[llvm] [RISCV] Merge shuffle sources if lanes are disjoint (PR #119401)

[Philip Reames via llvm-commits]

================
@@ -11,21 +11,18 @@ define void @deinterleave3_0_i8(ptr %in, ptr %out) {
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    vsetivli zero, 16, e8, m1, ta, ma
 ; CHECK-NEXT:    vle8.v v8, (a0)
+; CHECK-NEXT:    lui a0, %hi(.LCPI0_0)
+; CHECK-NEXT:    addi a0, a0, %lo(.LCPI0_0)
 ; CHECK-NEXT:    vsetivli zero, 8, e8, mf2, ta, ma
-; CHECK-NEXT:    vmv.v.i v9, -8
-; CHECK-NEXT:    vid.v v10
-; CHECK-NEXT:    li a0, 3
-; CHECK-NEXT:    vmadd.vx v10, a0, v9
+; CHECK-NEXT:    vle8.v v9, (a0)
 ; CHECK-NEXT:    li a0, 73
-; CHECK-NEXT:    vmv.s.x v9, a0
-; CHECK-NEXT:    li a0, 56
 ; CHECK-NEXT:    vmv.s.x v0, a0
-; CHECK-NEXT:    vcompress.vm v11, v8, v9
 ; CHECK-NEXT:    vsetivli zero, 8, e8, m1, ta, ma
-; CHECK-NEXT:    vslidedown.vi v8, v8, 8
-; CHECK-NEXT:    vsetivli zero, 8, e8, mf2, ta, mu
-; CHECK-NEXT:    vrgather.vv v11, v8, v10, v0.t
-; CHECK-NEXT:    vse8.v v11, (a1)
+; CHECK-NEXT:    vslidedown.vi v10, v8, 8
----------------
preames wrote:

This is a fascinating accident, I hadn't considered this would catch the general deinterleave case.  I'd been exploring a targeted change specific for deinterleave type cases, and this gets it generically.  Nice!

(The new codegen is fairly neutral on m1, but if the result is > m1, the vslidedown and smaller shuffle helps a ton.)

https://github.com/llvm/llvm-project/pull/119401