AMDGPU/GlobalISel/floor.f64.ll

2 …N: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=tahiti < %s | FileCheck -check-prefix=GFX6 %s
7 ; GFX6-LABEL: v_floor_f64_ieee:
8 ; GFX6:       ; %bb.0:
9 ; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
10 ; GFX6-NEXT:    s_mov_b32 s4, -1
11 ; GFX6-NEXT:    v_fract_f64_e32 v[2:3], v[0:1]
12 ; GFX6-NEXT:    s_mov_b32 s5, 0x3fefffff
13 ; GFX6-NEXT:    v_min_f64 v[2:3], v[2:3], s[4:5]
14 ; GFX6-NEXT:    v_cmp_o_f64_e32 vcc, v[0:1], v[0:1]
15 ; GFX6-NEXT:    v_cndmask_b32_e32 v2, v2, v0, vcc
16 ; GFX6-NEXT:    v_cndmask_b32_e32 v3, v3, v1, vcc
17 ; GFX6-NEXT:    v_add_f64 v[0:1], v[0:1], -v[2:3]
18 ; GFX6-NEXT:    s_setpc_b64 s[30:31]
30 ; GFX6-LABEL: v_floor_f64_ieee_nnan:
31 ; GFX6:       ; %bb.0:
32 ; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
33 ; GFX6-NEXT:    s_mov_b32 s4, -1
34 ; GFX6-NEXT:    v_fract_f64_e32 v[2:3], v[0:1]
35 ; GFX6-NEXT:    s_mov_b32 s5, 0x3fefffff
36 ; GFX6-NEXT:    v_min_f64 v[2:3], v[2:3], s[4:5]
37 ; GFX6-NEXT:    v_add_f64 v[0:1], v[0:1], -v[2:3]
38 ; GFX6-NEXT:    s_setpc_b64 s[30:31]
50 ; GFX6-LABEL: v_floor_f64_ieee_fneg:
51 ; GFX6:       ; %bb.0:
52 ; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
53 ; GFX6-NEXT:    s_mov_b32 s4, -1
54 ; GFX6-NEXT:    v_fract_f64_e64 v[2:3], -v[0:1]
55 ; GFX6-NEXT:    s_mov_b32 s5, 0x3fefffff
56 ; GFX6-NEXT:    v_min_f64 v[2:3], v[2:3], s[4:5]
57 ; GFX6-NEXT:    v_cmp_o_f64_e32 vcc, v[0:1], v[0:1]
58 ; GFX6-NEXT:    v_cndmask_b32_e32 v2, v2, v0, vcc
59 ; GFX6-NEXT:    v_cndmask_b32_e32 v3, v3, v1, vcc
60 ; GFX6-NEXT:    v_add_f64 v[0:1], -v[0:1], -v[2:3]
61 ; GFX6-NEXT:    s_setpc_b64 s[30:31]
74 ; GFX6-LABEL: v_floor_f64_nonieee:
75 ; GFX6:       ; %bb.0:
76 ; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
77 ; GFX6-NEXT:    s_mov_b32 s4, -1
78 ; GFX6-NEXT:    v_fract_f64_e32 v[2:3], v[0:1]
79 ; GFX6-NEXT:    s_mov_b32 s5, 0x3fefffff
80 ; GFX6-NEXT:    v_min_f64 v[2:3], v[2:3], s[4:5]
81 ; GFX6-NEXT:    v_cmp_o_f64_e32 vcc, v[0:1], v[0:1]
82 ; GFX6-NEXT:    v_cndmask_b32_e32 v2, v2, v0, vcc
83 ; GFX6-NEXT:    v_cndmask_b32_e32 v3, v3, v1, vcc
84 ; GFX6-NEXT:    v_add_f64 v[0:1], v[0:1], -v[2:3]
85 ; GFX6-NEXT:    s_setpc_b64 s[30:31]
97 ; GFX6-LABEL: v_floor_f64_nonieee_nnan:
98 ; GFX6:       ; %bb.0:
99 ; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
100 ; GFX6-NEXT:    s_mov_b32 s4, -1
101 ; GFX6-NEXT:    v_fract_f64_e32 v[2:3], v[0:1]
102 ; GFX6-NEXT:    s_mov_b32 s5, 0x3fefffff
103 ; GFX6-NEXT:    v_min_f64 v[2:3], v[2:3], s[4:5]
104 ; GFX6-NEXT:    v_add_f64 v[0:1], v[0:1], -v[2:3]
105 ; GFX6-NEXT:    s_setpc_b64 s[30:31]
117 ; GFX6-LABEL: v_floor_f64_non_ieee_fneg:
118 ; GFX6:       ; %bb.0:
119 ; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
120 ; GFX6-NEXT:    s_mov_b32 s4, -1
121 ; GFX6-NEXT:    v_fract_f64_e64 v[2:3], -v[0:1]
122 ; GFX6-NEXT:    s_mov_b32 s5, 0x3fefffff
123 ; GFX6-NEXT:    v_min_f64 v[2:3], v[2:3], s[4:5]
124 ; GFX6-NEXT:    v_cmp_o_f64_e32 vcc, v[0:1], v[0:1]
125 ; GFX6-NEXT:    v_cndmask_b32_e32 v2, v2, v0, vcc
126 ; GFX6-NEXT:    v_cndmask_b32_e32 v3, v3, v1, vcc
127 ; GFX6-NEXT:    v_add_f64 v[0:1], -v[0:1], -v[2:3]
128 ; GFX6-NEXT:    s_setpc_b64 s[30:31]
141 ; GFX6-LABEL: v_floor_f64_fabs:
142 ; GFX6:       ; %bb.0:
143 ; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
144 ; GFX6-NEXT:    s_mov_b32 s4, -1
145 ; GFX6-NEXT:    v_fract_f64_e64 v[2:3], |v[0:1]|
146 ; GFX6-NEXT:    s_mov_b32 s5, 0x3fefffff
147 ; GFX6-NEXT:    v_min_f64 v[2:3], v[2:3], s[4:5]
148 ; GFX6-NEXT:    v_cmp_o_f64_e32 vcc, v[0:1], v[0:1]
149 ; GFX6-NEXT:    v_cndmask_b32_e32 v2, v2, v0, vcc
150 ; GFX6-NEXT:    v_cndmask_b32_e32 v3, v3, v1, vcc
151 ; GFX6-NEXT:    v_add_f64 v[0:1], |v[0:1]|, -v[2:3]
152 ; GFX6-NEXT:    s_setpc_b64 s[30:31]
170 ; GFX6-LABEL: v_floor_f64_fneg_fabs:
171 ; GFX6:       ; %bb.0:
172 ; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
173 ; GFX6-NEXT:    s_mov_b32 s4, -1
174 ; GFX6-NEXT:    v_fract_f64_e64 v[2:3], -|v[0:1]|
175 ; GFX6-NEXT:    s_mov_b32 s5, 0x3fefffff
176 ; GFX6-NEXT:    v_min_f64 v[2:3], v[2:3], s[4:5]
177 ; GFX6-NEXT:    v_cmp_o_f64_e32 vcc, v[0:1], v[0:1]
178 ; GFX6-NEXT:    v_cndmask_b32_e32 v2, v2, v0, vcc
179 ; GFX6-NEXT:    v_cndmask_b32_e32 v3, v3, v1, vcc
180 ; GFX6-NEXT:    v_add_f64 v[0:1], -|v[0:1]|, -v[2:3]
181 ; GFX6-NEXT:    s_setpc_b64 s[30:31]
195 ; GFX6-LABEL: s_floor_f64:
196 ; GFX6:       ; %bb.0:
197 ; GFX6-NEXT:    s_mov_b32 s0, -1
198 ; GFX6-NEXT:    v_fract_f64_e32 v[0:1], s[2:3]
199 ; GFX6-NEXT:    s_mov_b32 s1, 0x3fefffff
200 ; GFX6-NEXT:    v_min_f64 v[0:1], v[0:1], s[0:1]
201 ; GFX6-NEXT:    v_cmp_o_f64_e64 vcc, s[2:3], s[2:3]
202 ; GFX6-NEXT:    v_mov_b32_e32 v2, s2
203 ; GFX6-NEXT:    v_mov_b32_e32 v3, s3
204 ; GFX6-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
205 ; GFX6-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
206 ; GFX6-NEXT:    v_add_f64 v[0:1], s[2:3], -v[0:1]
207 ; GFX6-NEXT:    ; return to shader part epilog
219 ; GFX6-LABEL: s_floor_f64_fneg:
220 ; GFX6:       ; %bb.0:
221 ; GFX6-NEXT:    s_mov_b32 s0, -1
222 ; GFX6-NEXT:    v_fract_f64_e64 v[0:1], -s[2:3]
223 ; GFX6-NEXT:    s_mov_b32 s1, 0x3fefffff
224 ; GFX6-NEXT:    v_min_f64 v[0:1], v[0:1], s[0:1]
225 ; GFX6-NEXT:    v_cmp_o_f64_e64 vcc, s[2:3], s[2:3]
226 ; GFX6-NEXT:    v_mov_b32_e32 v2, s2
227 ; GFX6-NEXT:    v_mov_b32_e32 v3, s3
228 ; GFX6-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
229 ; GFX6-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
230 ; GFX6-NEXT:    v_add_f64 v[0:1], -s[2:3], -v[0:1]
231 ; GFX6-NEXT:    ; return to shader part epilog
244 ; GFX6-LABEL: s_floor_f64_fabs:
245 ; GFX6:       ; %bb.0:
246 ; GFX6-NEXT:    s_mov_b32 s0, -1
247 ; GFX6-NEXT:    v_fract_f64_e64 v[0:1], |s[2:3]|
248 ; GFX6-NEXT:    s_mov_b32 s1, 0x3fefffff
249 ; GFX6-NEXT:    v_min_f64 v[0:1], v[0:1], s[0:1]
250 ; GFX6-NEXT:    v_cmp_o_f64_e64 vcc, s[2:3], s[2:3]
251 ; GFX6-NEXT:    v_mov_b32_e32 v2, s2
252 ; GFX6-NEXT:    v_mov_b32_e32 v3, s3
253 ; GFX6-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
254 ; GFX6-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
255 ; GFX6-NEXT:    v_add_f64 v[0:1], |s[2:3]|, -v[0:1]
256 ; GFX6-NEXT:    ; return to shader part epilog
269 ; GFX6-LABEL: s_floor_f64_fneg_fabs:
270 ; GFX6:       ; %bb.0:
271 ; GFX6-NEXT:    s_mov_b32 s0, -1
272 ; GFX6-NEXT:    v_fract_f64_e64 v[0:1], -|s[2:3]|
273 ; GFX6-NEXT:    s_mov_b32 s1, 0x3fefffff
274 ; GFX6-NEXT:    v_min_f64 v[0:1], v[0:1], s[0:1]
275 ; GFX6-NEXT:    v_cmp_o_f64_e64 vcc, s[2:3], s[2:3]
276 ; GFX6-NEXT:    v_mov_b32_e32 v2, s2
277 ; GFX6-NEXT:    v_mov_b32_e32 v3, s3
278 ; GFX6-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
279 ; GFX6-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
280 ; GFX6-NEXT:    v_add_f64 v[0:1], -|s[2:3]|, -v[0:1]
281 ; GFX6-NEXT:    ; return to shader part epilog