; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 4
; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=gfx90a < %s | FileCheck -check-prefixes=GFX90A %s
; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=gfx940 < %s | FileCheck -check-prefixes=GFX940 %s
; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=gfx1030 < %s | FileCheck -check-prefixes=GFX1030 %s
; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=gfx1100 < %s | FileCheck -check-prefixes=GFX1100 %s

define amdgpu_kernel void @test_insert_extract(i32 %p, i32 %q) {
; GFX90A-LABEL: test_insert_extract:
; GFX90A:       ; %bb.0: ; %entry
; GFX90A-NEXT:    s_load_dwordx2 s[0:1], s[8:9], 0x0
; GFX90A-NEXT:    s_mov_b32 s2, 0
; GFX90A-NEXT:    s_and_b64 vcc, exec, -1
; GFX90A-NEXT:    s_mov_b32 s3, 0
; GFX90A-NEXT:    s_mov_b32 s4, 0
; GFX90A-NEXT:    s_mov_b32 s5, 0
; GFX90A-NEXT:    s_mov_b32 s6, 0
; GFX90A-NEXT:  .LBB0_1: ; %for.body
; GFX90A-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    s_cmp_eq_u32 s1, 1
; GFX90A-NEXT:    s_cselect_b64 s[8:9], -1, 0
; GFX90A-NEXT:    s_and_b64 s[8:9], s[8:9], exec
; GFX90A-NEXT:    s_cselect_b32 s7, s4, s3
; GFX90A-NEXT:    s_cmp_eq_u32 s1, 2
; GFX90A-NEXT:    s_cselect_b64 s[8:9], -1, 0
; GFX90A-NEXT:    s_and_b64 s[8:9], s[8:9], exec
; GFX90A-NEXT:    s_cselect_b32 s7, s5, s7
; GFX90A-NEXT:    s_cmp_eq_u32 s1, 3
; GFX90A-NEXT:    s_cselect_b64 s[8:9], -1, 0
; GFX90A-NEXT:    s_and_b64 s[8:9], s[8:9], exec
; GFX90A-NEXT:    s_cselect_b32 s7, s6, s7
; GFX90A-NEXT:    s_or_b32 s7, s7, s0
; GFX90A-NEXT:    s_cmp_eq_u32 s1, 1
; GFX90A-NEXT:    s_cselect_b64 s[8:9], -1, 0
; GFX90A-NEXT:    s_and_b64 s[10:11], s[8:9], exec
; GFX90A-NEXT:    s_cselect_b32 s4, s7, s4
; GFX90A-NEXT:    s_cmp_eq_u32 s1, 3
; GFX90A-NEXT:    s_cselect_b64 s[10:11], -1, 0
; GFX90A-NEXT:    s_and_b64 s[12:13], s[10:11], exec
; GFX90A-NEXT:    s_cselect_b32 s6, s7, s6
; GFX90A-NEXT:    s_cmp_eq_u32 s1, 2
; GFX90A-NEXT:    s_cselect_b64 s[12:13], -1, 0
; GFX90A-NEXT:    s_and_b64 s[14:15], s[12:13], exec
; GFX90A-NEXT:    s_cselect_b32 s5, s7, s5
; GFX90A-NEXT:    s_cmp_eq_u32 s1, 0
; GFX90A-NEXT:    s_cselect_b32 s3, s7, s3
; GFX90A-NEXT:    s_or_b64 s[8:9], s[12:13], s[8:9]
; GFX90A-NEXT:    s_or_b64 s[8:9], s[10:11], s[8:9]
; GFX90A-NEXT:    s_and_b64 s[8:9], s[8:9], exec
; GFX90A-NEXT:    s_cselect_b32 s2, 0, s2
; GFX90A-NEXT:    s_mov_b64 vcc, vcc
; GFX90A-NEXT:    s_cbranch_vccnz .LBB0_1
; GFX90A-NEXT:  ; %bb.2: ; %DummyReturnBlock
; GFX90A-NEXT:    s_endpgm
;
; GFX940-LABEL: test_insert_extract:
; GFX940:       ; %bb.0: ; %entry
; GFX940-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
; GFX940-NEXT:    s_mov_b32 s2, 0
; GFX940-NEXT:    s_and_b64 vcc, exec, -1
; GFX940-NEXT:    s_mov_b32 s3, 0
; GFX940-NEXT:    s_mov_b32 s4, 0
; GFX940-NEXT:    s_mov_b32 s5, 0
; GFX940-NEXT:    s_mov_b32 s6, 0
; GFX940-NEXT:  .LBB0_1: ; %for.body
; GFX940-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
; GFX940-NEXT:    s_cmp_eq_u32 s1, 1
; GFX940-NEXT:    s_cselect_b64 s[8:9], -1, 0
; GFX940-NEXT:    s_and_b64 s[8:9], s[8:9], exec
; GFX940-NEXT:    s_cselect_b32 s7, s4, s3
; GFX940-NEXT:    s_cmp_eq_u32 s1, 2
; GFX940-NEXT:    s_cselect_b64 s[8:9], -1, 0
; GFX940-NEXT:    s_and_b64 s[8:9], s[8:9], exec
; GFX940-NEXT:    s_cselect_b32 s7, s5, s7
; GFX940-NEXT:    s_cmp_eq_u32 s1, 3
; GFX940-NEXT:    s_cselect_b64 s[8:9], -1, 0
; GFX940-NEXT:    s_and_b64 s[8:9], s[8:9], exec
; GFX940-NEXT:    s_cselect_b32 s7, s6, s7
; GFX940-NEXT:    s_or_b32 s7, s7, s0
; GFX940-NEXT:    s_cmp_eq_u32 s1, 1
; GFX940-NEXT:    s_cselect_b64 s[8:9], -1, 0
; GFX940-NEXT:    s_and_b64 s[10:11], s[8:9], exec
; GFX940-NEXT:    s_cselect_b32 s4, s7, s4
; GFX940-NEXT:    s_cmp_eq_u32 s1, 3
; GFX940-NEXT:    s_cselect_b64 s[10:11], -1, 0
; GFX940-NEXT:    s_and_b64 s[12:13], s[10:11], exec
; GFX940-NEXT:    s_cselect_b32 s6, s7, s6
; GFX940-NEXT:    s_cmp_eq_u32 s1, 2
; GFX940-NEXT:    s_cselect_b64 s[12:13], -1, 0
; GFX940-NEXT:    s_and_b64 s[14:15], s[12:13], exec
; GFX940-NEXT:    s_cselect_b32 s5, s7, s5
; GFX940-NEXT:    s_cmp_eq_u32 s1, 0
; GFX940-NEXT:    s_cselect_b32 s3, s7, s3
; GFX940-NEXT:    s_or_b64 s[8:9], s[12:13], s[8:9]
; GFX940-NEXT:    s_or_b64 s[8:9], s[10:11], s[8:9]
; GFX940-NEXT:    s_and_b64 s[8:9], s[8:9], exec
; GFX940-NEXT:    s_cselect_b32 s2, 0, s2
; GFX940-NEXT:    s_mov_b64 vcc, vcc
; GFX940-NEXT:    s_cbranch_vccnz .LBB0_1
; GFX940-NEXT:  ; %bb.2: ; %DummyReturnBlock
; GFX940-NEXT:    s_endpgm
;
; GFX1030-LABEL: test_insert_extract:
; GFX1030:       ; %bb.0: ; %entry
; GFX1030-NEXT:    s_load_dwordx2 s[0:1], s[8:9], 0x0
; GFX1030-NEXT:    s_mov_b32 s2, 0
; GFX1030-NEXT:    s_mov_b32 s3, 0
; GFX1030-NEXT:    s_mov_b32 s4, 0
; GFX1030-NEXT:    s_mov_b32 s5, 0
; GFX1030-NEXT:    s_mov_b32 s6, 0
; GFX1030-NEXT:    s_mov_b32 vcc_lo, exec_lo
; GFX1030-NEXT:    .p2align 6
; GFX1030-NEXT:  .LBB0_1: ; %for.body
; GFX1030-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX1030-NEXT:    s_waitcnt lgkmcnt(0)
; GFX1030-NEXT:    s_cmp_eq_u32 s1, 1
; GFX1030-NEXT:    s_cselect_b32 s7, -1, 0
; GFX1030-NEXT:    s_and_b32 s7, s7, exec_lo
; GFX1030-NEXT:    s_cselect_b32 s7, s4, s3
; GFX1030-NEXT:    s_cmp_eq_u32 s1, 2
; GFX1030-NEXT:    s_cselect_b32 s8, -1, 0
; GFX1030-NEXT:    s_and_b32 s8, s8, exec_lo
; GFX1030-NEXT:    s_cselect_b32 s7, s5, s7
; GFX1030-NEXT:    s_cmp_eq_u32 s1, 3
; GFX1030-NEXT:    s_cselect_b32 s8, -1, 0
; GFX1030-NEXT:    s_and_b32 s8, s8, exec_lo
; GFX1030-NEXT:    s_cselect_b32 s7, s6, s7
; GFX1030-NEXT:    s_or_b32 s7, s7, s0
; GFX1030-NEXT:    s_cmp_eq_u32 s1, 1
; GFX1030-NEXT:    s_cselect_b32 s8, -1, 0
; GFX1030-NEXT:    s_and_b32 s9, s8, exec_lo
; GFX1030-NEXT:    s_cselect_b32 s4, s7, s4
; GFX1030-NEXT:    s_cmp_eq_u32 s1, 3
; GFX1030-NEXT:    s_cselect_b32 s9, -1, 0
; GFX1030-NEXT:    s_and_b32 s10, s9, exec_lo
; GFX1030-NEXT:    s_cselect_b32 s6, s7, s6
; GFX1030-NEXT:    s_cmp_eq_u32 s1, 2
; GFX1030-NEXT:    s_cselect_b32 s10, -1, 0
; GFX1030-NEXT:    s_and_b32 s11, s10, exec_lo
; GFX1030-NEXT:    s_cselect_b32 s5, s7, s5
; GFX1030-NEXT:    s_cmp_eq_u32 s1, 0
; GFX1030-NEXT:    s_cselect_b32 s3, s7, s3
; GFX1030-NEXT:    s_or_b32 s7, s10, s8
; GFX1030-NEXT:    s_or_b32 s7, s9, s7
; GFX1030-NEXT:    s_and_b32 s7, s7, exec_lo
; GFX1030-NEXT:    s_cselect_b32 s2, 0, s2
; GFX1030-NEXT:    s_cbranch_vccnz .LBB0_1
; GFX1030-NEXT:  ; %bb.2: ; %DummyReturnBlock
; GFX1030-NEXT:    s_endpgm
;
; GFX1100-LABEL: test_insert_extract:
; GFX1100:       ; %bb.0: ; %entry
; GFX1100-NEXT:    s_load_b64 s[0:1], s[4:5], 0x0
; GFX1100-NEXT:    s_mov_b32 s2, 0
; GFX1100-NEXT:    s_mov_b32 s3, 0
; GFX1100-NEXT:    s_mov_b32 s4, 0
; GFX1100-NEXT:    s_mov_b32 s5, 0
; GFX1100-NEXT:    s_mov_b32 s6, 0
; GFX1100-NEXT:    s_mov_b32 vcc_lo, exec_lo
; GFX1100-NEXT:    .p2align 6
; GFX1100-NEXT:  .LBB0_1: ; %for.body
; GFX1100-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX1100-NEXT:    s_waitcnt lgkmcnt(0)
; GFX1100-NEXT:    s_cmp_eq_u32 s1, 1
; GFX1100-NEXT:    s_cselect_b32 s7, -1, 0
; GFX1100-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_3) | instid1(SALU_CYCLE_1)
; GFX1100-NEXT:    s_and_b32 s7, s7, exec_lo
; GFX1100-NEXT:    s_cselect_b32 s7, s4, s3
; GFX1100-NEXT:    s_cmp_eq_u32 s1, 2
; GFX1100-NEXT:    s_cselect_b32 s8, -1, 0
; GFX1100-NEXT:    s_and_b32 s8, s8, exec_lo
; GFX1100-NEXT:    s_cselect_b32 s7, s5, s7
; GFX1100-NEXT:    s_cmp_eq_u32 s1, 3
; GFX1100-NEXT:    s_cselect_b32 s8, -1, 0
; GFX1100-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX1100-NEXT:    s_and_b32 s8, s8, exec_lo
; GFX1100-NEXT:    s_cselect_b32 s7, s6, s7
; GFX1100-NEXT:    s_or_b32 s7, s7, s0
; GFX1100-NEXT:    s_cmp_eq_u32 s1, 1
; GFX1100-NEXT:    s_cselect_b32 s8, -1, 0
; GFX1100-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_3) | instid1(SALU_CYCLE_1)
; GFX1100-NEXT:    s_and_b32 s9, s8, exec_lo
; GFX1100-NEXT:    s_cselect_b32 s4, s7, s4
; GFX1100-NEXT:    s_cmp_eq_u32 s1, 3
; GFX1100-NEXT:    s_cselect_b32 s9, -1, 0
; GFX1100-NEXT:    s_and_b32 s10, s9, exec_lo
; GFX1100-NEXT:    s_cselect_b32 s6, s7, s6
; GFX1100-NEXT:    s_cmp_eq_u32 s1, 2
; GFX1100-NEXT:    s_cselect_b32 s10, -1, 0
; GFX1100-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_4) | instid1(SALU_CYCLE_1)
; GFX1100-NEXT:    s_and_b32 s11, s10, exec_lo
; GFX1100-NEXT:    s_cselect_b32 s5, s7, s5
; GFX1100-NEXT:    s_cmp_eq_u32 s1, 0
; GFX1100-NEXT:    s_cselect_b32 s3, s7, s3
; GFX1100-NEXT:    s_or_b32 s7, s10, s8
; GFX1100-NEXT:    s_or_b32 s7, s9, s7
; GFX1100-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX1100-NEXT:    s_and_b32 s7, s7, exec_lo
; GFX1100-NEXT:    s_cselect_b32 s2, 0, s2
; GFX1100-NEXT:    s_cbranch_vccnz .LBB0_1
; GFX1100-NEXT:  ; %bb.2: ; %DummyReturnBlock
; GFX1100-NEXT:    s_endpgm
entry:
  %init = insertelement <4 x i32> zeroinitializer, i32 0, i64 0
  br label %for.body

for.body:                                     ; preds = %for.body, %entry
  %x1 = phi <4 x i32> [ %init, %entry ], [ %i4, %for.body ]
  %x2 = phi <4 x i32> [ zeroinitializer, %entry ], [ %i2, %for.body ]
  %idxprom = zext i32 %q to i64
  %e1 = extractelement <4 x i32> %x2, i64 %idxprom
  %add = or i32 %e1, %p
  %i2 = insertelement <4 x i32> %x2, i32 %add, i64 %idxprom
  %e3 = extractelement <4 x i32> %x1, i64 %idxprom
  %i4 = insertelement <4 x i32> %x1, i32 %e3, i64 0
  br label %for.body
}