diff options
author | Bill Wendling <isanbard@gmail.com> | 2013-12-06 09:10:19 +0000 |
---|---|---|
committer | Bill Wendling <isanbard@gmail.com> | 2013-12-06 09:10:19 +0000 |
commit | 7f6926930f48234484167e9ecce90f627a030702 (patch) | |
tree | 0ef4e3c016b45cfbcdc394a3bbf1391ed521f2f2 | |
parent | aee5c3e1052b2e144fdc6461bd602cdc502a93cc (diff) | |
download | external_llvm-7f6926930f48234484167e9ecce90f627a030702.zip external_llvm-7f6926930f48234484167e9ecce90f627a030702.tar.gz external_llvm-7f6926930f48234484167e9ecce90f627a030702.tar.bz2 |
Merging r196508:
------------------------------------------------------------------------
r196508 | arnolds | 2013-12-05 07:14:40 -0800 (Thu, 05 Dec 2013) | 12 lines
SLPVectorizer: An in-tree vectorized entry cannot also be a scalar external use
We were creating external uses for scalar values in MustGather entries that also
had a ScalarToTreeEntry (they also are present in a vectorized tuple). This
meant we would keep a value 'alive' as a scalar and vectorized causing havoc.
This is not necessary because when we create a MustGather vector we explicitly
create external uses entries for the insertelement instructions of the
MustGather vector elements.
Fixes PR18129.
radar://15582184
------------------------------------------------------------------------
git-svn-id: https://llvm.org/svn/llvm-project/llvm/branches/release_34@196571 91177308-0d34-0410-b5e6-96231b3b80d8
-rw-r--r-- | lib/Transforms/Vectorize/SLPVectorizer.cpp | 6 | ||||
-rw-r--r-- | test/Transforms/SLPVectorizer/X86/external_user.ll | 35 |
2 files changed, 36 insertions, 5 deletions
diff --git a/lib/Transforms/Vectorize/SLPVectorizer.cpp b/lib/Transforms/Vectorize/SLPVectorizer.cpp index 9f18596..c72b51f 100644 --- a/lib/Transforms/Vectorize/SLPVectorizer.cpp +++ b/lib/Transforms/Vectorize/SLPVectorizer.cpp @@ -564,10 +564,8 @@ void BoUpSLP::buildTree(ArrayRef<Value *> Roots, ValueSet *Rdx) { UE = Scalar->use_end(); User != UE; ++User) { DEBUG(dbgs() << "SLP: Checking user:" << **User << ".\n"); - bool Gathered = MustGather.count(*User); - // Skip in-tree scalars that become vectors. - if (ScalarToTreeEntry.count(*User) && !Gathered) { + if (ScalarToTreeEntry.count(*User)) { DEBUG(dbgs() << "SLP: \tInternal user will be removed:" << **User << ".\n"); int Idx = ScalarToTreeEntry[*User]; (void) Idx; @@ -1638,8 +1636,6 @@ Value *BoUpSLP::vectorizeTree() { for (Value::use_iterator User = Scalar->use_begin(), UE = Scalar->use_end(); User != UE; ++User) { DEBUG(dbgs() << "SLP: \tvalidating user:" << **User << ".\n"); - assert(!MustGather.count(*User) && - "Replacing gathered value with undef"); assert((ScalarToTreeEntry.count(*User) || // It is legal to replace the reduction users by undef. diff --git a/test/Transforms/SLPVectorizer/X86/external_user.ll b/test/Transforms/SLPVectorizer/X86/external_user.ll index 22f0e64..6d09aa6 100644 --- a/test/Transforms/SLPVectorizer/X86/external_user.ll +++ b/test/Transforms/SLPVectorizer/X86/external_user.ll @@ -59,3 +59,38 @@ for.end: ; preds = %for.body ret double %mul3 } +; A need-to-gather entry cannot be an external use of the scalar element. +; Instead the insertelement instructions of the need-to-gather entry are the +; external users. +; This test would assert because we would keep the scalar fpext and fadd alive. +; PR18129 + +; CHECK-LABEL: needtogather +define i32 @needtogather(double *noalias %a, i32 *noalias %b, float * noalias %c, + i32 * noalias %d) { +entry: + %0 = load i32* %d, align 4 + %conv = sitofp i32 %0 to float + %1 = load float* %c + %sub = fsub float 0.000000e+00, %1 + %mul = fmul float %sub, 0.000000e+00 + %add = fadd float %conv, %mul + %conv1 = fpext float %add to double + %sub3 = fsub float 1.000000e+00, %1 + %mul4 = fmul float %sub3, 0.000000e+00 + %add5 = fadd float %conv, %mul4 + %conv6 = fpext float %add5 to double + %tobool = fcmp une float %add, 0.000000e+00 + br i1 %tobool, label %if.then, label %if.end + +if.then: + br label %if.end + +if.end: + %storemerge = phi double [ %conv6, %if.then ], [ %conv1, %entry ] + %e.0 = phi double [ %conv1, %if.then ], [ %conv6, %entry ] + store double %storemerge, double* %a, align 8 + %conv7 = fptosi double %e.0 to i32 + store i32 %conv7, i32* %b, align 4 + ret i32 undef +} |