1 ; RUN: llc -march=amdgcn -mcpu=tahiti -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=SI %s
4 ; Most SALU instructions ignore control flow, so we need to make sure
5 ; they don't overwrite values from other blocks.
7 ; If the branch decision is made based on a value in an SGPR then all
8 ; threads will execute the same code paths, so we don't need to worry
9 ; about instructions in different blocks overwriting each other.
10 ; SI-LABEL: {{^}}sgpr_if_else_salu_br:
16 define amdgpu_kernel void @sgpr_if_else_salu_br(i32 addrspace(1)* %out, i32 %a, i32 %b, i32 %c, i32 %d, i32 %e) {
18 %0 = icmp eq i32 %a, 0
19 br i1 %0, label %if, label %else
30 %3 = phi i32 [%1, %if], [%2, %else]
32 store i32 %4, i32 addrspace(1)* %out
36 ; SI-LABEL: {{^}}sgpr_if_else_salu_br_opt:
38 ; SI: s_cbranch_scc0 [[IF:BB[0-9]+_[0-9]+]]
40 ; SI: ; %bb.1: ; %else
41 ; SI: s_load_dword [[LOAD0:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0x2e
42 ; SI: s_load_dword [[LOAD1:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0x37
44 ; SI: s_branch [[ENDIF:BB[0-9]+_[0-9]+]]
47 ; SI: s_load_dword [[LOAD0]], s{{\[[0-9]+:[0-9]+\]}}, 0x1c
48 ; SI: s_load_dword [[LOAD1]], s{{\[[0-9]+:[0-9]+\]}}, 0x25
51 ; SI: [[ENDIF]]: ; %endif
52 ; SI: s_add_i32 s{{[0-9]+}}, [[LOAD0]], [[LOAD1]]
53 ; SI: buffer_store_dword
55 define amdgpu_kernel void @sgpr_if_else_salu_br_opt(i32 addrspace(1)* %out, [8 x i32], i32 %a, [8 x i32], i32 %b, [8 x i32], i32 %c, [8 x i32], i32 %d, [8 x i32], i32 %e) {
57 %cmp0 = icmp eq i32 %a, 0
58 br i1 %cmp0, label %if, label %else
61 %add0 = add i32 %b, %c
65 %add1 = add i32 %d, %e
69 %phi = phi i32 [%add0, %if], [%add1, %else]
70 %add2 = add i32 %phi, %a
71 store i32 %add2, i32 addrspace(1)* %out
75 ; The two S_ADD instructions should write to different registers, since
76 ; different threads will take different control flow paths.
78 ; SI-LABEL: {{^}}sgpr_if_else_valu_br:
79 ; SI: s_add_i32 [[SGPR:s[0-9]+]]
80 ; SI-NOT: s_add_i32 [[SGPR]]
82 define amdgpu_kernel void @sgpr_if_else_valu_br(i32 addrspace(1)* %out, float %a, i32 %b, i32 %c, i32 %d, i32 %e) {
84 %tid = call i32 @llvm.amdgcn.workitem.id.x() #0
85 %tid_f = uitofp i32 %tid to float
86 %tmp1 = fcmp ueq float %tid_f, 0.0
87 br i1 %tmp1, label %if, label %else
90 %tmp2 = add i32 %b, %c
94 %tmp3 = add i32 %d, %e
98 %tmp4 = phi i32 [%tmp2, %if], [%tmp3, %else]
99 store i32 %tmp4, i32 addrspace(1)* %out
103 ; SI-LABEL: {{^}}sgpr_if_else_valu_cmp_phi_br:
106 ; SI: buffer_load_dword [[AVAL:v[0-9]+]]
107 ; SI: v_cmp_gt_i32_e32 vcc, 0, [[AVAL]]
108 ; SI: s_and_b64 [[PHI:s\[[0-9]+:[0-9]+\]]], vcc, exec
111 ; SI: buffer_load_dword [[AVAL:v[0-9]+]]
112 ; SI-DAG: v_cmp_eq_u32_e32 [[CMP_ELSE:vcc]], 0, [[AVAL]]
113 ; SI-DAG: s_andn2_b64 [[PHI]], [[PHI]], exec
114 ; SI-DAG: s_and_b64 [[TMP:s\[[0-9]+:[0-9]+\]]], [[CMP_ELSE]], exec
115 ; SI: s_or_b64 [[PHI]], [[PHI]], [[TMP]]
118 ; SI: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], 0, -1, [[PHI]]
119 ; SI: buffer_store_dword [[RESULT]],
120 define amdgpu_kernel void @sgpr_if_else_valu_cmp_phi_br(i32 addrspace(1)* %out, i32 addrspace(1)* %a, i32 addrspace(1)* %b) {
122 %tid = call i32 @llvm.amdgcn.workitem.id.x() #0
123 %tmp1 = icmp eq i32 %tid, 0
124 br i1 %tmp1, label %if, label %else
127 %gep.if = getelementptr i32, i32 addrspace(1)* %a, i32 %tid
128 %a.val = load i32, i32 addrspace(1)* %gep.if
129 %cmp.if = icmp eq i32 %a.val, 0
133 %gep.else = getelementptr i32, i32 addrspace(1)* %b, i32 %tid
134 %b.val = load i32, i32 addrspace(1)* %gep.else
135 %cmp.else = icmp slt i32 %b.val, 0
139 %tmp4 = phi i1 [%cmp.if, %if], [%cmp.else, %else]
140 %ext = sext i1 %tmp4 to i32
141 store i32 %ext, i32 addrspace(1)* %out
145 declare i32 @llvm.amdgcn.workitem.id.x() #0
147 attributes #0 = { readnone }