[llvm] [RISCV] Use vrgather in llvm.experimental.vector.match (PR #117878)

[Luke Lau via llvm-commits]

================
@@ -434,69 +382,53 @@ define <8 x i1> @match_v8i16(<8 x i16> %op1, <8 x i16> %op2, <8 x i1> %mask) {
 define <8 x i1> @match_v8i8_v16i8(<8 x i8> %op1, <16 x i8> %op2, <8 x i1> %mask) {
 ; CHECK-LABEL: match_v8i8_v16i8:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetivli zero, 1, e8, m1, ta, ma
-; CHECK-NEXT:    vmv.x.s a0, v9
-; CHECK-NEXT:    vslidedown.vi v10, v9, 1
-; CHECK-NEXT:    vslidedown.vi v11, v9, 2
-; CHECK-NEXT:    vmv.x.s a1, v10
-; CHECK-NEXT:    vslidedown.vi v10, v9, 3
-; CHECK-NEXT:    vmv.x.s a2, v11
-; CHECK-NEXT:    vslidedown.vi v11, v9, 4
-; CHECK-NEXT:    vmv.x.s a3, v10
-; CHECK-NEXT:    vslidedown.vi v10, v9, 5
-; CHECK-NEXT:    vmv.x.s a4, v11
-; CHECK-NEXT:    vslidedown.vi v11, v9, 6
-; CHECK-NEXT:    vmv.x.s a5, v10
-; CHECK-NEXT:    vslidedown.vi v10, v9, 7
-; CHECK-NEXT:    vmv.x.s a6, v11
-; CHECK-NEXT:    vslidedown.vi v11, v9, 8
-; CHECK-NEXT:    vmv.x.s a7, v10
-; CHECK-NEXT:    vslidedown.vi v10, v9, 9
-; CHECK-NEXT:    vmv.x.s t0, v11
-; CHECK-NEXT:    vslidedown.vi v11, v9, 10
-; CHECK-NEXT:    vmv.x.s t1, v10
-; CHECK-NEXT:    vslidedown.vi v10, v9, 11
-; CHECK-NEXT:    vmv.x.s t2, v11
-; CHECK-NEXT:    vslidedown.vi v11, v9, 12
-; CHECK-NEXT:    vmv.x.s t3, v10
-; CHECK-NEXT:    vslidedown.vi v10, v9, 13
-; CHECK-NEXT:    vmv.x.s t4, v11
-; CHECK-NEXT:    vslidedown.vi v11, v9, 14
-; CHECK-NEXT:    vslidedown.vi v9, v9, 15
-; CHECK-NEXT:    vmv.x.s t5, v10
 ; CHECK-NEXT:    vsetivli zero, 8, e8, mf2, ta, ma
-; CHECK-NEXT:    vmseq.vx v10, v8, a0
-; CHECK-NEXT:    vmv.x.s a0, v11
-; CHECK-NEXT:    vmseq.vx v11, v8, a1
-; CHECK-NEXT:    vmv.x.s a1, v9
-; CHECK-NEXT:    vmseq.vx v9, v8, a2
+; CHECK-NEXT:    vrgather.vi v10, v9, 1
+; CHECK-NEXT:    vrgather.vi v11, v9, 0
+; CHECK-NEXT:    vmseq.vv v10, v8, v10
+; CHECK-NEXT:    vmseq.vv v11, v8, v11
+; CHECK-NEXT:    vmor.mm v10, v11, v10
+; CHECK-NEXT:    vrgather.vi v11, v9, 2
+; CHECK-NEXT:    vmseq.vv v11, v8, v11
+; CHECK-NEXT:    vmor.mm v10, v10, v11
+; CHECK-NEXT:    vrgather.vi v11, v9, 3
+; CHECK-NEXT:    vmseq.vv v11, v8, v11
+; CHECK-NEXT:    vmor.mm v10, v10, v11
+; CHECK-NEXT:    vrgather.vi v11, v9, 4
+; CHECK-NEXT:    vmseq.vv v11, v8, v11
+; CHECK-NEXT:    vmor.mm v10, v10, v11
+; CHECK-NEXT:    vrgather.vi v11, v9, 5
+; CHECK-NEXT:    vmseq.vv v11, v8, v11
 ; CHECK-NEXT:    vmor.mm v10, v10, v11
-; CHECK-NEXT:    vmseq.vx v11, v8, a3
+; CHECK-NEXT:    vrgather.vi v11, v9, 6
+; CHECK-NEXT:    vmseq.vv v11, v8, v11
+; CHECK-NEXT:    vmor.mm v10, v10, v11
+; CHECK-NEXT:    vrgather.vi v11, v9, 7
+; CHECK-NEXT:    vmseq.vv v11, v8, v11
+; CHECK-NEXT:    vmor.mm v10, v10, v11
+; CHECK-NEXT:    vrgather.vi v11, v9, 8
----------------
lukel97 wrote:

I think 8 here is past VLMAX for this vector, so I think this will just splat 0 per the spec:

> The source vector can be read at any index < VLMAX regardless of vl.

> If an element index is out of range ( vs1[i] ≥ VLMAX ) then zero is returned for the element value.

The minimum vlen is 128, the SEW is 8 and LMUL=MF2, so VLMAX = 128 / 8 / 2 = 8. I think we need to check that Vec fits inside VT and bail if it doesn't

https://github.com/llvm/llvm-project/pull/117878