[SimplifyCFG] Fix SimplifyCFG pass to skip folding when both blocks contain convergence loop/entry intrinsics. (#166452)

Fixes a bug #165642. [Similar
fix](#165643) is being made in
`IndVarSimplify` pass to account for convergence tokens.

[LLVM
Spec](https://llvm.org/docs/ConvergentOperations.html#llvm-experimental-convergence-loop)
states that only a single loop / entry convergence token can be included
in a basic block.

This PR fixes the issue in `SimplifyCFG` pass so that when a basic block
and its predecessor both contain such convergence intrinsics, it skips
merging the two blocks.

Author: luciechoi

llvm/lib/Transforms/Utils/BasicBlockUtils.cpp

@@ -92,6 +92,15 @@ emptyAndDetachBlock(BasicBlock *BB,
          "applying corresponding DTU updates.");
 }
 
+static bool HasLoopOrEntryConvergenceToken(const BasicBlock *BB) {
+  for (const Instruction &I : *BB) {
+    const ConvergenceControlInst *CCI = dyn_cast<ConvergenceControlInst>(&I);
+    if (CCI && (CCI->isLoop() || CCI->isEntry()))
+      return true;
+  }
+  return false;
+}
+
 void llvm::detachDeadBlocks(ArrayRef<BasicBlock *> BBs,
                             SmallVectorImpl<DominatorTree::UpdateType> *Updates,
                             bool KeepOneInputPHIs) {
@@ -259,6 +268,13 @@ bool llvm::MergeBlockIntoPredecessor(BasicBlock *BB, DomTreeUpdater *DTU,
     if (llvm::is_contained(PN.incoming_values(), &PN))
       return false;
 
+  // Don't break if both the basic block and the predecessor contain loop or
+  // entry convergent intrinsics, since there may only be one convergence token
+  // per block.
+  if (HasLoopOrEntryConvergenceToken(BB) &&
+      HasLoopOrEntryConvergenceToken(PredBB))
+    return false;
+
   LLVM_DEBUG(dbgs() << "Merging: " << BB->getName() << " into "
                     << PredBB->getName() << "\n");
 

llvm/test/Transforms/SimplifyCFG/skip-merging-duplicate-convergence-instrinsics.ll

@@ -0,0 +1,68 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -S -passes=simplifycfg | FileCheck %s
+
+declare token @llvm.experimental.convergence.entry() #0
+
+define void @nested(i32 %tidx, i32 %tidy, ptr %array) #0 {
+; CHECK-LABEL: @nested(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = tail call token @llvm.experimental.convergence.entry()
+; CHECK-NEXT:    [[TMP1:%.*]] = or i32 [[TIDY:%.*]], [[TIDX:%.*]]
+; CHECK-NEXT:    [[OR_COND_I:%.*]] = icmp eq i32 [[TMP1]], 0
+; CHECK-NEXT:    br label [[FOR_COND_I:%.*]]
+; CHECK:       for.cond.i:
+; CHECK-NEXT:    [[TMP2:%.*]] = call token @llvm.experimental.convergence.loop() [ "convergencectrl"(token [[TMP0]]) ]
+; CHECK-NEXT:    br label [[FOR_COND1_I:%.*]]
+; CHECK:       for.cond1.i:
+; CHECK-NEXT:    [[CMP2_I:%.*]] = phi i1 [ false, [[FOR_BODY4_I:%.*]] ], [ true, [[FOR_COND_I]] ]
+; CHECK-NEXT:    [[TMP3:%.*]] = call token @llvm.experimental.convergence.loop() [ "convergencectrl"(token [[TMP2]]) ]
+; CHECK-NEXT:    br i1 [[CMP2_I]], label [[FOR_BODY4_I]], label [[EXIT:%.*]]
+; CHECK:       for.body4.i:
+; CHECK-NEXT:    br i1 [[OR_COND_I]], label [[IF_THEN_I:%.*]], label [[FOR_COND1_I]]
+; CHECK:       if.then.i:
+; CHECK-NEXT:    [[TEST_VAL:%.*]] = call spir_func i32 @func_test(i32 0) [ "convergencectrl"(token [[TMP3]]) ]
+; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr inbounds i32, ptr [[ARRAY:%.*]], i32 0
+; CHECK-NEXT:    store i32 [[TEST_VAL]], ptr [[TMP4]], align 4
+; CHECK-NEXT:    br label [[EXIT]]
+; CHECK:       exit:
+; CHECK-NEXT:    ret void
+;
+entry:
+  %0 = tail call token @llvm.experimental.convergence.entry()
+  %2 = or i32 %tidy, %tidx
+  %or.cond.i = icmp eq i32 %2, 0
+  br label %for.cond.i
+
+for.cond.i:
+  %3 = call token @llvm.experimental.convergence.loop() [ "convergencectrl"(token %0) ]
+  br label %for.cond1.i
+
+for.cond1.i:
+  %cmp2.i = phi i1 [ false, %for.body4.i ], [ true, %for.cond.i ]
+  %4 = call token @llvm.experimental.convergence.loop() [ "convergencectrl"(token %3) ]
+  br i1 %cmp2.i, label %for.body4.i, label %cleanup.i.loopexit
+
+for.body4.i:
+  br i1 %or.cond.i, label %if.then.i, label %for.cond1.i
+
+if.then.i:
+  %test.val = call spir_func i32 @func_test(i32 0) [ "convergencectrl"(token %4) ]
+  %5 = getelementptr inbounds i32, ptr %array, i32 0
+  store i32 %test.val, ptr %5, align 4
+  br label %cleanup.i
+
+cleanup.i.loopexit:
+  br label %cleanup.i
+
+cleanup.i:
+  br label %exit
+
+exit:
+  ret void
+}
+
+declare token @llvm.experimental.convergence.loop() #0
+
+declare i32 @func_test(i32) #0
+
+attributes #0 = { convergent }