[gvn] PRE needs to skip convergent intrinsics/calls.

[hliao]

Summary

• As convergent intrinsics/calls could only be moved to control-equivalent blocks, or more precisely the same divergent branch, PRE needs to skip them.

[hliao]

Revise the comment.

[mkazantsev]

llvm/test/Transforms/GVN/pre-skip-convergent.ll

2

Please commit this test without your patch with auto-generated checks (utils/update_test_checks.py), then rebase on top of it to show what your patch changes.

[jdoerfert]

llvm/test/Transforms/GVN/pre-skip-convergent.ll

2

You probably also want to add a run line for the new pass manager -passes=..., assuming we run this code in the new PM. (Ignore this if we don't).

[hliao]

Revise the test following reviewers' comments.

[hliao]

llvm/test/Transforms/GVN/pre-skip-convergent.ll

2

with this patch, we won't PRE that call into convergent intrinsic @llvm.convergent in merge block into pre block as it won't function correctly.

[jdoerfert]

LGTM

llvm/lib/Transforms/Scalar/GVN.cpp

bool GVN::performScalarPRE(Instruction *CurInst) {
   // GEP's live range increases the register pressure, and therefore it can
   // introduce unnecessary spills.
   //
   // This doesn't prevent Load PRE. PHI translation will make the GEP available
   // to the load by moving it to the predecessor block if necessary.
   if (isa<GetElementPtrInst>(CurInst))
     return false;
 
+  if (auto *CallB = dyn_cast<CallBase>(CurInst)) {
     // We don't currently value number ANY inline asm calls.
-  if (auto *CallB = dyn_cast<CallBase>(CurInst))
     if (CallB->isInlineAsm())
       return false;
+    // Don't do PRE on convergent calls.
+    if (CallB->isConvergent())
+      return false;
+  }
 
   uint32_t ValNo = VN.lookup(CurInst);
 
   // Look for the predecessors for PRE opportunities.  We're
   // only trying to solve the basic diamond case, where
   // a value is computed in the successor and one predecessor,
   // but not the other.  We also explicitly disallow cases
   // where the successor is its own predecessor, because they're

llvm/test/Transforms/GVN/pre-skip-convergent.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -S -gvn -o - %s | FileCheck %s
+; RUN: opt -S -passes=gvn -o - %s | FileCheck %s
+
+define i32 @foo(i1 %cond, i32* %q, i32* %p) {
+; CHECK-LABEL: @foo(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[V0:%.*]] = call i32 @llvm.convergent(i32 0)
+; CHECK-NEXT:    store i32 [[V0]], i32* [[Q:%.*]], align 4
+; CHECK-NEXT:    br i1 [[COND:%.*]], label [[PRE:%.*]], label [[MERGE:%.*]]
+; CHECK:       pre:
+; CHECK-NEXT:    [[T0:%.*]] = load i32, i32* [[P:%.*]], align 4
+; CHECK-NEXT:    br label [[MERGE]]
+; CHECK:       merge:
+; CHECK-NEXT:    [[M0:%.*]] = phi i32 [ [[T0]], [[PRE]] ], [ 0, [[ENTRY:%.*]] ]
+; CHECK-NEXT:    [[R0:%.*]] = call i32 @llvm.convergent(i32 [[M0]])
+; CHECK-NEXT:    ret i32 [[R0]]
+;
+entry:
+  %v0 = call i32 @llvm.convergent(i32 0)
+  store i32 %v0, i32* %q
+  br i1 %cond, label %pre, label %merge
+
+pre:
+  %t0 = load i32, i32* %p
+  br label %merge
+
+merge:
+  %m0 = phi i32 [ %t0, %pre ], [ 0, %entry ]
+  %r0 = call i32 @llvm.convergent(i32 %m0)
+  ret i32 %r0
+}
+
+declare i32 @llvm.convergent(i32) #0
+
+attributes #0 = { convergent nounwind readnone }