1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc -mtriple=amdgcn -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck %s -check-prefix=GCN
4 define amdgpu_ps float @while_break(i32 %z, float %v, i32 %x, i32 %y) #0 {
5 ; GCN-LABEL: while_break:
6 ; GCN: ; %bb.0: ; %entry
7 ; GCN-NEXT: s_mov_b32 s1, -1
8 ; GCN-NEXT: s_mov_b32 s0, 0
9 ; GCN-NEXT: s_branch .LBB0_2
10 ; GCN-NEXT: .LBB0_1: ; %Flow2
11 ; GCN-NEXT: ; in Loop: Header=BB0_2 Depth=1
12 ; GCN-NEXT: s_or_b32 exec_lo, exec_lo, s4
13 ; GCN-NEXT: s_and_b32 s2, exec_lo, s3
14 ; GCN-NEXT: s_or_b32 s0, s2, s0
15 ; GCN-NEXT: s_andn2_b32 exec_lo, exec_lo, s0
16 ; GCN-NEXT: s_cbranch_execz .LBB0_8
17 ; GCN-NEXT: .LBB0_2: ; %header
18 ; GCN-NEXT: ; =>This Inner Loop Header: Depth=1
19 ; GCN-NEXT: s_add_i32 s1, s1, 1
20 ; GCN-NEXT: s_mov_b32 s2, 0
21 ; GCN-NEXT: v_cmp_ge_i32_e32 vcc_lo, s1, v2
22 ; GCN-NEXT: s_and_saveexec_b32 s3, vcc_lo
23 ; GCN-NEXT: s_xor_b32 s3, exec_lo, s3
24 ; GCN-NEXT: ; %bb.3: ; %else
25 ; GCN-NEXT: ; in Loop: Header=BB0_2 Depth=1
26 ; GCN-NEXT: v_cmp_lt_i32_e32 vcc_lo, s1, v3
27 ; GCN-NEXT: s_and_b32 s2, vcc_lo, exec_lo
28 ; GCN-NEXT: ; %bb.4: ; %Flow
29 ; GCN-NEXT: ; in Loop: Header=BB0_2 Depth=1
30 ; GCN-NEXT: s_andn2_saveexec_b32 s3, s3
31 ; GCN-NEXT: ; %bb.5: ; %if
32 ; GCN-NEXT: ; in Loop: Header=BB0_2 Depth=1
33 ; GCN-NEXT: v_add_f32_e32 v1, 1.0, v1
34 ; GCN-NEXT: s_or_b32 s2, s2, exec_lo
35 ; GCN-NEXT: ; %bb.6: ; %Flow1
36 ; GCN-NEXT: ; in Loop: Header=BB0_2 Depth=1
37 ; GCN-NEXT: s_or_b32 exec_lo, exec_lo, s3
38 ; GCN-NEXT: s_mov_b32 s3, -1
39 ; GCN-NEXT: s_and_saveexec_b32 s4, s2
40 ; GCN-NEXT: s_cbranch_execz .LBB0_1
41 ; GCN-NEXT: ; %bb.7: ; %latch
42 ; GCN-NEXT: ; in Loop: Header=BB0_2 Depth=1
43 ; GCN-NEXT: v_cmp_lt_i32_e32 vcc_lo, s1, v0
44 ; GCN-NEXT: s_orn2_b32 s3, vcc_lo, exec_lo
45 ; GCN-NEXT: s_branch .LBB0_1
46 ; GCN-NEXT: .LBB0_8: ; %end
47 ; GCN-NEXT: s_or_b32 exec_lo, exec_lo, s0
48 ; GCN-NEXT: v_mov_b32_e32 v0, v1
49 ; GCN-NEXT: ; return to shader part epilog
54 %v.1 = phi float [ %v, %entry ], [ %v.2, %latch ]
55 %ind = phi i32 [ 0, %entry], [ %ind.inc, %latch ]
56 %cc = icmp slt i32 %ind, %x
57 br i1 %cc, label %if, label %else
60 %v.if = fadd float %v.1, 1.0
64 %cc2 = icmp slt i32 %ind, %y
65 br i1 %cc2, label %latch, label %end
68 %v.2 = phi float [ %v.if, %if ], [ %v.1, %else ]
69 %ind.inc = add i32 %ind, 1
70 %cc3 = icmp slt i32 %ind, %z
71 br i1 %cc3, label %end, label %header
74 %r = phi float [ %v.2, %latch ], [ %v.1, %else ]
78 ; Just different dfs order from while_break.
79 define amdgpu_ps float @while_break2(i32 %z, float %v, i32 %x, i32 %y) #0 {
80 ; GCN-LABEL: while_break2:
81 ; GCN: ; %bb.0: ; %entry
82 ; GCN-NEXT: s_mov_b32 s1, -1
83 ; GCN-NEXT: s_mov_b32 s0, 0
84 ; GCN-NEXT: s_branch .LBB1_2
85 ; GCN-NEXT: .LBB1_1: ; %Flow2
86 ; GCN-NEXT: ; in Loop: Header=BB1_2 Depth=1
87 ; GCN-NEXT: s_or_b32 exec_lo, exec_lo, s4
88 ; GCN-NEXT: s_and_b32 s2, exec_lo, s3
89 ; GCN-NEXT: s_or_b32 s0, s2, s0
90 ; GCN-NEXT: s_andn2_b32 exec_lo, exec_lo, s0
91 ; GCN-NEXT: s_cbranch_execz .LBB1_8
92 ; GCN-NEXT: .LBB1_2: ; %header
93 ; GCN-NEXT: ; =>This Inner Loop Header: Depth=1
94 ; GCN-NEXT: s_add_i32 s1, s1, 1
95 ; GCN-NEXT: s_mov_b32 s2, 0
96 ; GCN-NEXT: v_cmp_ge_i32_e32 vcc_lo, s1, v2
97 ; GCN-NEXT: s_and_saveexec_b32 s3, vcc_lo
98 ; GCN-NEXT: s_xor_b32 s3, exec_lo, s3
99 ; GCN-NEXT: ; %bb.3: ; %if
100 ; GCN-NEXT: ; in Loop: Header=BB1_2 Depth=1
101 ; GCN-NEXT: v_add_f32_e32 v1, 1.0, v1
102 ; GCN-NEXT: s_mov_b32 s2, exec_lo
103 ; GCN-NEXT: ; %bb.4: ; %Flow
104 ; GCN-NEXT: ; in Loop: Header=BB1_2 Depth=1
105 ; GCN-NEXT: s_andn2_saveexec_b32 s3, s3
106 ; GCN-NEXT: ; %bb.5: ; %else
107 ; GCN-NEXT: ; in Loop: Header=BB1_2 Depth=1
108 ; GCN-NEXT: v_cmp_lt_i32_e32 vcc_lo, s1, v3
109 ; GCN-NEXT: s_andn2_b32 s2, s2, exec_lo
110 ; GCN-NEXT: s_and_b32 s4, vcc_lo, exec_lo
111 ; GCN-NEXT: s_or_b32 s2, s2, s4
112 ; GCN-NEXT: ; %bb.6: ; %Flow1
113 ; GCN-NEXT: ; in Loop: Header=BB1_2 Depth=1
114 ; GCN-NEXT: s_or_b32 exec_lo, exec_lo, s3
115 ; GCN-NEXT: s_mov_b32 s3, -1
116 ; GCN-NEXT: s_and_saveexec_b32 s4, s2
117 ; GCN-NEXT: s_cbranch_execz .LBB1_1
118 ; GCN-NEXT: ; %bb.7: ; %latch
119 ; GCN-NEXT: ; in Loop: Header=BB1_2 Depth=1
120 ; GCN-NEXT: v_cmp_lt_i32_e32 vcc_lo, s1, v0
121 ; GCN-NEXT: s_orn2_b32 s3, vcc_lo, exec_lo
122 ; GCN-NEXT: s_branch .LBB1_1
123 ; GCN-NEXT: .LBB1_8: ; %end
124 ; GCN-NEXT: s_or_b32 exec_lo, exec_lo, s0
125 ; GCN-NEXT: v_mov_b32_e32 v0, v1
126 ; GCN-NEXT: ; return to shader part epilog
131 %v.1 = phi float [ %v, %entry ], [ %v.2, %latch ]
132 %ind = phi i32 [ 0, %entry], [ %ind.inc, %latch ]
133 %cc = icmp slt i32 %ind, %x
134 br i1 %cc, label %else, label %if
137 %v.if = fadd float %v.1, 1.0
141 %cc2 = icmp slt i32 %ind, %y
142 br i1 %cc2, label %latch, label %end
145 %v.2 = phi float [ %v.if, %if ], [ %v.1, %else ]
146 %ind.inc = add i32 %ind, 1
147 %cc3 = icmp slt i32 %ind, %z
148 br i1 %cc3, label %end, label %header
151 %r = phi float [ %v.2, %latch ], [ %v.1, %else ]
155 ; Two chains of phi network that have the same value from %if block.
156 define amdgpu_ps < 2 x float> @while_break_two_chains_of_phi(float %v, i32 %x, i32 %y, i32 %z, ptr addrspace(1) %p) #0 {
157 ; GCN-LABEL: while_break_two_chains_of_phi:
158 ; GCN: ; %bb.0: ; %entry
159 ; GCN-NEXT: v_mov_b32_e32 v6, 0
160 ; GCN-NEXT: s_mov_b32 s2, 0
161 ; GCN-NEXT: s_mov_b32 s0, 0
162 ; GCN-NEXT: s_branch .LBB2_2
163 ; GCN-NEXT: .LBB2_1: ; %Flow1
164 ; GCN-NEXT: ; in Loop: Header=BB2_2 Depth=1
165 ; GCN-NEXT: s_or_b32 exec_lo, exec_lo, s4
166 ; GCN-NEXT: s_and_b32 s1, exec_lo, s1
167 ; GCN-NEXT: s_or_b32 s2, s1, s2
168 ; GCN-NEXT: s_andn2_b32 exec_lo, exec_lo, s2
169 ; GCN-NEXT: s_cbranch_execz .LBB2_6
170 ; GCN-NEXT: .LBB2_2: ; %header
171 ; GCN-NEXT: ; =>This Inner Loop Header: Depth=1
172 ; GCN-NEXT: v_cmp_ge_i32_e64 s3, s0, v1
173 ; GCN-NEXT: v_cmp_lt_i32_e32 vcc_lo, s0, v1
174 ; GCN-NEXT: s_and_saveexec_b32 s4, vcc_lo
175 ; GCN-NEXT: s_cbranch_execz .LBB2_4
176 ; GCN-NEXT: ; %bb.3: ; %if
177 ; GCN-NEXT: ; in Loop: Header=BB2_2 Depth=1
178 ; GCN-NEXT: s_ashr_i32 s1, s0, 31
179 ; GCN-NEXT: s_lshl_b64 s[6:7], s[0:1], 2
180 ; GCN-NEXT: s_andn2_b32 s1, s3, exec_lo
181 ; GCN-NEXT: v_add_co_u32 v6, vcc_lo, v4, s6
182 ; GCN-NEXT: v_add_co_ci_u32_e32 v7, vcc_lo, s7, v5, vcc_lo
183 ; GCN-NEXT: v_cmp_lt_i32_e32 vcc_lo, s0, v2
184 ; GCN-NEXT: global_load_dword v0, v[6:7], off
185 ; GCN-NEXT: s_and_b32 s3, vcc_lo, exec_lo
186 ; GCN-NEXT: s_or_b32 s3, s1, s3
187 ; GCN-NEXT: s_waitcnt vmcnt(0)
188 ; GCN-NEXT: v_add_f32_e32 v6, 1.0, v0
189 ; GCN-NEXT: v_mov_b32_e32 v0, v6
190 ; GCN-NEXT: .LBB2_4: ; %Flow
191 ; GCN-NEXT: ; in Loop: Header=BB2_2 Depth=1
192 ; GCN-NEXT: s_or_b32 exec_lo, exec_lo, s4
193 ; GCN-NEXT: s_mov_b32 s1, -1
194 ; GCN-NEXT: s_and_saveexec_b32 s4, s3
195 ; GCN-NEXT: s_cbranch_execz .LBB2_1
196 ; GCN-NEXT: ; %bb.5: ; %latch
197 ; GCN-NEXT: ; in Loop: Header=BB2_2 Depth=1
198 ; GCN-NEXT: v_cmp_lt_i32_e32 vcc_lo, s0, v3
199 ; GCN-NEXT: s_add_i32 s0, s0, 1
200 ; GCN-NEXT: s_orn2_b32 s1, vcc_lo, exec_lo
201 ; GCN-NEXT: s_branch .LBB2_1
202 ; GCN-NEXT: .LBB2_6: ; %end
203 ; GCN-NEXT: s_or_b32 exec_lo, exec_lo, s2
204 ; GCN-NEXT: v_mov_b32_e32 v1, v6
205 ; GCN-NEXT: ; return to shader part epilog
210 %v.1 = phi float [ %v, %entry ], [ %v.2, %latch ]
211 %v.copy = phi float [ 0.0, %entry ], [ %v.copy.2, %latch ]
212 %ind = phi i32 [ 0, %entry], [ %ind.inc, %latch ]
213 %cc = icmp slt i32 %ind, %x
214 br i1 %cc, label %if, label %latch
217 %v.ptr = getelementptr float, ptr addrspace(1) %p, i32 %ind
218 %v.load = load float, ptr addrspace(1) %v.ptr
219 %v.if = fadd float %v.load, 1.0
220 %cc2 = icmp slt i32 %ind, %y
221 br i1 %cc2, label %latch, label %end
224 %v.2 = phi float [ %v.1, %header ], [ %v.if, %if ]
225 %v.copy.2 = phi float [ %v.copy, %header ], [ %v.if, %if ]
226 %ind.inc = add i32 %ind, 1
227 %cc3 = icmp slt i32 %ind, %z
228 br i1 %cc3, label %end, label %header
231 %r = phi float [ %v.2, %latch ], [ %v.if, %if ]
232 %r2 = phi float [ %v.copy.2, %latch ], [ %v.if, %if ]
233 %packed0 = insertelement < 2 x float > poison, float %r, i32 0
234 %packed1 = insertelement < 2 x float > %packed0, float %r2, i32 1
235 ret < 2 x float> %packed1
238 attributes #0 = { nounwind }