[llvm-branch-commits] [llvm-branch] r196571 - Merging r196508:

Bill Wendling isanbard at gmail.com
Fri Dec 6 01:10:19 PST 2013


Author: void
Date: Fri Dec  6 03:10:19 2013
New Revision: 196571

URL: http://llvm.org/viewvc/llvm-project?rev=196571&view=rev
Log:
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
------------------------------------------------------------------------

Modified:
    llvm/branches/release_34/   (props changed)
    llvm/branches/release_34/lib/Transforms/Vectorize/SLPVectorizer.cpp
    llvm/branches/release_34/test/Transforms/SLPVectorizer/X86/external_user.ll

Propchange: llvm/branches/release_34/
------------------------------------------------------------------------------
--- svn:mergeinfo (original)
+++ svn:mergeinfo Fri Dec  6 03:10:19 2013
@@ -1,3 +1,3 @@
 /llvm/branches/Apple/Pertwee:110850,110961
 /llvm/branches/type-system-rewrite:133420-134817
-/llvm/trunk:155241,195092-195094,195100,195102-195103,195118,195129,195136,195138,195148,195152,195156-195157,195161-195162,195193,195272,195317-195318,195327,195330,195333,195339,195343,195355,195364,195379,195397-195399,195401,195408,195421,195423-195424,195432,195439,195444,195455-195456,195469,195476-195477,195479,195491-195493,195514,195528,195547,195567,195573-195576,195590-195591,195599,195632,195635-195636,195670,195677,195679,195682,195684,195713,195716,195769,195773,195779,195782,195787-195788,195791,195803,195812,195827,195834,195843-195844,195878-195881,195887,195903,195905,195912,195915,195932,195936-195943,195972-195973,195975-195976,196004,196044-196046,196069,196100,196104,196129,196151,196153,196267
+/llvm/trunk:155241,195092-195094,195100,195102-195103,195118,195129,195136,195138,195148,195152,195156-195157,195161-195162,195193,195272,195317-195318,195327,195330,195333,195339,195343,195355,195364,195379,195397-195399,195401,195408,195421,195423-195424,195432,195439,195444,195455-195456,195469,195476-195477,195479,195491-195493,195514,195528,195547,195567,195573-195576,195590-195591,195599,195632,195635-195636,195670,195677,195679,195682,195684,195713,195716,195769,195773,195779,195782,195787-195788,195791,195803,195812,195827,195834,195843-195844,195878-195881,195887,195903,195905,195912,195915,195932,195936-195943,195972-195973,195975-195976,196004,196044-196046,196069,196100,196104,196129,196151,196153,196267,196508

Modified: llvm/branches/release_34/lib/Transforms/Vectorize/SLPVectorizer.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/branches/release_34/lib/Transforms/Vectorize/SLPVectorizer.cpp?rev=196571&r1=196570&r2=196571&view=diff
==============================================================================
--- llvm/branches/release_34/lib/Transforms/Vectorize/SLPVectorizer.cpp (original)
+++ llvm/branches/release_34/lib/Transforms/Vectorize/SLPVectorizer.cpp Fri Dec  6 03:10:19 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;
@@ -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.

Modified: llvm/branches/release_34/test/Transforms/SLPVectorizer/X86/external_user.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/branches/release_34/test/Transforms/SLPVectorizer/X86/external_user.ll?rev=196571&r1=196570&r2=196571&view=diff
==============================================================================
--- llvm/branches/release_34/test/Transforms/SLPVectorizer/X86/external_user.ll (original)
+++ llvm/branches/release_34/test/Transforms/SLPVectorizer/X86/external_user.ll Fri Dec  6 03:10:19 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
+}





More information about the llvm-branch-commits mailing list