[llvm-branch-commits] [llvm-branch] r341523 - Merging r341416:

[Hans Wennborg via llvm-branch-commits]

Author: hans
Date: Thu Sep  6 01:16:34 2018
New Revision: 341523

URL: http://llvm.org/viewvc/llvm-project?rev=341523&view=rev
Log:
Merging r341416:
------------------------------------------------------------------------
r341416 | annat | 2018-09-05 00:12:23 +0200 (Wed, 05 Sep 2018) | 11 lines

[LV] First order recurrence phis should not be treated as uniform

This is fix for PR38786.
First order recurrence phis were incorrectly treated as uniform,
which caused them to be vectorized as uniform instructions.

Patch by Ayal Zaks and Orivej Desh!

Reviewed by: Anna

Differential Revision: https://reviews.llvm.org/D51639
------------------------------------------------------------------------

Modified:
    llvm/branches/release_70/   (props changed)
    llvm/branches/release_70/lib/Transforms/Vectorize/LoopVectorize.cpp
    llvm/branches/release_70/test/Transforms/LoopVectorize/X86/uniform-phi.ll

Propchange: llvm/branches/release_70/
------------------------------------------------------------------------------
--- svn:mergeinfo (original)
+++ svn:mergeinfo Thu Sep  6 01:16:34 2018
@@ -1,3 +1,3 @@
 /llvm/branches/Apple/Pertwee:110850,110961
 /llvm/branches/type-system-rewrite:133420-134817
-/llvm/trunk:155241,338552,338554,338569,338599,338610,338658,338665,338682,338703,338709,338716,338751,338762,338817,338841,338902,338915,338968,339073,339091,339166,339179,339184,339190,339225,339316,339319,339411,339492,339515,339533,339535-339536,339600,339636,339674,339769,339822,339883,339895-339896,339945,340158,340303,340416-340417,340455,340641,340691,340751,340820,340839,340900,340959,341094,341244
+/llvm/trunk:155241,338552,338554,338569,338599,338610,338658,338665,338682,338703,338709,338716,338751,338762,338817,338841,338902,338915,338968,339073,339091,339166,339179,339184,339190,339225,339316,339319,339411,339492,339515,339533,339535-339536,339600,339636,339674,339769,339822,339883,339895-339896,339945,340158,340303,340416-340417,340455,340641,340691,340751,340820,340839,340900,340959,341094,341244,341416

Modified: llvm/branches/release_70/lib/Transforms/Vectorize/LoopVectorize.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/branches/release_70/lib/Transforms/Vectorize/LoopVectorize.cpp?rev=341523&r1=341522&r2=341523&view=diff
==============================================================================
--- llvm/branches/release_70/lib/Transforms/Vectorize/LoopVectorize.cpp (original)
+++ llvm/branches/release_70/lib/Transforms/Vectorize/LoopVectorize.cpp Thu Sep  6 01:16:34 2018
@@ -4510,6 +4510,13 @@ void LoopVectorizationCostModel::collect
     for (auto OV : I->operand_values()) {
       if (isOutOfScope(OV))
         continue;
+      // First order recurrence Phi's should typically be considered
+      // non-uniform.
+      auto *OP = dyn_cast<PHINode>(OV);
+      if (OP && Legal->isFirstOrderRecurrence(OP))
+        continue;
+      // If all the users of the operand are uniform, then add the
+      // operand into the uniform worklist.
       auto *OI = cast<Instruction>(OV);
       if (llvm::all_of(OI->users(), [&](User *U) -> bool {
             auto *J = cast<Instruction>(U);

Modified: llvm/branches/release_70/test/Transforms/LoopVectorize/X86/uniform-phi.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/branches/release_70/test/Transforms/LoopVectorize/X86/uniform-phi.ll?rev=341523&r1=341522&r2=341523&view=diff
==============================================================================
--- llvm/branches/release_70/test/Transforms/LoopVectorize/X86/uniform-phi.ll (original)
+++ llvm/branches/release_70/test/Transforms/LoopVectorize/X86/uniform-phi.ll Thu Sep  6 01:16:34 2018
@@ -75,3 +75,25 @@ for.end:
   ret i64 %retval
 }
 
+; CHECK-LABEL: PR38786
+; Check that first order recurrence phis (%phi32 and %phi64) are not uniform.
+; CHECK-NOT: LV: Found uniform instruction:   %phi
+define void @PR38786(double* %y, double* %x, i64 %n) {
+entry:
+  br label %for.body
+
+for.body:
+  %phi32 = phi i32 [ 0, %entry ], [ %i32next, %for.body ]
+  %phi64 = phi i64 [ 0, %entry ], [ %i64next, %for.body ]
+  %i32next = add i32 %phi32, 1
+  %i64next = zext i32 %i32next to i64
+  %xip = getelementptr inbounds double, double* %x, i64 %i64next
+  %yip = getelementptr inbounds double, double* %y, i64 %phi64
+  %xi = load double, double* %xip, align 8
+  store double %xi, double* %yip, align 8
+  %cmp = icmp slt i64 %i64next, %n
+  br i1 %cmp, label %for.body, label %for.end
+
+for.end:
+  ret void
+}