[clang] [llvm] [mlir] [AMDGPU] add clamp immediate operand to WMMA iu8 intrinsic (PR #171069)

[Matt Arsenault via llvm-commits]

================
@@ -4504,6 +4511,21 @@ static Value *upgradeARMIntrinsicCall(StringRef Name, CallBase *CI, Function *F,
 //
 static Value *upgradeAMDGCNIntrinsicCall(StringRef Name, CallBase *CI,
                                          Function *F, IRBuilder<> &Builder) {
+  if (Name.starts_with("wmma.i32.16x16x64.iu8")) {
+    // Legacy WMMA IU8 intrinsic lacked the optional clamp operand. Append
+    // clamp=false for compatibility.
+    if (CI->arg_size() != 7)
+      return nullptr;
+
+    SmallVector<Value *, 8> Args(CI->args().begin(), CI->args().end());
+    Args.push_back(Builder.getFalse());
+
+    Function *NewDecl = Intrinsic::getOrInsertDeclaration(
+      F->getParent(), Intrinsic::amdgcn_wmma_i32_16x16x64_iu8,
+      {CI->getArgOperand(4)->getType(), CI->getArgOperand(1)->getType()});
+    return Builder.CreateCall(NewDecl, Args);
----------------
arsenm wrote:

Do you really need to create a new call, and not mutate the one in place? As it is this is losing callsite attributes and metadata 

https://github.com/llvm/llvm-project/pull/171069