[llvm] r196508 - SLPVectorizer: An in-tree vectorized entry cannot also be a scalar external use

[Arnold Schwaighofer]

Author: arnolds
Date: Thu Dec  5 09:14:40 2013
New Revision: 196508

URL: http://llvm.org/viewvc/llvm-project?rev=196508&view=rev
Log:
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

Modified:
    llvm/trunk/lib/Transforms/Vectorize/SLPVectorizer.cpp
    llvm/trunk/test/Transforms/SLPVectorizer/X86/external_user.ll

Modified: llvm/trunk/lib/Transforms/Vectorize/SLPVectorizer.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/Vectorize/SLPVectorizer.cpp?rev=196508&r1=196507&r2=196508&view=diff
==============================================================================
--- llvm/trunk/lib/Transforms/Vectorize/SLPVectorizer.cpp (original)
+++ llvm/trunk/lib/Transforms/Vectorize/SLPVectorizer.cpp Thu Dec  5 09:14:40 2013
@@ -564,10 +564,8 @@ void BoUpSLP::buildTree(ArrayRef<Value *
            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;
@@ -1641,8 +1639,6 @@ 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.

Modified: llvm/trunk/test/Transforms/SLPVectorizer/X86/external_user.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/SLPVectorizer/X86/external_user.ll?rev=196508&r1=196507&r2=196508&view=diff
==============================================================================
--- llvm/trunk/test/Transforms/SLPVectorizer/X86/external_user.ll (original)
+++ llvm/trunk/test/Transforms/SLPVectorizer/X86/external_user.ll Thu Dec  5 09:14:40 2013
@@ -59,3 +59,38 @@ for.end:
   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
+}