[PATCH] D23513: Fix assert in pipeliner due to an invalid Phi in the epilog

Brendon Cahoon via llvm-commits llvm-commits at lists.llvm.org
Tue Aug 16 07:37:30 PDT 2016


This revision was automatically updated to reflect the committed changes.
Closed by commit rL278805: [Pipeliner] Fix an asssert due to invalid Phi in the epilog (authored by bcahoon).

Changed prior to commit:
  https://reviews.llvm.org/D23513?vs=68041&id=68175#toc

Repository:
  rL LLVM

https://reviews.llvm.org/D23513

Files:
  llvm/trunk/lib/CodeGen/MachinePipeliner.cpp
  llvm/trunk/test/CodeGen/Hexagon/swp-epilog-reuse-1.ll

Index: llvm/trunk/lib/CodeGen/MachinePipeliner.cpp
===================================================================
--- llvm/trunk/lib/CodeGen/MachinePipeliner.cpp
+++ llvm/trunk/lib/CodeGen/MachinePipeliner.cpp
@@ -2648,8 +2648,7 @@
       // references another Phi, and the other Phi is scheduled in an
       // earlier stage. We can try to reuse an existing Phi up until the last
       // stage of the current Phi.
-      if (LoopDefIsPhi && VRMap[CurStageNum].count(LoopVal) &&
-          LoopValStage >= (int)(CurStageNum - LastStageNum)) {
+      if (LoopDefIsPhi && (int)PrologStage >= StageScheduled) {
         int LVNumStages = Schedule.getStagesForPhi(LoopVal);
         int StageDiff = (StageScheduled - LoopValStage);
         LVNumStages -= StageDiff;
Index: llvm/trunk/test/CodeGen/Hexagon/swp-epilog-reuse-1.ll
===================================================================
--- llvm/trunk/test/CodeGen/Hexagon/swp-epilog-reuse-1.ll
+++ llvm/trunk/test/CodeGen/Hexagon/swp-epilog-reuse-1.ll
@@ -0,0 +1,44 @@
+; RUN: llc -march=hexagon -mcpu=hexagonv60 < %s
+; REQUIRES: asserts
+
+; Test that the pipeliner reuses an existing Phi when generating the epilog
+; block. In this case, the original loops has a Phi whose operand is another
+; Phi. When the loop is pipelined, the Phi that generates the operand value
+; is used in two stages. This means the the Phi for the second stage can
+; be reused. The bug causes an assert due to an invalid virtual register error
+; in the live variable analysis.
+
+define void @test(i8* %a, i8* %b)  #0 {
+entry:
+  br label %for.body6.us.prol
+
+for.body6.us.prol:
+  %i.065.us.prol = phi i32 [ 0, %entry ], [ %inc.us.prol, %for.body6.us.prol ]
+  %im1.064.us.prol = phi i32 [ undef, %entry ], [ %i.065.us.prol, %for.body6.us.prol ]
+  %prol.iter = phi i32 [ undef, %entry ], [ %prol.iter.sub, %for.body6.us.prol ]
+  %arrayidx8.us.prol = getelementptr inbounds i8, i8* %b, i32 %im1.064.us.prol
+  %0 = load i8, i8* %arrayidx8.us.prol, align 1
+  %conv9.us.prol = sext i8 %0 to i32
+  %add.us.prol = add nsw i32 %conv9.us.prol, 0
+  %add12.us.prol = add nsw i32 %add.us.prol, 0
+  %mul.us.prol = mul nsw i32 %add12.us.prol, 3
+  %conv13.us.prol = trunc i32 %mul.us.prol to i8
+  %arrayidx14.us.prol = getelementptr inbounds i8, i8* %a, i32 %i.065.us.prol
+  store i8 %conv13.us.prol, i8* %arrayidx14.us.prol, align 1
+  %inc.us.prol = add nuw nsw i32 %i.065.us.prol, 1
+  %prol.iter.sub = add i32 %prol.iter, -1
+  %prol.iter.cmp = icmp eq i32 %prol.iter.sub, 0
+  br i1 %prol.iter.cmp, label %for.body6.us, label %for.body6.us.prol
+
+for.body6.us:
+  %im2.063.us = phi i32 [ undef, %for.body6.us ], [ %im1.064.us.prol, %for.body6.us.prol ]
+  %arrayidx10.us = getelementptr inbounds i8, i8* %b, i32 %im2.063.us
+  %1 = load i8, i8* %arrayidx10.us, align 1
+  %conv11.us = sext i8 %1 to i32
+  %add12.us = add nsw i32 0, %conv11.us
+  %mul.us = mul nsw i32 %add12.us, 3
+  %conv13.us = trunc i32 %mul.us to i8
+  store i8 %conv13.us, i8* undef, align 1
+  br label %for.body6.us
+}
+


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D23513.68175.patch
Type: text/x-patch
Size: 3049 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20160816/59998d50/attachment.bin>


More information about the llvm-commits mailing list