[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})
+          .legalFor({F32, F64});
+  auto &StrictFPOpActions =
+      getActionDefinitionsBuilder({G_STRICT_FADD, G_STRICT_FMUL, G_STRICT_FMA})
           .legalFor({F32, F64});
   auto &TrigActions =
       getActionDefinitionsBuilder({G_FSIN, G_FCOS}).customFor({F32, F64});
   auto &FDIVActions = getActionDefinitionsBuilder(G_FDIV).customFor({F32, F64});
 
   if (ST.has16BitInsts()) {
-    if (ST.hasVOP3PInsts())
+    if (ST.hasVOP3PInsts()) {
       FPOpActions.legalFor({F16, V2F16});
-    else
+      StrictFPOpActions.legalFor({F16, V2F16});
+    } else {
       FPOpActions.legalFor({F16});
+      StrictFPOpActions.legalFor({F16});
+    }
 
     TrigActions.customFor({F16});
     FDIVActions.customFor({F16});
   }
 
+  FPOpActions.widenScalarFor({BF16}, changeElementTo(0, F32));
----------------
petar-avramovic wrote:

this should work for strict opcodes as well I assume

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