[llvm] [AArch64][SVE] Fold zero-extend into add reduction. (PR #102325)

[David Sherwood via llvm-commits]

================
@@ -103,17 +103,12 @@ define i32 @add_i32(<vscale x 8 x i32> %a, <vscale x 4 x i32> %b) {
 define i16 @add_ext_i16(<vscale x 16 x i8> %a, <vscale x 16 x i8> %b) {
 ; CHECK-LABEL: add_ext_i16:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    uunpkhi z2.h, z0.b
-; CHECK-NEXT:    uunpklo z0.h, z0.b
-; CHECK-NEXT:    uunpkhi z3.h, z1.b
-; CHECK-NEXT:    uunpklo z1.h, z1.b
-; CHECK-NEXT:    ptrue p0.h
-; CHECK-NEXT:    add z0.h, z0.h, z2.h
-; CHECK-NEXT:    add z1.h, z1.h, z3.h
-; CHECK-NEXT:    add z0.h, z0.h, z1.h
-; CHECK-NEXT:    uaddv d0, p0, z0.h
-; CHECK-NEXT:    fmov x0, d0
-; CHECK-NEXT:    // kill: def $w0 killed $w0 killed $x0
+; CHECK-NEXT:    ptrue p0.b
+; CHECK-NEXT:    uaddv d0, p0, z0.b
+; CHECK-NEXT:    uaddv d1, p0, z1.b
+; CHECK-NEXT:    fmov w8, s0
----------------
david-arm wrote:

Not for this patch, but I wonder if we can improve this further with:

```
  uaddv d0, p0, z0.b
  uaddv d1, p0, z1.b
  add v0.4s, v0.4s, v1.4s
  fmov w0, s0
```

The throughput of the NEON add is much higher than that of fmov and the latency is about the same.

https://github.com/llvm/llvm-project/pull/102325