[llvm-branch-commits] [clang] [llvm] AMDGPU: Support v_cvt_scalef32_pk32_{bf|f}6_{bf|fp}16 for gfx950 (PR #117592)

[Shilei Tian via llvm-branch-commits]

================
@@ -408,11 +408,23 @@ def FeatureFP6BF6ConversionScaleInsts : SubtargetFeature<"fp6bf6-cvt-scale-insts
   "Has fp6 and bf6 conversion scale instructions"
 >;
 
+def FeatureF16BF16ToFP6BF6ConversionScaleInsts : SubtargetFeature<"f16bf16-to-fp6bf6-cvt-scale-insts",
+  "HasF16BF16ToFP6BF6ConversionScaleInsts",
+  "true",
+  "Has f16bf16 to fp6bf6 conversion scale instructions"
+>;
+
 def FeatureGFX950Insts : SubtargetFeature<"gfx950-insts",
   "GFX950Insts",
   "true",
   "Additional instructions for GFX950+",
-  [FeaturePermlane16Swap, FeaturePermlane32Swap, FeatureFP8ConversionScaleInsts, FeatureBF8ConversionScaleInsts, FeatureFP4ConversionScaleInsts, FeatureFP6BF6ConversionScaleInsts]
+  [FeaturePermlane16Swap,
+  FeaturePermlane32Swap,
----------------
shiltian wrote:

the alignment is off here but that can be fixed later

https://github.com/llvm/llvm-project/pull/117592