• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1; RUN: llc < %s -march=amdgcn -mcpu=verde -verify-machineinstrs | FileCheck --check-prefix=SI --check-prefix=FUNC %s
2; RUN: llc < %s -march=amdgcn -mcpu=tonga -verify-machineinstrs | FileCheck --check-prefix=SI --check-prefix=FUNC %s
3
4; FUNC-LABEL: {{^}}break_inserted_outside_of_loop:
5
6; SI: [[LOOP_LABEL:[A-Z0-9]+]]:
7; Lowered break instructin:
8; SI: s_or_b64
9; Lowered Loop instruction:
10; SI: s_andn2_b64
11; s_cbranch_execnz [[LOOP_LABEL]]
12; SI: s_endpgm
13define void @break_inserted_outside_of_loop(i32 addrspace(1)* %out, i32 %a) {
14main_body:
15  %tid = call i32 @llvm.amdgcn.mbcnt.lo(i32 -1, i32 0) #0
16  %0 = and i32 %a, %tid
17  %1 = trunc i32 %0 to i1
18  br label %ENDIF
19
20ENDLOOP:
21  store i32 0, i32 addrspace(1)* %out
22  ret void
23
24ENDIF:
25  br i1 %1, label %ENDLOOP, label %ENDIF
26}
27
28
29; FUNC-LABEL: {{^}}phi_cond_outside_loop:
30; FIXME: This could be folded into the s_or_b64 instruction
31; SI: s_mov_b64 [[ZERO:s\[[0-9]+:[0-9]+\]]], 0
32; SI: [[LOOP_LABEL:[A-Z0-9]+]]
33; SI: v_cmp_ne_i32_e32 vcc, 0, v{{[0-9]+}}
34
35; SI_IF_BREAK instruction:
36; SI: s_or_b64 [[BREAK:s\[[0-9]+:[0-9]+\]]], vcc, [[ZERO]]
37
38; SI_LOOP instruction:
39; SI: s_andn2_b64 exec, exec, [[BREAK]]
40; SI: s_cbranch_execnz [[LOOP_LABEL]]
41; SI: s_endpgm
42
43define void @phi_cond_outside_loop(i32 %b) {
44entry:
45  %tid = call i32 @llvm.amdgcn.mbcnt.lo(i32 -1, i32 0) #0
46  %0 = icmp eq i32 %tid , 0
47  br i1 %0, label %if, label %else
48
49if:
50  br label %endif
51
52else:
53  %1 = icmp eq i32 %b, 0
54  br label %endif
55
56endif:
57  %2 = phi i1 [0, %if], [%1, %else]
58  br label %loop
59
60loop:
61  br i1 %2, label %exit, label %loop
62
63exit:
64  ret void
65}
66
67; FIXME: should emit s_endpgm
68; CHECK-LABEL: {{^}}switch_unreachable:
69; CHECK-NOT: s_endpgm
70; CHECK: .Lfunc_end2
71define void @switch_unreachable(i32 addrspace(1)* %g, i8 addrspace(3)* %l, i32 %x) nounwind {
72centry:
73  switch i32 %x, label %sw.default [
74    i32 0, label %sw.bb
75    i32 60, label %sw.bb
76  ]
77
78sw.bb:
79  unreachable
80
81sw.default:
82  unreachable
83
84sw.epilog:
85  ret void
86}
87
88declare float @llvm.fabs.f32(float) nounwind readnone
89
90; This broke the old AMDIL cfg structurizer
91; FUNC-LABEL: {{^}}loop_land_info_assert:
92; SI: s_cmp_gt_i32
93; SI-NEXT: s_cbranch_scc0 [[ENDPGM:BB[0-9]+_[0-9]+]]
94
95; SI: s_cmp_gt_i32
96; SI-NEXT: s_cbranch_scc1 [[ENDPGM]]
97
98; SI: [[INFLOOP:BB[0-9]+_[0-9]+]]
99; SI: s_branch [[INFLOOP]]
100
101; SI: [[ENDPGM]]:
102; SI: s_endpgm
103define void @loop_land_info_assert(i32 %c0, i32 %c1, i32 %c2, i32 %c3, i32 %x, i32 %y, i1 %arg) nounwind {
104entry:
105  %cmp = icmp sgt i32 %c0, 0
106  br label %while.cond.outer
107
108while.cond.outer:
109  %tmp = load float, float addrspace(1)* undef
110  br label %while.cond
111
112while.cond:
113  %cmp1 = icmp slt i32 %c1, 4
114  br i1 %cmp1, label %convex.exit, label %for.cond
115
116convex.exit:
117  %or = or i1 %cmp, %cmp1
118  br i1 %or, label %return, label %if.end
119
120if.end:
121  %tmp3 = call float @llvm.fabs.f32(float %tmp) nounwind readnone
122  %cmp2 = fcmp olt float %tmp3, 0x3E80000000000000
123  br i1 %cmp2, label %if.else, label %while.cond.outer
124
125if.else:
126  store volatile i32 3, i32 addrspace(1)* undef, align 4
127  br label %while.cond
128
129for.cond:
130  %cmp3 = icmp slt i32 %c3, 1000
131  br i1 %cmp3, label %for.body, label %return
132
133for.body:
134  br i1 %cmp3, label %self.loop, label %if.end.2
135
136if.end.2:
137  %or.cond2 = or i1 %cmp3, %arg
138  br i1 %or.cond2, label %return, label %for.cond
139
140self.loop:
141 br label %self.loop
142
143return:
144  ret void
145}
146
147
148declare i32 @llvm.amdgcn.mbcnt.lo(i32, i32) #0
149
150attributes #0 = { nounwind readnone }
151