[PATCH] D87434: [CodeGenPrepare] Fix zapping dead operands of assume

Yevgeny Rouban via Phabricator via llvm-commits llvm-commits at lists.llvm.org
Wed Sep 9 22:46:09 PDT 2020


yrouban created this revision.
yrouban added reviewers: bkramer, lattner, Carrot.
Herald added a subscriber: hiraditya.
Herald added a project: LLVM.
yrouban requested review of this revision.

This patch fixes a problem of the commit
52cc97a0 <https://github.com/llvm/llvm-project/commit/52cc97a0> Zap the argument of llvm.assume when deleting it
A test case is created to demonstrate the crash caused by the instruction iterator invalidated by the recursive removal of dead operands of assume. The patch fixes //CurInstIterator// by taking its value from eraseFromParent().
Another solution would be to restart from the first instruction in case //CurInstIterator// is invalidated by //RecursivelyDeleteTriviallyDeadInstructions()// as it is done at the end of //optimizeMemoryInst()// (see the commit af1bcce <https://github.com/llvm/llvm-project/commit/af1bcce>).


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D87434

Files:
  llvm/lib/CodeGen/CodeGenPrepare.cpp
  llvm/test/Transforms/CodeGenPrepare/X86/recursively-delete-dead-instructions.ll


Index: llvm/test/Transforms/CodeGenPrepare/X86/recursively-delete-dead-instructions.ll
===================================================================
--- /dev/null
+++ llvm/test/Transforms/CodeGenPrepare/X86/recursively-delete-dead-instructions.ll
@@ -0,0 +1,27 @@
+; RUN: opt -codegenprepare -S -mtriple=x86_64-linux < %s | FileCheck %s
+
+declare void @llvm.assume(i1 noundef) nounwind willreturn
+
+; Recursively deleting dead operands of assume() may result in its next
+; instruction deleted and the iterator pointing to the next instruction
+; invalidated. This prevents the following simple loop in
+; CodeGenPrepare::optimizeBlock() unless CurInstIterator is fixed:
+;
+;   CurInstIterator = BB.begin();
+;   while (CurInstIterator != BB.end())
+;     optimizeInst(&*CurInstIterator++, ModifiedDT);
+;
+define i32 @test_assume_in_loop(i1 %cond1, i1 %cond2) {
+; CHECK-LABEL: @test_assume_in_loop(
+; CHECK-NEXT:  entry:
+entry:
+  br label %loop
+
+; CHECK: loop:
+; CHECK-NEXT:  br label %loop
+loop:
+  %cond3 = phi i1 [%cond1, %entry], [%cond4, %loop]
+  call void @llvm.assume(i1 %cond3)
+  %cond4 = icmp ult i1 %cond1, %cond2
+  br label %loop
+}
Index: llvm/lib/CodeGen/CodeGenPrepare.cpp
===================================================================
--- llvm/lib/CodeGen/CodeGenPrepare.cpp
+++ llvm/lib/CodeGen/CodeGenPrepare.cpp
@@ -2045,11 +2045,14 @@
     default: break;
     case Intrinsic::assume: {
       Value *Operand = II->getOperand(0);
-      II->eraseFromParent();
+      II->setOperand(0, nullptr);
       // Prune the operand, it's most likely dead.
       RecursivelyDeleteTriviallyDeadInstructions(
-          Operand, TLInfo, nullptr,
-          [&](Value *V) { removeAllAssertingVHReferences(V); });
+          Operand, TLInfo, nullptr, [&](Value *V) {
+            assert(V != II && "initial assume must persist");
+            removeAllAssertingVHReferences(V);
+          });
+      CurInstIterator = II->eraseFromParent();
       return true;
     }
 


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D87434.290873.patch
Type: text/x-patch
Size: 2002 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20200910/ad7b28cc/attachment.bin>


More information about the llvm-commits mailing list