[PATCH] D41467: PR35710: Nary reassociation falls into infinite loop

Evgeny Stupachenko via Phabricator via llvm-commits llvm-commits at lists.llvm.org
Wed Dec 20 14:50:02 PST 2017


evstupac created this revision.
evstupac added reviewers: jingyue, marcello.maggioni.

The following example:
define i8 @foo(i8 %v) local_unnamed_addr #0 { 
region.0:

  %0 = mul nsw i8 16, %v 
  %1 = mul nsw i8 0, %0 
  %2 = mul nsw i8 1, %1 
  ret i8 %2 

} 
Causes nary reassociation to fall into infinite loop.
The root cause is in overflowed SCEV expression (16*v)*(16*v) - which is 0 for "i8" type.

The patch inserts a check to avoid the case.


Repository:
  rL LLVM

https://reviews.llvm.org/D41467

Files:
  lib/Transforms/Scalar/NaryReassociate.cpp
  test/Transforms/NaryReassociate/pr35710.ll


Index: lib/Transforms/Scalar/NaryReassociate.cpp
===================================================================
--- lib/Transforms/Scalar/NaryReassociate.cpp
+++ lib/Transforms/Scalar/NaryReassociate.cpp
@@ -463,6 +463,12 @@
 Instruction *NaryReassociatePass::tryReassociatedBinaryOp(const SCEV *LHSExpr,
                                                           Value *RHS,
                                                           BinaryOperator *I) {
+  // Need to avoid special case (C1*v)*(C2*v) = 0 because of overflow and
+  // there is Instruction "mul x, 0", as it leads to infinite loop
+  if (I->getOpcode() == Instruction::Mul && LHSExpr->isZero() &&
+      SE->getSCEV(RHS)->isZero())
+    return nullptr;
+
   // Look for the closest dominator LHS of I that computes LHSExpr, and replace
   // I with LHS op RHS.
   auto *LHS = findClosestMatchingDominator(LHSExpr, I);
Index: test/Transforms/NaryReassociate/pr35710.ll
===================================================================
--- test/Transforms/NaryReassociate/pr35710.ll
+++ test/Transforms/NaryReassociate/pr35710.ll
@@ -0,0 +1,19 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -nary-reassociate -S | FileCheck %s
+
+; The test check that compilation does not fall into infinite loop.
+
+define i8 @foo(i8 %v) local_unnamed_addr #0 {
+; CHECK-LABEL: @foo(
+; CHECK-NEXT:  region.0:
+; CHECK-NEXT:    [[TMP0:%.*]] = mul nsw i8 16, [[V:%.*]]
+; CHECK-NEXT:    [[TMP1:%.*]] = mul nsw i8 0, [[TMP0]]
+; CHECK-NEXT:    [[TMP2:%.*]] = mul i8 [[TMP1]], [[TMP0]]
+; CHECK-NEXT:    ret i8 [[TMP2]]
+;
+region.0:
+  %0 = mul nsw i8 16, %v
+  %1 = mul nsw i8 0, %0
+  %2 = mul nsw i8 1, %1
+  ret i8 %2
+}


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D41467.127789.patch
Type: text/x-patch
Size: 1729 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20171220/dcbbfbbb/attachment.bin>


More information about the llvm-commits mailing list