[llvm] AMDGPU/GlobalISel: Legalize BF16 for FP math opcodes (PR #214059)

[Petar Avramovic via llvm-commits]

================
@@ -976,36 +974,42 @@ AMDGPULegalizerInfo::AMDGPULegalizerInfo(const GCNSubtarget &ST_,
   getActionDefinitionsBuilder(G_BLOCK_ADDR).legalFor({CodePtr});
 
   auto &FPOpActions =
-      getActionDefinitionsBuilder({G_FADD, G_FMUL, G_FMA, G_FCANONICALIZE,
-                                   G_STRICT_FADD, G_STRICT_FMUL, G_STRICT_FMA})
+      getActionDefinitionsBuilder({G_FADD, G_FMUL, G_FMA, G_FCANONICALIZE})
----------------
petar-avramovic wrote:

I assume it would be best to remove G_FCANONICALIZE from this group and put it in same group as min/max, but again not exactly sure which kind of min/max. In the end could do G_FCANONICALIZE alone.


https://github.com/llvm/llvm-project/pull/214059