1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --function-signature --check-globals
2; RUN: opt -S -passes=openmp-opt < %s | FileCheck %s
3; RUN: opt -S -passes=openmp-opt -openmp-opt-disable-spmdization < %s | FileCheck %s --check-prefix=CHECK-DISABLED
4;
5;    void pure(void) __attribute__((pure, assume("ompx_spmd_amenable")));
6;    int no_openmp(int *) __attribute__((assume("omp_no_openmp","ompx_spmd_amenable")));
7;
8;    void sequential_loop(int *x, int N) {
9;    #pragma omp target teams
10;      {
11;        x[0] = 0;
12;        x[1] = 1;
13;        x[N] = N;
14;        for (int i = 2; i < N - 1; ++i)
15;          x[i] = i - 1;
16;        #pragma omp parallel
17;        {}
18;        int r1 = no_openmp(x);
19;        x[r1] = r1;
20;        int r2 = no_openmp(x);
21;        x[r2] = r2;
22;        int r3 = no_openmp(x);
23;        x[r3] = r3;
24;
25;        no_openmp(x);
26;        pure();
27;        no_openmp(x);
28;        pure();
29;        no_openmp(x);
30;        pure();
31;      }
32;    }
33;
34target datalayout = "e-i64:64-i128:128-v16:16-v32:32-n16:32:64"
35target triple = "nvptx64"
36
37%struct.ident_t = type { i32, i32, i32, i32, i8* }
38
39@0 = private unnamed_addr constant [23 x i8] c";unknown;unknown;0;0;;\00", align 1
40@1 = private unnamed_addr constant %struct.ident_t { i32 0, i32 2, i32 0, i32 0, i8* getelementptr inbounds ([23 x i8], [23 x i8]* @0, i32 0, i32 0) }, align 8
41@__omp_offloading_2a_fbfa7a_sequential_loop_l6_exec_mode = weak constant i8 1
42@llvm.compiler.used = appending global [1 x i8*] [i8* @__omp_offloading_2a_fbfa7a_sequential_loop_l6_exec_mode], section "llvm.metadata"
43
44; Function Attrs: convergent norecurse nounwind
45;.
46; CHECK: @[[GLOB0:[0-9]+]] = private unnamed_addr constant [23 x i8] c"
47; CHECK: @[[GLOB1:[0-9]+]] = private unnamed_addr constant [[STRUCT_IDENT_T:%.*]] { i32 0, i32 2, i32 0, i32 0, i8* getelementptr inbounds ([23 x i8], [23 x i8]* @[[GLOB0]], i32 0, i32 0) }, align 8
48; CHECK: @[[__OMP_OFFLOADING_2A_FBFA7A_SEQUENTIAL_LOOP_L6_EXEC_MODE:[a-zA-Z0-9_$"\\.-]+]] = weak constant i8 3
49; CHECK: @[[LLVM_COMPILER_USED:[a-zA-Z0-9_$"\\.-]+]] = appending global [1 x i8*] [i8* @__omp_offloading_2a_fbfa7a_sequential_loop_l6_exec_mode], section "llvm.metadata"
50;.
51; CHECK-DISABLED: @[[GLOB0:[0-9]+]] = private unnamed_addr constant [23 x i8] c"
52; CHECK-DISABLED: @[[GLOB1:[0-9]+]] = private unnamed_addr constant [[STRUCT_IDENT_T:%.*]] { i32 0, i32 2, i32 0, i32 0, i8* getelementptr inbounds ([23 x i8], [23 x i8]* @[[GLOB0]], i32 0, i32 0) }, align 8
53; CHECK-DISABLED: @[[__OMP_OFFLOADING_2A_FBFA7A_SEQUENTIAL_LOOP_L6_EXEC_MODE:[a-zA-Z0-9_$"\\.-]+]] = weak constant i8 1
54; CHECK-DISABLED: @[[LLVM_COMPILER_USED:[a-zA-Z0-9_$"\\.-]+]] = appending global [1 x i8*] [i8* @__omp_offloading_2a_fbfa7a_sequential_loop_l6_exec_mode], section "llvm.metadata"
55; CHECK-DISABLED: @[[__OMP_OUTLINED__1_WRAPPER_ID:[a-zA-Z0-9_$"\\.-]+]] = private constant i8 undef
56;.
57define weak void @__omp_offloading_2a_fbfa7a_sequential_loop_l6(i32* %x, i64 %N) #0 {
58; CHECK-LABEL: define {{[^@]+}}@__omp_offloading_2a_fbfa7a_sequential_loop_l6
59; CHECK-SAME: (i32* [[X:%.*]], i64 [[N:%.*]]) #[[ATTR0:[0-9]+]] {
60; CHECK-NEXT:  entry:
61; CHECK-NEXT:    [[N_ADDR_SROA_0_0_EXTRACT_TRUNC:%.*]] = trunc i64 [[N]] to i32
62; CHECK-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_target_init(%struct.ident_t* nonnull @[[GLOB1]], i8 2, i1 false, i1 false) #[[ATTR4:[0-9]+]]
63; CHECK-NEXT:    [[EXEC_USER_CODE:%.*]] = icmp eq i32 [[TMP0]], -1
64; CHECK-NEXT:    br i1 [[EXEC_USER_CODE]], label [[USER_CODE_ENTRY:%.*]], label [[WORKER_EXIT:%.*]]
65; CHECK:       user_code.entry:
66; CHECK-NEXT:    [[TMP1:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* nonnull @[[GLOB1]]) #[[ATTR4]]
67; CHECK-NEXT:    [[ARRAYIDX1_I:%.*]] = getelementptr inbounds i32, i32* [[X]], i64 1
68; CHECK-NEXT:    [[SEXT:%.*]] = shl i64 [[N]], 32
69; CHECK-NEXT:    [[IDXPROM_I:%.*]] = ashr exact i64 [[SEXT]], 32
70; CHECK-NEXT:    [[ARRAYIDX2_I:%.*]] = getelementptr inbounds i32, i32* [[X]], i64 [[IDXPROM_I]]
71; CHECK-NEXT:    br label [[REGION_CHECK_TID:%.*]]
72; CHECK:       region.check.tid:
73; CHECK-NEXT:    [[TMP2:%.*]] = call i32 @__kmpc_get_hardware_thread_id_in_block()
74; CHECK-NEXT:    [[TMP3:%.*]] = icmp eq i32 [[TMP2]], 0
75; CHECK-NEXT:    br i1 [[TMP3]], label [[REGION_GUARDED:%.*]], label [[REGION_BARRIER:%.*]]
76; CHECK:       region.guarded:
77; CHECK-NEXT:    store i32 0, i32* [[X]], align 4, !noalias !8
78; CHECK-NEXT:    store i32 1, i32* [[ARRAYIDX1_I]], align 4, !noalias !8
79; CHECK-NEXT:    store i32 [[N_ADDR_SROA_0_0_EXTRACT_TRUNC]], i32* [[ARRAYIDX2_I]], align 4, !noalias !8
80; CHECK-NEXT:    br label [[REGION_GUARDED_END:%.*]]
81; CHECK:       region.guarded.end:
82; CHECK-NEXT:    br label [[REGION_BARRIER]]
83; CHECK:       region.barrier:
84; CHECK-NEXT:    call void @__kmpc_barrier_simple_spmd(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]])
85; CHECK-NEXT:    br label [[REGION_EXIT:%.*]]
86; CHECK:       region.exit:
87; CHECK-NEXT:    br label [[FOR_COND_I:%.*]]
88; CHECK:       for.cond.i:
89; CHECK-NEXT:    [[I_0_I:%.*]] = phi i32 [ 2, [[REGION_EXIT]] ], [ [[INC_I:%.*]], [[REGION_EXIT3:%.*]] ]
90; CHECK-NEXT:    [[SUB_I:%.*]] = add nsw i32 [[N_ADDR_SROA_0_0_EXTRACT_TRUNC]], -1
91; CHECK-NEXT:    [[CMP_I:%.*]] = icmp slt i32 [[I_0_I]], [[SUB_I]]
92; CHECK-NEXT:    br i1 [[CMP_I]], label [[FOR_BODY_I:%.*]], label [[__OMP_OUTLINED___EXIT:%.*]]
93; CHECK:       for.body.i:
94; CHECK-NEXT:    [[SUB3_I:%.*]] = add nsw i32 [[I_0_I]], -1
95; CHECK-NEXT:    [[IDXPROM4_I:%.*]] = zext i32 [[I_0_I]] to i64
96; CHECK-NEXT:    [[ARRAYIDX5_I:%.*]] = getelementptr inbounds i32, i32* [[X]], i64 [[IDXPROM4_I]]
97; CHECK-NEXT:    br label [[REGION_CHECK_TID5:%.*]]
98; CHECK:       region.check.tid5:
99; CHECK-NEXT:    [[TMP4:%.*]] = call i32 @__kmpc_get_hardware_thread_id_in_block()
100; CHECK-NEXT:    [[TMP5:%.*]] = icmp eq i32 [[TMP4]], 0
101; CHECK-NEXT:    br i1 [[TMP5]], label [[REGION_GUARDED4:%.*]], label [[REGION_BARRIER2:%.*]]
102; CHECK:       region.guarded4:
103; CHECK-NEXT:    store i32 [[SUB3_I]], i32* [[ARRAYIDX5_I]], align 4, !noalias !8
104; CHECK-NEXT:    br label [[REGION_GUARDED_END1:%.*]]
105; CHECK:       region.guarded.end1:
106; CHECK-NEXT:    br label [[REGION_BARRIER2]]
107; CHECK:       region.barrier2:
108; CHECK-NEXT:    call void @__kmpc_barrier_simple_spmd(%struct.ident_t* @[[GLOB1]], i32 [[TMP4]])
109; CHECK-NEXT:    br label [[REGION_EXIT3]]
110; CHECK:       region.exit3:
111; CHECK-NEXT:    [[INC_I]] = add nuw nsw i32 [[I_0_I]], 1
112; CHECK-NEXT:    br label [[FOR_COND_I]], !llvm.loop [[LOOP11:![0-9]+]]
113; CHECK:       __omp_outlined__.exit:
114; CHECK-NEXT:    call void @__kmpc_parallel_51(%struct.ident_t* null, i32 0, i32 1, i32 -1, i32 -1, i8* bitcast (void (i32*, i32*)* @__omp_outlined__1 to i8*), i8* bitcast (void (i16, i32)* @__omp_outlined__1_wrapper to i8*), i8** null, i64 0)
115; CHECK-NEXT:    [[CALL_I:%.*]] = call i32 @no_openmp(i32* nonnull [[X]]) #[[ATTR7:[0-9]+]], !noalias !8
116; CHECK-NEXT:    [[IDXPROM6_I:%.*]] = sext i32 [[CALL_I]] to i64
117; CHECK-NEXT:    [[ARRAYIDX7_I:%.*]] = getelementptr inbounds i32, i32* [[X]], i64 [[IDXPROM6_I]]
118; CHECK-NEXT:    br label [[REGION_CHECK_TID10:%.*]]
119; CHECK:       region.check.tid10:
120; CHECK-NEXT:    [[TMP6:%.*]] = call i32 @__kmpc_get_hardware_thread_id_in_block()
121; CHECK-NEXT:    [[TMP7:%.*]] = icmp eq i32 [[TMP6]], 0
122; CHECK-NEXT:    br i1 [[TMP7]], label [[REGION_GUARDED9:%.*]], label [[REGION_BARRIER7:%.*]]
123; CHECK:       region.guarded9:
124; CHECK-NEXT:    store i32 [[CALL_I]], i32* [[ARRAYIDX7_I]], align 4, !noalias !8
125; CHECK-NEXT:    br label [[REGION_GUARDED_END6:%.*]]
126; CHECK:       region.guarded.end6:
127; CHECK-NEXT:    br label [[REGION_BARRIER7]]
128; CHECK:       region.barrier7:
129; CHECK-NEXT:    call void @__kmpc_barrier_simple_spmd(%struct.ident_t* @[[GLOB1]], i32 [[TMP6]])
130; CHECK-NEXT:    br label [[REGION_EXIT8:%.*]]
131; CHECK:       region.exit8:
132; CHECK-NEXT:    [[CALL8_I:%.*]] = call i32 @no_openmp(i32* nonnull [[X]]) #[[ATTR7]], !noalias !8
133; CHECK-NEXT:    [[IDXPROM9_I:%.*]] = sext i32 [[CALL8_I]] to i64
134; CHECK-NEXT:    [[ARRAYIDX10_I:%.*]] = getelementptr inbounds i32, i32* [[X]], i64 [[IDXPROM9_I]]
135; CHECK-NEXT:    br label [[REGION_CHECK_TID15:%.*]]
136; CHECK:       region.check.tid15:
137; CHECK-NEXT:    [[TMP8:%.*]] = call i32 @__kmpc_get_hardware_thread_id_in_block()
138; CHECK-NEXT:    [[TMP9:%.*]] = icmp eq i32 [[TMP8]], 0
139; CHECK-NEXT:    br i1 [[TMP9]], label [[REGION_GUARDED14:%.*]], label [[REGION_BARRIER12:%.*]]
140; CHECK:       region.guarded14:
141; CHECK-NEXT:    store i32 [[CALL8_I]], i32* [[ARRAYIDX10_I]], align 4, !noalias !8
142; CHECK-NEXT:    br label [[REGION_GUARDED_END11:%.*]]
143; CHECK:       region.guarded.end11:
144; CHECK-NEXT:    br label [[REGION_BARRIER12]]
145; CHECK:       region.barrier12:
146; CHECK-NEXT:    call void @__kmpc_barrier_simple_spmd(%struct.ident_t* @[[GLOB1]], i32 [[TMP8]])
147; CHECK-NEXT:    br label [[REGION_EXIT13:%.*]]
148; CHECK:       region.exit13:
149; CHECK-NEXT:    [[CALL11_I:%.*]] = call i32 @no_openmp(i32* nonnull [[X]]) #[[ATTR7]], !noalias !8
150; CHECK-NEXT:    [[IDXPROM12_I:%.*]] = sext i32 [[CALL11_I]] to i64
151; CHECK-NEXT:    [[ARRAYIDX13_I:%.*]] = getelementptr inbounds i32, i32* [[X]], i64 [[IDXPROM12_I]]
152; CHECK-NEXT:    br label [[REGION_CHECK_TID20:%.*]]
153; CHECK:       region.check.tid20:
154; CHECK-NEXT:    [[TMP10:%.*]] = call i32 @__kmpc_get_hardware_thread_id_in_block()
155; CHECK-NEXT:    [[TMP11:%.*]] = icmp eq i32 [[TMP10]], 0
156; CHECK-NEXT:    br i1 [[TMP11]], label [[REGION_GUARDED19:%.*]], label [[REGION_BARRIER17:%.*]]
157; CHECK:       region.guarded19:
158; CHECK-NEXT:    store i32 [[CALL11_I]], i32* [[ARRAYIDX13_I]], align 4, !noalias !8
159; CHECK-NEXT:    br label [[REGION_GUARDED_END16:%.*]]
160; CHECK:       region.guarded.end16:
161; CHECK-NEXT:    br label [[REGION_BARRIER17]]
162; CHECK:       region.barrier17:
163; CHECK-NEXT:    call void @__kmpc_barrier_simple_spmd(%struct.ident_t* @[[GLOB1]], i32 [[TMP10]])
164; CHECK-NEXT:    br label [[REGION_EXIT18:%.*]]
165; CHECK:       region.exit18:
166; CHECK-NEXT:    [[CALL14_I:%.*]] = call i32 @no_openmp(i32* nonnull [[X]]) #[[ATTR7]], !noalias !8
167; CHECK-NEXT:    [[CALL15_I:%.*]] = call i32 @no_openmp(i32* nonnull [[X]]) #[[ATTR7]], !noalias !8
168; CHECK-NEXT:    [[CALL16_I:%.*]] = call i32 @no_openmp(i32* nonnull [[X]]) #[[ATTR7]], !noalias !8
169; CHECK-NEXT:    call void @__kmpc_target_deinit(%struct.ident_t* nonnull @[[GLOB1]], i8 2, i1 false) #[[ATTR4]]
170; CHECK-NEXT:    ret void
171; CHECK:       worker.exit:
172; CHECK-NEXT:    ret void
173;
174; CHECK-DISABLED-LABEL: define {{[^@]+}}@__omp_offloading_2a_fbfa7a_sequential_loop_l6
175; CHECK-DISABLED-SAME: (i32* [[X:%.*]], i64 [[N:%.*]]) #[[ATTR0:[0-9]+]] {
176; CHECK-DISABLED-NEXT:  entry:
177; CHECK-DISABLED-NEXT:    [[WORKER_WORK_FN_ADDR:%.*]] = alloca i8*, align 8
178; CHECK-DISABLED-NEXT:    [[N_ADDR_SROA_0_0_EXTRACT_TRUNC:%.*]] = trunc i64 [[N]] to i32
179; CHECK-DISABLED-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_target_init(%struct.ident_t* nonnull @[[GLOB1]], i8 1, i1 false, i1 true) #[[ATTR4:[0-9]+]]
180; CHECK-DISABLED-NEXT:    [[BLOCK_HW_SIZE:%.*]] = call i32 @__kmpc_get_hardware_num_threads_in_block()
181; CHECK-DISABLED-NEXT:    [[WARP_SIZE:%.*]] = call i32 @__kmpc_get_warp_size()
182; CHECK-DISABLED-NEXT:    [[BLOCK_SIZE:%.*]] = sub i32 [[BLOCK_HW_SIZE]], [[WARP_SIZE]]
183; CHECK-DISABLED-NEXT:    [[THREAD_IS_MAIN_OR_WORKER:%.*]] = icmp slt i32 [[TMP0]], [[BLOCK_SIZE]]
184; CHECK-DISABLED-NEXT:    br i1 [[THREAD_IS_MAIN_OR_WORKER]], label [[IS_WORKER_CHECK:%.*]], label [[WORKER_STATE_MACHINE_FINISHED:%.*]]
185; CHECK-DISABLED:       is_worker_check:
186; CHECK-DISABLED-NEXT:    [[THREAD_IS_WORKER:%.*]] = icmp ne i32 [[TMP0]], -1
187; CHECK-DISABLED-NEXT:    br i1 [[THREAD_IS_WORKER]], label [[WORKER_STATE_MACHINE_BEGIN:%.*]], label [[THREAD_USER_CODE_CHECK:%.*]]
188; CHECK-DISABLED:       worker_state_machine.begin:
189; CHECK-DISABLED-NEXT:    call void @__kmpc_barrier_simple_generic(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]])
190; CHECK-DISABLED-NEXT:    [[WORKER_IS_ACTIVE:%.*]] = call i1 @__kmpc_kernel_parallel(i8** [[WORKER_WORK_FN_ADDR]])
191; CHECK-DISABLED-NEXT:    [[WORKER_WORK_FN:%.*]] = load i8*, i8** [[WORKER_WORK_FN_ADDR]], align 8
192; CHECK-DISABLED-NEXT:    [[WORKER_WORK_FN_ADDR_CAST:%.*]] = bitcast i8* [[WORKER_WORK_FN]] to void (i16, i32)*
193; CHECK-DISABLED-NEXT:    [[WORKER_IS_DONE:%.*]] = icmp eq i8* [[WORKER_WORK_FN]], null
194; CHECK-DISABLED-NEXT:    br i1 [[WORKER_IS_DONE]], label [[WORKER_STATE_MACHINE_FINISHED]], label [[WORKER_STATE_MACHINE_IS_ACTIVE_CHECK:%.*]]
195; CHECK-DISABLED:       worker_state_machine.finished:
196; CHECK-DISABLED-NEXT:    ret void
197; CHECK-DISABLED:       worker_state_machine.is_active.check:
198; CHECK-DISABLED-NEXT:    br i1 [[WORKER_IS_ACTIVE]], label [[WORKER_STATE_MACHINE_PARALLEL_REGION_CHECK:%.*]], label [[WORKER_STATE_MACHINE_DONE_BARRIER:%.*]]
199; CHECK-DISABLED:       worker_state_machine.parallel_region.check:
200; CHECK-DISABLED-NEXT:    br i1 true, label [[WORKER_STATE_MACHINE_PARALLEL_REGION_EXECUTE:%.*]], label [[WORKER_STATE_MACHINE_PARALLEL_REGION_CHECK1:%.*]]
201; CHECK-DISABLED:       worker_state_machine.parallel_region.execute:
202; CHECK-DISABLED-NEXT:    call void @__omp_outlined__1_wrapper(i16 0, i32 [[TMP0]])
203; CHECK-DISABLED-NEXT:    br label [[WORKER_STATE_MACHINE_PARALLEL_REGION_END:%.*]]
204; CHECK-DISABLED:       worker_state_machine.parallel_region.check1:
205; CHECK-DISABLED-NEXT:    br label [[WORKER_STATE_MACHINE_PARALLEL_REGION_END]]
206; CHECK-DISABLED:       worker_state_machine.parallel_region.end:
207; CHECK-DISABLED-NEXT:    call void @__kmpc_kernel_end_parallel()
208; CHECK-DISABLED-NEXT:    br label [[WORKER_STATE_MACHINE_DONE_BARRIER]]
209; CHECK-DISABLED:       worker_state_machine.done.barrier:
210; CHECK-DISABLED-NEXT:    call void @__kmpc_barrier_simple_generic(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]])
211; CHECK-DISABLED-NEXT:    br label [[WORKER_STATE_MACHINE_BEGIN]]
212; CHECK-DISABLED:       thread.user_code.check:
213; CHECK-DISABLED-NEXT:    [[EXEC_USER_CODE:%.*]] = icmp eq i32 [[TMP0]], -1
214; CHECK-DISABLED-NEXT:    br i1 [[EXEC_USER_CODE]], label [[USER_CODE_ENTRY:%.*]], label [[WORKER_EXIT:%.*]]
215; CHECK-DISABLED:       user_code.entry:
216; CHECK-DISABLED-NEXT:    [[TMP1:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* nonnull @[[GLOB1]]) #[[ATTR4]]
217; CHECK-DISABLED-NEXT:    store i32 0, i32* [[X]], align 4, !noalias !8
218; CHECK-DISABLED-NEXT:    [[ARRAYIDX1_I:%.*]] = getelementptr inbounds i32, i32* [[X]], i64 1
219; CHECK-DISABLED-NEXT:    store i32 1, i32* [[ARRAYIDX1_I]], align 4, !noalias !8
220; CHECK-DISABLED-NEXT:    [[SEXT:%.*]] = shl i64 [[N]], 32
221; CHECK-DISABLED-NEXT:    [[IDXPROM_I:%.*]] = ashr exact i64 [[SEXT]], 32
222; CHECK-DISABLED-NEXT:    [[ARRAYIDX2_I:%.*]] = getelementptr inbounds i32, i32* [[X]], i64 [[IDXPROM_I]]
223; CHECK-DISABLED-NEXT:    store i32 [[N_ADDR_SROA_0_0_EXTRACT_TRUNC]], i32* [[ARRAYIDX2_I]], align 4, !noalias !8
224; CHECK-DISABLED-NEXT:    br label [[FOR_COND_I:%.*]]
225; CHECK-DISABLED:       for.cond.i:
226; CHECK-DISABLED-NEXT:    [[I_0_I:%.*]] = phi i32 [ 2, [[USER_CODE_ENTRY]] ], [ [[INC_I:%.*]], [[FOR_BODY_I:%.*]] ]
227; CHECK-DISABLED-NEXT:    [[SUB_I:%.*]] = add nsw i32 [[N_ADDR_SROA_0_0_EXTRACT_TRUNC]], -1
228; CHECK-DISABLED-NEXT:    [[CMP_I:%.*]] = icmp slt i32 [[I_0_I]], [[SUB_I]]
229; CHECK-DISABLED-NEXT:    br i1 [[CMP_I]], label [[FOR_BODY_I]], label [[__OMP_OUTLINED___EXIT:%.*]]
230; CHECK-DISABLED:       for.body.i:
231; CHECK-DISABLED-NEXT:    [[SUB3_I:%.*]] = add nsw i32 [[I_0_I]], -1
232; CHECK-DISABLED-NEXT:    [[IDXPROM4_I:%.*]] = zext i32 [[I_0_I]] to i64
233; CHECK-DISABLED-NEXT:    [[ARRAYIDX5_I:%.*]] = getelementptr inbounds i32, i32* [[X]], i64 [[IDXPROM4_I]]
234; CHECK-DISABLED-NEXT:    store i32 [[SUB3_I]], i32* [[ARRAYIDX5_I]], align 4, !noalias !8
235; CHECK-DISABLED-NEXT:    [[INC_I]] = add nuw nsw i32 [[I_0_I]], 1
236; CHECK-DISABLED-NEXT:    br label [[FOR_COND_I]], !llvm.loop [[LOOP11:![0-9]+]]
237; CHECK-DISABLED:       __omp_outlined__.exit:
238; CHECK-DISABLED-NEXT:    call void @__kmpc_parallel_51(%struct.ident_t* null, i32 0, i32 1, i32 -1, i32 -1, i8* bitcast (void (i32*, i32*)* @__omp_outlined__1 to i8*), i8* @__omp_outlined__1_wrapper.ID, i8** null, i64 0)
239; CHECK-DISABLED-NEXT:    [[CALL_I:%.*]] = call i32 @no_openmp(i32* nonnull [[X]]) #[[ATTR7:[0-9]+]], !noalias !8
240; CHECK-DISABLED-NEXT:    [[IDXPROM6_I:%.*]] = sext i32 [[CALL_I]] to i64
241; CHECK-DISABLED-NEXT:    [[ARRAYIDX7_I:%.*]] = getelementptr inbounds i32, i32* [[X]], i64 [[IDXPROM6_I]]
242; CHECK-DISABLED-NEXT:    store i32 [[CALL_I]], i32* [[ARRAYIDX7_I]], align 4, !noalias !8
243; CHECK-DISABLED-NEXT:    [[CALL8_I:%.*]] = call i32 @no_openmp(i32* nonnull [[X]]) #[[ATTR7]], !noalias !8
244; CHECK-DISABLED-NEXT:    [[IDXPROM9_I:%.*]] = sext i32 [[CALL8_I]] to i64
245; CHECK-DISABLED-NEXT:    [[ARRAYIDX10_I:%.*]] = getelementptr inbounds i32, i32* [[X]], i64 [[IDXPROM9_I]]
246; CHECK-DISABLED-NEXT:    store i32 [[CALL8_I]], i32* [[ARRAYIDX10_I]], align 4, !noalias !8
247; CHECK-DISABLED-NEXT:    [[CALL11_I:%.*]] = call i32 @no_openmp(i32* nonnull [[X]]) #[[ATTR7]], !noalias !8
248; CHECK-DISABLED-NEXT:    [[IDXPROM12_I:%.*]] = sext i32 [[CALL11_I]] to i64
249; CHECK-DISABLED-NEXT:    [[ARRAYIDX13_I:%.*]] = getelementptr inbounds i32, i32* [[X]], i64 [[IDXPROM12_I]]
250; CHECK-DISABLED-NEXT:    store i32 [[CALL11_I]], i32* [[ARRAYIDX13_I]], align 4, !noalias !8
251; CHECK-DISABLED-NEXT:    [[CALL14_I:%.*]] = call i32 @no_openmp(i32* nonnull [[X]]) #[[ATTR7]], !noalias !8
252; CHECK-DISABLED-NEXT:    [[CALL15_I:%.*]] = call i32 @no_openmp(i32* nonnull [[X]]) #[[ATTR7]], !noalias !8
253; CHECK-DISABLED-NEXT:    [[CALL16_I:%.*]] = call i32 @no_openmp(i32* nonnull [[X]]) #[[ATTR7]], !noalias !8
254; CHECK-DISABLED-NEXT:    call void @__kmpc_target_deinit(%struct.ident_t* nonnull @[[GLOB1]], i8 1, i1 true) #[[ATTR4]]
255; CHECK-DISABLED-NEXT:    ret void
256; CHECK-DISABLED:       worker.exit:
257; CHECK-DISABLED-NEXT:    ret void
258;
259entry:
260  %N.addr.sroa.0.0.extract.trunc = trunc i64 %N to i32
261  %0 = call i32 @__kmpc_target_init(%struct.ident_t* nonnull @1, i8 1, i1 true, i1 true) #3
262  %exec_user_code = icmp eq i32 %0, -1
263  br i1 %exec_user_code, label %user_code.entry, label %worker.exit
264
265user_code.entry:                                  ; preds = %entry
266  %1 = call i32 @__kmpc_global_thread_num(%struct.ident_t* nonnull @1)
267  store i32 0, i32* %x, align 4, !noalias !8
268  %arrayidx1.i = getelementptr inbounds i32, i32* %x, i64 1
269  store i32 1, i32* %arrayidx1.i, align 4, !noalias !8
270  %sext = shl i64 %N, 32
271  %idxprom.i = ashr exact i64 %sext, 32
272  %arrayidx2.i = getelementptr inbounds i32, i32* %x, i64 %idxprom.i
273  store i32 %N.addr.sroa.0.0.extract.trunc, i32* %arrayidx2.i, align 4, !noalias !8
274  br label %for.cond.i
275
276for.cond.i:                                       ; preds = %for.body.i, %user_code.entry
277  %i.0.i = phi i32 [ 2, %user_code.entry ], [ %inc.i, %for.body.i ]
278  %sub.i = add nsw i32 %N.addr.sroa.0.0.extract.trunc, -1
279  %cmp.i = icmp slt i32 %i.0.i, %sub.i
280  br i1 %cmp.i, label %for.body.i, label %__omp_outlined__.exit
281
282for.body.i:                                       ; preds = %for.cond.i
283  %sub3.i = add nsw i32 %i.0.i, -1
284  %idxprom4.i = zext i32 %i.0.i to i64
285  %arrayidx5.i = getelementptr inbounds i32, i32* %x, i64 %idxprom4.i
286  store i32 %sub3.i, i32* %arrayidx5.i, align 4, !noalias !8
287  %inc.i = add nuw nsw i32 %i.0.i, 1
288  br label %for.cond.i, !llvm.loop !11
289
290__omp_outlined__.exit:                            ; preds = %for.cond.i
291  call void @__kmpc_parallel_51(%struct.ident_t* null, i32 0, i32 1, i32 -1, i32 -1, i8* bitcast (void (i32*, i32*)* @__omp_outlined__1 to i8*), i8* bitcast (void (i16, i32)* @__omp_outlined__1_wrapper to i8*), i8** null, i64 0)
292  %call.i = call i32 @no_openmp(i32* nonnull %x) #5, !noalias !8
293  %idxprom6.i = sext i32 %call.i to i64
294  %arrayidx7.i = getelementptr inbounds i32, i32* %x, i64 %idxprom6.i
295  store i32 %call.i, i32* %arrayidx7.i, align 4, !noalias !8
296  %call8.i = call i32 @no_openmp(i32* nonnull %x) #5, !noalias !8
297  %idxprom9.i = sext i32 %call8.i to i64
298  %arrayidx10.i = getelementptr inbounds i32, i32* %x, i64 %idxprom9.i
299  store i32 %call8.i, i32* %arrayidx10.i, align 4, !noalias !8
300  %call11.i = call i32 @no_openmp(i32* nonnull %x) #5, !noalias !8
301  %idxprom12.i = sext i32 %call11.i to i64
302  %arrayidx13.i = getelementptr inbounds i32, i32* %x, i64 %idxprom12.i
303  store i32 %call11.i, i32* %arrayidx13.i, align 4, !noalias !8
304  %call14.i = call i32 @no_openmp(i32* nonnull %x) #5, !noalias !8
305  %call15.i = call i32 @no_openmp(i32* nonnull %x) #5, !noalias !8
306  %call16.i = call i32 @no_openmp(i32* nonnull %x) #5, !noalias !8
307  call void @__kmpc_target_deinit(%struct.ident_t* nonnull @1, i8 1, i1 true) #3
308  ret void
309
310worker.exit:                                      ; preds = %entry
311  ret void
312}
313
314define internal void @__omp_outlined__1(i32* noalias %.global_tid., i32* noalias %.bound_tid.) {
315; CHECK-LABEL: define {{[^@]+}}@__omp_outlined__1
316; CHECK-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) {
317; CHECK-NEXT:    ret void
318;
319; CHECK-DISABLED-LABEL: define {{[^@]+}}@__omp_outlined__1
320; CHECK-DISABLED-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) {
321; CHECK-DISABLED-NEXT:    ret void
322;
323  ret void
324}
325define internal void @__omp_outlined__1_wrapper(i16 zeroext %0, i32 %1) {
326; CHECK-LABEL: define {{[^@]+}}@__omp_outlined__1_wrapper
327; CHECK-SAME: (i16 zeroext [[TMP0:%.*]], i32 [[TMP1:%.*]]) {
328; CHECK-NEXT:    ret void
329;
330; CHECK-DISABLED-LABEL: define {{[^@]+}}@__omp_outlined__1_wrapper
331; CHECK-DISABLED-SAME: (i16 zeroext [[TMP0:%.*]], i32 [[TMP1:%.*]]) {
332; CHECK-DISABLED-NEXT:    ret void
333;
334  ret void
335}
336
337declare void @__kmpc_parallel_51(%struct.ident_t*, i32, i32, i32, i32, i8*, i8*, i8**, i64)
338
339declare i32 @__kmpc_target_init(%struct.ident_t*, i8, i1, i1)
340
341; Function Attrs: convergent
342declare i32 @no_openmp(i32*) #1
343
344; Function Attrs: convergent nounwind readonly willreturn
345declare void @pure() #2
346
347; Function Attrs: nounwind
348declare i32 @__kmpc_global_thread_num(%struct.ident_t*) #3
349
350declare void @__kmpc_target_deinit(%struct.ident_t*, i8, i1)
351
352; Function Attrs: inaccessiblememonly nofree nosync nounwind willreturn
353declare void @llvm.experimental.noalias.scope.decl(metadata) #4
354
355attributes #0 = { convergent norecurse nounwind "frame-pointer"="all" "min-legal-vector-width"="0" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="sm_53" "target-features"="+ptx32,+sm_53" }
356attributes #1 = { convergent "frame-pointer"="all" "llvm.assume"="omp_no_openmp,ompx_spmd_amenable" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="sm_53" "target-features"="+ptx32,+sm_53" }
357attributes #2 = { convergent nounwind readonly willreturn "frame-pointer"="all" "llvm.assume"="ompx_spmd_amenable" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="sm_53" "target-features"="+ptx32,+sm_53" }
358attributes #3 = { nounwind }
359attributes #4 = { inaccessiblememonly nofree nosync nounwind willreturn }
360attributes #5 = { convergent nounwind "llvm.assume"="omp_no_openmp,ompx_spmd_amenable" }
361
362!omp_offload.info = !{!0}
363!nvvm.annotations = !{!1}
364!llvm.module.flags = !{!2, !3, !4, !5, !6}
365!llvm.ident = !{!7}
366
367!0 = !{i32 0, i32 42, i32 16513658, !"sequential_loop", i32 6, i32 0}
368!1 = !{void (i32*, i64)* @__omp_offloading_2a_fbfa7a_sequential_loop_l6, !"kernel", i32 1}
369!2 = !{i32 1, !"wchar_size", i32 4}
370!3 = !{i32 7, !"openmp", i32 50}
371!4 = !{i32 7, !"openmp-device", i32 50}
372!5 = !{i32 7, !"PIC Level", i32 2}
373!6 = !{i32 7, !"frame-pointer", i32 2}
374!7 = !{!"clang version 14.0.0"}
375!8 = !{!9}
376!9 = distinct !{!9, !10, !"__omp_outlined__: %__context"}
377!10 = distinct !{!10, !"__omp_outlined__"}
378!11 = distinct !{!11, !12}
379!12 = !{!"llvm.loop.mustprogress"}
380;.
381; CHECK: attributes #[[ATTR0]] = { convergent norecurse nounwind "frame-pointer"="all" "min-legal-vector-width"="0" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="sm_53" "target-features"="+ptx32,+sm_53" }
382; CHECK: attributes #[[ATTR1:[0-9]+]] = { alwaysinline }
383; CHECK: attributes #[[ATTR2:[0-9]+]] = { convergent "frame-pointer"="all" "llvm.assume"="omp_no_openmp,ompx_spmd_amenable" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="sm_53" "target-features"="+ptx32,+sm_53" }
384; CHECK: attributes #[[ATTR3:[0-9]+]] = { convergent nounwind readonly willreturn "frame-pointer"="all" "llvm.assume"="ompx_spmd_amenable" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="sm_53" "target-features"="+ptx32,+sm_53" }
385; CHECK: attributes #[[ATTR4]] = { nounwind }
386; CHECK: attributes #[[ATTR5:[0-9]+]] = { inaccessiblememonly nofree nosync nounwind willreturn }
387; CHECK: attributes #[[ATTR6:[0-9]+]] = { convergent nounwind }
388; CHECK: attributes #[[ATTR7]] = { convergent nounwind "llvm.assume"="omp_no_openmp,ompx_spmd_amenable" }
389;.
390; CHECK-DISABLED: attributes #[[ATTR0]] = { convergent norecurse nounwind "frame-pointer"="all" "min-legal-vector-width"="0" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="sm_53" "target-features"="+ptx32,+sm_53" }
391; CHECK-DISABLED: attributes #[[ATTR1:[0-9]+]] = { alwaysinline }
392; CHECK-DISABLED: attributes #[[ATTR2:[0-9]+]] = { convergent "frame-pointer"="all" "llvm.assume"="omp_no_openmp,ompx_spmd_amenable" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="sm_53" "target-features"="+ptx32,+sm_53" }
393; CHECK-DISABLED: attributes #[[ATTR3:[0-9]+]] = { convergent nounwind readonly willreturn "frame-pointer"="all" "llvm.assume"="ompx_spmd_amenable" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="sm_53" "target-features"="+ptx32,+sm_53" }
394; CHECK-DISABLED: attributes #[[ATTR4]] = { nounwind }
395; CHECK-DISABLED: attributes #[[ATTR5:[0-9]+]] = { inaccessiblememonly nofree nosync nounwind willreturn }
396; CHECK-DISABLED: attributes #[[ATTR6:[0-9]+]] = { convergent nounwind }
397; CHECK-DISABLED: attributes #[[ATTR7]] = { convergent nounwind "llvm.assume"="omp_no_openmp,ompx_spmd_amenable" }
398;.
399; CHECK: [[META0:![0-9]+]] = !{i32 0, i32 42, i32 16513658, !"sequential_loop", i32 6, i32 0}
400; CHECK: [[META1:![0-9]+]] = !{void (i32*, i64)* @__omp_offloading_2a_fbfa7a_sequential_loop_l6, !"kernel", i32 1}
401; CHECK: [[META2:![0-9]+]] = !{i32 1, !"wchar_size", i32 4}
402; CHECK: [[META3:![0-9]+]] = !{i32 7, !"openmp", i32 50}
403; CHECK: [[META4:![0-9]+]] = !{i32 7, !"openmp-device", i32 50}
404; CHECK: [[META5:![0-9]+]] = !{i32 7, !"PIC Level", i32 2}
405; CHECK: [[META6:![0-9]+]] = !{i32 7, !"frame-pointer", i32 2}
406; CHECK: [[META7:![0-9]+]] = !{!"clang version 14.0.0"}
407; CHECK: [[META8:![0-9]+]] = !{!9}
408; CHECK: [[META9:![0-9]+]] = distinct !{!9, !10, !"__omp_outlined__: %__context"}
409; CHECK: [[META10:![0-9]+]] = distinct !{!10, !"__omp_outlined__"}
410; CHECK: [[LOOP11]] = distinct !{!11, !12}
411; CHECK: [[META12:![0-9]+]] = !{!"llvm.loop.mustprogress"}
412;.
413; CHECK-DISABLED: [[META0:![0-9]+]] = !{i32 0, i32 42, i32 16513658, !"sequential_loop", i32 6, i32 0}
414; CHECK-DISABLED: [[META1:![0-9]+]] = !{void (i32*, i64)* @__omp_offloading_2a_fbfa7a_sequential_loop_l6, !"kernel", i32 1}
415; CHECK-DISABLED: [[META2:![0-9]+]] = !{i32 1, !"wchar_size", i32 4}
416; CHECK-DISABLED: [[META3:![0-9]+]] = !{i32 7, !"openmp", i32 50}
417; CHECK-DISABLED: [[META4:![0-9]+]] = !{i32 7, !"openmp-device", i32 50}
418; CHECK-DISABLED: [[META5:![0-9]+]] = !{i32 7, !"PIC Level", i32 2}
419; CHECK-DISABLED: [[META6:![0-9]+]] = !{i32 7, !"frame-pointer", i32 2}
420; CHECK-DISABLED: [[META7:![0-9]+]] = !{!"clang version 14.0.0"}
421; CHECK-DISABLED: [[META8:![0-9]+]] = !{!9}
422; CHECK-DISABLED: [[META9:![0-9]+]] = distinct !{!9, !10, !"__omp_outlined__: %__context"}
423; CHECK-DISABLED: [[META10:![0-9]+]] = distinct !{!10, !"__omp_outlined__"}
424; CHECK-DISABLED: [[LOOP11]] = distinct !{!11, !12}
425; CHECK-DISABLED: [[META12:![0-9]+]] = !{!"llvm.loop.mustprogress"}
426;.
427