fix bug for useless malloc elimination in CodeGenPrepare

[drcut]

Summary

Put AllocationFn check before I->willReturn can allow CodeGenPrepare to remove useless malloc instruction

[aeubanks]

llvm/test/CodeGen/Generic/dead-allocation.ll

19

add a newline at the end

[nikic]

Please put the test in llvm/test/Transforms/CodeGenPrepare instead and use update_test_checks.py. Omit the target triple if possible, which it likely is.

[drcut]

Reformat the test file with update_test_checks.py

[drcut]

In D130126#3663981, @nikic wrote:

Please put the test in llvm/test/Transforms/CodeGenPrepare instead and use update_test_checks.py. Omit the target triple if possible, which it likely is.

Thanks for the suggestions. The target triple is required for codegenprepare pass.

[drcut]

This comment was removed by drcut.

[craig.topper]

llvm/test/Transforms/CodeGenPrepare/dead-allocation.ll

1 ↗

(On Diff #446178)

This test needs REQUIRES: arm-registered-target otherwise it fails if the ARM target isn't built

[dyung]

The test you added is failing on the PS4 Windows/Linux bots likely for the reason that Craig pointed out above. Please take a look and revert if you need time to investigate.

https://lab.llvm.org/buildbot/#/builders/139/builds/25268
https://lab.llvm.org/staging/#/builders/204/builds/1677

[craig.topper]

In D130126#3666262, @dyung wrote:

The test you added is failing on the PS4 Windows/Linux bots likely for the reason that Craig pointed out above. Please take a look and revert if you need time to investigate.

https://lab.llvm.org/buildbot/#/builders/139/builds/25268
https://lab.llvm.org/staging/#/builders/204/builds/1677

I'll go ahead and push the needed REQUIRES directive in a couple minutes

[aeubanks]

sorry for that, fixed in https://github.com/llvm/llvm-project/commit/19d4f5e6495ce7e03b803235b648b915d5745314

llvm/lib/Transforms/Utils/Local.cpp

if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(I)) {
     return true;
   }
   if (DbgLabelInst *DLI = dyn_cast<DbgLabelInst>(I)) {
     if (DLI->getLabel())
       return false;
     return true;
   }
 
+  if (isAllocationFn(I, TLI) && isAllocRemovable(cast<CallBase>(I), TLI))
+    return true;
+
   if (!I->willReturn())
     return false;
 
   if (!I->mayHaveSideEffects())
     return true;
 
   // Special case intrinsics that "may have side effects" but can be deleted
   // when dead.
if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
     }
 
     if (auto *FPI = dyn_cast<ConstrainedFPIntrinsic>(I)) {
       Optional<fp::ExceptionBehavior> ExBehavior = FPI->getExceptionBehavior();
       return *ExBehavior != fp::ebStrict;
     }
   }
 
-  if (isAllocationFn(I, TLI) && isAllocRemovable(cast<CallBase>(I), TLI))
-    return true;
-
   if (CallInst *CI = isFreeCall(I, TLI))
     if (Constant *C = dyn_cast<Constant>(CI->getArgOperand(0)))
       return C->isNullValue() || isa<UndefValue>(C);
 
   if (auto *Call = dyn_cast<CallBase>(I))
     if (isMathLibCallNoop(Call, TLI))
       return true;
 

llvm/test/CodeGen/Generic/dead-allocation.ll

+; Eliminate the dead allocation instruction
+; RUN: opt -mtriple arm64-apple-darwin -codegenprepare < %s -S | FileCheck %s
+
+@.str = private unnamed_addr constant [2 x i8] c"a\00", align 1
+
+define void @func() {
+; CHECK-LABEL: @func
+; CHECK-LABEL: entry:
+; CHECK-NEXT:  ret void
+entry:
+  %0 = tail call noalias noundef nonnull dereferenceable(32) ptr @_Znwm(i64 noundef 32)
+  %1 = icmp ugt ptr %0, @.str
+  tail call void @llvm.assume(i1 %1)
+  ret void
+}
+
+declare noundef nonnull ptr @_Znwm(i64 noundef) local_unnamed_addr
+
+declare void @llvm.assume(i1 noundef)
 No newline at end of file