[llvm] [GlobalISel] Combine away G_UNMERGE(G_IMPLICITDEF). (PR #119183)

David Green via llvm-commits llvm-commits at lists.llvm.org
Sun Dec 15 15:02:51 PST 2024


davemgreen wrote:

This now passes the type of the Reg required through to the ArtifactFinder and
includes https://github.com/llvm/llvm-project/pull/119850.

Most of the tests are just changing because they use IMPLICIT_DEF for the
input. Most look OK to me, in that they are just changing `<complex
legalization of implicitdef> + operations` to `implicitdef + operations`, as in
```
     ; VI-LABEL: name: test_ffloor_v3s16
-    ; VI: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
-    ; VI-NEXT: [[UV:%[0-9]+]]:_(<2 x s16>), [[UV1:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[DEF]](<4 x s16>)
-    ; VI-NEXT: [[BITCAST:%[0-9]+]]:_(s32) = G_BITCAST [[UV]](<2 x s16>)
-    ; VI-NEXT: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST]](s32)
-    ; VI-NEXT: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
-    ; VI-NEXT: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C]](s32)
-    ; VI-NEXT: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
-    ; VI-NEXT: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[UV1]](<2 x s16>)
-    ; VI-NEXT: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST1]](s32)
-    ; VI-NEXT: [[FFLOOR:%[0-9]+]]:_(s16) = G_FFLOOR [[TRUNC]]
-    ; VI-NEXT: [[FFLOOR1:%[0-9]+]]:_(s16) = G_FFLOOR [[TRUNC1]]
-    ; VI-NEXT: [[FFLOOR2:%[0-9]+]]:_(s16) = G_FFLOOR [[TRUNC2]]
+    ; VI: [[DEF:%[0-9]+]]:_(s16) = G_IMPLICIT_DEF
+    ; VI-NEXT: [[FFLOOR:%[0-9]+]]:_(s16) = G_FFLOOR [[DEF]]
+    ; VI-NEXT: [[FFLOOR1:%[0-9]+]]:_(s16) = G_FFLOOR [[DEF]]
+    ; VI-NEXT: [[FFLOOR2:%[0-9]+]]:_(s16) = G_FFLOOR [[DEF]]
```

If any of these do look like a problem then let me know. I can try and define
the inputs better if we can come up with a way to do that (many are
G_IMPLICIT_DEF presumably because of the types involved).

One Mips test has an extra mov in one of the tests from a COPY of undef. Some
of the tests are now folding ZExt(IMPLICIT_DEF) to 0 which loses information
about the undef bits but would likely have been done anyway.

https://github.com/llvm/llvm-project/pull/119183


More information about the llvm-commits mailing list