totally unaligned. (This shouldn't make any difference // it's totally unaligned. (This shouldn't make any difference
// vs the proper alignment in practice.) // vs the proper alignment in practice.)
align = 1; align = 1;
ptr = lCorrectLocalPtr(this, ptr); /* NVPTX */
llvm::Instruction *inst = new llvm::LoadInst(ptr, name, llvm::Instruction *inst = new llvm::LoadInst(ptr, name,
false /* not volatile */, false /* not volatile */,
align, bblock); align, bblock);

4
ptxtestcc.sh
View File

@@ -5,8 +5,8 @@ DIS=$HOME/usr/local/llvm/bin-3.2/bin/llvm-dis
ISPC=ispc ISPC=ispc
PTXCC=ptxcc PTXCC=ptxcc
PTXGEN=~/ptxgen PTXGEN=~/ptxgen
$(cat $1 |grep -v 'width'|$ISPC --target=nvptx --emit-llvm -o -|$LLC -march=nvptx64 -mcpu=sm_35 -o $1.ptx) && \ #$(cat $1 |grep -v 'width'|$ISPC --target=nvptx --emit-llvm -o -|$LLC -march=nvptx64 -mcpu=sm_35 -o $1.ptx) && \
#$(cat $1 |grep -v 'width'|$ISPC --target=nvptx --emit-llvm -o -|$DIS -o $1_32_ptx.ll && $PTXGEN $1_32_ptx.ll > $1.ptx) && \ $(cat $1 |grep -v 'width'|$ISPC --target=nvptx --emit-llvm -o -|$DIS -o $1_32_ptx.ll && $PTXGEN $1_32_ptx.ll > $1.ptx) && \
$($PTXCC $1.ptx -Xptxas=-v -o $1.ptx.o) && \ $($PTXCC $1.ptx -Xptxas=-v -o $1.ptx.o) && \
nvcc -o test_nvptx test_static_nvptx.cpp examples_ptx/nvcc_helpers.cu examples_ptx/ispc_malloc.cpp $1.ptx.o -arch=sm_35 -Iexamples_ptx/ -D_CUDA_ -lcudadevrt -DTEST_SIG=$2 nvcc -o test_nvptx test_static_nvptx.cpp examples_ptx/nvcc_helpers.cu examples_ptx/ispc_malloc.cpp $1.ptx.o -arch=sm_35 -Iexamples_ptx/ -D_CUDA_ -lcudadevrt -DTEST_SIG=$2