1 ; RUN: llc -march=amdgcn -mcpu=tahiti -verify-machineinstrs -amdgpu-s-branch-bits=4 -simplifycfg-require-and-preserve-domtree=1 < %s | FileCheck -enable-var-scope -check-prefix=GCN %s
4 ; FIXME: We should use llvm-mc for this, but we can't even parse our own output.
6 ; RUN: llc -march=amdgcn -verify-machineinstrs -amdgpu-s-branch-bits=4 -o %t.o -filetype=obj -simplifycfg-require-and-preserve-domtree=1 %s
7 ; RUN: llvm-readobj -r %t.o | FileCheck --check-prefix=OBJ %s
12 ; Restrict maximum branch to between +7 and -8 dwords
14 ; Used to emit an always 4 byte instruction. Inline asm always assumes
15 ; each instruction is the maximum size.
16 declare void @llvm.amdgcn.s.sleep(i32) #0
18 declare i32 @llvm.amdgcn.workitem.id.x() #1
21 ; GCN-LABEL: {{^}}uniform_conditional_max_short_forward_branch:
22 ; GCN: s_load_dword [[CND:s[0-9]+]]
23 ; GCN: s_cmp_eq_u32 [[CND]], 0
24 ; GCN-NEXT: s_cbranch_scc1 [[BB3:BB[0-9]+_[0-9]+]]
27 ; GCN-NEXT: ; %bb.1: ; %bb2
28 ; GCN-NEXT: ;;#ASMSTART
35 ; GCN-NEXT: [[BB3]]: ; %bb3
36 ; GCN: v_mov_b32_e32 [[V_CND:v[0-9]+]], [[CND]]
37 ; GCN: buffer_store_dword [[V_CND]]
39 define amdgpu_kernel void @uniform_conditional_max_short_forward_branch(i32 addrspace(1)* %arg, i32 %cnd) #0 {
41 %cmp = icmp eq i32 %cnd, 0
42 br i1 %cmp, label %bb3, label %bb2 ; +8 dword branch
46 call void asm sideeffect
50 call void @llvm.amdgcn.s.sleep(i32 0)
54 store volatile i32 %cnd, i32 addrspace(1)* %arg
58 ; GCN-LABEL: {{^}}uniform_conditional_min_long_forward_branch:
59 ; GCN: s_load_dword [[CND:s[0-9]+]]
60 ; GCN: s_cmp_eq_u32 [[CND]], 0
61 ; GCN-NEXT: s_cbranch_scc0 [[LONGBB:BB[0-9]+_[0-9]+]]
63 ; GCN-NEXT: {{BB[0-9]+_[0-9]+}}: ; %bb0
64 ; GCN-NEXT: s_getpc_b64 s{{\[}}[[PC_LO:[0-9]+]]:[[PC_HI:[0-9]+]]{{\]}}
65 ; GCN-NEXT: [[POST_GETPC:.Lpost_getpc[0-9]+]]:{{$}}
66 ; GCN-NEXT: s_add_u32 s[[PC_LO]], s[[PC_LO]], ([[ENDBB:BB[0-9]+_[0-9]+]]-[[POST_GETPC]])&4294967295
67 ; GCN-NEXT: s_addc_u32 s[[PC_HI]], s[[PC_HI]], ([[ENDBB]]-[[POST_GETPC]])>>32
68 ; GCN-NEXT: s_setpc_b64 s{{\[}}[[PC_LO]]:[[PC_HI]]{{\]}}
70 ; GCN-NEXT: [[LONGBB]]:
71 ; GCN-NEXT: ;;#ASMSTART
78 ; GCN-NEXT: [[ENDBB]]:
79 ; GCN: v_mov_b32_e32 [[V_CND:v[0-9]+]], [[CND]]
80 ; GCN: buffer_store_dword [[V_CND]]
82 define amdgpu_kernel void @uniform_conditional_min_long_forward_branch(i32 addrspace(1)* %arg, i32 %cnd) #0 {
84 %cmp = icmp eq i32 %cnd, 0
85 br i1 %cmp, label %bb3, label %bb2 ; +9 dword branch
89 call void asm sideeffect
97 store volatile i32 %cnd, i32 addrspace(1)* %arg
101 ; GCN-LABEL: {{^}}uniform_conditional_min_long_forward_vcnd_branch:
102 ; GCN: s_load_dword [[CND:s[0-9]+]]
104 ; GCN-DAG: v_cmp_eq_f32_e64 [[UNMASKED:s\[[0-9]+:[0-9]+\]]], [[CND]], 0
105 ; GCN-DAG: s_and_b64 vcc, exec, [[UNMASKED]]
106 ; GCN: s_cbranch_vccz [[LONGBB:BB[0-9]+_[0-9]+]]
108 ; GCN-NEXT: {{BB[0-9]+_[0-9]+}}: ; %bb0
109 ; GCN-NEXT: s_getpc_b64 s{{\[}}[[PC_LO:[0-9]+]]:[[PC_HI:[0-9]+]]{{\]}}
110 ; GCN-NEXT: [[POST_GETPC:.Lpost_getpc[0-9]+]]:{{$}}
111 ; GCN-NEXT: s_add_u32 s[[PC_LO]], s[[PC_LO]], ([[ENDBB:BB[0-9]+_[0-9]+]]-[[POST_GETPC]])&4294967295
112 ; GCN-NEXT: s_addc_u32 s[[PC_HI]], s[[PC_HI]], ([[ENDBB]]-[[POST_GETPC]])>>32
113 ; GCN-NEXT: s_setpc_b64 s{{\[}}[[PC_LO]]:[[PC_HI]]{{\]}}
115 ; GCN-NEXT: [[LONGBB]]:
122 ; GCN: v_mov_b32_e32 [[V_CND:v[0-9]+]], [[CND]]
123 ; GCN: buffer_store_dword [[V_CND]]
125 define amdgpu_kernel void @uniform_conditional_min_long_forward_vcnd_branch(float addrspace(1)* %arg, float %cnd) #0 {
127 %cmp = fcmp oeq float %cnd, 0.0
128 br i1 %cmp, label %bb3, label %bb2 ; + 8 dword branch
131 call void asm sideeffect " ; 32 bytes
139 store volatile float %cnd, float addrspace(1)* %arg
143 ; GCN-LABEL: {{^}}min_long_forward_vbranch:
145 ; GCN: buffer_load_dword
146 ; GCN: v_cmp_ne_u32_e32 vcc, 0, v{{[0-9]+}}
147 ; GCN: s_and_saveexec_b64 [[SAVE:s\[[0-9]+:[0-9]+\]]], vcc
154 ; GCN: s_or_b64 exec, exec, [[SAVE]]
155 ; GCN: buffer_store_dword
157 define amdgpu_kernel void @min_long_forward_vbranch(i32 addrspace(1)* %arg) #0 {
159 %tid = call i32 @llvm.amdgcn.workitem.id.x()
160 %tid.ext = zext i32 %tid to i64
161 %gep = getelementptr inbounds i32, i32 addrspace(1)* %arg, i64 %tid.ext
162 %load = load volatile i32, i32 addrspace(1)* %gep
163 %cmp = icmp eq i32 %load, 0
164 br i1 %cmp, label %bb3, label %bb2 ; + 8 dword branch
167 call void asm sideeffect " ; 32 bytes
175 store volatile i32 %load, i32 addrspace(1)* %gep
179 ; GCN-LABEL: {{^}}long_backward_sbranch:
180 ; GCN: s_mov_b32 [[LOOPIDX:s[0-9]+]], 0{{$}}
182 ; GCN: [[LOOPBB:BB[0-9]+_[0-9]+]]: ; %bb2
183 ; GCN-NEXT: ; =>This Inner Loop Header: Depth=1
184 ; GCN-NEXT: s_add_i32 [[INC:s[0-9]+]], [[LOOPIDX]], 1
185 ; GCN-NEXT: s_cmp_lt_i32 [[INC]], 10
187 ; GCN-NEXT: ;;#ASMSTART
188 ; GCN-NEXT: v_nop_e64
189 ; GCN-NEXT: v_nop_e64
190 ; GCN-NEXT: v_nop_e64
191 ; GCN-NEXT: ;;#ASMEND
193 ; GCN-NEXT: s_cbranch_scc0 [[ENDBB:BB[0-9]+_[0-9]+]]
195 ; GCN-NEXT: {{BB[0-9]+_[0-9]+}}: ; %bb2
196 ; GCN-NEXT: ; in Loop: Header=[[LOOPBB]] Depth=1
198 ; GCN-NEXT: s_getpc_b64 s{{\[}}[[PC_LO:[0-9]+]]:[[PC_HI:[0-9]+]]{{\]}}
199 ; GCN-NEXT: [[POST_GETPC:.Lpost_getpc[0-9]+]]:{{$}}
200 ; GCN-NEXT: s_add_u32 s[[PC_LO]], s[[PC_LO]], ([[LOOPBB]]-[[POST_GETPC]])&4294967295
201 ; GCN-NEXT: s_addc_u32 s[[PC_HI]], s[[PC_HI]], ([[LOOPBB]]-[[POST_GETPC]])>>32
202 ; GCN-NEXT: s_setpc_b64 s{{\[}}[[PC_LO]]:[[PC_HI]]{{\]}}
204 ; GCN-NEXT: [[ENDBB]]:
206 define amdgpu_kernel void @long_backward_sbranch(i32 addrspace(1)* %arg) #0 {
211 %loop.idx = phi i32 [ 0, %bb ], [ %inc, %bb2 ]
213 call void asm sideeffect
217 %inc = add nsw i32 %loop.idx, 1 ; add cost 4
218 %cmp = icmp slt i32 %inc, 10 ; condition cost = 8
219 br i1 %cmp, label %bb2, label %bb3 ; -
225 ; Requires expansion of unconditional branch from %bb2 to %bb4 (and
226 ; expansion of conditional branch from %bb to %bb3.
228 ; GCN-LABEL: {{^}}uniform_unconditional_min_long_forward_branch:
230 ; GCN: s_cbranch_scc{{[0-1]}} [[BB2:BB[0-9]+_[0-9]+]]
232 ; GCN-NEXT: {{BB[0-9]+_[0-9]+}}: ; %bb0
233 ; GCN-NEXT: s_getpc_b64 s{{\[}}[[PC0_LO:[0-9]+]]:[[PC0_HI:[0-9]+]]{{\]}}
234 ; GCN-NEXT: [[POST_GETPC:.Lpost_getpc[0-9]+]]:{{$}}
235 ; GCN-NEXT: s_add_u32 s[[PC0_LO]], s[[PC0_LO]], ([[BB3:BB[0-9]_[0-9]+]]-[[POST_GETPC]])&4294967295
236 ; GCN-NEXT: s_addc_u32 s[[PC0_HI]], s[[PC0_HI]], ([[BB3:BB[0-9]_[0-9]+]]-[[POST_GETPC]])>>32
237 ; GCN-NEXT: s_setpc_b64 s{{\[}}[[PC0_LO]]:[[PC0_HI]]{{\]}}
239 ; GCN: [[BB2]]: ; %bb3
247 ; GCN: v_mov_b32_e32 [[BB2_K:v[0-9]+]], 17
248 ; GCN: buffer_store_dword [[BB2_K]]
250 ; GCN: v_mov_b32_e32 [[BB4_K:v[0-9]+]], 63
251 ; GCN: buffer_store_dword [[BB4_K]]
253 ; GCN-NEXT: .Lfunc_end{{[0-9]+}}:
254 define amdgpu_kernel void @uniform_unconditional_min_long_forward_branch(i32 addrspace(1)* %arg, i32 %arg1) {
256 %tmp = icmp ne i32 %arg1, 0
257 br i1 %tmp, label %bb2, label %bb3
260 store volatile i32 17, i32 addrspace(1)* undef
265 call void asm sideeffect
273 store volatile i32 63, i32 addrspace(1)* %arg
277 ; GCN-LABEL: {{^}}uniform_unconditional_min_long_backward_branch:
278 ; GCN-NEXT: ; %bb.0: ; %entry
280 ; GCN-NEXT: [[LOOP:BB[0-9]_[0-9]+]]: ; %loop
281 ; GCN-NEXT: ; =>This Inner Loop Header: Depth=1
282 ; GCN-NEXT: ;;#ASMSTART
283 ; GCN-NEXT: v_nop_e64
284 ; GCN-NEXT: v_nop_e64
285 ; GCN-NEXT: v_nop_e64
286 ; GCN-NEXT: v_nop_e64
287 ; GCN-NEXT: ;;#ASMEND
289 ; GCN-NEXT: {{BB[0-9]+_[0-9]+}}: ; %loop
290 ; GCN-NEXT: ; in Loop: Header=[[LOOP]] Depth=1
292 ; GCN-NEXT: s_getpc_b64 s{{\[}}[[PC_LO:[0-9]+]]:[[PC_HI:[0-9]+]]{{\]}}
293 ; GCN-NEXT: [[POST_GETPC:.Lpost_getpc[0-9]+]]:{{$}}
294 ; GCN-NEXT: s_add_u32 s[[PC_LO]], s[[PC_LO]], ([[LOOP]]-[[POST_GETPC]])&4294967295
295 ; GCN-NEXT: s_addc_u32 s[[PC_HI]], s[[PC_HI]], ([[LOOP]]-[[POST_GETPC]])>>32
296 ; GCN-NEXT: s_setpc_b64 s{{\[}}[[PC_LO]]:[[PC_HI]]{{\]}}
297 ; GCN-NEXT: .Lfunc_end{{[0-9]+}}:
298 define amdgpu_kernel void @uniform_unconditional_min_long_backward_branch(i32 addrspace(1)* %arg, i32 %arg1) {
304 call void asm sideeffect
312 ; Expansion of branch from %bb1 to %bb3 introduces need to expand
313 ; branch from %bb0 to %bb2
315 ; GCN-LABEL: {{^}}expand_requires_expand:
316 ; GCN-NEXT: ; %bb.0: ; %bb0
318 ; GCN: {{s|v}}_cmp_lt_i32
322 ; GCN-NEXT: s_waitcnt lgkmcnt(0)
323 ; GCN-NEXT: v_cmp_{{eq|ne}}_u32_e64
324 ; GCN: s_cbranch_vccz [[BB2:BB[0-9]_[0-9]+]]
326 ; GCN-NEXT: {{BB[0-9]+_[0-9]+}}:
327 ; GCN-NEXT: s_getpc_b64 s{{\[}}[[PC1_LO:[0-9]+]]:[[PC1_HI:[0-9]+]]{{\]}}
328 ; GCN-NEXT: [[POST_GETPC:.Lpost_getpc[0-9]+]]:{{$}}
329 ; GCN-NEXT: s_add_u32 s[[PC1_LO]], s[[PC1_LO]], ([[BB3:BB[0-9]+_[0-9]+]]-[[POST_GETPC]])&4294967295
330 ; GCN-NEXT: s_addc_u32 s[[PC1_HI]], s[[PC1_HI]], ([[BB3:BB[0-9]+_[0-9]+]]-[[POST_GETPC]])>>32
331 ; GCN-NEXT: s_setpc_b64 s{{\[}}[[PC1_LO]]:[[PC1_HI]]{{\]}}
333 ; GCN-NEXT: [[BB2]]: ; %bb2
334 ; GCN-NEXT: ;;#ASMSTART
335 ; GCN-NEXT: v_nop_e64
336 ; GCN-NEXT: v_nop_e64
337 ; GCN-NEXT: v_nop_e64
338 ; GCN-NEXT: v_nop_e64
339 ; GCN-NEXT: ;;#ASMEND
341 ; GCN-NEXT: [[BB3]]: ; %bb3
342 ; GCN-NEXT: ;;#ASMSTART
343 ; GCN-NEXT: v_nop_e64
344 ; GCN-NEXT: ;;#ASMEND
345 ; GCN-NEXT: ;;#ASMSTART
346 ; GCN-NEXT: v_nop_e64
347 ; GCN-NEXT: ;;#ASMEND
349 define amdgpu_kernel void @expand_requires_expand(i32 %cond0) #0 {
351 %tmp = tail call i32 @llvm.amdgcn.workitem.id.x() #0
352 %cmp0 = icmp slt i32 %cond0, 0
353 br i1 %cmp0, label %bb2, label %bb1
356 %val = load volatile i32, i32 addrspace(4)* undef
357 %cmp1 = icmp eq i32 %val, 3
358 br i1 %cmp1, label %bb3, label %bb2
361 call void asm sideeffect
369 ; These NOPs prevent tail-duplication-based outlining
370 ; from firing, which defeats the need to expand the branches and this test.
371 call void asm sideeffect
373 call void asm sideeffect
378 ; Requires expanding of required skip branch.
380 ; GCN-LABEL: {{^}}uniform_inside_divergent:
381 ; GCN: v_cmp_gt_u32_e32 vcc, 16, v{{[0-9]+}}
382 ; GCN-NEXT: s_and_saveexec_b64 [[MASK:s\[[0-9]+:[0-9]+\]]], vcc
383 ; GCN-NEXT: s_cbranch_execnz [[IF:BB[0-9]+_[0-9]+]]
385 ; GCN-NEXT: {{BB[0-9]+_[0-9]+}}: ; %entry
386 ; GCN-NEXT: s_getpc_b64 s{{\[}}[[PC_LO:[0-9]+]]:[[PC_HI:[0-9]+]]{{\]}}
387 ; GCN-NEXT: [[POST_GETPC:.Lpost_getpc[0-9]+]]:{{$}}
388 ; GCN-NEXT: s_add_u32 s[[PC_LO]], s[[PC_LO]], ([[BB2:BB[0-9]_[0-9]+]]-[[POST_GETPC]])&4294967295
389 ; GCN-NEXT: s_addc_u32 s[[PC_HI]], s[[PC_HI]], ([[BB2:BB[0-9]_[0-9]+]]-[[POST_GETPC]])>>32
390 ; GCN-NEXT: s_setpc_b64 s{{\[}}[[PC_LO]]:[[PC_HI]]{{\]}}
392 ; GCN-NEXT: [[IF]]: ; %if
393 ; GCN: buffer_store_dword
395 ; GCN: s_cbranch_scc1 [[ENDIF:BB[0-9]+_[0-9]+]]
397 ; GCN-NEXT: ; %bb.2: ; %if_uniform
398 ; GCN: buffer_store_dword
400 ; GCN-NEXT: [[ENDIF]]: ; %endif
401 ; GCN-NEXT: s_or_b64 exec, exec, [[MASK]]
402 ; GCN-NEXT: s_sleep 5
404 define amdgpu_kernel void @uniform_inside_divergent(i32 addrspace(1)* %out, i32 %cond) #0 {
406 %tid = call i32 @llvm.amdgcn.workitem.id.x()
407 %d_cmp = icmp ult i32 %tid, 16
408 br i1 %d_cmp, label %if, label %endif
411 store i32 0, i32 addrspace(1)* %out
412 %u_cmp = icmp eq i32 %cond, 0
413 br i1 %u_cmp, label %if_uniform, label %endif
416 store i32 1, i32 addrspace(1)* %out
420 ; layout can remove the split branch if it can copy the return block.
421 ; This call makes the return block long enough that it doesn't get copied.
422 call void @llvm.amdgcn.s.sleep(i32 5);
428 ; GCN-LABEL: {{^}}analyze_mask_branch:
429 ; GCN: v_cmp_nlt_f32_e32 vcc
430 ; GCN-NEXT: s_and_saveexec_b64 [[TEMP_MASK:s\[[0-9]+:[0-9]+\]]], vcc
431 ; GCN-NEXT: s_xor_b64 [[MASK:s\[[0-9]+:[0-9]+\]]], exec, [[TEMP_MASK]]
433 ; GCN: BB{{[0-9]+_[0-9]+}}: ; %Flow
434 ; GCN-NEXT: s_or_saveexec_b64 [[TEMP_MASK1:s\[[0-9]+:[0-9]+\]]], [[MASK]]
435 ; GCN-NEXT: s_xor_b64 exec, exec, [[TEMP_MASK1]]
437 ; GCN: [[LOOP_BODY:BB[0-9]+_[0-9]+]]: ; %loop{{$}}
446 ; GCN: s_cbranch_{{vccz|vccnz}} [[RET:BB[0-9]+_[0-9]+]]
448 ; GCN-NEXT: {{BB[0-9]+_[0-9]+}}: ; %loop
449 ; GCN-NEXT: ; in Loop: Header=[[LOOP_BODY]] Depth=1
450 ; GCN-NEXT: s_getpc_b64 s{{\[}}[[PC_LO:[0-9]+]]:[[PC_HI:[0-9]+]]{{\]}}
451 ; GCN-NEXT: [[POST_GETPC:.Lpost_getpc[0-9]+]]:{{$}}
452 ; GCN-NEXT: s_add_u32 s[[PC_LO]], s[[PC_LO]], ([[LOOP_BODY]]-[[POST_GETPC]])&4294967295
453 ; GCN-NEXT: s_addc_u32 s[[PC_HI]], s[[PC_HI]], ([[LOOP_BODY]]-[[POST_GETPC]])>>32
454 ; GCN-NEXT: s_setpc_b64 s{{\[}}[[PC_LO]]:[[PC_HI]]{{\]}}
456 ; GCN-NEXT: [[RET]]: ; %UnifiedReturnBlock
458 define amdgpu_kernel void @analyze_mask_branch() #0 {
460 %reg = call float asm sideeffect "v_mov_b32_e64 $0, 0", "=v"()
461 %cmp0 = fcmp ogt float %reg, 0.000000e+00
462 br i1 %cmp0, label %loop, label %ret
465 %phi = phi float [ 0.000000e+00, %loop_body ], [ 1.000000e+00, %entry ]
466 call void asm sideeffect
469 %cmp1 = fcmp olt float %phi, 8.0
470 br i1 %cmp1, label %loop_body, label %ret
473 call void asm sideeffect
481 store volatile i32 7, i32 addrspace(1)* undef
485 ; GCN-LABEL: {{^}}long_branch_hang:
486 ; GCN: s_cmp_lt_i32 s{{[0-9]+}}, 6
487 ; GCN: s_cbranch_scc{{[0-1]}} [[LONG_BR_0:BB[0-9]+_[0-9]+]]
488 ; GCN-NEXT: BB{{[0-9]+_[0-9]+}}:
490 ; GCN: s_getpc_b64 s{{\[}}[[PC_LO:[0-9]+]]:[[PC_HI:[0-9]+]]{{\]}}
491 ; GCN-NEXT: [[POST_GETPC:.Lpost_getpc[0-9]+]]:{{$}}
492 ; GCN-NEXT: s_add_u32 s[[PC_LO]], s[[PC_LO]], ([[LONG_BR_DEST0:BB[0-9]+_[0-9]+]]-[[POST_GETPC]])&4294967295
493 ; GCN-NEXT: s_addc_u32 s[[PC_HI]], s[[PC_HI]], ([[LONG_BR_DEST0]]-[[POST_GETPC]])>>32
494 ; GCN-NEXT: s_setpc_b64 s{{\[}}[[PC_LO]]:[[PC_HI]]{{\]}}
495 ; GCN-NEXT: [[LONG_BR_0]]:
497 ; GCN: [[LONG_BR_DEST0]]:
499 ; GCN-DAG: v_cmp_lt_i32
500 ; GCN-DAG: v_cmp_ge_i32
502 ; GCN: s_cbranch_vccz
506 define amdgpu_kernel void @long_branch_hang(i32 addrspace(1)* nocapture %arg, i32 %arg1, i32 %arg2, i32 %arg3, i32 %arg4, i64 %arg5) #0 {
508 %tmp = icmp slt i32 %arg2, 9
509 %tmp6 = icmp eq i32 %arg1, 0
510 %tmp7 = icmp sgt i32 %arg4, 0
511 %tmp8 = icmp sgt i32 %arg4, 5
512 br i1 %tmp8, label %bb9, label %bb13
515 %tmp10 = and i1 %tmp7, %tmp
516 %tmp11 = icmp slt i32 %arg3, %arg4
517 %tmp12 = or i1 %tmp11, %tmp7
518 br i1 %tmp12, label %bb19, label %bb14
521 call void asm sideeffect
526 br i1 %tmp6, label %bb19, label %bb14
528 bb14: ; preds = %bb13, %bb9
529 %tmp15 = icmp slt i32 %arg3, %arg4
530 %tmp16 = or i1 %tmp15, %tmp
531 %tmp17 = and i1 %tmp6, %tmp16
532 %tmp18 = zext i1 %tmp17 to i32
535 bb19: ; preds = %bb14, %bb13, %bb9
536 %tmp20 = phi i32 [ undef, %bb9 ], [ undef, %bb13 ], [ %tmp18, %bb14 ]
537 %tmp21 = getelementptr inbounds i32, i32 addrspace(1)* %arg, i64 %arg5
538 store i32 %tmp20, i32 addrspace(1)* %tmp21, align 4
542 attributes #0 = { nounwind }
543 attributes #1 = { nounwind readnone }