1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc < %s -mtriple=riscv64 -mattr=+v,+zvfh,+zvfbfmin | FileCheck %s --check-prefixes=CHECK,ZVFH
3 ; RUN: llc < %s -mtriple=riscv64 -mattr=+v,+zvfhmin,+zvfbfmin | FileCheck %s --check-prefixes=CHECK,ZVFHMIN
5 define <vscale x 2 x float> @vfwadd_same_operand_nxv2bf16(<vscale x 2 x bfloat> %arg, i32 signext %vl) {
6 ; CHECK-LABEL: vfwadd_same_operand_nxv2bf16:
7 ; CHECK: # %bb.0: # %bb
8 ; CHECK-NEXT: slli a0, a0, 32
9 ; CHECK-NEXT: srli a0, a0, 32
10 ; CHECK-NEXT: vsetvli zero, a0, e16, mf2, ta, ma
11 ; CHECK-NEXT: vfwcvtbf16.f.f.v v9, v8
12 ; CHECK-NEXT: vsetvli zero, zero, e32, m1, ta, ma
13 ; CHECK-NEXT: vfadd.vv v8, v9, v9
16 %tmp = call <vscale x 2 x float> @llvm.vp.fpext.nxv2f32.nxv2bf16(<vscale x 2 x bfloat> %arg, <vscale x 2 x i1> splat (i1 true), i32 %vl)
17 %tmp2 = call <vscale x 2 x float> @llvm.vp.fadd.nxv2f32(<vscale x 2 x float> %tmp, <vscale x 2 x float> %tmp, <vscale x 2 x i1> splat (i1 true), i32 %vl)
18 ret <vscale x 2 x float> %tmp2
21 ; Make sure we don't widen vfmadd.vv -> vfwmaccvbf16.vv if there's other
23 define <vscale x 2 x float> @vfwadd_same_operand_nxv2bf16_multiuse(<vscale x 2 x bfloat> %arg, <vscale x 2 x float> %acc, i32 signext %vl, ptr %p) {
24 ; CHECK-LABEL: vfwadd_same_operand_nxv2bf16_multiuse:
25 ; CHECK: # %bb.0: # %bb
26 ; CHECK-NEXT: slli a0, a0, 32
27 ; CHECK-NEXT: srli a0, a0, 32
28 ; CHECK-NEXT: vsetvli zero, a0, e16, mf2, ta, ma
29 ; CHECK-NEXT: vfwcvtbf16.f.f.v v10, v8
30 ; CHECK-NEXT: vsetvli zero, zero, e32, m1, ta, ma
31 ; CHECK-NEXT: vfadd.vv v8, v10, v10
32 ; CHECK-NEXT: vfmadd.vv v10, v10, v9
33 ; CHECK-NEXT: vs1r.v v10, (a1)
36 %tmp = call <vscale x 2 x float> @llvm.vp.fpext.nxv2f32.nxv2bf16(<vscale x 2 x bfloat> %arg, <vscale x 2 x i1> splat (i1 true), i32 %vl)
37 %tmp2 = call <vscale x 2 x float> @llvm.vp.fadd.nxv2f32(<vscale x 2 x float> %tmp, <vscale x 2 x float> %tmp, <vscale x 2 x i1> splat (i1 true), i32 %vl)
38 %tmp3 = call <vscale x 2 x float> @llvm.vp.fma.nxv2f32(<vscale x 2 x float> %tmp, <vscale x 2 x float> %tmp, <vscale x 2 x float> %acc, <vscale x 2 x i1> splat (i1 true), i32 %vl)
39 store <vscale x 2 x float> %tmp3, ptr %p
40 ret <vscale x 2 x float> %tmp2
43 define <vscale x 2 x float> @vfwadd_same_operand(<vscale x 2 x half> %arg, i32 signext %vl) {
44 ; ZVFH-LABEL: vfwadd_same_operand:
45 ; ZVFH: # %bb.0: # %bb
46 ; ZVFH-NEXT: slli a0, a0, 32
47 ; ZVFH-NEXT: srli a0, a0, 32
48 ; ZVFH-NEXT: vsetvli zero, a0, e16, mf2, ta, ma
49 ; ZVFH-NEXT: vfwadd.vv v9, v8, v8
50 ; ZVFH-NEXT: vmv1r.v v8, v9
53 ; ZVFHMIN-LABEL: vfwadd_same_operand:
54 ; ZVFHMIN: # %bb.0: # %bb
55 ; ZVFHMIN-NEXT: slli a0, a0, 32
56 ; ZVFHMIN-NEXT: srli a0, a0, 32
57 ; ZVFHMIN-NEXT: vsetvli zero, a0, e16, mf2, ta, ma
58 ; ZVFHMIN-NEXT: vfwcvt.f.f.v v9, v8
59 ; ZVFHMIN-NEXT: vsetvli zero, zero, e32, m1, ta, ma
60 ; ZVFHMIN-NEXT: vfadd.vv v8, v9, v9
63 %tmp = call <vscale x 2 x float> @llvm.vp.fpext.nxv2f32.nxv2f16(<vscale x 2 x half> %arg, <vscale x 2 x i1> splat (i1 true), i32 %vl)
64 %tmp2 = call <vscale x 2 x float> @llvm.vp.fadd.nxv2f32(<vscale x 2 x float> %tmp, <vscale x 2 x float> %tmp, <vscale x 2 x i1> splat (i1 true), i32 %vl)
65 ret <vscale x 2 x float> %tmp2
68 define <vscale x 2 x float> @vfwadd_tu(<vscale x 2 x half> %arg, <vscale x 2 x float> %arg1, i32 signext %arg2) {
69 ; ZVFH-LABEL: vfwadd_tu:
70 ; ZVFH: # %bb.0: # %bb
71 ; ZVFH-NEXT: slli a0, a0, 32
72 ; ZVFH-NEXT: srli a0, a0, 32
73 ; ZVFH-NEXT: vsetvli zero, a0, e16, mf2, tu, ma
74 ; ZVFH-NEXT: vfwadd.wv v9, v9, v8
75 ; ZVFH-NEXT: vmv1r.v v8, v9
78 ; ZVFHMIN-LABEL: vfwadd_tu:
79 ; ZVFHMIN: # %bb.0: # %bb
80 ; ZVFHMIN-NEXT: slli a0, a0, 32
81 ; ZVFHMIN-NEXT: srli a0, a0, 32
82 ; ZVFHMIN-NEXT: vsetvli zero, a0, e16, mf2, ta, ma
83 ; ZVFHMIN-NEXT: vfwcvt.f.f.v v10, v8
84 ; ZVFHMIN-NEXT: vsetvli zero, zero, e32, m1, tu, ma
85 ; ZVFHMIN-NEXT: vfadd.vv v9, v9, v10
86 ; ZVFHMIN-NEXT: vmv1r.v v8, v9
89 %tmp = call <vscale x 2 x float> @llvm.vp.fpext.nxv2f32.nxv2f16(<vscale x 2 x half> %arg, <vscale x 2 x i1> splat (i1 true), i32 %arg2)
90 %tmp3 = call <vscale x 2 x float> @llvm.vp.fadd.nxv2f32(<vscale x 2 x float> %arg1, <vscale x 2 x float> %tmp, <vscale x 2 x i1> splat (i1 true), i32 %arg2)
91 %tmp4 = call <vscale x 2 x float> @llvm.vp.merge.nxv2f32(<vscale x 2 x i1> splat (i1 true), <vscale x 2 x float> %tmp3, <vscale x 2 x float> %arg1, i32 %arg2)
92 ret <vscale x 2 x float> %tmp4
95 declare <vscale x 2 x float> @llvm.vp.fpext.nxv2f32.nxv2f16(<vscale x 2 x half>, <vscale x 2 x i1>, i32)
96 declare <vscale x 2 x float> @llvm.vp.fadd.nxv2f32(<vscale x 2 x float>, <vscale x 2 x float>, <vscale x 2 x i1>, i32)
97 declare <vscale x 2 x float> @llvm.vp.merge.nxv2f32(<vscale x 2 x i1>, <vscale x 2 x float>, <vscale x 2 x float>, i32)