llvm/test/CodeGen/AMDGPU/structurize-hoist.ll

; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
; RUN: llc -mtriple=amdgpu9.00 < %s | FileCheck -check-prefix=GFX900 %s


%pair = type { i32, i32 }

define void @test_extractvalue_then_else(ptr %ptr, i1 %cond) {
; GFX900-LABEL: test_extractvalue_then_else:
; GFX900:       ; %bb.0: ; %if
; GFX900-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX900-NEXT:    flat_load_dword v3, v[0:1]
; GFX900-NEXT:    v_and_b32_e32 v2, 1, v2
; GFX900-NEXT:    v_cmp_ne_u32_e32 vcc, 1, v2
; GFX900-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX900-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX900-NEXT:    s_cbranch_execz .LBB0_2
; GFX900-NEXT:  ; %bb.1: ; %else
; GFX900-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX900-NEXT:    v_add_u32_e32 v3, 1, v3
; GFX900-NEXT:  .LBB0_2: ; %Flow
; GFX900-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX900-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX900-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX900-NEXT:    flat_store_dword v[0:1], v3
; GFX900-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX900-NEXT:    s_setpc_b64 s[30:31]
if:
  %load_then = load %pair, ptr %ptr
  br i1 %cond, label %then, label %else

then:
  %a_then = extractvalue %pair %load_then, 0
  br label %merge

else:
  %a_else = extractvalue %pair %load_then, 0
  %sum_else = add i32 %a_else, 1
  br label %merge

merge:
  %phi = phi i32  [ %a_then, %then ], [ %sum_else, %else ]
  store i32 %phi, ptr  %ptr
  ret void
}

define void @test_extractvalue_else_then(ptr %ptr, i1 %cond) {
; GFX900-LABEL: test_extractvalue_else_then:
; GFX900:       ; %bb.0: ; %if
; GFX900-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX900-NEXT:    flat_load_dword v3, v[0:1]
; GFX900-NEXT:    v_and_b32_e32 v2, 1, v2
; GFX900-NEXT:    v_cmp_ne_u32_e32 vcc, 1, v2
; GFX900-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX900-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX900-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX900-NEXT:    s_cbranch_execz .LBB1_2
; GFX900-NEXT:  ; %bb.1: ; %else
; GFX900-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX900-NEXT:    v_add_u32_e32 v3, 1, v3
; GFX900-NEXT:  .LBB1_2: ; %merge
; GFX900-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX900-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX900-NEXT:    flat_store_dword v[0:1], v3
; GFX900-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX900-NEXT:    s_setpc_b64 s[30:31]
if:
  %load_then = load %pair, ptr %ptr
  br i1 %cond, label %else, label %then

else:
  %a_else = extractvalue %pair %load_then, 0
  %sum_else = add i32 %a_else, 1
  br label %merge

then:
  %a_then = extractvalue %pair %load_then, 0
  br label %merge

merge:
  %phi = phi i32  [ %a_then, %then ], [ %sum_else, %else ]
  store i32 %phi, ptr  %ptr
  ret void
}

define amdgpu_kernel void @test_loop_with_if( ptr %ptr, i1 %cond) #0 {
; GFX900-LABEL: test_loop_with_if:
; GFX900:       ; %bb.0: ; %entry
; GFX900-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GFX900-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x24
; GFX900-NEXT:    s_mov_b64 s[0:1], 0
; GFX900-NEXT:    v_mov_b32_e32 v5, 0
; GFX900-NEXT:    s_movk_i32 s10, 0xfe
; GFX900-NEXT:    s_waitcnt lgkmcnt(0)
; GFX900-NEXT:    s_bitcmp1_b32 s2, 0
; GFX900-NEXT:    s_cselect_b64 s[2:3], -1, 0
; GFX900-NEXT:    s_xor_b64 s[4:5], s[2:3], -1
; GFX900-NEXT:    v_mov_b32_e32 v1, s6
; GFX900-NEXT:    v_mov_b32_e32 v2, s7
; GFX900-NEXT:    s_branch .LBB2_2
; GFX900-NEXT:  .LBB2_1: ; %latch
; GFX900-NEXT:    ; in Loop: Header=BB2_2 Depth=1
; GFX900-NEXT:    s_or_b64 exec, exec, s[8:9]
; GFX900-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX900-NEXT:    v_add_u32_e32 v5, 20, v3
; GFX900-NEXT:    v_cmp_lt_i32_e32 vcc, s10, v5
; GFX900-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
; GFX900-NEXT:    flat_store_dword v[1:2], v3
; GFX900-NEXT:    s_andn2_b64 exec, exec, s[0:1]
; GFX900-NEXT:    s_cbranch_execz .LBB2_8
; GFX900-NEXT:  .LBB2_2: ; %loop
; GFX900-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX900-NEXT:    flat_load_dwordx2 v[3:4], v[1:2]
; GFX900-NEXT:    s_and_b64 s[6:7], s[2:3], exec
; GFX900-NEXT:    s_mov_b64 s[8:9], s[4:5]
; GFX900-NEXT:    s_mov_b64 s[6:7], 0
; GFX900-NEXT:    s_cbranch_scc0 .LBB2_4
; GFX900-NEXT:  ; %bb.3: ; %if
; GFX900-NEXT:    ; in Loop: Header=BB2_2 Depth=1
; GFX900-NEXT:    v_cmp_gt_i32_e32 vcc, 11, v5
; GFX900-NEXT:    s_andn2_b64 s[8:9], s[4:5], exec
; GFX900-NEXT:    s_mov_b64 s[6:7], -1
; GFX900-NEXT:    s_or_b64 s[8:9], s[8:9], vcc
; GFX900-NEXT:  .LBB2_4: ; %Flow
; GFX900-NEXT:    ; in Loop: Header=BB2_2 Depth=1
; GFX900-NEXT:    s_and_saveexec_b64 s[12:13], s[8:9]
; GFX900-NEXT:    s_xor_b64 s[8:9], exec, s[12:13]
; GFX900-NEXT:    s_cbranch_execz .LBB2_6
; GFX900-NEXT:  ; %bb.5: ; %else
; GFX900-NEXT:    ; in Loop: Header=BB2_2 Depth=1
; GFX900-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX900-NEXT:    v_add_u32_e32 v3, v3, v4
; GFX900-NEXT:    s_andn2_b64 s[6:7], s[6:7], exec
; GFX900-NEXT:  .LBB2_6: ; %Flow1
; GFX900-NEXT:    ; in Loop: Header=BB2_2 Depth=1
; GFX900-NEXT:    s_or_b64 exec, exec, s[8:9]
; GFX900-NEXT:    s_and_saveexec_b64 s[8:9], s[6:7]
; GFX900-NEXT:    s_cbranch_execz .LBB2_1
; GFX900-NEXT:  ; %bb.7: ; %then
; GFX900-NEXT:    ; in Loop: Header=BB2_2 Depth=1
; GFX900-NEXT:    flat_store_dword v[1:2], v0
; GFX900-NEXT:    s_branch .LBB2_1
; GFX900-NEXT:  .LBB2_8: ; %end
; GFX900-NEXT:    s_endpgm
entry:
  %a = tail call i32 @llvm.amdgcn.workitem.id.x()
  br label %loop

loop:
  %entry_phi = phi i32 [ 0, %entry ], [ %a15, %latch ]
  %load = load %pair, ptr %ptr
  br i1 %cond, label %if, label %else

if:
  %cmp = icmp sgt i32 %entry_phi, 10
  br i1 %cmp, label %then, label %else

then:
  %a_then = extractvalue %pair %load, 0
  store i32 %a, ptr %ptr, align 4
  br label %latch

else:
  %a2 = extractvalue %pair %load, 1
  %y = extractvalue %pair %load, 0
  %a_else = add i32 %y, %a2
  br label %latch

latch:
  %a_test = phi i32 [ %a_then, %then ], [ %a_else, %else ]
  store i32 %a_test, ptr  %ptr
  %a15 = add nsw i32 %a_test, 20
  %a16 = icmp slt i32  %a15, 255
  br i1 %a16, label %loop, label %end

end:
  ret void
}

define void @test_nested_if(ptr %ptr, i32 %val, i1 %cond) {
; GFX900-LABEL: test_nested_if:
; GFX900:       ; %bb.0: ; %entry
; GFX900-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX900-NEXT:    flat_load_dword v4, v[0:1]
; GFX900-NEXT:    v_and_b32_e32 v3, 1, v3
; GFX900-NEXT:    v_cmp_eq_u32_e64 s[6:7], 1, v3
; GFX900-NEXT:    s_mov_b64 s[10:11], -1
; GFX900-NEXT:    s_xor_b64 s[4:5], s[6:7], -1
; GFX900-NEXT:    s_mov_b64 s[12:13], s[6:7]
; GFX900-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX900-NEXT:    v_mov_b32_e32 v3, v4
; GFX900-NEXT:    s_and_saveexec_b64 s[8:9], s[4:5]
; GFX900-NEXT:    s_cbranch_execz .LBB3_4
; GFX900-NEXT:  ; %bb.1: ; %if
; GFX900-NEXT:    s_and_saveexec_b64 s[12:13], s[4:5]
; GFX900-NEXT:    s_cbranch_execz .LBB3_3
; GFX900-NEXT:  ; %bb.2: ; %if_2
; GFX900-NEXT:    flat_load_dword v3, v[0:1]
; GFX900-NEXT:    s_xor_b64 s[10:11], exec, -1
; GFX900-NEXT:  .LBB3_3: ; %Flow3
; GFX900-NEXT:    s_or_b64 exec, exec, s[12:13]
; GFX900-NEXT:    s_andn2_b64 s[12:13], s[6:7], exec
; GFX900-NEXT:    s_and_b64 s[10:11], s[10:11], exec
; GFX900-NEXT:    s_or_b64 s[12:13], s[12:13], s[10:11]
; GFX900-NEXT:  .LBB3_4: ; %Flow2
; GFX900-NEXT:    s_or_b64 exec, exec, s[8:9]
; GFX900-NEXT:    s_and_saveexec_b64 s[8:9], s[12:13]
; GFX900-NEXT:    s_or_b64 exec, exec, s[8:9]
; GFX900-NEXT:    s_and_saveexec_b64 s[8:9], s[6:7]
; GFX900-NEXT:    s_cbranch_execz .LBB3_8
; GFX900-NEXT:  ; %bb.5: ; %if_3
; GFX900-NEXT:    s_movk_i32 s6, 0xfe
; GFX900-NEXT:    v_cmp_lt_i32_e32 vcc, s6, v2
; GFX900-NEXT:    s_mov_b64 s[6:7], -1
; GFX900-NEXT:    s_and_saveexec_b64 s[10:11], vcc
; GFX900-NEXT:    s_cbranch_execz .LBB3_7
; GFX900-NEXT:  ; %bb.6: ; %if_4
; GFX900-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX900-NEXT:    v_add_u32_e32 v4, 1, v3
; GFX900-NEXT:    s_xor_b64 s[6:7], exec, -1
; GFX900-NEXT:  .LBB3_7: ; %Flow1
; GFX900-NEXT:    s_or_b64 exec, exec, s[10:11]
; GFX900-NEXT:    s_andn2_b64 s[4:5], s[4:5], exec
; GFX900-NEXT:    s_and_b64 s[6:7], s[6:7], exec
; GFX900-NEXT:    s_or_b64 s[4:5], s[4:5], s[6:7]
; GFX900-NEXT:  .LBB3_8: ; %Flow
; GFX900-NEXT:    s_or_b64 exec, exec, s[8:9]
; GFX900-NEXT:    s_and_saveexec_b64 s[6:7], s[4:5]
; GFX900-NEXT:    s_or_b64 exec, exec, s[6:7]
; GFX900-NEXT:    flat_store_dword v[0:1], v4
; GFX900-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX900-NEXT:    s_setpc_b64 s[30:31]
entry:
  %load = load %pair, ptr %ptr
  br i1 %cond, label %else, label %if

if:
  %a16 = icmp slt i32  %val, 255
  br i1 %cond, label %else, label %if_2

if_2:
  %loaded = load i32, ptr  %ptr
  br label %merge

else:
  %a_else = extractvalue %pair %load, 0
  br label %merge

merge:
  %phi = phi i32  [ %loaded, %if_2 ], [ %a_else, %else ]
  br i1 %cond, label %if_3, label %else_2

if_3:
  %a17 = icmp slt i32  %val, 255
  br i1 %a17, label %else_2, label %if_4

if_4:
  %sum_load = add i32 %phi, 1
  br label %merge_2

else_2:
  %a_else_2 = extractvalue %pair %load, 0
  br label %merge_2

merge_2:
  %phi_2 = phi i32  [ %sum_load, %if_4 ], [ %a_else_2, %else_2 ]
  store i32 %phi_2, ptr  %ptr
  ret void
}