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

Fixed a number of issues related to memory alignment; a number of places

Browse Source 
were expecting vector-width-aligned pointers where in point of fact,
there's no guarantee that they would have been in general.

Removed the aligned memory allocation routines from some of the examples;
they're no longer needed.

No perf. difference on Core2/Core i5 CPUs; older CPUs may see some
regressions.

Still need to update the documentation for this change and finish reviewing
alignment issues in Load/Store instructions generated by .cpp files.
This commit is contained in:
Matt Pharr
2011-06-23 18:18:33 -07:00
parent d340dcbfcc
commit b84167dddd
11 changed files with 45 additions and 112 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
opt.cpp
View File

@@ -1131,10 +1131,17 @@ MaskedStoreOptPass::runOnBasicBlock(llvm::BasicBlock &bb) {
}
else if (maskAsInt == allOnMask) {
// The mask is all on, so turn this into a regular store
const llvm::Type *ptrType = llvm::PointerType::get(rvalue->getType(), 0);
const llvm::Type *rvalueType = rvalue->getType();
const llvm::Type *ptrType = llvm::PointerType::get(rvalueType, 0);
// Need to update this when int8/int16 are added
int align = (called == pms32Func || called == pms64Func ||
called == msb32Func) ? 4 : 8;
lvalue = new llvm::BitCastInst(lvalue, ptrType, "lvalue_to_ptr_type", callInst);
lCopyMetadata(lvalue, callInst);
llvm::Instruction *store = new llvm::StoreInst(rvalue, lvalue);
llvm::Instruction *store =
new llvm::StoreInst(rvalue, lvalue, false /* not volatile */,
align);
lCopyMetadata(store, callInst);
llvm::ReplaceInstWithInst(callInst, store);