AMDGPU: Legalize fminimum and fmaximum f32 for gfx950 (llvm#117634)

Select to minimum3/maximum3. Leave f16/v2f16 for later
since it's complicated by only having the vector version.

Author: arsenm

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

@@ -855,6 +855,10 @@ SITargetLowering::SITargetLowering(const TargetMachine &TM,
     setOperationAction({ISD::FMINIMUM, ISD::FMAXIMUM},
                        {MVT::v4f16, MVT::v8f16, MVT::v16f16, MVT::v32f16},
                        Custom);
+  } else {
+    // FIXME: For nnan fmaximum, emit the fmaximum3 instead of fmaxnum
+    if (Subtarget->hasMinimum3Maximum3F32())
+      setOperationAction({ISD::FMAXIMUM, ISD::FMINIMUM}, MVT::f32, Legal);
   }
 
   setOperationAction(ISD::INTRINSIC_WO_CHAIN,

llvm/lib/Target/AMDGPU/VOP3Instructions.td

@@ -1234,6 +1234,23 @@ def : IntClampPat<V_MQSAD_PK_U16_U8_e64, int_amdgcn_mqsad_pk_u16_u8>;
 def : IntClampPat<V_QSAD_PK_U16_U8_e64, int_amdgcn_qsad_pk_u16_u8>;
 def : IntClampPat<V_MQSAD_U32_U8_e64, int_amdgcn_mqsad_u32_u8>;
 
+//===----------------------------------------------------------------------===//
+// Floating-point operation Patterns
+//===----------------------------------------------------------------------===//
+
+// Implement fminimum(x, y) by using minimum3(x, y, y)
+class MinimumMaximumByMinimum3Maximum3<SDPatternOperator node, ValueType vt,
+                                       Instruction inst> : GCNPat<
+  (vt (node (VOP3Mods vt:$src0, i32:$src0_mods), (VOP3Mods vt:$src1, i32:$src1_mods))),
+  (inst $src0_mods, $src0, $src1_mods, $src1, $src1_mods, $src1)
+>;
+
+// Prefer the real 2 operand form if legal
+let SubtargetPredicate = HasMinimum3Maximum3F32, AddedComplexity = -1000 in {
+def : MinimumMaximumByMinimum3Maximum3<fminimum, f32, V_MINIMUM3_F32_e64>;
+def : MinimumMaximumByMinimum3Maximum3<fmaximum, f32, V_MAXIMUM3_F32_e64>;
+}
+
 //===----------------------------------------------------------------------===//
 // Target-specific instruction encodings.
 //===----------------------------------------------------------------------===//