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

[foad]

Summary

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.

[arsenm]

In the future could check for matching convergence tokens

llvm/lib/Transforms/Scalar/GVN.cpp

uint32_t GVNPass::ValueTable::lookupOrAddCall(CallInst *C) {
   // 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;
     return e;
   }
 
   if (MD && AA->onlyReadsMemory(C)) {
bool GVNPass::performScalarPRE(Instruction *CurInst) {
   // 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 (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

llvm/lib/Transforms/Scalar/NewGVN.cpp

NewGVN::ExprResult NewGVN::performSymbolicCallEvaluation(Instruction *I) const {
   // 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 (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()));
   } else if (AA->onlyReadsMemory(CI)) {
     if (auto *MA = MSSA->getMemoryAccess(CI)) {
       auto *DefiningAccess = MSSAWalker->getClobberingMemoryAccess(MA);
       return ExprResult::some(createCallExpression(CI, DefiningAccess));
     } else // MSSA determined that CI does not access memory.

llvm/test/Transforms/GVN/convergent.ll

 ; 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 i32 [[V0]]
+; 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

 ; 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 i32 [[V0]]
+; 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