[GVN] Do not combine convergent calls in GVN/NewGVN

Note that this is very conservative since it will not even combine
convergent calls that appear in the same basic block, but EarlyCSE will
handle that case.

Differential Revision: https://reviews.llvm.org/D150974

Change-Id: I3cad9a8750dc2d007e6e37ec4182e68d3fcd49a9

Author: jayfoad

llvm/lib/Transforms/Scalar/GVN.cpp

@@ -461,15 +461,27 @@ void GVNPass::ValueTable::add(Value *V, uint32_t num) {
 }
 
 uint32_t GVNPass::ValueTable::lookupOrAddCall(CallInst *C) {
-  if (AA->doesNotAccessMemory(C) &&
-      // FIXME: Currently the calls which may access the thread id may
-      // be considered as not accessing the memory. But this is
-      // problematic for coroutines, since coroutines may resume in a
-      // different thread. So we disable the optimization here for the
-      // correctness. However, it may block many other correct
-      // optimizations. Revert this one when we detect the memory
-      // accessing kind more precisely.
-      !C->getFunction()->isPresplitCoroutine()) {
+  // FIXME: Currently the calls which may access the thread id may
+  // be considered as not accessing the memory. But this is
+  // problematic for coroutines, since coroutines may resume in a
+  // different thread. So we disable the optimization here for the
+  // correctness. However, it may block many other correct
+  // optimizations. Revert this one when we detect the memory
+  // accessing kind more precisely.
+  if (C->getFunction()->isPresplitCoroutine()) {
+    valueNumbering[C] = nextValueNumber;
+    return nextValueNumber++;
+  }
+
+  // Do not combine convergent calls since they implicitly depend on the set of
+  // threads that is currently executing, and they might be in different basic
+  // blocks.
+  if (C->isConvergent()) {
+    valueNumbering[C] = nextValueNumber;
+    return nextValueNumber++;
+  }
+
+  if (AA->doesNotAccessMemory(C)) {
     Expression exp = createExpr(C);
     uint32_t e = assignExpNewValueNum(exp).first;
     valueNumbering[C] = e;
@@ -2768,9 +2780,6 @@ bool GVNPass::performScalarPRE(Instruction *CurInst) {
     // We don't currently value number ANY inline asm calls.
     if (CallB->isInlineAsm())
       return false;
-    // Don't do PRE on convergent calls.
-    if (CallB->isConvergent())
-      return false;
   }
 
   uint32_t ValNo = VN.lookup(CurInst);

llvm/lib/Transforms/Scalar/NewGVN.cpp

@@ -1617,6 +1617,12 @@ NewGVN::ExprResult NewGVN::performSymbolicCallEvaluation(Instruction *I) const {
   if (CI->getFunction()->isPresplitCoroutine())
     return ExprResult::none();
 
+  // Do not combine convergent calls since they implicitly depend on the set of
+  // threads that is currently executing, and they might be in different basic
+  // blocks.
+  if (CI->isConvergent())
+    return ExprResult::none();
+
   if (AA->doesNotAccessMemory(CI)) {
     return ExprResult::some(
         createCallExpression(CI, TOPClass->getMemoryLeader()));

llvm/test/Transforms/GVN/convergent.ll

@@ -0,0 +1,29 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -S -passes=gvn -o - %s | FileCheck %s
+
+define i32 @foo(i1 %cond) {
+; CHECK-LABEL: @foo(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[V0:%.*]] = call i32 @llvm.convergent()
+; CHECK-NEXT:    br i1 [[COND:%.*]], label [[BODY:%.*]], label [[END:%.*]]
+; CHECK:       body:
+; CHECK-NEXT:    [[V1:%.*]] = call i32 @llvm.convergent()
+; CHECK-NEXT:    br label [[END]]
+; CHECK:       end:
+; CHECK-NEXT:    [[RET:%.*]] = phi i32 [ [[V0]], [[ENTRY:%.*]] ], [ [[V1]], [[BODY]] ]
+; CHECK-NEXT:    ret i32 [[RET]]
+;
+entry:
+  %v0 = call i32 @llvm.convergent()
+  br i1 %cond, label %body, label %end
+
+body:
+  %v1 = call i32 @llvm.convergent()
+  br label %end
+
+end:
+  %ret = phi i32 [ %v0, %entry ], [ %v1, %body ]
+  ret i32 %ret
+}
+
+declare i32 @llvm.convergent() convergent nounwind readnone willreturn

llvm/test/Transforms/NewGVN/convergent.ll

@@ -0,0 +1,29 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -S -passes=newgvn -o - %s | FileCheck %s
+
+define i32 @foo(i1 %cond) {
+; CHECK-LABEL: @foo(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[V0:%.*]] = call i32 @llvm.convergent()
+; CHECK-NEXT:    br i1 [[COND:%.*]], label [[BODY:%.*]], label [[END:%.*]]
+; CHECK:       body:
+; CHECK-NEXT:    [[V1:%.*]] = call i32 @llvm.convergent()
+; CHECK-NEXT:    br label [[END]]
+; CHECK:       end:
+; CHECK-NEXT:    [[RET:%.*]] = phi i32 [ [[V0]], [[ENTRY:%.*]] ], [ [[V1]], [[BODY]] ]
+; CHECK-NEXT:    ret i32 [[RET]]
+;
+entry:
+  %v0 = call i32 @llvm.convergent()
+  br i1 %cond, label %body, label %end
+
+body:
+  %v1 = call i32 @llvm.convergent()
+  br label %end
+
+end:
+  %ret = phi i32 [ %v0, %entry ], [ %v1, %body ]
+  ret i32 %ret
+}
+
+declare i32 @llvm.convergent() convergent nounwind readnone willreturn