[llvm] r193437 - Fix LSR: don't normalize quadratic recurrences.

[Andrew Trick]

Author: atrick
Date: Fri Oct 25 16:35:52 2013
New Revision: 193437

URL: http://llvm.org/viewvc/llvm-project?rev=193437&view=rev
Log:
Fix LSR: don't normalize quadratic recurrences.

Partial fix for PR17459: wrong code at -O3 on x86_64-linux-gnu
(affecting trunk and 3.3)

ScalarEvolutionNormalization was attempting to normalize by adding and
subtracting strides. Chained recurrences don't work that way.

Modified:
    llvm/trunk/lib/Analysis/ScalarEvolutionNormalization.cpp
    llvm/trunk/test/Transforms/LoopStrengthReduce/quadradic-exit-value.ll

Modified: llvm/trunk/lib/Analysis/ScalarEvolutionNormalization.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Analysis/ScalarEvolutionNormalization.cpp?rev=193437&r1=193436&r2=193437&view=diff
==============================================================================
--- llvm/trunk/lib/Analysis/ScalarEvolutionNormalization.cpp (original)
+++ llvm/trunk/lib/Analysis/ScalarEvolutionNormalization.cpp Fri Oct 25 16:35:52 2013
@@ -119,11 +119,19 @@ TransformImpl(const SCEV *S, Instruction
     const SCEV *Result = SE.getAddRecExpr(Operands, L, SCEV::FlagAnyWrap);
     switch (Kind) {
     case NormalizeAutodetect:
-      if (IVUseShouldUsePostIncValue(User, OperandValToReplace, L, &DT)) {
-        const SCEV *TransformedStep =
-          TransformSubExpr(AR->getStepRecurrence(SE),
-                           User, OperandValToReplace);
-        Result = SE.getMinusSCEV(Result, TransformedStep);
+      // Normalize this SCEV by subtracting the expression for the final step.
+      // We only allow affine AddRecs to be normalized, otherwise we would not
+      // be able to correctly denormalize.
+      // e.g. {1,+,3,+,2} == {-2,+,1,+,2} + {3,+,2}
+      // Normalized form:   {-2,+,1,+,2}
+      // Denormalized form: {1,+,3,+,2}
+      //
+      // However, denormalization would use the a different step expression than
+      // normalization (see getPostIncExpr), generating the wrong final
+      // expression: {-2,+,1,+,2} + {1,+,2} => {-1,+,3,+,2}
+      if (AR->isAffine() &&
+          IVUseShouldUsePostIncValue(User, OperandValToReplace, L, &DT)) {
+        Result = SE.getMinusSCEV(Result, AR->getStepRecurrence(SE));
         Loops.insert(L);
       }
 #if 0

Modified: llvm/trunk/test/Transforms/LoopStrengthReduce/quadradic-exit-value.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/LoopStrengthReduce/quadradic-exit-value.ll?rev=193437&r1=193436&r2=193437&view=diff
==============================================================================
--- llvm/trunk/test/Transforms/LoopStrengthReduce/quadradic-exit-value.ll (original)
+++ llvm/trunk/test/Transforms/LoopStrengthReduce/quadradic-exit-value.ll Fri Oct 25 16:35:52 2013
@@ -1,18 +1,50 @@
-; RUN: opt < %s -analyze -iv-users | grep "{1,+,3,+,2}<%loop> (post-inc with loop %loop)"
+; RUN: opt < %s -analyze -iv-users | FileCheck %s
 
 ; The value of %r is dependent on a polynomial iteration expression.
-
+;
+; CHECK-LABEL: IV Users for loop %foo.loop
+; CHECK: {1,+,3,+,2}<%foo.loop>
 define i64 @foo(i64 %n) {
 entry:
-  br label %loop
+  br label %foo.loop
 
-loop:
-  %indvar = phi i64 [ 0, %entry ], [ %indvar.next, %loop ]
+foo.loop:
+  %indvar = phi i64 [ 0, %entry ], [ %indvar.next, %foo.loop ]
   %indvar.next = add i64 %indvar, 1
   %c = icmp eq i64 %indvar.next, %n
-  br i1 %c, label %exit, label %loop
+  br i1 %c, label %exit, label %foo.loop
 
 exit:
   %r = mul i64 %indvar.next, %indvar.next
   ret i64 %r
 }
+
+; PR15470: LSR miscompile. The test2 function should return '1'.
+;
+; SCEV does not know how to denormalize chained recurrences, so make
+; sure they aren't marked as post-inc users.
+;
+; CHECK-LABEL: IV Users for loop %test2.loop
+; CHECK: %sext.us = {0,+,(16777216 + (-16777216 * %sub.us)),+,33554432}<%test2.loop> in %f = ashr i32 %sext.us, 24
+define i32 @test2() {
+entry:
+  br label %test2.loop
+
+test2.loop:
+  %inc1115.us = phi i32 [ 0, %entry ], [ %inc11.us, %test2.loop ]
+  %inc11.us = add nsw i32 %inc1115.us, 1
+  %cmp.us = icmp slt i32 %inc11.us, 2
+  br i1 %cmp.us, label %test2.loop, label %for.end
+
+for.end:
+  %tobool.us = icmp eq i32 %inc1115.us, 0
+  %sub.us = select i1 %tobool.us, i32 0, i32 0
+  %mul.us = shl i32 %inc1115.us, 24
+  %sub.cond.us = sub nsw i32 %inc1115.us, %sub.us
+  %sext.us = mul i32 %mul.us, %sub.cond.us
+  %f = ashr i32 %sext.us, 24
+  br label %exit
+
+exit:
+  ret i32 %f
+}