[llvm] [AMDGPU] Adjust amdgpu fmax/fmin legalization for DAG (PR #203150)

[Matt Arsenault via llvm-commits]

================
@@ -817,20 +837,32 @@ SITargetLowering::SITargetLowering(const TargetMachine &TM,
                         MVT::v8f16, MVT::v8bf16, MVT::v16f16, MVT::v16bf16,
                         MVT::v32f16, MVT::v32bf16},
                        Custom);
+    if (Subtarget->hasIEEEMinimumMaximumInsts() &&
+        Subtarget->hasSALUMinimumMaximumInsts()) {
----------------
arsenm wrote:

Same. We probably should merge these blocks together. Either we should have a unique feature per fp-type, or fully lean into the fact that the new instructions were introduced for all types (though I think that's not true in the gfx950 case)

https://github.com/llvm/llvm-project/pull/203150