[llvm] [AMDGPU] Elide bitcast fold i64 imm to build_vector (PR #154115)

[Janek van Oirschot via llvm-commits]

================
@@ -5318,6 +5325,8 @@ SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
     if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Src)) {
       SDLoc SL(N);
       uint64_t CVal = C->getZExtValue();
+      if (canMov64b(CVal))
+        break;
----------------
JanekvO wrote:

The original patch didn't add any tests and only affected about 4 tests; I've checked what tests are affected by removing this combine altogether locally and spawned #154363 with tests for which I enable gfx942 (I did cut a bunch of tests for gfx942 enabling that seemed to be more bug/crash tests rather than actual codegen tests)

https://github.com/llvm/llvm-project/pull/154115