[llvm] [SelectionDAG] Scalarize binary ops of splats before legal types (PR #100749)

[Luke Lau via llvm-commits]

================
@@ -864,21 +864,19 @@ define <vscale x 8 x i64> @vmul_vi_nxv8i64_2(<vscale x 8 x i64> %va) {
 }
 
 define <vscale x 8 x i64> @vmul_xx_nxv8i64(i64 %a, i64 %b) nounwind {
-; RV32-LABEL: vmul_xx_nxv8i64:
-; RV32:       # %bb.0:
-; RV32-NEXT:    addi sp, sp, -16
-; RV32-NEXT:    sw a1, 12(sp)
-; RV32-NEXT:    sw a0, 8(sp)
-; RV32-NEXT:    addi a0, sp, 8
-; RV32-NEXT:    vsetvli a1, zero, e64, m8, ta, ma
-; RV32-NEXT:    vlse64.v v8, (a0), zero
-; RV32-NEXT:    sw a3, 4(sp)
-; RV32-NEXT:    sw a2, 0(sp)
-; RV32-NEXT:    mv a0, sp
-; RV32-NEXT:    vlse64.v v16, (a0), zero
-; RV32-NEXT:    vmul.vv v8, v8, v16
-; RV32-NEXT:    addi sp, sp, 16
-; RV32-NEXT:    ret
+; RV32NOM-LABEL: vmul_xx_nxv8i64:
+; RV32NOM:       # %bb.0:
+; RV32NOM-NEXT:    addi sp, sp, -16
+; RV32NOM-NEXT:    sw ra, 12(sp) # 4-byte Folded Spill
+; RV32NOM-NEXT:    call __muldi3
----------------
lukel97 wrote:

I think this patch is doing the right thing by checking if the operation is custom or legal, but I'm still surprised to see that mul is scalarized here despite it being a libcall. Do you know why that is?

https://github.com/llvm/llvm-project/pull/100749