[llvm] [NVVMReflect] Force dead branch elimination in NVVMReflect (PR #81189)

[Joseph Huber via llvm-commits]

================
@@ -171,13 +175,71 @@ static bool runNVVMReflect(Function &F, unsigned SmVersion) {
     } else if (ReflectArg == "__CUDA_ARCH") {
       ReflectVal = SmVersion * 10;
     }
+
+    // If the immediate user is a simple comparison we want to simplify it.
+    // TODO: This currently does not handle switch instructions.
+    for (User *U : Call->users())
+      if (ICmpInst *I = dyn_cast<ICmpInst>(U))
+        ToSimplify.push_back(I);
+
     Call->replaceAllUsesWith(ConstantInt::get(Call->getType(), ReflectVal));
     ToRemove.push_back(Call);
   }
 
   for (Instruction *I : ToRemove)
     I->eraseFromParent();
 
+  // The code guarded by __nvvm_reflect may be invalid for the target machine.
+  // We need to do some basic dead code elimination to trim invalid code before
+  // it reaches the backend at all optimization levels.
+  SmallVector<BranchInst *> Simplified;
+  for (ICmpInst *Cmp : ToSimplify) {
+    Constant *LHS = dyn_cast<Constant>(Cmp->getOperand(0));
+    Constant *RHS = dyn_cast<Constant>(Cmp->getOperand(1));
+
+    if (!LHS || !RHS)
+      continue;
+
+    // If the comparison is a compile time constant we simply propagate it.
+    Constant *C = ConstantFoldCompareInstOperands(
+        Cmp->getPredicate(), LHS, RHS, Cmp->getModule()->getDataLayout());
+
+    if (!C)
+      continue;
+
+    for (User *U : Cmp->users())
+      if (BranchInst *I = dyn_cast<BranchInst>(U))
+        Simplified.push_back(I);
+
+    Cmp->replaceAllUsesWith(C);
+    Cmp->eraseFromParent();
+  }
+
+  // Each instruction here is a conditional branch off of a constant true or
+  // false value. Simply replace it with an unconditional branch to the
+  // appropriate basic block and delete the rest if it is trivially dead.
+  DenseSet<Instruction *> Removed;
+  for (BranchInst *Branch : Simplified) {
----------------
jhuber6 wrote:

Missed that, thanks. I'll probably make an updated version that also handles the Switch case since this function makes that trivial.

Side note, how do we handle the ROCm-DL constants? I remember we have some mandatory constant prop + DCE in a similar case.

https://github.com/llvm/llvm-project/pull/81189