xref: /llvm-project/llvm/test/CodeGen/RISCV/rvv/vp-combine-reverse-load.ll (revision fc7a1ed0ba5f437bc7f262f562e83488225f0152) 
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc -mtriple=riscv64 -mattr=+f,+v -verify-machineinstrs < %s | FileCheck %s
3
4define <vscale x 2 x float> @test_reverse_load_combiner(<vscale x 2 x float>* %ptr, i32 zeroext %evl) {
5; CHECK-LABEL: test_reverse_load_combiner:
6; CHECK:       # %bb.0:
7; CHECK-NEXT:    slli a2, a1, 2
8; CHECK-NEXT:    add a0, a2, a0
9; CHECK-NEXT:    addi a0, a0, -4
10; CHECK-NEXT:    li a2, -4
11; CHECK-NEXT:    vsetvli zero, a1, e32, m1, ta, ma
12; CHECK-NEXT:    vlse32.v v8, (a0), a2
13; CHECK-NEXT:    ret
14  %load = call <vscale x 2 x float> @llvm.vp.load.nxv2f32.p0nxv2f32(<vscale x 2 x float>* %ptr, <vscale x 2 x i1> splat (i1 true), i32 %evl)
15  %rev = call <vscale x 2 x float> @llvm.experimental.vp.reverse.nxv2f32(<vscale x 2 x float> %load, <vscale x 2 x i1> splat (i1 true), i32 %evl)
16  ret <vscale x 2 x float> %rev
17}
18
19define <vscale x 2 x float> @test_load_mask_is_vp_reverse(<vscale x 2 x float>* %ptr, <vscale x 2 x i1> %mask, i32 zeroext %evl) {
20; CHECK-LABEL: test_load_mask_is_vp_reverse:
21; CHECK:       # %bb.0:
22; CHECK-NEXT:    slli a2, a1, 2
23; CHECK-NEXT:    add a0, a2, a0
24; CHECK-NEXT:    addi a0, a0, -4
25; CHECK-NEXT:    li a2, -4
26; CHECK-NEXT:    vsetvli zero, a1, e32, m1, ta, ma
27; CHECK-NEXT:    vlse32.v v8, (a0), a2, v0.t
28; CHECK-NEXT:    ret
29  %loadmask = call <vscale x 2 x i1> @llvm.experimental.vp.reverse.nxv2i1(<vscale x 2 x i1> %mask, <vscale x 2 x i1> splat (i1 true), i32 %evl)
30  %load = call <vscale x 2 x float> @llvm.vp.load.nxv2f32.p0nxv2f32(<vscale x 2 x float>* %ptr, <vscale x 2 x i1> %loadmask, i32 %evl)
31  %rev = call <vscale x 2 x float> @llvm.experimental.vp.reverse.nxv2f32(<vscale x 2 x float> %load, <vscale x 2 x i1> splat (i1 true), i32 %evl)
32  ret <vscale x 2 x float> %rev
33}
34
35define <vscale x 2 x float> @test_load_mask_not_all_one(<vscale x 2 x float>* %ptr, <vscale x 2 x i1> %notallones, i32 zeroext %evl) {
36; CHECK-LABEL: test_load_mask_not_all_one:
37; CHECK:       # %bb.0:
38; CHECK-NEXT:    vsetvli zero, a1, e32, m1, ta, ma
39; CHECK-NEXT:    vle32.v v9, (a0), v0.t
40; CHECK-NEXT:    vid.v v8, v0.t
41; CHECK-NEXT:    addi a1, a1, -1
42; CHECK-NEXT:    vrsub.vx v10, v8, a1, v0.t
43; CHECK-NEXT:    vrgather.vv v8, v9, v10, v0.t
44; CHECK-NEXT:    ret
45  %load = call <vscale x 2 x float> @llvm.vp.load.nxv2f32.p0nxv2f32(<vscale x 2 x float>* %ptr, <vscale x 2 x i1> %notallones, i32 %evl)
46  %rev = call <vscale x 2 x float> @llvm.experimental.vp.reverse.nxv2f32(<vscale x 2 x float> %load, <vscale x 2 x i1> %notallones, i32 %evl)
47  ret <vscale x 2 x float> %rev
48}
49
50define <vscale x 2 x float> @test_different_evl(<vscale x 2 x float>* %ptr, <vscale x 2 x i1> %mask, i32 zeroext %evl1, i32 zeroext %evl2) {
51; CHECK-LABEL: test_different_evl:
52; CHECK:       # %bb.0:
53; CHECK-NEXT:    addi a3, a1, -1
54; CHECK-NEXT:    vsetvli zero, a1, e16, mf2, ta, ma
55; CHECK-NEXT:    vid.v v8
56; CHECK-NEXT:    vsetvli zero, zero, e8, mf4, ta, ma
57; CHECK-NEXT:    vmv.v.i v9, 0
58; CHECK-NEXT:    vsetvli zero, zero, e16, mf2, ta, ma
59; CHECK-NEXT:    vrsub.vx v8, v8, a3
60; CHECK-NEXT:    vsetvli zero, zero, e8, mf4, ta, ma
61; CHECK-NEXT:    vmerge.vim v9, v9, 1, v0
62; CHECK-NEXT:    vrgatherei16.vv v10, v9, v8
63; CHECK-NEXT:    vmsne.vi v0, v10, 0
64; CHECK-NEXT:    vsetvli zero, a2, e32, m1, ta, ma
65; CHECK-NEXT:    vle32.v v9, (a0), v0.t
66; CHECK-NEXT:    addi a2, a2, -1
67; CHECK-NEXT:    vid.v v8
68; CHECK-NEXT:    vrsub.vx v10, v8, a2
69; CHECK-NEXT:    vrgather.vv v8, v9, v10
70; CHECK-NEXT:    ret
71  %loadmask = call <vscale x 2 x i1> @llvm.experimental.vp.reverse.nxv2i1(<vscale x 2 x i1> %mask, <vscale x 2 x i1> splat (i1 true), i32 %evl1)
72  %load = call <vscale x 2 x float> @llvm.vp.load.nxv2f32.p0nxv2f32(<vscale x 2 x float>* %ptr, <vscale x 2 x i1> %loadmask, i32 %evl2)
73  %rev = call <vscale x 2 x float> @llvm.experimental.vp.reverse.nxv2f32(<vscale x 2 x float> %load, <vscale x 2 x i1> splat (i1 true), i32 %evl2)
74  ret <vscale x 2 x float> %rev
75}
76
77declare <vscale x 2 x float> @llvm.vp.load.nxv2f32.p0nxv2f32(<vscale x 2 x float>* nocapture, <vscale x 2 x i1>, i32)
78declare <vscale x 2 x float> @llvm.experimental.vp.reverse.nxv2f32(<vscale x 2 x float>, <vscale x 2 x i1>, i32)
79declare <vscale x 2 x i1> @llvm.experimental.vp.reverse.nxv2i1(<vscale x 2 x i1>, <vscale x 2 x i1>, i32)
80