1 ; RUN: llc -march=amdgcn -verify-machineinstrs -amdgpu-s-branch-bits=4 < %s | FileCheck -check-prefix=GCN %s
2 ; Restrict maximum branch to between +7 and -8 dwords
4 ; Used to emit an always 4 byte instruction. Inline asm always assumes
5 ; each instruction is the maximum size.
6 declare void @llvm.amdgcn.s.sleep(i32) #0
8 declare i32 @llvm.amdgcn.workitem.id.x() #1
11 ; GCN-LABEL: {{^}}uniform_conditional_max_short_forward_branch:
12 ; GCN: s_load_dword [[CND:s[0-9]+]]
13 ; GCN: s_cmp_eq_u32 [[CND]], 0
14 ; GCN-NEXT: s_cbranch_scc1 [[BB3:BB[0-9]+_[0-9]+]]
17 ; GCN-NEXT: ; BB#1: ; %bb2
18 ; GCN-NEXT: ;;#ASMSTART
25 ; GCN-NEXT: [[BB3]]: ; %bb3
26 ; GCN: v_mov_b32_e32 [[V_CND:v[0-9]+]], [[CND]]
27 ; GCN: buffer_store_dword [[V_CND]]
29 define amdgpu_kernel void @uniform_conditional_max_short_forward_branch(i32 addrspace(1)* %arg, i32 %cnd) #0 {
31 %cmp = icmp eq i32 %cnd, 0
32 br i1 %cmp, label %bb3, label %bb2 ; +8 dword branch
36 call void asm sideeffect
40 call void @llvm.amdgcn.s.sleep(i32 0)
44 store volatile i32 %cnd, i32 addrspace(1)* %arg
48 ; GCN-LABEL: {{^}}uniform_conditional_min_long_forward_branch:
49 ; GCN: s_load_dword [[CND:s[0-9]+]]
50 ; GCN: s_cmp_eq_u32 [[CND]], 0
51 ; GCN-NEXT: s_cbranch_scc0 [[LONGBB:BB[0-9]+_[0-9]+]]
53 ; GCN-NEXT: [[LONG_JUMP:BB[0-9]+_[0-9]+]]: ; %bb0
54 ; GCN-NEXT: s_getpc_b64 vcc
55 ; GCN-NEXT: s_add_u32 vcc_lo, vcc_lo, [[ENDBB:BB[0-9]+_[0-9]+]]-([[LONG_JUMP]]+4)
56 ; GCN-NEXT: s_addc_u32 vcc_hi, vcc_hi, 0
57 ; GCN-NEXT: s_setpc_b64 vcc
59 ; GCN-NEXT: [[LONGBB]]:
60 ; GCN-NEXT: ;;#ASMSTART
67 ; GCN-NEXT: [[ENDBB]]:
68 ; GCN: v_mov_b32_e32 [[V_CND:v[0-9]+]], [[CND]]
69 ; GCN: buffer_store_dword [[V_CND]]
71 define amdgpu_kernel void @uniform_conditional_min_long_forward_branch(i32 addrspace(1)* %arg, i32 %cnd) #0 {
73 %cmp = icmp eq i32 %cnd, 0
74 br i1 %cmp, label %bb3, label %bb2 ; +9 dword branch
78 call void asm sideeffect
86 store volatile i32 %cnd, i32 addrspace(1)* %arg
90 ; GCN-LABEL: {{^}}uniform_conditional_min_long_forward_vcnd_branch:
91 ; GCN: s_load_dword [[CND:s[0-9]+]]
92 ; GCN-DAG: v_mov_b32_e32 [[V_CND:v[0-9]+]], [[CND]]
93 ; GCN-DAG: v_cmp_eq_f32_e64 vcc, [[CND]], 0
94 ; GCN: s_cbranch_vccz [[LONGBB:BB[0-9]+_[0-9]+]]
96 ; GCN-NEXT: [[LONG_JUMP:BB[0-9]+_[0-9]+]]: ; %bb0
97 ; GCN-NEXT: s_getpc_b64 vcc
98 ; GCN-NEXT: s_add_u32 vcc_lo, vcc_lo, [[ENDBB:BB[0-9]+_[0-9]+]]-([[LONG_JUMP]]+4)
99 ; GCN-NEXT: s_addc_u32 vcc_hi, vcc_hi, 0
100 ; GCN-NEXT: s_setpc_b64 vcc
102 ; GCN-NEXT: [[LONGBB]]:
109 ; GCN: buffer_store_dword [[V_CND]]
111 define amdgpu_kernel void @uniform_conditional_min_long_forward_vcnd_branch(float addrspace(1)* %arg, float %cnd) #0 {
113 %cmp = fcmp oeq float %cnd, 0.0
114 br i1 %cmp, label %bb3, label %bb2 ; + 8 dword branch
117 call void asm sideeffect " ; 32 bytes
125 store volatile float %cnd, float addrspace(1)* %arg
129 ; GCN-LABEL: {{^}}min_long_forward_vbranch:
131 ; GCN: buffer_load_dword
132 ; GCN: v_cmp_ne_u32_e32 vcc, 0, v{{[0-9]+}}
133 ; GCN: s_and_saveexec_b64 [[SAVE:s\[[0-9]+:[0-9]+\]]], vcc
134 ; GCN: s_xor_b64 [[SAVE]], exec, [[SAVE]]
141 ; GCN: s_or_b64 exec, exec, [[SAVE]]
142 ; GCN: buffer_store_dword
144 define amdgpu_kernel void @min_long_forward_vbranch(i32 addrspace(1)* %arg) #0 {
146 %tid = call i32 @llvm.amdgcn.workitem.id.x()
147 %tid.ext = zext i32 %tid to i64
148 %gep = getelementptr inbounds i32, i32 addrspace(1)* %arg, i64 %tid.ext
149 %load = load volatile i32, i32 addrspace(1)* %gep
150 %cmp = icmp eq i32 %load, 0
151 br i1 %cmp, label %bb3, label %bb2 ; + 8 dword branch
154 call void asm sideeffect " ; 32 bytes
162 store volatile i32 %load, i32 addrspace(1)* %gep
166 ; GCN-LABEL: {{^}}long_backward_sbranch:
167 ; GCN: s_mov_b32 [[LOOPIDX:s[0-9]+]], 0{{$}}
169 ; GCN: [[LOOPBB:BB[0-9]+_[0-9]+]]: ; %bb2
170 ; GCN-NEXT: ; =>This Inner Loop Header: Depth=1
171 ; GCN-NEXT: s_add_i32 [[INC:s[0-9]+]], [[LOOPIDX]], 1
172 ; GCN-NEXT: s_cmp_lt_i32 [[INC]], 10
174 ; GCN-NEXT: ;;#ASMSTART
175 ; GCN-NEXT: v_nop_e64
176 ; GCN-NEXT: v_nop_e64
177 ; GCN-NEXT: v_nop_e64
178 ; GCN-NEXT: ;;#ASMEND
180 ; GCN-NEXT: s_cbranch_scc0 [[ENDBB:BB[0-9]+_[0-9]+]]
182 ; GCN-NEXT: [[LONG_JUMP:BB[0-9]+_[0-9]+]]: ; %bb2
183 ; GCN-NEXT: ; in Loop: Header=[[LOOPBB]] Depth=1
184 ; GCN-NEXT: s_getpc_b64 vcc
185 ; GCN-NEXT: s_sub_u32 vcc_lo, vcc_lo, ([[LONG_JUMP]]+4)-[[LOOPBB]]
186 ; GCN-NEXT: s_subb_u32 vcc_hi, vcc_hi, 0
187 ; GCN-NEXT: s_setpc_b64 vcc
189 ; GCN-NEXT: [[ENDBB]]:
191 define amdgpu_kernel void @long_backward_sbranch(i32 addrspace(1)* %arg) #0 {
196 %loop.idx = phi i32 [ 0, %bb ], [ %inc, %bb2 ]
198 call void asm sideeffect
202 %inc = add nsw i32 %loop.idx, 1 ; add cost 4
203 %cmp = icmp slt i32 %inc, 10 ; condition cost = 8
204 br i1 %cmp, label %bb2, label %bb3 ; -
210 ; Requires expansion of unconditional branch from %bb2 to %bb4 (and
211 ; expansion of conditional branch from %bb to %bb3.
213 ; GCN-LABEL: {{^}}uniform_unconditional_min_long_forward_branch:
215 ; GCN-NEXT: s_cbranch_scc0 [[BB2:BB[0-9]+_[0-9]+]]
217 ; GCN-NEXT: [[LONG_JUMP0:BB[0-9]+_[0-9]+]]: ; %bb0
218 ; GCN-NEXT: s_getpc_b64 vcc
219 ; GCN-NEXT: s_add_u32 vcc_lo, vcc_lo, [[BB3:BB[0-9]_[0-9]+]]-([[LONG_JUMP0]]+4)
220 ; GCN-NEXT: s_addc_u32 vcc_hi, vcc_hi, 0{{$}}
221 ; GCN-NEXT: s_setpc_b64 vcc
223 ; GCN-NEXT: [[BB2]]: ; %bb2
224 ; GCN: v_mov_b32_e32 [[BB2_K:v[0-9]+]], 17
225 ; GCN: buffer_store_dword [[BB2_K]]
227 ; GCN-NEXT: [[LONG_JUMP1:BB[0-9]+_[0-9]+]]: ; %bb2
228 ; GCN-NEXT: s_getpc_b64 vcc
229 ; GCN-NEXT: s_add_u32 vcc_lo, vcc_lo, [[BB4:BB[0-9]_[0-9]+]]-([[LONG_JUMP1]]+4)
230 ; GCN-NEXT: s_addc_u32 vcc_hi, vcc_hi, 0{{$}}
231 ; GCN-NEXT: s_setpc_b64 vcc
233 ; GCN: [[BB3]]: ; %bb3
240 ; GCN-NEXT: [[BB4]]: ; %bb4
241 ; GCN: v_mov_b32_e32 [[BB4_K:v[0-9]+]], 63
242 ; GCN: buffer_store_dword [[BB4_K]]
244 ; GCN-NEXT: .Lfunc_end{{[0-9]+}}:
245 define amdgpu_kernel void @uniform_unconditional_min_long_forward_branch(i32 addrspace(1)* %arg, i32 %arg1) {
247 %tmp = icmp ne i32 %arg1, 0
248 br i1 %tmp, label %bb2, label %bb3
251 store volatile i32 17, i32 addrspace(1)* undef
256 call void asm sideeffect
264 store volatile i32 63, i32 addrspace(1)* %arg
268 ; GCN-LABEL: {{^}}uniform_unconditional_min_long_backward_branch:
269 ; GCN-NEXT: ; BB#0: ; %entry
271 ; GCN-NEXT: [[LOOP:BB[0-9]_[0-9]+]]: ; %loop
272 ; GCN-NEXT: ; =>This Inner Loop Header: Depth=1
273 ; GCN-NEXT: ;;#ASMSTART
274 ; GCN-NEXT: v_nop_e64
275 ; GCN-NEXT: v_nop_e64
276 ; GCN-NEXT: v_nop_e64
277 ; GCN-NEXT: v_nop_e64
278 ; GCN-NEXT: ;;#ASMEND
280 ; GCN-NEXT: [[LONGBB:BB[0-9]+_[0-9]+]]: ; %loop
281 ; GCN-NEXT: ; in Loop: Header=[[LOOP]] Depth=1
282 ; GCN-NEXT: s_getpc_b64 vcc
283 ; GCN-NEXT: s_sub_u32 vcc_lo, vcc_lo, ([[LONGBB]]+4)-[[LOOP]]
284 ; GCN-NEXT: s_subb_u32 vcc_hi, vcc_hi, 0{{$}}
285 ; GCN-NEXT: s_setpc_b64 vcc
286 ; GCN-NEXT .Lfunc_end{{[0-9]+}}:
287 define amdgpu_kernel void @uniform_unconditional_min_long_backward_branch(i32 addrspace(1)* %arg, i32 %arg1) {
293 call void asm sideeffect
301 ; Expansion of branch from %bb1 to %bb3 introduces need to expand
302 ; branch from %bb0 to %bb2
304 ; GCN-LABEL: {{^}}expand_requires_expand:
305 ; GCN-NEXT: ; BB#0: ; %bb0
307 ; GCN: s_cmp_lt_i32 s{{[0-9]+}}, 0{{$}}
308 ; GCN-NEXT: s_cbranch_scc0 [[BB1:BB[0-9]+_[0-9]+]]
310 ; GCN-NEXT: [[LONGBB0:BB[0-9]+_[0-9]+]]: ; %bb0
311 ; GCN-NEXT: s_getpc_b64 vcc
312 ; GCN-NEXT: s_add_u32 vcc_lo, vcc_lo, [[BB2:BB[0-9]_[0-9]+]]-([[LONGBB0]]+4)
313 ; GCN-NEXT: s_addc_u32 vcc_hi, vcc_hi, 0{{$}}
314 ; GCN-NEXT: s_setpc_b64 vcc
316 ; GCN-NEXT: [[BB1]]: ; %bb1
317 ; GCN-NEXT: s_load_dword
318 ; GCN-NEXT: s_waitcnt lgkmcnt(0)
319 ; GCN-NEXT: s_cmp_eq_u32 s{{[0-9]+}}, 3{{$}}
320 ; GCN-NEXT: s_cbranch_scc0 [[BB2:BB[0-9]_[0-9]+]]
322 ; GCN-NEXT: [[LONGBB1:BB[0-9]+_[0-9]+]]: ; %bb1
323 ; GCN-NEXT: s_getpc_b64 vcc
324 ; GCN-NEXT: s_add_u32 vcc_lo, vcc_lo, [[BB3:BB[0-9]+_[0-9]+]]-([[LONGBB1]]+4)
325 ; GCN-NEXT: s_addc_u32 vcc_hi, vcc_hi, 0{{$}}
326 ; GCN-NEXT: s_setpc_b64 vcc
328 ; GCN-NEXT: [[BB2]]: ; %bb2
329 ; GCN-NEXT: ;;#ASMSTART
330 ; GCN-NEXT: v_nop_e64
331 ; GCN-NEXT: v_nop_e64
332 ; GCN-NEXT: v_nop_e64
333 ; GCN-NEXT: v_nop_e64
334 ; GCN-NEXT: ;;#ASMEND
336 ; GCN-NEXT: [[BB3]]: ; %bb3
337 ; GCN-NEXT: ;;#ASMSTART
338 ; GCN-NEXT: v_nop_e64
339 ; GCN-NEXT: ;;#ASMEND
340 ; GCN-NEXT: ;;#ASMSTART
341 ; GCN-NEXT: v_nop_e64
342 ; GCN-NEXT: ;;#ASMEND
344 define amdgpu_kernel void @expand_requires_expand(i32 %cond0) #0 {
346 %tmp = tail call i32 @llvm.amdgcn.workitem.id.x() #0
347 %cmp0 = icmp slt i32 %cond0, 0
348 br i1 %cmp0, label %bb2, label %bb1
351 %val = load volatile i32, i32 addrspace(2)* undef
352 %cmp1 = icmp eq i32 %val, 3
353 br i1 %cmp1, label %bb3, label %bb2
356 call void asm sideeffect
364 ; These NOPs prevent tail-duplication-based outlining
365 ; from firing, which defeats the need to expand the branches and this test.
366 call void asm sideeffect
368 call void asm sideeffect
373 ; Requires expanding of required skip branch.
375 ; GCN-LABEL: {{^}}uniform_inside_divergent:
376 ; GCN: v_cmp_gt_u32_e32 vcc, 16, v{{[0-9]+}}
377 ; GCN-NEXT: s_and_saveexec_b64 [[MASK:s\[[0-9]+:[0-9]+\]]], vcc
378 ; GCN-NEXT: s_xor_b64 [[MASK1:s\[[0-9]+:[0-9]+\]]], exec, [[MASK]]
379 ; GCN-NEXT: ; mask branch [[ENDIF:BB[0-9]+_[0-9]+]]
380 ; GCN-NEXT: s_cbranch_execnz [[IF:BB[0-9]+_[0-9]+]]
382 ; GCN-NEXT: [[LONGBB:BB[0-9]+_[0-9]+]]: ; %entry
383 ; GCN-NEXT: s_getpc_b64 vcc
384 ; GCN-NEXT: s_add_u32 vcc_lo, vcc_lo, [[BB2:BB[0-9]_[0-9]+]]-([[LONGBB]]+4)
385 ; GCN-NEXT: s_addc_u32 vcc_hi, vcc_hi, 0{{$}}
386 ; GCN-NEXT: s_setpc_b64 vcc
388 ; GCN-NEXT: [[IF]]: ; %if
389 ; GCN: buffer_store_dword
391 ; GCN: s_cbranch_scc1 [[ENDIF]]
393 ; GCN-NEXT: ; BB#2: ; %if_uniform
394 ; GCN: buffer_store_dword
396 ; GCN-NEXT: [[ENDIF]]: ; %endif
397 ; GCN-NEXT: s_or_b64 exec, exec, [[MASK]]
398 ; GCN-NEXT: s_sleep 5
400 define amdgpu_kernel void @uniform_inside_divergent(i32 addrspace(1)* %out, i32 %cond) #0 {
402 %tid = call i32 @llvm.amdgcn.workitem.id.x()
403 %d_cmp = icmp ult i32 %tid, 16
404 br i1 %d_cmp, label %if, label %endif
407 store i32 0, i32 addrspace(1)* %out
408 %u_cmp = icmp eq i32 %cond, 0
409 br i1 %u_cmp, label %if_uniform, label %endif
412 store i32 1, i32 addrspace(1)* %out
416 ; layout can remove the split branch if it can copy the return block.
417 ; This call makes the return block long enough that it doesn't get copied.
418 call void @llvm.amdgcn.s.sleep(i32 5);
426 ; GCN-LABEL: {{^}}analyze_mask_branch:
427 ; GCN: v_cmp_lt_f32_e32 vcc
428 ; GCN-NEXT: s_and_saveexec_b64 [[MASK:s\[[0-9]+:[0-9]+\]]], vcc
429 ; GCN-NEXT: s_xor_b64 [[MASK]], exec, [[MASK]]
430 ; GCN-NEXT: ; mask branch [[RET:BB[0-9]+_[0-9]+]]
431 ; GCN-NEXT: s_cbranch_execz [[BRANCH_SKIP:BB[0-9]+_[0-9]+]]
432 ; GCN-NEXT: s_branch [[LOOP_BODY:BB[0-9]+_[0-9]+]]
434 ; GCN-NEXT: [[BRANCH_SKIP]]: ; %entry
435 ; GCN-NEXT: s_getpc_b64 vcc
436 ; GCN-NEXT: s_add_u32 vcc_lo, vcc_lo, [[RET]]-([[BRANCH_SKIP]]+4)
437 ; GCN-NEXT: s_addc_u32 vcc_hi, vcc_hi, 0
438 ; GCN-NEXT: s_setpc_b64 vcc
440 ; GCN-NEXT: [[LOOP_BODY]]: ; %loop_body
441 ; GCN: s_mov_b64 vcc, -1{{$}}
450 ; GCN-NEXT: s_cbranch_vccz [[RET]]
452 ; GCN-NEXT: [[LONGBB:BB[0-9]+_[0-9]+]]: ; %loop_body
453 ; GCN-NEXT: ; in Loop: Header=[[LOOP_BODY]] Depth=1
454 ; GCN-NEXT: s_getpc_b64 vcc
455 ; GCN-NEXT: s_sub_u32 vcc_lo, vcc_lo, ([[LONGBB]]+4)-[[LOOP_BODY]]
456 ; GCN-NEXT: s_subb_u32 vcc_hi, vcc_hi, 0
457 ; GCN-NEXT: s_setpc_b64 vcc
459 ; GCN-NEXT: [[RET]]: ; %Flow
460 ; GCN-NEXT: s_or_b64 exec, exec, [[MASK]]
461 ; GCN: buffer_store_dword
463 define amdgpu_kernel void @analyze_mask_branch() #0 {
465 %reg = call float asm sideeffect "v_mov_b32_e64 $0, 0", "=v"()
466 %cmp0 = fcmp ogt float %reg, 0.000000e+00
467 br i1 %cmp0, label %loop, label %ret
470 %phi = phi float [ 0.000000e+00, %loop_body ], [ 1.000000e+00, %entry ]
471 call void asm sideeffect
474 %cmp1 = fcmp olt float %phi, 8.0
475 br i1 %cmp1, label %loop_body, label %ret
478 call void asm sideeffect
486 store volatile i32 7, i32 addrspace(1)* undef
490 ; GCN-LABEL: {{^}}long_branch_hang:
491 ; GCN: s_cmp_lt_i32 s{{[0-9]+}}, 6
492 ; GCN-NEXT: s_cbranch_scc1 {{BB[0-9]+_[0-9]+}}
493 ; GCN-NEXT: s_branch [[LONG_BR_0:BB[0-9]+_[0-9]+]]
494 ; GCN-NEXT: BB{{[0-9]+_[0-9]+}}:
496 ; GCN: s_add_u32 vcc_lo, vcc_lo, [[LONG_BR_DEST0:BB[0-9]+_[0-9]+]]-(
499 ; GCN-NEXT: [[LONG_BR_0]]:
500 ; GCN-DAG: v_cmp_lt_i32
501 ; GCN-DAG: v_cmp_gt_i32
502 ; GCN: s_cbranch_vccnz
507 ; GCN: [[LONG_BR_DEST0]]
508 ; GCN: v_cmp_ne_u32_e32
509 ; GCN-NEXT: s_cbranch_vccz
513 define amdgpu_kernel void @long_branch_hang(i32 addrspace(1)* nocapture %arg, i32 %arg1, i32 %arg2, i32 %arg3, i32 %arg4, i64 %arg5) #0 {
515 %tmp = icmp slt i32 %arg2, 9
516 %tmp6 = icmp eq i32 %arg1, 0
517 %tmp7 = icmp sgt i32 %arg4, 0
518 %tmp8 = icmp sgt i32 %arg4, 5
519 br i1 %tmp8, label %bb9, label %bb13
522 %tmp10 = and i1 %tmp7, %tmp
523 %tmp11 = icmp slt i32 %arg3, %arg4
524 %tmp12 = or i1 %tmp11, %tmp7
525 br i1 %tmp12, label %bb19, label %bb14
528 br i1 %tmp6, label %bb19, label %bb14
530 bb14: ; preds = %bb13, %bb9
531 %tmp15 = icmp slt i32 %arg3, %arg4
532 %tmp16 = or i1 %tmp15, %tmp
533 %tmp17 = and i1 %tmp6, %tmp16
534 %tmp18 = zext i1 %tmp17 to i32
537 bb19: ; preds = %bb14, %bb13, %bb9
538 %tmp20 = phi i32 [ undef, %bb9 ], [ undef, %bb13 ], [ %tmp18, %bb14 ]
539 %tmp21 = getelementptr inbounds i32, i32 addrspace(1)* %arg, i64 %arg5
540 store i32 %tmp20, i32 addrspace(1)* %tmp21, align 4
544 attributes #0 = { nounwind }
545 attributes #1 = { nounwind readnone }