diff options
author | Tom Stellard <tstellar@redhat.com> | 2018-04-10 02:39:11 +0000 |
---|---|---|
committer | Tom Stellard <tstellar@redhat.com> | 2018-04-10 02:39:11 +0000 |
commit | 54d8b1a9f397a6ead6442391bc86ef147aaad378 (patch) | |
tree | 538103c1949e003c4ce262b78b6c3f6added5589 | |
parent | 983d8003779bab5477054d7f83a2a1b9e46c355d (diff) |
Merging r328748:
------------------------------------------------------------------------
r328748 | gbiv | 2018-03-28 17:54:39 -0700 (Wed, 28 Mar 2018) | 12 lines
[MemorySSA] Consider callsite args for hashing and equality.
We use a `DenseMap<MemoryLocOrCall, MemlocStackInfo>` to keep track of
prior work when optimizing uses in MemorySSA. Because we weren't
accounting for callsite arguments in either the hash code or equality
tests for `MemoryLocOrCall`s, we optimized uses too aggressively in
some rare cases.
Fix by Daniel Berlin.
Should fix PR36883.
------------------------------------------------------------------------
git-svn-id: https://llvm.org/svn/llvm-project/llvm/branches/release_60@329663 91177308-0d34-0410-b5e6-96231b3b80d8
-rw-r--r-- | lib/Analysis/MemorySSA.cpp | 29 | ||||
-rw-r--r-- | test/Analysis/MemorySSA/pr36883.ll | 38 |
2 files changed, 58 insertions, 9 deletions
diff --git a/lib/Analysis/MemorySSA.cpp b/lib/Analysis/MemorySSA.cpp index 6e9368c49d6..137b2022e84 100644 --- a/lib/Analysis/MemorySSA.cpp +++ b/lib/Analysis/MemorySSA.cpp @@ -153,9 +153,14 @@ public: if (IsCall != Other.IsCall) return false; - if (IsCall) - return CS.getCalledValue() == Other.CS.getCalledValue(); - return Loc == Other.Loc; + if (!IsCall) + return Loc == Other.Loc; + + if (CS.getCalledValue() != Other.CS.getCalledValue()) + return false; + + assert(CS.arg_size() == Other.CS.arg_size()); + return std::equal(CS.arg_begin(), CS.arg_end(), Other.CS.arg_begin()); } private: @@ -179,12 +184,18 @@ template <> struct DenseMapInfo<MemoryLocOrCall> { } static unsigned getHashValue(const MemoryLocOrCall &MLOC) { - if (MLOC.IsCall) - return hash_combine(MLOC.IsCall, - DenseMapInfo<const Value *>::getHashValue( - MLOC.getCS().getCalledValue())); - return hash_combine( - MLOC.IsCall, DenseMapInfo<MemoryLocation>::getHashValue(MLOC.getLoc())); + if (!MLOC.IsCall) + return hash_combine( + MLOC.IsCall, + DenseMapInfo<MemoryLocation>::getHashValue(MLOC.getLoc())); + + hash_code hash = + hash_combine(MLOC.IsCall, DenseMapInfo<const Value *>::getHashValue( + MLOC.getCS().getCalledValue())); + + for (const Value *Arg : MLOC.getCS().args()) + hash = hash_combine(hash, DenseMapInfo<const Value *>::getHashValue(Arg)); + return hash; } static bool isEqual(const MemoryLocOrCall &LHS, const MemoryLocOrCall &RHS) { diff --git a/test/Analysis/MemorySSA/pr36883.ll b/test/Analysis/MemorySSA/pr36883.ll new file mode 100644 index 00000000000..8411b0c228b --- /dev/null +++ b/test/Analysis/MemorySSA/pr36883.ll @@ -0,0 +1,38 @@ +; RUN: opt -basicaa -memoryssa -analyze < %s 2>&1 -S | FileCheck %s +; RUN: opt -aa-pipeline=basic-aa -passes='print<memoryssa>,verify<memoryssa>' -S < %s 2>&1 | FileCheck %s +; +; We weren't properly considering the args in callsites in equality or hashing. + +target triple = "armv7-dcg-linux-gnueabi" + +; CHECK-LABEL: define <8 x i16> @vpx_idct32_32_neon +define <8 x i16> @vpx_idct32_32_neon(i8* %p, <8 x i16> %v) { +entry: +; CHECK: MemoryUse(liveOnEntry) + %load1 = call <8 x i16> @llvm.arm.neon.vld1.v8i16.p0i8(i8* %p, i32 2) #4 ; load CSE replacement + +; CHECK: 1 = MemoryDef(liveOnEntry) + call void @llvm.arm.neon.vst1.p0i8.v8i16(i8* %p, <8 x i16> %v, i32 2) #4 ; clobber + + %p_next = getelementptr inbounds i8, i8* %p, i32 16 +; CHECK: MemoryUse(liveOnEntry) + %load2 = call <8 x i16> @llvm.arm.neon.vld1.v8i16.p0i8(i8* %p_next, i32 2) #4 ; non-aliasing load needed to trigger bug + +; CHECK: MemoryUse(1) + %load3 = call <8 x i16> @llvm.arm.neon.vld1.v8i16.p0i8(i8* %p, i32 2) #4 ; load CSE removed + + %add = add <8 x i16> %load1, %load2 + %ret = add <8 x i16> %add, %load3 + ret <8 x i16> %ret +} + +; Function Attrs: argmemonly nounwind readonly +declare <8 x i16> @llvm.arm.neon.vld1.v8i16.p0i8(i8*, i32) #2 + +; Function Attrs: argmemonly nounwind +declare void @llvm.arm.neon.vst1.p0i8.v8i16(i8*, <8 x i16>, i32) #1 + +attributes #1 = { argmemonly nounwind } +attributes #2 = { argmemonly nounwind readonly } +attributes #3 = { nounwind readnone } +attributes #4 = { nounwind } |