[llvm] [AMDGPU][InstCombine] Fold unused m0 operand to poison for sendmsg intrinsics (PR #183755)

[Matt Arsenault via llvm-commits]

================
@@ -1448,6 +1448,33 @@ GCNTTIImpl::instCombineIntrinsic(InstCombiner &IC, IntrinsicInst &II) const {
     // amdgcn.kill(i1 1) is a no-op
     return IC.eraseInstFromFunction(II);
   }
+  case Intrinsic::amdgcn_s_sendmsg:
+  case Intrinsic::amdgcn_s_sendmsghalt: {
+    // The second operand is copied to m0, but is only actually used for
+    // certain message types. For message types that are known to not use m0,
+    // fold it to poison.
+    using namespace AMDGPU::SendMsg;
+
+    Value *M0Val = II.getArgOperand(1);
+    if (isa<PoisonValue>(M0Val))
+      break;
+
+    auto *MsgImm = cast<ConstantInt>(II.getArgOperand(0));
+    uint64_t Msg = MsgImm->getZExtValue();
+    // Extract the message ID. Pre-GFX11 uses the lower 4 bits, GFX11+ uses
+    // the lower 8 bits. Use 4-bit mask for extracting base message ID since
+    // all message types we handle fit in 4 bits, and the upper bits encode
+    // stream ID or other parameters for some message types (e.g., MSG_GS).
+    uint64_t MsgId = Msg & ID_MASK_PreGFX11_;
+
+    if (!msgDoesNotUseM0(MsgId, *ST))
+      break;
+
+    // Drop noundef attribute since we're replacing with poison.
+    II.removeParamAttr(1, Attribute::NoUndef);
----------------
arsenm wrote:

Maybe safer to use dropUBImplyingAttrsAndMetadata, in case that nopoison attribute ever gets added or something else 

https://github.com/llvm/llvm-project/pull/183755