|
150
|
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
|
|
|
2 ; RUN: llc < %s -mtriple=amdgcn-amd-mesa3d -mcpu=fiji -verify-machineinstrs | FileCheck -check-prefix=VI %s
|
|
|
3 ; RUN: llc < %s -mtriple=amdgcn-amd-mesa3d -mcpu=gfx900 -verify-machineinstrs | FileCheck -check-prefix=GFX9 %s
|
|
|
4 ; RUN: llc < %s -mtriple=amdgcn-amd-mesa3d -mcpu=gfx1010 -verify-machineinstrs | FileCheck -check-prefix=GFX10 %s
|
|
|
5
|
|
|
6 ; ===================================================================================
|
|
|
7 ; V_ADD3_U32
|
|
|
8 ; ===================================================================================
|
|
|
9
|
|
|
10 define amdgpu_ps float @add3(i32 %a, i32 %b, i32 %c) {
|
|
|
11 ; VI-LABEL: add3:
|
|
|
12 ; VI: ; %bb.0:
|
|
|
13 ; VI-NEXT: v_add_u32_e32 v0, vcc, v0, v1
|
|
|
14 ; VI-NEXT: v_add_u32_e32 v0, vcc, v0, v2
|
|
|
15 ; VI-NEXT: ; return to shader part epilog
|
|
|
16 ;
|
|
|
17 ; GFX9-LABEL: add3:
|
|
|
18 ; GFX9: ; %bb.0:
|
|
|
19 ; GFX9-NEXT: v_add3_u32 v0, v0, v1, v2
|
|
|
20 ; GFX9-NEXT: ; return to shader part epilog
|
|
|
21 ;
|
|
|
22 ; GFX10-LABEL: add3:
|
|
|
23 ; GFX10: ; %bb.0:
|
|
|
24 ; GFX10-NEXT: v_add3_u32 v0, v0, v1, v2
|
|
|
25 ; GFX10-NEXT: ; return to shader part epilog
|
|
|
26 %x = add i32 %a, %b
|
|
|
27 %result = add i32 %x, %c
|
|
|
28 %bc = bitcast i32 %result to float
|
|
|
29 ret float %bc
|
|
|
30 }
|
|
|
31
|
|
|
32 ; V_MAD_U32_U24 is given higher priority.
|
|
|
33 define amdgpu_ps float @mad_no_add3(i32 %a, i32 %b, i32 %c, i32 %d, i32 %e) {
|
|
|
34 ; VI-LABEL: mad_no_add3:
|
|
|
35 ; VI: ; %bb.0:
|
|
|
36 ; VI-NEXT: v_mad_u32_u24 v0, v0, v1, v4
|
|
|
37 ; VI-NEXT: v_mad_u32_u24 v0, v2, v3, v0
|
|
|
38 ; VI-NEXT: ; return to shader part epilog
|
|
|
39 ;
|
|
|
40 ; GFX9-LABEL: mad_no_add3:
|
|
|
41 ; GFX9: ; %bb.0:
|
|
|
42 ; GFX9-NEXT: v_mad_u32_u24 v0, v0, v1, v4
|
|
|
43 ; GFX9-NEXT: v_mad_u32_u24 v0, v2, v3, v0
|
|
|
44 ; GFX9-NEXT: ; return to shader part epilog
|
|
|
45 ;
|
|
|
46 ; GFX10-LABEL: mad_no_add3:
|
|
|
47 ; GFX10: ; %bb.0:
|
|
|
48 ; GFX10-NEXT: v_mad_u32_u24 v0, v0, v1, v4
|
|
|
49 ; GFX10-NEXT: v_mad_u32_u24 v0, v2, v3, v0
|
|
|
50 ; GFX10-NEXT: ; return to shader part epilog
|
|
|
51 %a0 = shl i32 %a, 8
|
|
|
52 %a1 = lshr i32 %a0, 8
|
|
|
53 %b0 = shl i32 %b, 8
|
|
|
54 %b1 = lshr i32 %b0, 8
|
|
|
55 %mul1 = mul i32 %a1, %b1
|
|
|
56
|
|
|
57 %c0 = shl i32 %c, 8
|
|
|
58 %c1 = lshr i32 %c0, 8
|
|
|
59 %d0 = shl i32 %d, 8
|
|
|
60 %d1 = lshr i32 %d0, 8
|
|
|
61 %mul2 = mul i32 %c1, %d1
|
|
|
62
|
|
|
63 %add0 = add i32 %e, %mul1
|
|
|
64 %add1 = add i32 %mul2, %add0
|
|
|
65
|
|
|
66 %bc = bitcast i32 %add1 to float
|
|
|
67 ret float %bc
|
|
|
68 }
|
|
|
69
|
|
|
70 ; ThreeOp instruction variant not used due to Constant Bus Limitations
|
|
|
71 ; TODO: with reassociation it is possible to replace a v_add_u32_e32 with a s_add_i32
|
|
|
72 define amdgpu_ps float @add3_vgpr_b(i32 inreg %a, i32 %b, i32 inreg %c) {
|
|
|
73 ; VI-LABEL: add3_vgpr_b:
|
|
|
74 ; VI: ; %bb.0:
|
|
|
75 ; VI-NEXT: s_add_i32 s3, s3, s2
|
|
|
76 ; VI-NEXT: v_add_u32_e32 v0, vcc, s3, v0
|
|
|
77 ; VI-NEXT: ; return to shader part epilog
|
|
|
78 ;
|
|
|
79 ; GFX9-LABEL: add3_vgpr_b:
|
|
|
80 ; GFX9: ; %bb.0:
|
|
|
81 ; GFX9-NEXT: s_add_i32 s3, s3, s2
|
|
|
82 ; GFX9-NEXT: v_add_u32_e32 v0, s3, v0
|
|
|
83 ; GFX9-NEXT: ; return to shader part epilog
|
|
|
84 ;
|
|
|
85 ; GFX10-LABEL: add3_vgpr_b:
|
|
|
86 ; GFX10: ; %bb.0:
|
|
|
87 ; GFX10-NEXT: v_add3_u32 v0, s3, s2, v0
|
|
|
88 ; GFX10-NEXT: ; return to shader part epilog
|
|
|
89 %x = add i32 %a, %b
|
|
|
90 %result = add i32 %x, %c
|
|
|
91 %bc = bitcast i32 %result to float
|
|
|
92 ret float %bc
|
|
|
93 }
|
|
|
94
|
|
|
95 define amdgpu_ps float @add3_vgpr_all2(i32 %a, i32 %b, i32 %c) {
|
|
|
96 ; VI-LABEL: add3_vgpr_all2:
|
|
|
97 ; VI: ; %bb.0:
|
|
|
98 ; VI-NEXT: v_add_u32_e32 v1, vcc, v1, v2
|
|
|
99 ; VI-NEXT: v_add_u32_e32 v0, vcc, v0, v1
|
|
|
100 ; VI-NEXT: ; return to shader part epilog
|
|
|
101 ;
|
|
|
102 ; GFX9-LABEL: add3_vgpr_all2:
|
|
|
103 ; GFX9: ; %bb.0:
|
|
|
104 ; GFX9-NEXT: v_add3_u32 v0, v1, v2, v0
|
|
|
105 ; GFX9-NEXT: ; return to shader part epilog
|
|
|
106 ;
|
|
|
107 ; GFX10-LABEL: add3_vgpr_all2:
|
|
|
108 ; GFX10: ; %bb.0:
|
|
|
109 ; GFX10-NEXT: v_add3_u32 v0, v1, v2, v0
|
|
|
110 ; GFX10-NEXT: ; return to shader part epilog
|
|
|
111 %x = add i32 %b, %c
|
|
|
112 %result = add i32 %a, %x
|
|
|
113 %bc = bitcast i32 %result to float
|
|
|
114 ret float %bc
|
|
|
115 }
|
|
|
116
|
|
|
117 define amdgpu_ps float @add3_vgpr_bc(i32 inreg %a, i32 %b, i32 %c) {
|
|
|
118 ; VI-LABEL: add3_vgpr_bc:
|
|
|
119 ; VI: ; %bb.0:
|
|
|
120 ; VI-NEXT: v_add_u32_e32 v0, vcc, s2, v0
|
|
|
121 ; VI-NEXT: v_add_u32_e32 v0, vcc, v0, v1
|
|
|
122 ; VI-NEXT: ; return to shader part epilog
|
|
|
123 ;
|
|
|
124 ; GFX9-LABEL: add3_vgpr_bc:
|
|
|
125 ; GFX9: ; %bb.0:
|
|
|
126 ; GFX9-NEXT: v_add3_u32 v0, s2, v0, v1
|
|
|
127 ; GFX9-NEXT: ; return to shader part epilog
|
|
|
128 ;
|
|
|
129 ; GFX10-LABEL: add3_vgpr_bc:
|
|
|
130 ; GFX10: ; %bb.0:
|
|
|
131 ; GFX10-NEXT: v_add3_u32 v0, s2, v0, v1
|
|
|
132 ; GFX10-NEXT: ; return to shader part epilog
|
|
|
133 %x = add i32 %a, %b
|
|
|
134 %result = add i32 %x, %c
|
|
|
135 %bc = bitcast i32 %result to float
|
|
|
136 ret float %bc
|
|
|
137 }
|
|
|
138
|
|
|
139 define amdgpu_ps float @add3_vgpr_const(i32 %a, i32 %b) {
|
|
|
140 ; VI-LABEL: add3_vgpr_const:
|
|
|
141 ; VI: ; %bb.0:
|
|
|
142 ; VI-NEXT: v_add_u32_e32 v0, vcc, v0, v1
|
|
|
143 ; VI-NEXT: v_add_u32_e32 v0, vcc, 16, v0
|
|
|
144 ; VI-NEXT: ; return to shader part epilog
|
|
|
145 ;
|
|
|
146 ; GFX9-LABEL: add3_vgpr_const:
|
|
|
147 ; GFX9: ; %bb.0:
|
|
|
148 ; GFX9-NEXT: v_add3_u32 v0, v0, v1, 16
|
|
|
149 ; GFX9-NEXT: ; return to shader part epilog
|
|
|
150 ;
|
|
|
151 ; GFX10-LABEL: add3_vgpr_const:
|
|
|
152 ; GFX10: ; %bb.0:
|
|
|
153 ; GFX10-NEXT: v_add3_u32 v0, v0, v1, 16
|
|
|
154 ; GFX10-NEXT: ; return to shader part epilog
|
|
|
155 %x = add i32 %a, %b
|
|
|
156 %result = add i32 %x, 16
|
|
|
157 %bc = bitcast i32 %result to float
|
|
|
158 ret float %bc
|
|
|
159 }
|
|
|
160
|
|
|
161 define amdgpu_ps <2 x float> @add3_multiuse_outer(i32 %a, i32 %b, i32 %c, i32 %x) {
|
|
|
162 ; VI-LABEL: add3_multiuse_outer:
|
|
|
163 ; VI: ; %bb.0:
|
|
|
164 ; VI-NEXT: v_add_u32_e32 v0, vcc, v0, v1
|
|
|
165 ; VI-NEXT: v_add_u32_e32 v0, vcc, v0, v2
|
|
|
166 ; VI-NEXT: v_mul_lo_u32 v1, v0, v3
|
|
|
167 ; VI-NEXT: ; return to shader part epilog
|
|
|
168 ;
|
|
|
169 ; GFX9-LABEL: add3_multiuse_outer:
|
|
|
170 ; GFX9: ; %bb.0:
|
|
|
171 ; GFX9-NEXT: v_add3_u32 v0, v0, v1, v2
|
|
|
172 ; GFX9-NEXT: v_mul_lo_u32 v1, v0, v3
|
|
|
173 ; GFX9-NEXT: ; return to shader part epilog
|
|
|
174 ;
|
|
|
175 ; GFX10-LABEL: add3_multiuse_outer:
|
|
|
176 ; GFX10: ; %bb.0:
|
|
|
177 ; GFX10-NEXT: v_add3_u32 v0, v0, v1, v2
|
|
|
178 ; GFX10-NEXT: v_mul_lo_u32 v1, v0, v3
|
|
|
179 ; GFX10-NEXT: ; return to shader part epilog
|
|
|
180 %inner = add i32 %a, %b
|
|
|
181 %outer = add i32 %inner, %c
|
|
|
182 %x1 = mul i32 %outer, %x
|
|
|
183 %r1 = insertelement <2 x i32> undef, i32 %outer, i32 0
|
|
|
184 %r0 = insertelement <2 x i32> %r1, i32 %x1, i32 1
|
|
|
185 %bc = bitcast <2 x i32> %r0 to <2 x float>
|
|
|
186 ret <2 x float> %bc
|
|
|
187 }
|
|
|
188
|
|
|
189 define amdgpu_ps <2 x float> @add3_multiuse_inner(i32 %a, i32 %b, i32 %c) {
|
|
|
190 ; VI-LABEL: add3_multiuse_inner:
|
|
|
191 ; VI: ; %bb.0:
|
|
|
192 ; VI-NEXT: v_add_u32_e32 v0, vcc, v0, v1
|
|
|
193 ; VI-NEXT: v_add_u32_e32 v1, vcc, v0, v2
|
|
|
194 ; VI-NEXT: ; return to shader part epilog
|
|
|
195 ;
|
|
|
196 ; GFX9-LABEL: add3_multiuse_inner:
|
|
|
197 ; GFX9: ; %bb.0:
|
|
|
198 ; GFX9-NEXT: v_add_u32_e32 v0, v0, v1
|
|
|
199 ; GFX9-NEXT: v_add_u32_e32 v1, v0, v2
|
|
|
200 ; GFX9-NEXT: ; return to shader part epilog
|
|
|
201 ;
|
|
|
202 ; GFX10-LABEL: add3_multiuse_inner:
|
|
|
203 ; GFX10: ; %bb.0:
|
|
|
204 ; GFX10-NEXT: v_add_nc_u32_e32 v0, v0, v1
|
|
|
205 ; GFX10-NEXT: v_add_nc_u32_e32 v1, v0, v2
|
|
|
206 ; GFX10-NEXT: ; return to shader part epilog
|
|
|
207 %inner = add i32 %a, %b
|
|
|
208 %outer = add i32 %inner, %c
|
|
|
209 %r1 = insertelement <2 x i32> undef, i32 %inner, i32 0
|
|
|
210 %r0 = insertelement <2 x i32> %r1, i32 %outer, i32 1
|
|
|
211 %bc = bitcast <2 x i32> %r0 to <2 x float>
|
|
|
212 ret <2 x float> %bc
|
|
|
213 }
|
|
|
214
|
|
|
215 ; A case where uniform values end up in VGPRs -- we could use v_add3_u32 here,
|
|
|
216 ; but we don't.
|
|
|
217 define amdgpu_ps float @add3_uniform_vgpr(float inreg %a, float inreg %b, float inreg %c) {
|
|
|
218 ; VI-LABEL: add3_uniform_vgpr:
|
|
|
219 ; VI: ; %bb.0:
|
|
|
220 ; VI-NEXT: v_mov_b32_e32 v2, 0x40400000
|
|
|
221 ; VI-NEXT: v_add_f32_e64 v0, s2, 1.0
|
|
|
222 ; VI-NEXT: v_add_f32_e64 v1, s3, 2.0
|
|
|
223 ; VI-NEXT: v_add_f32_e32 v2, s4, v2
|
|
|
224 ; VI-NEXT: v_add_u32_e32 v0, vcc, v1, v0
|
|
|
225 ; VI-NEXT: v_add_u32_e32 v0, vcc, v2, v0
|
|
|
226 ; VI-NEXT: ; return to shader part epilog
|
|
|
227 ;
|
|
|
228 ; GFX9-LABEL: add3_uniform_vgpr:
|
|
|
229 ; GFX9: ; %bb.0:
|
|
|
230 ; GFX9-NEXT: v_mov_b32_e32 v2, 0x40400000
|
|
|
231 ; GFX9-NEXT: v_add_f32_e64 v0, s2, 1.0
|
|
|
232 ; GFX9-NEXT: v_add_f32_e64 v1, s3, 2.0
|
|
|
233 ; GFX9-NEXT: v_add_f32_e32 v2, s4, v2
|
|
|
234 ; GFX9-NEXT: v_add_u32_e32 v0, v0, v1
|
|
|
235 ; GFX9-NEXT: v_add_u32_e32 v0, v0, v2
|
|
|
236 ; GFX9-NEXT: ; return to shader part epilog
|
|
|
237 ;
|
|
|
238 ; GFX10-LABEL: add3_uniform_vgpr:
|
|
|
239 ; GFX10: ; %bb.0:
|
|
173
|
240 ; GFX10-NEXT: v_add_f32_e64 v0, s2, 1.0
|
|
150
|
241 ; GFX10-NEXT: v_add_f32_e64 v1, s3, 2.0
|
|
173
|
242 ; GFX10-NEXT: v_add_f32_e64 v2, 0x40400000, s4
|
|
|
243 ; GFX10-NEXT: v_add_nc_u32_e32 v0, v0, v1
|
|
|
244 ; GFX10-NEXT: v_add_nc_u32_e32 v0, v0, v2
|
|
150
|
245 ; GFX10-NEXT: ; return to shader part epilog
|
|
|
246 %a1 = fadd float %a, 1.0
|
|
|
247 %b2 = fadd float %b, 2.0
|
|
|
248 %c3 = fadd float %c, 3.0
|
|
|
249 %bc.a = bitcast float %a1 to i32
|
|
|
250 %bc.b = bitcast float %b2 to i32
|
|
|
251 %bc.c = bitcast float %c3 to i32
|
|
|
252 %x = add i32 %bc.a, %bc.b
|
|
|
253 %result = add i32 %x, %bc.c
|
|
|
254 %bc = bitcast i32 %result to float
|
|
|
255 ret float %bc
|
|
|
256 }
|
