[PATCH] D154528: [AMDGPU][GlobalISel] Generate fast fp64-to-fp16 conversions in unsafe mode.

[Jay Foad via Phabricator via llvm-commits]

foad added inline comments.


================
Comment at: llvm/lib/Target/AMDGPU/VOP1Instructions.td:519-522
+def : GCNPat<
+    (f16 (fpround f64:$src)),
+    (V_CVT_F16_F32_e64 SRCMODS.NONE, (V_CVT_F32_F64_e64 SRCMODS.NONE, $src))
+>;
----------------
kosarev wrote:
> arsenm wrote:
> > Should just let the expansion split them up, this is missing the source modifiers
> Updated to match modifiers. The expansion would combine back to a single fpround f64, so doesn't work.
"The expansion would combine back to a single fpround f64" - that sounds like the combine is broken then? It should respect legality.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D154528/new/

https://reviews.llvm.org/D154528