[llvm] AMDGPU: Custom lower fptrunc vectors for f32 -> f16 (PR #141883)

[Shilei Tian via llvm-commits]

================
@@ -1061,9 +1061,10 @@ AMDGPULegalizerInfo::AMDGPULegalizerInfo(const GCNSubtarget &ST_,
   }
 
   auto &FPTruncActions = getActionDefinitionsBuilder(G_FPTRUNC);
-  if (ST.hasCvtPkF16F32Inst())
-    FPTruncActions.legalFor({{S32, S64}, {S16, S32}, {V2S16, V2S32}});
-  else
+  if (ST.hasCvtPkF16F32Inst()) {
+    FPTruncActions.legalFor({{S32, S64}, {S16, S32}, {V2S16, V2S32}})
+                  .customFor({{V4S16, V4S32}, {V8S16, V8S32}});
+  } else
     FPTruncActions.legalFor({{S32, S64}, {S16, S32}});
----------------
shiltian wrote:

```suggestion
  if (ST.hasCvtPkF16F32Inst()) {
    FPTruncActions.legalFor({{S32, S64}, {S16, S32}, {V2S16, V2S32}})
                  .customFor({{V4S16, V4S32}, {V8S16, V8S32}});
  } else {
    FPTruncActions.legalFor({{S32, S64}, {S16, S32}});
  }
```

Also it looks like the format is off

https://github.com/llvm/llvm-project/pull/141883