[llvm] [AMDGPU] Support v_lshl_add_u64 with non-constant shift amount (PR #179904)

[Frederik Harwath via llvm-commits]

================
@@ -638,19 +638,13 @@ class ThreeOpFrag<SDPatternOperator op1, SDPatternOperator op2> : ThreeOpFragSDA
 }
 
 def shl_0_to_4 : PatFrag<
-  (ops node:$src0, node:$src1), (shl node:$src0, node:$src1),
-  [{
-     if (auto *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
-       return C->getZExtValue() <= 4;
-     }
-     return false;
-   }]> {
+  (ops node:$src0, node:$src1), (shl node:$src0, node:$src1), [{
+    KnownBits KB = CurDAG->computeKnownBits(N->getOperand(1));
+    return KB.getMaxValue().getZExtValue() <= 4;
+  }]> {
   let GISelPredicateCode = [{
----------------
frederik-h wrote:

There seems to be no test coverage for the GlobalISel path. I compiled the existing test with -globa-isel=1 and there was only a single use of v_lshl_add_u64 in the output (in the gep test case).

https://github.com/llvm/llvm-project/pull/179904