[llvm] [GlobalISel][AMDGPU] Import patterns with multiple defs (PR #84171)

[Pierre van Houtryve via llvm-commits]

================
@@ -728,25 +728,34 @@ def : OpSelBinOpClampPat<saddsat, V_ADD_I16_e64>;
 def : OpSelBinOpClampPat<ssubsat, V_SUB_I16_e64>;
 } // End SubtargetPredicate = isGFX9Plus
 
-// FIXME: GlobalISel in general does not handle instructions with 2 results,
-// so it cannot use these patterns.
 multiclass IMAD32_Pats <VOP3_Pseudo inst> {
   def : GCNPat <
         (ThreeOpFrag<mul, add> i32:$src0, i32:$src1, i32:$src2),
-        (EXTRACT_SUBREG (inst $src0, $src1,
+        (EXTRACT_SUBREG (inst i32:$src0, i32:$src1,
                               (REG_SEQUENCE SReg_64, // Use scalar and let it be legalized
                                             $src2, sub0,
                                             (i32 (IMPLICIT_DEF)), sub1),
                                             0 /* clamp */),
                         sub0)
         >;
+
+  // GISel-specific pattern that avoids creating a SGPR->VGPR copy if
+  // $src2 is a VGPR.
+  def : GCNPat <
+        (ThreeOpFrag<mul, add> i32:$src0, i32:$src1, VGPR_32:$src2),
+        (EXTRACT_SUBREG (inst i32:$src0, i32:$src1,
+                              (REG_SEQUENCE VReg_64,
+                                            $src2, sub0,
+                                            (i32 (IMPLICIT_DEF)), sub1),
+                                            0 /* clamp */),
+                        sub0)
+        >;
----------------
Pierre-vh wrote:

I didn't find a worthwhile way to do it, we save 2 lines in each pattern but add like 3-4 lines to declare the fragment so it's more or less the same

https://github.com/llvm/llvm-project/pull/84171