[llvm] [AMDGPU] Omit umin on ctlz/cttz if operand is non-zero. (PR #79127)

[Leon Clark via llvm-commits]

================
@@ -3114,8 +3114,13 @@ SDValue AMDGPUTargetLowering::LowerCTLZ_CTTZ(SDValue Op, SelectionDAG &DAG) cons
     // (cttz hi:lo) -> (umin (S_FF1_I32_B64 src), 64)
     // (ctlz_zero_undef src) -> (S_FLBIT_I32_B64 src)
     // (cttz_zero_undef src) -> (S_FF1_I32_B64 src)
+
+    // umin can be omitted if the operand is known to be non-zero.
+    auto KB = DAG.computeKnownBits(Src);
+    auto const IsNonZero = KB.countMinPopulation() > 0u;
+
     SDValue NewOpr = DAG.getNode(NewOpc, SL, MVT::i32, Src);
-    if (!ZeroUndef) {
+    if (!ZeroUndef && !IsNonZero) {
----------------
PeddleSpam wrote:

Thanks, I'll fix that now.

https://github.com/llvm/llvm-project/pull/79127