[llvm] AMDGPU: Make v2f64 -> v2f16 conversion Legal only when unsafe fast math is set (PR #134738)

[Changpeng Fang via llvm-commits]

================
@@ -1053,10 +1053,10 @@ AMDGPULegalizerInfo::AMDGPULegalizerInfo(const GCNSubtarget &ST_,
   }
 
   auto &FPTruncActions = getActionDefinitionsBuilder(G_FPTRUNC);
-  if (ST.hasCvtPkF16F32Inst())
-    FPTruncActions.legalFor(
-        {{S32, S64}, {S16, S32}, {V2S16, V2S32}, {V2S16, V2S64}});
-  else
+  if (ST.hasCvtPkF16F32Inst()) {
+    FPTruncActions.legalFor({{S32, S64}, {S16, S32}, {V2S16, V2S32}})
+                  .customFor({V2S16, V2S64});
----------------
changpeng wrote:

> It's suspect that we need custom handling for the v2 case, but not the regular scalar

Previously only v2f16 was made Legal in order to generate v_cvt_pk_f16_f32, I think.

https://github.com/llvm/llvm-project/pull/134738