xref: /llvm-project/llvm/test/CodeGen/AMDGPU/GlobalISel/divergence-structurizer.ll (revision 6c2eec5ceadf26ce8d732d718a8906d075a7d6c7)

; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 3
; RUN: llc -global-isel -mtriple=amdgcn-amd-amdpal -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck -check-prefix=GFX10 %s

; Simples case, if - then, that requires lane mask merging,
; %phi lane mask will hold %val_A at %A. Lanes that are active in %B
; will overwrite its own lane bit in lane mask with val_B
define amdgpu_ps void @divergent_i1_phi_if_then(ptr addrspace(1) %out, i32 %tid, i32 %cond) {
; GFX10-LABEL: divergent_i1_phi_if_then:
; GFX10:       ; %bb.0: ; %A
; GFX10-NEXT:    v_cmp_le_u32_e64 s0, 6, v2
; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc_lo, 0, v3
; GFX10-NEXT:    s_and_saveexec_b32 s1, vcc_lo
; GFX10-NEXT:  ; %bb.1: ; %B
; GFX10-NEXT:    v_cmp_gt_u32_e32 vcc_lo, 1, v2
; GFX10-NEXT:    s_andn2_b32 s0, s0, exec_lo
; GFX10-NEXT:    s_and_b32 s2, exec_lo, vcc_lo
; GFX10-NEXT:    s_or_b32 s0, s0, s2
; GFX10-NEXT:  ; %bb.2: ; %exit
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s1
; GFX10-NEXT:    v_cndmask_b32_e64 v2, 0, -1, s0
; GFX10-NEXT:    v_add_nc_u32_e32 v2, 2, v2
; GFX10-NEXT:    global_store_dword v[0:1], v2, off
; GFX10-NEXT:    s_endpgm
A:
  %val_A = icmp uge i32 %tid, 6
  %cmp = icmp eq i32 %cond, 0
  br i1 %cmp, label %B, label %exit

B:
  %val_B = icmp ult i32 %tid, 1
  br label %exit

exit:
  %phi = phi i1 [ %val_A, %A ], [ %val_B, %B ]
  %sel = select i1 %phi, i32 1, i32 2
  store i32 %sel, ptr addrspace(1) %out
  ret void
}

; if - else
define amdgpu_ps void @divergent_i1_phi_if_else(ptr addrspace(1) %out, i32 %tid, i32 %cond) {
; GFX10-LABEL: divergent_i1_phi_if_else:
; GFX10:       ; %bb.0: ; %entry
; GFX10-NEXT:    s_and_b32 s0, 1, s0
; GFX10-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v3
; GFX10-NEXT:    v_cmp_ne_u32_e64 s0, 0, s0
; GFX10-NEXT:    s_and_saveexec_b32 s1, vcc_lo
; GFX10-NEXT:    s_xor_b32 s1, exec_lo, s1
; GFX10-NEXT:  ; %bb.1: ; %B
; GFX10-NEXT:    v_cmp_gt_u32_e32 vcc_lo, 2, v2
; GFX10-NEXT:    s_andn2_b32 s0, s0, exec_lo
; GFX10-NEXT:    ; implicit-def: $vgpr2
; GFX10-NEXT:    s_and_b32 s2, exec_lo, vcc_lo
; GFX10-NEXT:    s_or_b32 s0, s0, s2
; GFX10-NEXT:  ; %bb.2: ; %Flow
; GFX10-NEXT:    s_andn2_saveexec_b32 s1, s1
; GFX10-NEXT:  ; %bb.3: ; %A
; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, 1, v2
; GFX10-NEXT:    s_andn2_b32 s0, s0, exec_lo
; GFX10-NEXT:    s_and_b32 s2, exec_lo, vcc_lo
; GFX10-NEXT:    s_or_b32 s0, s0, s2
; GFX10-NEXT:  ; %bb.4: ; %exit
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s1
; GFX10-NEXT:    v_cndmask_b32_e64 v2, 0, -1, s0
; GFX10-NEXT:    v_add_nc_u32_e32 v2, 2, v2
; GFX10-NEXT:    global_store_dword v[0:1], v2, off
; GFX10-NEXT:    s_endpgm
entry:
  %cmp = icmp eq i32 %cond, 0
  br i1 %cmp, label %A, label %B

A:
  %val_A = icmp uge i32 %tid, 1
  br label %exit

B:
  %val_B = icmp ult i32 %tid, 2
  br label %exit

exit:
  %phi = phi i1 [ %val_A, %A ], [ %val_B, %B ]
  %sel = select i1 %phi, i32 1, i32 2
  store i32 %sel, ptr addrspace(1) %out
  ret void
}

; if - break;

;  counter = 0;
;  do {
;    if (a[counter] == 0)
;      break;
;    if (b[counter] == 0)
;      break;
;    if (c[counter] == 0)
;      break;
;    x[counter++]+=1;
;  } while (counter<100);

; Tests with multiple break conditions. Divergent phis will be used to track
; if any of the break conditions was reached. We only need to do simple lane
; mask merging (for current loop iteration only). There is an intrinsic,
; if_break, that will merge lane masks across all iterations of the loop.

define amdgpu_cs void @loop_with_1break(ptr addrspace(1) %x, ptr addrspace(1) %a) {
; GFX10-LABEL: loop_with_1break:
; GFX10:       ; %bb.0: ; %entry
; GFX10-NEXT:    s_mov_b32 s0, 0
; GFX10-NEXT:    ; implicit-def: $sgpr1
; GFX10-NEXT:    v_mov_b32_e32 v4, s0
; GFX10-NEXT:    s_branch .LBB2_2
; GFX10-NEXT:  .LBB2_1: ; %Flow
; GFX10-NEXT:    ; in Loop: Header=BB2_2 Depth=1
; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s2
; GFX10-NEXT:    s_and_b32 s2, exec_lo, s1
; GFX10-NEXT:    s_or_b32 s0, s2, s0
; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s0
; GFX10-NEXT:    s_cbranch_execz .LBB2_4
; GFX10-NEXT:  .LBB2_2: ; %A
; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT:    v_ashrrev_i32_e32 v5, 31, v4
; GFX10-NEXT:    s_andn2_b32 s1, s1, exec_lo
; GFX10-NEXT:    s_and_b32 s2, exec_lo, -1
; GFX10-NEXT:    s_or_b32 s1, s1, s2
; GFX10-NEXT:    v_lshlrev_b64 v[5:6], 2, v[4:5]
; GFX10-NEXT:    v_add_co_u32 v7, vcc_lo, v2, v5
; GFX10-NEXT:    v_add_co_ci_u32_e32 v8, vcc_lo, v3, v6, vcc_lo
; GFX10-NEXT:    global_load_dword v7, v[7:8], off
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v7
; GFX10-NEXT:    s_and_saveexec_b32 s2, vcc_lo
; GFX10-NEXT:    s_cbranch_execz .LBB2_1
; GFX10-NEXT:  ; %bb.3: ; %loop.body
; GFX10-NEXT:    ; in Loop: Header=BB2_2 Depth=1
; GFX10-NEXT:    v_add_co_u32 v5, vcc_lo, v0, v5
; GFX10-NEXT:    v_add_co_ci_u32_e32 v6, vcc_lo, v1, v6, vcc_lo
; GFX10-NEXT:    v_add_nc_u32_e32 v8, 1, v4
; GFX10-NEXT:    v_cmp_gt_u32_e32 vcc_lo, 0x64, v4
; GFX10-NEXT:    s_andn2_b32 s1, s1, exec_lo
; GFX10-NEXT:    global_load_dword v7, v[5:6], off
; GFX10-NEXT:    v_mov_b32_e32 v4, v8
; GFX10-NEXT:    s_and_b32 s3, exec_lo, vcc_lo
; GFX10-NEXT:    s_or_b32 s1, s1, s3
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    v_add_nc_u32_e32 v7, 1, v7
; GFX10-NEXT:    global_store_dword v[5:6], v7, off
; GFX10-NEXT:    s_branch .LBB2_1
; GFX10-NEXT:  .LBB2_4: ; %exit
; GFX10-NEXT:    s_endpgm
entry:
  br label %A

A:
  %counter = phi i32 [ %counter.plus.1, %loop.body ], [ 0, %entry ]
  %a.plus.counter = getelementptr inbounds i32, ptr addrspace(1) %a, i32 %counter
  %a.val = load i32, ptr addrspace(1) %a.plus.counter
  %a.cond = icmp eq i32 %a.val, 0
  br i1 %a.cond, label %exit, label %loop.body

loop.body:
  %x.plus.counter = getelementptr inbounds i32, ptr addrspace(1) %x, i32 %counter
  %x.val = load i32, ptr addrspace(1) %x.plus.counter
  %x.val.plus.1 = add i32 %x.val, 1
  store i32 %x.val.plus.1, ptr addrspace(1) %x.plus.counter
  %counter.plus.1 = add i32 %counter, 1
  %x.cond = icmp ult i32 %counter, 100
  br i1 %x.cond, label %exit, label %A

exit:
  ret void
}

define amdgpu_cs void @loop_with_2breaks(ptr addrspace(1) %x, ptr addrspace(1) %a, ptr addrspace(1) %b) {
; GFX10-LABEL: loop_with_2breaks:
; GFX10:       ; %bb.0: ; %entry
; GFX10-NEXT:    s_mov_b32 s0, 0
; GFX10-NEXT:    ; implicit-def: $sgpr1
; GFX10-NEXT:    v_mov_b32_e32 v6, s0
; GFX10-NEXT:    s_branch .LBB3_3
; GFX10-NEXT:  .LBB3_1: ; %Flow3
; GFX10-NEXT:    ; in Loop: Header=BB3_3 Depth=1
; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s3
; GFX10-NEXT:    s_andn2_b32 s1, s1, exec_lo
; GFX10-NEXT:    s_and_b32 s3, exec_lo, s4
; GFX10-NEXT:    s_or_b32 s1, s1, s3
; GFX10-NEXT:  .LBB3_2: ; %Flow
; GFX10-NEXT:    ; in Loop: Header=BB3_3 Depth=1
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s2
; GFX10-NEXT:    s_and_b32 s2, exec_lo, s1
; GFX10-NEXT:    s_or_b32 s0, s2, s0
; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s0
; GFX10-NEXT:    s_cbranch_execz .LBB3_6
; GFX10-NEXT:  .LBB3_3: ; %A
; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT:    v_ashrrev_i32_e32 v7, 31, v6
; GFX10-NEXT:    s_andn2_b32 s1, s1, exec_lo
; GFX10-NEXT:    s_and_b32 s2, exec_lo, -1
; GFX10-NEXT:    s_or_b32 s1, s1, s2
; GFX10-NEXT:    v_lshlrev_b64 v[7:8], 2, v[6:7]
; GFX10-NEXT:    v_add_co_u32 v9, vcc_lo, v2, v7
; GFX10-NEXT:    v_add_co_ci_u32_e32 v10, vcc_lo, v3, v8, vcc_lo
; GFX10-NEXT:    global_load_dword v9, v[9:10], off
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v9
; GFX10-NEXT:    s_and_saveexec_b32 s2, vcc_lo
; GFX10-NEXT:    s_cbranch_execz .LBB3_2
; GFX10-NEXT:  ; %bb.4: ; %B
; GFX10-NEXT:    ; in Loop: Header=BB3_3 Depth=1
; GFX10-NEXT:    v_add_co_u32 v9, vcc_lo, v4, v7
; GFX10-NEXT:    v_add_co_ci_u32_e32 v10, vcc_lo, v5, v8, vcc_lo
; GFX10-NEXT:    s_mov_b32 s4, -1
; GFX10-NEXT:    global_load_dword v9, v[9:10], off
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v9
; GFX10-NEXT:    s_and_saveexec_b32 s3, vcc_lo
; GFX10-NEXT:    s_cbranch_execz .LBB3_1
; GFX10-NEXT:  ; %bb.5: ; %loop.body
; GFX10-NEXT:    ; in Loop: Header=BB3_3 Depth=1
; GFX10-NEXT:    v_add_co_u32 v7, vcc_lo, v0, v7
; GFX10-NEXT:    v_add_co_ci_u32_e32 v8, vcc_lo, v1, v8, vcc_lo
; GFX10-NEXT:    v_add_nc_u32_e32 v10, 1, v6
; GFX10-NEXT:    v_cmp_gt_u32_e32 vcc_lo, 0x64, v6
; GFX10-NEXT:    s_andn2_b32 s4, -1, exec_lo
; GFX10-NEXT:    global_load_dword v9, v[7:8], off
; GFX10-NEXT:    v_mov_b32_e32 v6, v10
; GFX10-NEXT:    s_and_b32 s5, exec_lo, vcc_lo
; GFX10-NEXT:    s_or_b32 s4, s4, s5
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    v_add_nc_u32_e32 v9, 1, v9
; GFX10-NEXT:    global_store_dword v[7:8], v9, off
; GFX10-NEXT:    s_branch .LBB3_1
; GFX10-NEXT:  .LBB3_6: ; %exit
; GFX10-NEXT:    s_endpgm
entry:
  br label %A

A:
  %counter = phi i32 [ %counter.plus.1, %loop.body ], [ 0, %entry ]
  %a.plus.counter = getelementptr inbounds i32, ptr addrspace(1) %a, i32 %counter
  %a.val = load i32, ptr addrspace(1) %a.plus.counter
  %a.cond = icmp eq i32 %a.val, 0
  br i1 %a.cond, label %exit, label %B

B:
  %b.plus.counter = getelementptr inbounds i32, ptr addrspace(1) %b, i32 %counter
  %b.val = load i32, ptr addrspace(1) %b.plus.counter
  %b.cond = icmp eq i32 %b.val, 0
  br i1 %b.cond, label %exit, label %loop.body

loop.body:
  %x.plus.counter = getelementptr inbounds i32, ptr addrspace(1) %x, i32 %counter
  %x.val = load i32, ptr addrspace(1) %x.plus.counter
  %x.val.plus.1 = add i32 %x.val, 1
  store i32 %x.val.plus.1, ptr addrspace(1) %x.plus.counter
  %counter.plus.1 = add i32 %counter, 1
  %x.cond = icmp ult i32 %counter, 100
  br i1 %x.cond, label %exit, label %A

exit:
  ret void
}

define amdgpu_cs void @loop_with_3breaks(ptr addrspace(1) %x, ptr addrspace(1) %a, ptr addrspace(1) %b, ptr addrspace(1) %c) {
; GFX10-LABEL: loop_with_3breaks:
; GFX10:       ; %bb.0: ; %entry
; GFX10-NEXT:    s_mov_b32 s0, 0
; GFX10-NEXT:    ; implicit-def: $sgpr1
; GFX10-NEXT:    v_mov_b32_e32 v8, s0
; GFX10-NEXT:    s_branch .LBB4_4
; GFX10-NEXT:  .LBB4_1: ; %Flow5
; GFX10-NEXT:    ; in Loop: Header=BB4_4 Depth=1
; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_andn2_b32 s4, -1, exec_lo
; GFX10-NEXT:    s_and_b32 s5, exec_lo, s5
; GFX10-NEXT:    s_or_b32 s4, s4, s5
; GFX10-NEXT:  .LBB4_2: ; %Flow4
; GFX10-NEXT:    ; in Loop: Header=BB4_4 Depth=1
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s3
; GFX10-NEXT:    s_andn2_b32 s1, s1, exec_lo
; GFX10-NEXT:    s_and_b32 s3, exec_lo, s4
; GFX10-NEXT:    s_or_b32 s1, s1, s3
; GFX10-NEXT:  .LBB4_3: ; %Flow
; GFX10-NEXT:    ; in Loop: Header=BB4_4 Depth=1
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s2
; GFX10-NEXT:    s_and_b32 s2, exec_lo, s1
; GFX10-NEXT:    s_or_b32 s0, s2, s0
; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s0
; GFX10-NEXT:    s_cbranch_execz .LBB4_8
; GFX10-NEXT:  .LBB4_4: ; %A
; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT:    v_ashrrev_i32_e32 v9, 31, v8
; GFX10-NEXT:    s_andn2_b32 s1, s1, exec_lo
; GFX10-NEXT:    s_and_b32 s2, exec_lo, -1
; GFX10-NEXT:    s_or_b32 s1, s1, s2
; GFX10-NEXT:    v_lshlrev_b64 v[9:10], 2, v[8:9]
; GFX10-NEXT:    v_add_co_u32 v11, vcc_lo, v2, v9
; GFX10-NEXT:    v_add_co_ci_u32_e32 v12, vcc_lo, v3, v10, vcc_lo
; GFX10-NEXT:    global_load_dword v11, v[11:12], off
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v11
; GFX10-NEXT:    s_and_saveexec_b32 s2, vcc_lo
; GFX10-NEXT:    s_cbranch_execz .LBB4_3
; GFX10-NEXT:  ; %bb.5: ; %B
; GFX10-NEXT:    ; in Loop: Header=BB4_4 Depth=1
; GFX10-NEXT:    v_add_co_u32 v11, vcc_lo, v4, v9
; GFX10-NEXT:    v_add_co_ci_u32_e32 v12, vcc_lo, v5, v10, vcc_lo
; GFX10-NEXT:    s_mov_b32 s4, -1
; GFX10-NEXT:    global_load_dword v11, v[11:12], off
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v11
; GFX10-NEXT:    s_and_saveexec_b32 s3, vcc_lo
; GFX10-NEXT:    s_cbranch_execz .LBB4_2
; GFX10-NEXT:  ; %bb.6: ; %C
; GFX10-NEXT:    ; in Loop: Header=BB4_4 Depth=1
; GFX10-NEXT:    v_add_co_u32 v11, vcc_lo, v6, v9
; GFX10-NEXT:    v_add_co_ci_u32_e32 v12, vcc_lo, v7, v10, vcc_lo
; GFX10-NEXT:    s_mov_b32 s5, -1
; GFX10-NEXT:    global_load_dword v11, v[11:12], off
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v11
; GFX10-NEXT:    s_and_saveexec_b32 s4, vcc_lo
; GFX10-NEXT:    s_cbranch_execz .LBB4_1
; GFX10-NEXT:  ; %bb.7: ; %loop.body
; GFX10-NEXT:    ; in Loop: Header=BB4_4 Depth=1
; GFX10-NEXT:    v_add_co_u32 v9, vcc_lo, v0, v9
; GFX10-NEXT:    v_add_co_ci_u32_e32 v10, vcc_lo, v1, v10, vcc_lo
; GFX10-NEXT:    v_add_nc_u32_e32 v12, 1, v8
; GFX10-NEXT:    v_cmp_gt_u32_e32 vcc_lo, 0x64, v8
; GFX10-NEXT:    s_andn2_b32 s5, -1, exec_lo
; GFX10-NEXT:    global_load_dword v11, v[9:10], off
; GFX10-NEXT:    v_mov_b32_e32 v8, v12
; GFX10-NEXT:    s_and_b32 s6, exec_lo, vcc_lo
; GFX10-NEXT:    s_or_b32 s5, s5, s6
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    v_add_nc_u32_e32 v11, 1, v11
; GFX10-NEXT:    global_store_dword v[9:10], v11, off
; GFX10-NEXT:    s_branch .LBB4_1
; GFX10-NEXT:  .LBB4_8: ; %exit
; GFX10-NEXT:    s_endpgm
entry:
  br label %A

A:
  %counter = phi i32 [ %counter.plus.1, %loop.body ], [ 0, %entry ]
  %a.plus.counter = getelementptr inbounds i32, ptr addrspace(1) %a, i32 %counter
  %a.val = load i32, ptr addrspace(1) %a.plus.counter
  %a.cond = icmp eq i32 %a.val, 0
  br i1 %a.cond, label %exit, label %B

B:
  %b.plus.counter = getelementptr inbounds i32, ptr addrspace(1) %b, i32 %counter
  %b.val = load i32, ptr addrspace(1) %b.plus.counter
  %b.cond = icmp eq i32 %b.val, 0
  br i1 %b.cond, label %exit, label %C

C:
  %c.plus.counter = getelementptr inbounds i32, ptr addrspace(1) %c, i32 %counter
  %c.val = load i32, ptr addrspace(1) %c.plus.counter
  %c.cond = icmp eq i32 %c.val, 0
  br i1 %c.cond, label %exit, label %loop.body

loop.body:
  %x.plus.counter = getelementptr inbounds i32, ptr addrspace(1) %x, i32 %counter
  %x.val = load i32, ptr addrspace(1) %x.plus.counter
  %x.val.plus.1 = add i32 %x.val, 1
  store i32 %x.val.plus.1, ptr addrspace(1) %x.plus.counter
  %counter.plus.1 = add i32 %counter, 1
  %x.cond = icmp ult i32 %counter, 100
  br i1 %x.cond, label %exit, label %A

exit:
  ret void
}

; Divergent condition if with body, ending with break. This is loop with two
; exits but structurizer will create phi that will track exit from break
; and move break.body after the loop. Loop will then have one exit and phi
; used outside of the loop by condition used to enter the break.body.
define amdgpu_cs void @loop_with_div_break_with_body(ptr addrspace(1) %x, ptr addrspace(1) %a, ptr addrspace(1) %a.break) {
; GFX10-LABEL: loop_with_div_break_with_body:
; GFX10:       ; %bb.0: ; %entry
; GFX10-NEXT:    s_mov_b32 s0, 0
; GFX10-NEXT:    ; implicit-def: $sgpr1
; GFX10-NEXT:    ; implicit-def: $sgpr2
; GFX10-NEXT:    ; implicit-def: $sgpr3
; GFX10-NEXT:    v_mov_b32_e32 v6, s0
; GFX10-NEXT:    s_branch .LBB5_2
; GFX10-NEXT:  .LBB5_1: ; %Flow
; GFX10-NEXT:    ; in Loop: Header=BB5_2 Depth=1
; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_and_b32 s4, exec_lo, s2
; GFX10-NEXT:    s_or_b32 s0, s4, s0
; GFX10-NEXT:    s_andn2_b32 s1, s1, exec_lo
; GFX10-NEXT:    s_and_b32 s4, exec_lo, s3
; GFX10-NEXT:    s_or_b32 s1, s1, s4
; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s0
; GFX10-NEXT:    s_cbranch_execz .LBB5_4
; GFX10-NEXT:  .LBB5_2: ; %A
; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT:    v_ashrrev_i32_e32 v7, 31, v6
; GFX10-NEXT:    s_andn2_b32 s3, s3, exec_lo
; GFX10-NEXT:    s_and_b32 s4, exec_lo, -1
; GFX10-NEXT:    s_andn2_b32 s2, s2, exec_lo
; GFX10-NEXT:    s_or_b32 s3, s3, s4
; GFX10-NEXT:    v_lshlrev_b64 v[7:8], 2, v[6:7]
; GFX10-NEXT:    s_or_b32 s2, s2, s4
; GFX10-NEXT:    v_add_co_u32 v9, vcc_lo, v2, v7
; GFX10-NEXT:    v_add_co_ci_u32_e32 v10, vcc_lo, v3, v8, vcc_lo
; GFX10-NEXT:    global_load_dword v9, v[9:10], off
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v9
; GFX10-NEXT:    s_and_saveexec_b32 s4, vcc_lo
; GFX10-NEXT:    s_cbranch_execz .LBB5_1
; GFX10-NEXT:  ; %bb.3: ; %loop.body
; GFX10-NEXT:    ; in Loop: Header=BB5_2 Depth=1
; GFX10-NEXT:    v_add_co_u32 v7, vcc_lo, v0, v7
; GFX10-NEXT:    v_add_co_ci_u32_e32 v8, vcc_lo, v1, v8, vcc_lo
; GFX10-NEXT:    v_add_nc_u32_e32 v10, 1, v6
; GFX10-NEXT:    v_cmp_gt_u32_e32 vcc_lo, 0x64, v6
; GFX10-NEXT:    s_andn2_b32 s3, s3, exec_lo
; GFX10-NEXT:    global_load_dword v9, v[7:8], off
; GFX10-NEXT:    s_and_b32 s5, exec_lo, 0
; GFX10-NEXT:    v_mov_b32_e32 v6, v10
; GFX10-NEXT:    s_andn2_b32 s2, s2, exec_lo
; GFX10-NEXT:    s_and_b32 s6, exec_lo, vcc_lo
; GFX10-NEXT:    s_or_b32 s3, s3, s5
; GFX10-NEXT:    s_or_b32 s2, s2, s6
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    v_add_nc_u32_e32 v9, 1, v9
; GFX10-NEXT:    global_store_dword v[7:8], v9, off
; GFX10-NEXT:    s_branch .LBB5_1
; GFX10-NEXT:  .LBB5_4: ; %loop.exit.guard
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX10-NEXT:    s_and_saveexec_b32 s0, s1
; GFX10-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX10-NEXT:    s_cbranch_execz .LBB5_6
; GFX10-NEXT:  ; %bb.5: ; %break.body
; GFX10-NEXT:    v_mov_b32_e32 v0, 10
; GFX10-NEXT:    global_store_dword v[4:5], v0, off
; GFX10-NEXT:  .LBB5_6: ; %exit
; GFX10-NEXT:    s_endpgm
entry:
  br label %A

A:
  %counter = phi i32 [ %counter.plus.1, %loop.body ], [ 0, %entry ]
  %a.plus.counter = getelementptr inbounds i32, ptr addrspace(1) %a, i32 %counter
  %a.val = load i32, ptr addrspace(1) %a.plus.counter
  %a.cond = icmp eq i32 %a.val, 0
  br i1 %a.cond, label %break.body, label %loop.body

break.body:
  store i32 10, ptr addrspace(1) %a.break
  br label %exit


loop.body:
  %x.plus.counter = getelementptr inbounds i32, ptr addrspace(1) %x, i32 %counter
  %x.val = load i32, ptr addrspace(1) %x.plus.counter
  %x.val.plus.1 = add i32 %x.val, 1
  store i32 %x.val.plus.1, ptr addrspace(1) %x.plus.counter
  %counter.plus.1 = add i32 %counter, 1
  %x.cond = icmp ult i32 %counter, 100
  br i1 %x.cond, label %exit, label %A

exit:
  ret void
}

; Snippet from test generated by the GraphicsFuzz tool, frontend generates ir
; with irreducible control flow graph. FixIrreducible converts it into natural
; loop and in the process creates i1 phi with three incoming values.

; int loop(int x, int y, int a0, int a1, int a2, int a3, int a4) {
;   do {
;     if (y < a2) {
;       do {
;       } while (x < a2);
;     }
;     if (x < a3) {
;       return a1;
;     }
;   } while (y < a2);
;   return a0;
; }

; This test is also interesting because it has phi with three incomings
;define amdgpu_ps i32 @irreducible_cfg(i32 %x, i32 %y, i32 %a0, i32 %a1, i32 %a2, i32 %a3) {
;.entry:
; %.y_lt_a2 = icmp sgt i32 %a2, %y
; %.x_lt_a2 = icmp sgt i32 %a2, %x
; %.x_lt_a3 = icmp sgt i32 %a3, %x
; br i1 %.y_lt_a2, label %.preheader, label %.loopexit ; first iteration, jump to inner loop if 'y < a2' or start with 'if (x < a3)'
;
;.preheader: ; if (y < a2),
; br label %.inner_loop
;
;.inner_loop: ; do while x < a2
; br i1 %.x_lt_a2, label %.inner_loop, label %.loopexit
;
;.loopexit: ; if x < a3
; %not.inner_loop = xor i1 %.y_lt_a2, true
; %brmerge = select i1 %.x_lt_a3, i1 true, i1 %not.inner_loop ; exit loop if 'x < a3' or 'loop ends since !(y < a2)'
; %.ret = select i1 %.x_lt_a3, i32 %a1, i32 %a0               ; select retrun value a1 'x < a3' or a0 'loop ends'
; br i1 %brmerge, label %.exit, label %.preheader
;
;.exit:
; ret i32 %.ret
;}