[llvm] AMDGPU/GlobalISel: Switch some FP opcodes to extended LLTs (PR #212664)

[Petar Avramovic via llvm-commits]

================
@@ -1057,11 +1058,9 @@ AMDGPULegalizerInfo::AMDGPULegalizerInfo(const GCNSubtarget &ST_,
   }
 
   if (ST.hasVOP3PInsts())
-    FPOpActions.clampMaxNumElementsStrict(0, S16, 2);
+    FPOpActions.clampMaxNumElementsStrict(0, F16, 2);
 
-  FPOpActions
-    .scalarize(0)
-    .clampScalar(0, ST.has16BitInsts() ? S16 : S32, S64);
+  FPOpActions.scalarize(0).clampScalar(0, ST.has16BitInsts() ? F16 : F32, F64);
----------------
petar-avramovic wrote:

think `clampScalar(0, ST.has16BitInsts() ? F16 : F32, F64)` is not appropriate here. Only real action is for targets without 16bit instrs to `widen f16 to f32`.
There is no smaller type then f16 to widen to f16, and similarly no larger type then f64 to narrow to f64

https://github.com/llvm/llvm-project/pull/212664