[PATCH] D78191: [GlobalISel] Fix invalid combine of unmerge(merge) with intermediate cast

[Matt Arsenault via Phabricator via llvm-commits]

arsenm added inline comments.


================
Comment at: llvm/include/llvm/CodeGen/GlobalISel/LegalizationArtifactCombiner.h:528-530
+        // The following combine is only valid for vectors
+        if (!OpTy.isVector())
+          return false;
----------------
canFoldMergeOpcode should have done this legality check


================
Comment at: llvm/test/CodeGen/AMDGPU/GlobalISel/artifact-combiner-unmerge-values.mir:915
+    %0:_(s64) = COPY $vgpr0_vgpr1
+    %1:_(s64) = COPY $vgpr0_vgpr1
+    %2:_(s128) = G_MERGE_VALUES %0, %1
----------------
It doesn't really matter, but this should use different source registers for the two copies. The second should be $vgpr2_vgpr3


================
Comment at: llvm/test/CodeGen/AMDGPU/GlobalISel/artifact-combiner-unmerge-values.mir:938
+    %0:_(s32) = COPY $vgpr0
+    %1:_(s32) = COPY $vgpr0
+    %2:_(s64) = G_MERGE_VALUES %0, %1
----------------
$vgpr1


================
Comment at: llvm/test/CodeGen/AMDGPU/GlobalISel/cvt_f32_ubyte-xfail.ll:4
+
+; ERROR: Segmentation fault
+define <3 x float> @v_uitofp_v3i8_to_v3f32(i32 %arg0) nounwind {
----------------
This will only work on some linuxes sometimes, there’s not much point in checking the error though 


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D78191/new/

https://reviews.llvm.org/D78191