1 ; RUN: llc -march=amdgcn -mcpu=gfx1010 -mattr=-nsa-encoding -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,NONSA %s
2 ; RUN: llc -march=amdgcn -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX1010,NSA %s
3 ; RUN: llc -march=amdgcn -mcpu=gfx1030 -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX1030,NSA %s
5 ; GCN-LABEL: {{^}}sample_2d:
8 ; GCN: v_mov_b32_e32 v2, v0
10 ; GCN: image_sample v[0:3], v[1:2],
11 define amdgpu_ps <4 x float> @sample_2d(<8 x i32> inreg %rsrc, <4 x i32> inreg %samp, float %t, float %s) {
13 %v = call <4 x float> @llvm.amdgcn.image.sample.2d.v4f32.f32(i32 15, float %s, float %t, <8 x i32> %rsrc, <4 x i32> %samp, i1 0, i32 0, i32 0)
17 ; GCN-LABEL: {{^}}sample_3d:
18 ; NONSA: v_mov_b32_e32 v3, v0
19 ; NONSA: image_sample v[0:3], v[1:3],
20 ; NSA: image_sample v[0:3], [v1, v2, v0],
21 define amdgpu_ps <4 x float> @sample_3d(<8 x i32> inreg %rsrc, <4 x i32> inreg %samp, float %r, float %s, float %t) {
23 %v = call <4 x float> @llvm.amdgcn.image.sample.3d.v4f32.f32(i32 15, float %s, float %t, float %r, <8 x i32> %rsrc, <4 x i32> %samp, i1 0, i32 0, i32 0)
27 ; GCN-LABEL: {{^}}sample_d_3d:
28 ; GFX1010: image_sample_d v[0:3], v[7:22],
29 ; GFX1030: image_sample_d v[0:3], [v3, v8, v7, v5, v4, v6, v0, v2, v1],
30 define amdgpu_ps <4 x float> @sample_d_3d(<8 x i32> inreg %rsrc, <4 x i32> inreg %samp, float %s, float %r, float %t, float %dsdh, float %dtdv, float %dsdv, float %drdv, float %drdh, float %dtdh) {
32 %v = call <4 x float> @llvm.amdgcn.image.sample.d.3d.v4f32.f32(i32 15, float %dsdh, float %dtdh, float %drdh, float %dsdv, float %dtdv, float %drdv, float %s, float %t, float %r, <8 x i32> %rsrc, <4 x i32> %samp, i1 0, i32 0, i32 0)
36 ; GCN-LABEL: {{^}}sample_contig_nsa:
37 ; NONSA: image_sample_c_l v5, v[0:4],
38 ; NSA: image_sample_c_l v{{[0-9]+}}, v[0:4],
39 ; NSA: image_sample v{{[0-9]+}}, [v6, v7, v5],
40 define amdgpu_ps <2 x float> @sample_contig_nsa(<8 x i32> inreg %rsrc, <4 x i32> inreg %samp, float %zcompare, float %s1, float %t1, float %r1, float %lod, float %r2, float %s2, float %t2) {
42 %v1 = call float @llvm.amdgcn.image.sample.c.l.3d.f32.f32(i32 1, float %zcompare, float %s1, float %t1, float %r1, float %lod, <8 x i32> %rsrc, <4 x i32> %samp, i1 0, i32 0, i32 0)
43 %v2 = call float @llvm.amdgcn.image.sample.3d.f32.f32(i32 1, float %s2, float %t2, float %r2, <8 x i32> %rsrc, <4 x i32> %samp, i1 0, i32 0, i32 0)
44 %r.0 = insertelement <2 x float> undef, float %v1, i32 0
45 %r = insertelement <2 x float> %r.0, float %v2, i32 1
49 ; GCN-LABEL: {{^}}sample_nsa_nsa:
50 ; NSA: image_sample_c_l v{{[0-9]+}}, [v1, v2, v3, v4, v0],
51 ; NSA: image_sample v{{[0-9]+}}, [v6, v7, v5],
52 define amdgpu_ps <2 x float> @sample_nsa_nsa(<8 x i32> inreg %rsrc, <4 x i32> inreg %samp, float %lod, float %zcompare, float %s1, float %t1, float %r1, float %r2, float %s2, float %t2) {
54 %v1 = call float @llvm.amdgcn.image.sample.c.l.3d.f32.f32(i32 1, float %zcompare, float %s1, float %t1, float %r1, float %lod, <8 x i32> %rsrc, <4 x i32> %samp, i1 0, i32 0, i32 0)
55 %v2 = call float @llvm.amdgcn.image.sample.3d.f32.f32(i32 1, float %s2, float %t2, float %r2, <8 x i32> %rsrc, <4 x i32> %samp, i1 0, i32 0, i32 0)
56 %r.0 = insertelement <2 x float> undef, float %v1, i32 0
57 %r = insertelement <2 x float> %r.0, float %v2, i32 1
61 ; GCN-LABEL: {{^}}sample_nsa_contig:
62 ; NSA: image_sample_c_l v{{[0-9]+}}, [v1, v2, v3, v4, v0],
63 ; NSA: image_sample v{{[0-9]+}}, v[5:7],
64 define amdgpu_ps <2 x float> @sample_nsa_contig(<8 x i32> inreg %rsrc, <4 x i32> inreg %samp, float %lod, float %zcompare, float %s1, float %t1, float %r1, float %s2, float %t2, float %r2) {
66 %v1 = call float @llvm.amdgcn.image.sample.c.l.3d.f32.f32(i32 1, float %zcompare, float %s1, float %t1, float %r1, float %lod, <8 x i32> %rsrc, <4 x i32> %samp, i1 0, i32 0, i32 0)
67 %v2 = call float @llvm.amdgcn.image.sample.3d.f32.f32(i32 1, float %s2, float %t2, float %r2, <8 x i32> %rsrc, <4 x i32> %samp, i1 0, i32 0, i32 0)
68 %r.0 = insertelement <2 x float> undef, float %v1, i32 0
69 %r = insertelement <2 x float> %r.0, float %v2, i32 1
73 ; GCN-LABEL: {{^}}sample_contig_contig:
74 ; NSA: image_sample_c_l v{{[0-9]+}}, v[0:4],
75 ; NSA: image_sample v{{[0-9]+}}, v[5:7],
76 ; NONSA: image_sample_c_l v{{[0-9]+}}, v[0:4],
77 ; NONSA: image_sample v{{[0-9]+}}, v[5:7],
78 define amdgpu_ps <2 x float> @sample_contig_contig(<8 x i32> inreg %rsrc, <4 x i32> inreg %samp, float %zcompare, float %s1, float %t1, float %r1, float %lod, float %s2, float %t2, float %r2) {
80 %v1 = call float @llvm.amdgcn.image.sample.c.l.3d.f32.f32(i32 1, float %zcompare, float %s1, float %t1, float %r1, float %lod, <8 x i32> %rsrc, <4 x i32> %samp, i1 0, i32 0, i32 0)
81 %v2 = call float @llvm.amdgcn.image.sample.3d.f32.f32(i32 1, float %s2, float %t2, float %r2, <8 x i32> %rsrc, <4 x i32> %samp, i1 0, i32 0, i32 0)
82 %r.0 = insertelement <2 x float> undef, float %v1, i32 0
83 %r = insertelement <2 x float> %r.0, float %v2, i32 1
87 ; Test that undef inputs with NSA are handled safely; these tests used to crash.
89 ; GCN-LABEL: {{^}}sample_undef_undef_undef_undef:
90 ; GCN: image_sample_c_b v0, v[0:3], s[0:7], s[8:11] dmask:0x1 dim:SQ_RSRC_IMG_1D_ARRAY
91 define amdgpu_ps float @sample_undef_undef_undef_undef(<8 x i32> inreg %rsrc, <4 x i32> inreg %samp) {
92 %r = call float @llvm.amdgcn.image.sample.c.b.1darray.f32.f32.f32(i32 1, float undef, float undef, float undef, float undef, <8 x i32> %rsrc, <4 x i32> %samp, i1 false, i32 0, i32 0)
96 ; GCN-LABEL: {{^}}sample_undef_undef_undef_def:
97 ; NONSA: v_mov_b32_e32 v3, v0
98 ; NONSA: image_sample_c_b v0, v[0:3], s[0:7], s[8:11] dmask:0x1 dim:SQ_RSRC_IMG_1D_ARRAY
99 ; NSA: image_sample_c_b v0, [v0, v0, v0, v0], s[0:7], s[8:11] dmask:0x1 dim:SQ_RSRC_IMG_1D_ARRAY
100 define amdgpu_ps float @sample_undef_undef_undef_def(<8 x i32> inreg %rsrc, <4 x i32> inreg %samp, float %layer) {
101 %r = call float @llvm.amdgcn.image.sample.c.b.1darray.f32.f32.f32(i32 1, float undef, float undef, float undef, float %layer, <8 x i32> %rsrc, <4 x i32> %samp, i1 false, i32 0, i32 0)
105 ; GCN-LABEL: {{^}}sample_undef_undef_undef_def_rnd:
106 ; GCN: v_rndne_f32_e32 v3, v0
107 ; GCN: image_sample_c_b v0, v[0:3], s[0:7], s[8:11] dmask:0x1 dim:SQ_RSRC_IMG_1D_ARRAY
108 define amdgpu_ps float @sample_undef_undef_undef_def_rnd(<8 x i32> inreg %rsrc, <4 x i32> inreg %samp, float %layer) {
109 %layer_rnd = call float @llvm.rint.f32(float %layer)
110 %r = call float @llvm.amdgcn.image.sample.c.b.1darray.f32.f32.f32(i32 1, float undef, float undef, float undef, float %layer_rnd, <8 x i32> %rsrc, <4 x i32> %samp, i1 false, i32 0, i32 0)
114 ; GCN-LABEL: {{^}}sample_def_undef_undef_undef:
115 ; GCN: v_add_f32_e32 v0, 1.0, v0
116 ; GCN: image_sample_c_b v0, v[0:3], s[0:7], s[8:11] dmask:0x1 dim:SQ_RSRC_IMG_1D_ARRAY
117 define amdgpu_ps float @sample_def_undef_undef_undef(<8 x i32> inreg %rsrc, <4 x i32> inreg %samp, float %z0) {
118 ; The NSA reassign pass is conservative (quite reasonably!) when one of the operands
119 ; comes directly from a function argument (via COPY). To test that NSA can be
120 ; eliminated in the presence of undef, just add an arbitrary intermediate
122 %c0 = fadd float %z0, 1.0
123 %r = call float @llvm.amdgcn.image.sample.c.b.1darray.f32.f32.f32(i32 1, float %c0, float undef, float undef, float undef, <8 x i32> %rsrc, <4 x i32> %samp, i1 false, i32 0, i32 0)
127 declare <4 x float> @llvm.amdgcn.image.sample.2d.v4f32.f32(i32, float, float, <8 x i32>, <4 x i32>, i1, i32, i32) #1
128 declare <4 x float> @llvm.amdgcn.image.sample.3d.v4f32.f32(i32, float, float, float, <8 x i32>, <4 x i32>, i1, i32, i32) #1
129 declare <4 x float> @llvm.amdgcn.image.sample.d.3d.v4f32.f32(i32, float, float, float, float, float, float, float, float, float, <8 x i32>, <4 x i32>, i1, i32, i32) #1
131 declare float @llvm.amdgcn.image.sample.3d.f32.f32(i32, float, float, float, <8 x i32>, <4 x i32>, i1, i32, i32) #1
132 declare float @llvm.amdgcn.image.sample.c.l.3d.f32.f32(i32, float, float, float, float, float, <8 x i32>, <4 x i32>, i1, i32, i32) #1
134 declare float @llvm.rint.f32(float) #2
135 declare float @llvm.amdgcn.image.sample.c.b.1darray.f32.f32.f32(i32 immarg, float, float, float, float, <8 x i32>, <4 x i32>, i1 immarg, i32 immarg, i32 immarg) #1
137 attributes #1 = { nounwind readonly }
138 attributes #2 = { nounwind readnone speculatable willreturn }