[llvm] [RISCV] Generaize reduction tree matching to fp sum reductions (PR #68599)

[Luke Lau via llvm-commits]

================
@@ -764,6 +764,165 @@ define i32 @reduce_umin_16xi32_prefix5(ptr %p) {
   %umin3 = call i32 @llvm.umin.i32(i32 %umin2, i32 %e4)
   ret i32 %umin3
 }
+
+define float @reduce_fadd_16xf32_prefix2(ptr %p) {
+; CHECK-LABEL: reduce_fadd_16xf32_prefix2:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetivli zero, 2, e32, mf2, ta, ma
+; CHECK-NEXT:    vle32.v v8, (a0)
+; CHECK-NEXT:    vmv.s.x v9, zero
+; CHECK-NEXT:    vfredusum.vs v8, v8, v9
+; CHECK-NEXT:    vfmv.f.s fa0, v8
+; CHECK-NEXT:    ret
+  %v = load <16 x float>, ptr %p, align 256
+  %e0 = extractelement <16 x float> %v, i32 0
+  %e1 = extractelement <16 x float> %v, i32 1
+  %fadd0 = fadd fast float %e0, %e1
+  ret float %fadd0
+}
+
+define float @reduce_fadd_16xi32_prefix5(ptr %p) {
+; CHECK-LABEL: reduce_fadd_16xi32_prefix5:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    lui a1, 524288
+; CHECK-NEXT:    vsetivli zero, 8, e32, m2, ta, ma
+; CHECK-NEXT:    vle32.v v8, (a0)
+; CHECK-NEXT:    vmv.s.x v10, a1
+; CHECK-NEXT:    vsetivli zero, 6, e32, m2, tu, ma
+; CHECK-NEXT:    vslideup.vi v8, v10, 5
+; CHECK-NEXT:    vsetivli zero, 7, e32, m2, tu, ma
+; CHECK-NEXT:    vslideup.vi v8, v10, 6
+; CHECK-NEXT:    vsetivli zero, 8, e32, m2, ta, ma
+; CHECK-NEXT:    vslideup.vi v8, v10, 7
+; CHECK-NEXT:    vfredusum.vs v8, v8, v10
----------------
lukel97 wrote:

I wonder if we could improve the legalisation here by using VP nodes (something like https://reviews.llvm.org/D148523 but for reductions) to avoid having to pad out the vector with zeroes. Or it looks like there could also be a combine to replace all these inserts with a splat and single slide.

https://github.com/llvm/llvm-project/pull/68599