diff options
author | Hal Finkel <hfinkel@anl.gov> | 2016-04-28 23:00:04 +0000 |
---|---|---|
committer | Hal Finkel <hfinkel@anl.gov> | 2016-04-28 23:00:04 +0000 |
commit | 50316d95a935161b8626e637dbc29f14bc813b77 (patch) | |
tree | 7709f674664ccf3332824a1589e45eebf7ad0a1a /llvm/lib/Transforms/Utils/InlineFunction.cpp | |
parent | 70e70e6eb93cf8e99d4284326c520463f2b880d1 (diff) | |
download | bcm5719-llvm-50316d95a935161b8626e637dbc29f14bc813b77.tar.gz bcm5719-llvm-50316d95a935161b8626e637dbc29f14bc813b77.zip |
[Inliner] Preserve llvm.mem.parallel_loop_access metadata
When inlining a call site with llvm.mem.parallel_loop_access metadata, this
metadata needs to be propagated to all cloned memory-accessing instructions.
Otherwise, inlining parts of the loop body will invalidate the annotation.
With this functionality, we now vectorize the following as expected:
void Body(int *res, int *c, int *d, int *p, int i) {
res[i] = (p[i] == 0) ? res[i] : res[i] + d[i];
}
void Test(int *res, int *c, int *d, int *p, int n) {
int i;
#pragma clang loop vectorize(assume_safety)
for (i = 0; i < 1600; i++) {
Body(res, c, d, p, i);
}
}
llvm-svn: 267949
Diffstat (limited to 'llvm/lib/Transforms/Utils/InlineFunction.cpp')
-rw-r--r-- | llvm/lib/Transforms/Utils/InlineFunction.cpp | 31 |
1 files changed, 31 insertions, 0 deletions
diff --git a/llvm/lib/Transforms/Utils/InlineFunction.cpp b/llvm/lib/Transforms/Utils/InlineFunction.cpp index 0c4ffd93dbf..5b0640c7237 100644 --- a/llvm/lib/Transforms/Utils/InlineFunction.cpp +++ b/llvm/lib/Transforms/Utils/InlineFunction.cpp @@ -688,6 +688,34 @@ static void HandleInlinedEHPad(InvokeInst *II, BasicBlock *FirstNewBlock, UnwindDest->removePredecessor(InvokeBB); } +/// When inlining a call site that has !llvm.mem.parallel_loop_access metadata, +/// that metadata should be propagated to all memory-accessing cloned +/// instructions. +static void PropagateParallelLoopAccessMetadata(CallSite CS, + ValueToValueMapTy &VMap) { + MDNode *M = + CS.getInstruction()->getMetadata(LLVMContext::MD_mem_parallel_loop_access); + if (!M) + return; + + for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end(); + VMI != VMIE; ++VMI) { + if (!VMI->second) + continue; + + Instruction *NI = dyn_cast<Instruction>(VMI->second); + if (!NI) + continue; + + if (MDNode *PM = NI->getMetadata(LLVMContext::MD_mem_parallel_loop_access)) { + M = MDNode::concatenate(PM, M); + NI->setMetadata(LLVMContext::MD_mem_parallel_loop_access, M); + } else if (NI->mayReadOrWriteMemory()) { + NI->setMetadata(LLVMContext::MD_mem_parallel_loop_access, M); + } + } +} + /// When inlining a function that contains noalias scope metadata, /// this metadata needs to be cloned so that the inlined blocks /// have different "unqiue scopes" at every call site. Were this not done, then @@ -1574,6 +1602,9 @@ bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI, // Add noalias metadata if necessary. AddAliasScopeMetadata(CS, VMap, DL, CalleeAAR); + // Propagate llvm.mem.parallel_loop_access if necessary. + PropagateParallelLoopAccessMetadata(CS, VMap); + // FIXME: We could register any cloned assumptions instead of clearing the // whole function's cache. if (IFI.ACT) |