; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 2
; RUN: llc -mtriple=amdgcn -mcpu=gfx1250 < %s | FileCheck --check-prefixes=GCN,SDAG %s
; RUN: llc -global-isel -mtriple=amdgcn -mcpu=gfx1250 < %s | FileCheck --check-prefixes=GCN,GISEL %s

define amdgpu_ps float @flat_load_b32_idxprom(ptr align 4 inreg %p, i32 %idx) {
; GCN-LABEL: flat_load_b32_idxprom:
; GCN:       ; %bb.0: ; %entry
; GCN-NEXT:    flat_load_b32 v0, v0, s[0:1] scale_offset
; GCN-NEXT:    s_wait_loadcnt_dscnt 0x0
; GCN-NEXT:    ; return to shader part epilog
entry:
  %idxprom = sext i32 %idx to i64
  %arrayidx = getelementptr inbounds float, ptr %p, i64 %idxprom
  %ret = load float, ptr %arrayidx, align 4
  ret float %ret
}

define amdgpu_ps float @flat_load_b32_idx32(ptr align 4 inreg %p, i32 %idx) {
; GCN-LABEL: flat_load_b32_idx32:
; GCN:       ; %bb.0: ; %entry
; GCN-NEXT:    flat_load_b32 v0, v0, s[0:1] scale_offset
; GCN-NEXT:    s_wait_loadcnt_dscnt 0x0
; GCN-NEXT:    ; return to shader part epilog
entry:
  %arrayidx = getelementptr inbounds float, ptr %p, i32 %idx
  %ret = load float, ptr %arrayidx, align 4
  ret float %ret
}

define amdgpu_ps float @flat_load_b32_idxprom_wrong_stride(ptr align 4 inreg %p, i32 %idx) {
; SDAG-LABEL: flat_load_b32_idxprom_wrong_stride:
; SDAG:       ; %bb.0: ; %entry
; SDAG-NEXT:    v_ashrrev_i32_e32 v1, 31, v0
; SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; SDAG-NEXT:    v_lshl_add_u64 v[0:1], v[0:1], 3, s[0:1]
; SDAG-NEXT:    flat_load_b32 v0, v[0:1]
; SDAG-NEXT:    s_wait_loadcnt_dscnt 0x0
; SDAG-NEXT:    ; return to shader part epilog
;
; GISEL-LABEL: flat_load_b32_idxprom_wrong_stride:
; GISEL:       ; %bb.0: ; %entry
; GISEL-NEXT:    v_ashrrev_i32_e32 v1, 31, v0
; GISEL-NEXT:    v_mov_b64_e32 v[2:3], s[0:1]
; GISEL-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GISEL-NEXT:    v_lshlrev_b64_e32 v[0:1], 3, v[0:1]
; GISEL-NEXT:    v_add_co_u32 v0, vcc_lo, v2, v0
; GISEL-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GISEL-NEXT:    v_add_co_ci_u32_e64 v1, null, v3, v1, vcc_lo
; GISEL-NEXT:    flat_load_b32 v0, v[0:1]
; GISEL-NEXT:    s_wait_loadcnt_dscnt 0x0
; GISEL-NEXT:    ; return to shader part epilog
entry:
  %idxprom = sext i32 %idx to i64
  %arrayidx = getelementptr inbounds <2 x float>, ptr %p, i64 %idxprom
  %ret = load float, ptr %arrayidx, align 4
  ret float %ret
}

define amdgpu_ps float @flat_load_b16_idxprom_ioffset(ptr align 4 inreg %p, i32 %idx) {
; GCN-LABEL: flat_load_b16_idxprom_ioffset:
; GCN:       ; %bb.0: ; %entry
; GCN-NEXT:    flat_load_u16 v0, v0, s[0:1] offset:32 scale_offset
; GCN-NEXT:    s_wait_loadcnt_dscnt 0x0
; GCN-NEXT:    ; return to shader part epilog
entry:
  %idxprom = sext i32 %idx to i64
  %idxadd = add i64 %idxprom, 16
  %arrayidx = getelementptr inbounds i16, ptr %p, i64 %idxadd
  %ld = load i16, ptr %arrayidx, align 2
  %ret.i32 = zext i16 %ld to i32
  %ret = bitcast i32 %ret.i32 to float
  ret float %ret
}

define amdgpu_ps <2 x float> @flat_load_b64_idxprom(ptr align 4 inreg %p, i32 %idx) {
; GCN-LABEL: flat_load_b64_idxprom:
; GCN:       ; %bb.0: ; %entry
; GCN-NEXT:    flat_load_b64 v[0:1], v0, s[0:1] scale_offset
; GCN-NEXT:    s_wait_loadcnt_dscnt 0x0
; GCN-NEXT:    ; return to shader part epilog
entry:
  %idxprom = sext i32 %idx to i64
  %arrayidx = getelementptr inbounds <2 x float>, ptr %p, i64 %idxprom
  %ret = load <2 x float>, ptr %arrayidx, align 4
  ret <2 x float> %ret
}

define amdgpu_ps <3 x float> @flat_load_b96_idxprom(ptr align 4 inreg %p, i32 %idx) {
; GCN-LABEL: flat_load_b96_idxprom:
; GCN:       ; %bb.0: ; %entry
; GCN-NEXT:    flat_load_b96 v[0:2], v0, s[0:1] scale_offset
; GCN-NEXT:    s_wait_loadcnt_dscnt 0x0
; GCN-NEXT:    ; return to shader part epilog
entry:
  %idxprom = sext i32 %idx to i64
  %arrayidx = getelementptr inbounds [3 x float], ptr %p, i64 %idxprom
  %ret = load <3 x float>, ptr %arrayidx, align 4
  ret <3 x float> %ret
}

define amdgpu_ps <3 x float> @flat_load_b96_idxpromi_ioffset(ptr align 4 inreg %p, i32 %idx) {
; GCN-LABEL: flat_load_b96_idxpromi_ioffset:
; GCN:       ; %bb.0: ; %entry
; GCN-NEXT:    flat_load_b96 v[0:2], v0, s[0:1] offset:192 scale_offset
; GCN-NEXT:    s_wait_loadcnt_dscnt 0x0
; GCN-NEXT:    ; return to shader part epilog
entry:
  %idxprom = sext i32 %idx to i64
  %idxadd = add i64 %idxprom, 16
  %arrayidx = getelementptr inbounds [3 x float], ptr %p, i64 %idxadd
  %ret = load <3 x float>, ptr %arrayidx, align 4
  ret <3 x float> %ret
}

define amdgpu_ps <4 x float> @flat_load_b128_idxprom(ptr align 4 inreg %p, i32 %idx) {
; GCN-LABEL: flat_load_b128_idxprom:
; GCN:       ; %bb.0: ; %entry
; GCN-NEXT:    flat_load_b128 v[0:3], v0, s[0:1] scale_offset
; GCN-NEXT:    s_wait_loadcnt_dscnt 0x0
; GCN-NEXT:    ; return to shader part epilog
entry:
  %idxprom = sext i32 %idx to i64
  %arrayidx = getelementptr inbounds <4 x float>, ptr %p, i64 %idxprom
  %ret = load <4 x float>, ptr %arrayidx, align 4
  ret <4 x float> %ret
}

define amdgpu_ps float @flat_load_b32_idxprom_range(ptr align 4 inreg %p, ptr align 4 %pp) {
; GCN-LABEL: flat_load_b32_idxprom_range:
; GCN:       ; %bb.0: ; %entry
; GCN-NEXT:    flat_load_b32 v0, v[0:1]
; GCN-NEXT:    s_wait_loadcnt_dscnt 0x0
; GCN-NEXT:    flat_load_b32 v0, v0, s[0:1] scale_offset
; GCN-NEXT:    s_wait_loadcnt_dscnt 0x0
; GCN-NEXT:    ; return to shader part epilog
entry:
  %idx = load i32, ptr %pp, align 4, !range !0
  %idxprom = sext i32 %idx to i64
  %arrayidx = getelementptr inbounds float, ptr %p, i64 %idxprom
  %ret = load float, ptr %arrayidx, align 4
  ret float %ret
}

define amdgpu_ps float @flat_load_b32_idxprom_range_ioffset(ptr align 4 inreg %p, ptr align 4 %pp) {
; GCN-LABEL: flat_load_b32_idxprom_range_ioffset:
; GCN:       ; %bb.0: ; %entry
; GCN-NEXT:    flat_load_b32 v0, v[0:1]
; GCN-NEXT:    s_wait_loadcnt_dscnt 0x0
; GCN-NEXT:    flat_load_b32 v0, v0, s[0:1] offset:64 scale_offset
; GCN-NEXT:    s_wait_loadcnt_dscnt 0x0
; GCN-NEXT:    ; return to shader part epilog
entry:
  %idx = load i32, ptr %pp, align 4, !range !0
  %idxprom = sext i32 %idx to i64
  %idxadd = add i64 %idxprom, 16
  %arrayidx = getelementptr inbounds float, ptr %p, i64 %idxadd
  %ret = load float, ptr %arrayidx, align 4
  ret float %ret
}

; Note: this is a byte load, there is nothing to scale

define amdgpu_ps float @flat_load_b8_idxprom_range_ioffset(ptr align 4 inreg %p, ptr align 4 %pp) {
; GCN-LABEL: flat_load_b8_idxprom_range_ioffset:
; GCN:       ; %bb.0: ; %entry
; GCN-NEXT:    flat_load_b32 v0, v[0:1]
; GCN-NEXT:    s_wait_loadcnt_dscnt 0x0
; GCN-NEXT:    flat_load_u8 v0, v0, s[0:1] offset:16
; GCN-NEXT:    s_wait_loadcnt_dscnt 0x0
; GCN-NEXT:    ; return to shader part epilog
entry:
  %idx = load i32, ptr %pp, align 4, !range !0
  %idxprom = sext i32 %idx to i64
  %idxadd = add i64 %idxprom, 16
  %arrayidx = getelementptr inbounds i8, ptr %p, i64 %idxadd
  %ld = load i8, ptr %arrayidx
  %ret.i32 = zext i8 %ld to i32
  %ret = bitcast i32 %ret.i32 to float
  ret float %ret
}

define amdgpu_ps float @flat_load_b16_idxprom_range(ptr align 4 inreg %p, ptr align 4 %pp) {
; GCN-LABEL: flat_load_b16_idxprom_range:
; GCN:       ; %bb.0: ; %entry
; GCN-NEXT:    flat_load_b32 v0, v[0:1]
; GCN-NEXT:    s_wait_loadcnt_dscnt 0x0
; GCN-NEXT:    flat_load_u16 v0, v0, s[0:1] scale_offset
; GCN-NEXT:    s_wait_loadcnt_dscnt 0x0
; GCN-NEXT:    ; return to shader part epilog
entry:
  %idx = load i32, ptr %pp, align 4, !range !0
  %idxprom = sext i32 %idx to i64
  %arrayidx = getelementptr inbounds i16, ptr %p, i64 %idxprom
  %ld = load i16, ptr %arrayidx, align 2
  %ret.i32 = zext i16 %ld to i32
  %ret = bitcast i32 %ret.i32 to float
  ret float %ret
}

define amdgpu_ps float @flat_load_b16_idxprom_range_ioffset(ptr align 4 inreg %p, ptr align 4 %pp) {
; GCN-LABEL: flat_load_b16_idxprom_range_ioffset:
; GCN:       ; %bb.0: ; %entry
; GCN-NEXT:    flat_load_b32 v0, v[0:1]
; GCN-NEXT:    s_wait_loadcnt_dscnt 0x0
; GCN-NEXT:    flat_load_u16 v0, v0, s[0:1] offset:32 scale_offset
; GCN-NEXT:    s_wait_loadcnt_dscnt 0x0
; GCN-NEXT:    ; return to shader part epilog
entry:
  %idx = load i32, ptr %pp, align 4, !range !0
  %idxprom = sext i32 %idx to i64
  %idxadd = add i64 %idxprom, 16
  %arrayidx = getelementptr inbounds i16, ptr %p, i64 %idxadd
  %ld = load i16, ptr %arrayidx, align 2
  %ret.i32 = zext i16 %ld to i32
  %ret = bitcast i32 %ret.i32 to float
  ret float %ret
}

define amdgpu_ps <2 x float> @flat_load_b64_idxprom_range(ptr align 4 inreg %p, ptr align 4 %pp) {
; GCN-LABEL: flat_load_b64_idxprom_range:
; GCN:       ; %bb.0: ; %entry
; GCN-NEXT:    flat_load_b32 v0, v[0:1]
; GCN-NEXT:    s_wait_loadcnt_dscnt 0x0
; GCN-NEXT:    flat_load_b64 v[0:1], v0, s[0:1] scale_offset
; GCN-NEXT:    s_wait_loadcnt_dscnt 0x0
; GCN-NEXT:    ; return to shader part epilog
entry:
  %idx = load i32, ptr %pp, align 4, !range !0
  %idxprom = sext i32 %idx to i64
  %arrayidx = getelementptr inbounds <2 x float>, ptr %p, i64 %idxprom
  %ret = load <2 x float>, ptr %arrayidx, align 4
  ret <2 x float> %ret
}

define amdgpu_ps <3 x float> @flat_load_b96_idxprom_range(ptr align 4 inreg %p, ptr align 4 %pp) {
; GCN-LABEL: flat_load_b96_idxprom_range:
; GCN:       ; %bb.0: ; %entry
; GCN-NEXT:    flat_load_b32 v0, v[0:1]
; GCN-NEXT:    s_wait_loadcnt_dscnt 0x0
; GCN-NEXT:    flat_load_b96 v[0:2], v0, s[0:1] scale_offset
; GCN-NEXT:    s_wait_loadcnt_dscnt 0x0
; GCN-NEXT:    ; return to shader part epilog
entry:
  %idx = load i32, ptr %pp, align 4, !range !0
  %idxprom = sext i32 %idx to i64
  %arrayidx = getelementptr inbounds [3 x float], ptr %p, i64 %idxprom
  %ret = load <3 x float>, ptr %arrayidx, align 4
  ret <3 x float> %ret
}

define amdgpu_ps <3 x float> @flat_load_b96_idxprom_range_ioffset(ptr align 4 inreg %p, ptr align 4 %pp) {
; GCN-LABEL: flat_load_b96_idxprom_range_ioffset:
; GCN:       ; %bb.0: ; %entry
; GCN-NEXT:    flat_load_b32 v0, v[0:1]
; GCN-NEXT:    s_wait_loadcnt_dscnt 0x0
; GCN-NEXT:    flat_load_b96 v[0:2], v0, s[0:1] offset:192 scale_offset
; GCN-NEXT:    s_wait_loadcnt_dscnt 0x0
; GCN-NEXT:    ; return to shader part epilog
entry:
  %idx = load i32, ptr %pp, align 4, !range !0
  %idxprom = sext i32 %idx to i64
  %idxadd = add i64 %idxprom, 16
  %arrayidx = getelementptr inbounds [3 x float], ptr %p, i64 %idxadd
  %ret = load <3 x float>, ptr %arrayidx, align 4
  ret <3 x float> %ret
}

define amdgpu_ps <4 x float> @flat_load_b128_idxprom_range(ptr align 4 inreg %p, ptr align 4 %pp) {
; GCN-LABEL: flat_load_b128_idxprom_range:
; GCN:       ; %bb.0: ; %entry
; GCN-NEXT:    flat_load_b32 v0, v[0:1]
; GCN-NEXT:    s_wait_loadcnt_dscnt 0x0
; GCN-NEXT:    flat_load_b128 v[0:3], v0, s[0:1] scale_offset
; GCN-NEXT:    s_wait_loadcnt_dscnt 0x0
; GCN-NEXT:    ; return to shader part epilog
entry:
  %idx = load i32, ptr %pp, align 4, !range !0
  %idxprom = sext i32 %idx to i64
  %arrayidx = getelementptr inbounds <4 x float>, ptr %p, i64 %idxprom
  %ret = load <4 x float>, ptr %arrayidx, align 4
  ret <4 x float> %ret
}

define amdgpu_ps void @flat_store_b32_idxprom(ptr align 4 inreg %p, i32 %idx) {
; GCN-LABEL: flat_store_b32_idxprom:
; GCN:       ; %bb.0: ; %entry
; GCN-NEXT:    v_mov_b32_e32 v1, 1.0
; GCN-NEXT:    flat_store_b32 v0, v1, s[0:1] scale_offset
; GCN-NEXT:    s_endpgm
entry:
  %idxprom = sext i32 %idx to i64
  %arrayidx = getelementptr inbounds float, ptr %p, i64 %idxprom
  store float 1.0, ptr %arrayidx, align 4
  ret void
}

define amdgpu_ps void @flat_store_b16_idxprom(ptr align 2 inreg %p, i32 %idx) {
; GCN-LABEL: flat_store_b16_idxprom:
; GCN:       ; %bb.0: ; %entry
; GCN-NEXT:    v_mov_b32_e32 v1, 1
; GCN-NEXT:    flat_store_b16 v0, v1, s[0:1] scale_offset
; GCN-NEXT:    s_endpgm
entry:
  %idxprom = sext i32 %idx to i64
  %arrayidx = getelementptr inbounds i16, ptr %p, i64 %idxprom
  store i16 1, ptr %arrayidx, align 2
  ret void
}

define amdgpu_ps void @flat_store_b64_idxprom(ptr align 4 inreg %p, i32 %idx) {
; GCN-LABEL: flat_store_b64_idxprom:
; GCN:       ; %bb.0: ; %entry
; GCN-NEXT:    v_mov_b64_e32 v[2:3], 1.0
; GCN-NEXT:    flat_store_b64 v0, v[2:3], s[0:1] scale_offset
; GCN-NEXT:    s_endpgm
entry:
  %idxprom = sext i32 %idx to i64
  %arrayidx = getelementptr inbounds double, ptr %p, i64 %idxprom
  store double 1.0, ptr %arrayidx, align 4
  ret void
}

define amdgpu_ps void @flat_atomicrmw_b32_idxprom(ptr align 4 inreg %p, i32 %idx) {
; GCN-LABEL: flat_atomicrmw_b32_idxprom:
; GCN:       ; %bb.0: ; %entry
; GCN-NEXT:    v_mov_b32_e32 v1, 1
; GCN-NEXT:    flat_atomic_add_u32 v0, v1, s[0:1] scale_offset scope:SCOPE_SYS
; GCN-NEXT:    s_endpgm
entry:
  %idxprom = sext i32 %idx to i64
  %arrayidx = getelementptr inbounds i32, ptr %p, i64 %idxprom
  atomicrmw add ptr %arrayidx, i32 1 monotonic
  ret void
}

define amdgpu_ps <2 x float> @flat_atomicrmw_b64_rtn_idxprom(ptr align 8 inreg %p, i32 %idx) {
; SDAG-LABEL: flat_atomicrmw_b64_rtn_idxprom:
; SDAG:       ; %bb.0: ; %entry
; SDAG-NEXT:    v_ashrrev_i32_e32 v1, 31, v0
; SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
; SDAG-NEXT:    v_lshl_add_u64 v[2:3], v[0:1], 3, s[0:1]
; SDAG-NEXT:    s_mov_b64 s[0:1], src_private_base
; SDAG-NEXT:    s_mov_b32 s0, exec_lo
; SDAG-NEXT:    ; implicit-def: $vgpr0_vgpr1
; SDAG-NEXT:    s_wait_alu 0xfffe
; SDAG-NEXT:    v_cmpx_ne_u32_e64 s1, v3
; SDAG-NEXT:    s_xor_b32 s0, exec_lo, s0
; SDAG-NEXT:    s_cbranch_execnz .LBB21_3
; SDAG-NEXT:  ; %bb.1: ; %Flow
; SDAG-NEXT:    s_wait_alu 0xfffe
; SDAG-NEXT:    s_and_not1_saveexec_b32 s0, s0
; SDAG-NEXT:    s_cbranch_execnz .LBB21_4
; SDAG-NEXT:  .LBB21_2: ; %atomicrmw.phi
; SDAG-NEXT:    s_wait_alu 0xfffe
; SDAG-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; SDAG-NEXT:    s_wait_loadcnt_dscnt 0x0
; SDAG-NEXT:    s_branch .LBB21_5
; SDAG-NEXT:  .LBB21_3: ; %atomicrmw.global
; SDAG-NEXT:    v_mov_b64_e32 v[0:1], 1
; SDAG-NEXT:    flat_atomic_add_u64 v[0:1], v[2:3], v[0:1] th:TH_ATOMIC_RETURN scope:SCOPE_SYS
; SDAG-NEXT:    ; implicit-def: $vgpr2_vgpr3
; SDAG-NEXT:    s_wait_xcnt 0x0
; SDAG-NEXT:    s_wait_alu 0xfffe
; SDAG-NEXT:    s_and_not1_saveexec_b32 s0, s0
; SDAG-NEXT:    s_cbranch_execz .LBB21_2
; SDAG-NEXT:  .LBB21_4: ; %atomicrmw.private
; SDAG-NEXT:    v_cmp_ne_u64_e32 vcc_lo, 0, v[2:3]
; SDAG-NEXT:    v_cndmask_b32_e32 v4, -1, v2, vcc_lo
; SDAG-NEXT:    s_wait_loadcnt_dscnt 0x0
; SDAG-NEXT:    scratch_load_b64 v[0:1], v4, off
; SDAG-NEXT:    s_wait_loadcnt 0x0
; SDAG-NEXT:    v_lshl_add_u64 v[2:3], v[0:1], 0, 1
; SDAG-NEXT:    scratch_store_b64 v4, v[2:3], off
; SDAG-NEXT:    s_wait_xcnt 0x0
; SDAG-NEXT:    s_wait_alu 0xfffe
; SDAG-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; SDAG-NEXT:    s_branch .LBB21_5
; SDAG-NEXT:  .LBB21_5:
;
; GISEL-LABEL: flat_atomicrmw_b64_rtn_idxprom:
; GISEL:       ; %bb.0: ; %entry
; GISEL-NEXT:    v_mov_b32_e32 v2, v0
; GISEL-NEXT:    v_mov_b64_e32 v[4:5], s[0:1]
; GISEL-NEXT:    s_mov_b64 s[2:3], src_private_base
; GISEL-NEXT:    s_mov_b32 s2, exec_lo
; GISEL-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GISEL-NEXT:    v_ashrrev_i32_e32 v3, 31, v2
; GISEL-NEXT:    v_lshlrev_b64_e32 v[0:1], 3, v[2:3]
; GISEL-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GISEL-NEXT:    v_add_co_u32 v4, vcc_lo, v4, v0
; GISEL-NEXT:    v_add_co_ci_u32_e64 v5, null, v5, v1, vcc_lo
; GISEL-NEXT:    ; implicit-def: $vgpr0_vgpr1
; GISEL-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GISEL-NEXT:    v_cmpx_ne_u32_e64 s3, v5
; GISEL-NEXT:    s_xor_b32 s2, exec_lo, s2
; GISEL-NEXT:    s_cbranch_execnz .LBB21_3
; GISEL-NEXT:  ; %bb.1: ; %Flow
; GISEL-NEXT:    s_wait_alu 0xfffe
; GISEL-NEXT:    s_and_not1_saveexec_b32 s0, s2
; GISEL-NEXT:    s_cbranch_execnz .LBB21_4
; GISEL-NEXT:  .LBB21_2: ; %atomicrmw.phi
; GISEL-NEXT:    s_wait_alu 0xfffe
; GISEL-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GISEL-NEXT:    s_wait_loadcnt_dscnt 0x0
; GISEL-NEXT:    s_branch .LBB21_5
; GISEL-NEXT:  .LBB21_3: ; %atomicrmw.global
; GISEL-NEXT:    v_mov_b64_e32 v[0:1], 1
; GISEL-NEXT:    ; implicit-def: $vgpr4_vgpr5
; GISEL-NEXT:    flat_atomic_add_u64 v[0:1], v2, v[0:1], s[0:1] scale_offset th:TH_ATOMIC_RETURN scope:SCOPE_SYS
; GISEL-NEXT:    s_wait_xcnt 0x0
; GISEL-NEXT:    s_wait_alu 0xfffe
; GISEL-NEXT:    s_and_not1_saveexec_b32 s0, s2
; GISEL-NEXT:    s_cbranch_execz .LBB21_2
; GISEL-NEXT:  .LBB21_4: ; %atomicrmw.private
; GISEL-NEXT:    v_cmp_ne_u64_e32 vcc_lo, 0, v[4:5]
; GISEL-NEXT:    s_wait_alu 0xfffd
; GISEL-NEXT:    v_cndmask_b32_e32 v4, -1, v4, vcc_lo
; GISEL-NEXT:    s_wait_loadcnt_dscnt 0x0
; GISEL-NEXT:    scratch_load_b64 v[0:1], v4, off
; GISEL-NEXT:    s_wait_loadcnt 0x0
; GISEL-NEXT:    v_lshl_add_u64 v[2:3], v[0:1], 0, 1
; GISEL-NEXT:    scratch_store_b64 v4, v[2:3], off
; GISEL-NEXT:    s_wait_xcnt 0x0
; GISEL-NEXT:    s_wait_alu 0xfffe
; GISEL-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GISEL-NEXT:    s_branch .LBB21_5
; GISEL-NEXT:  .LBB21_5:
entry:
  %idxprom = sext i32 %idx to i64
  %arrayidx = getelementptr inbounds i64, ptr %p, i64 %idxprom
  %ret = atomicrmw add ptr %arrayidx, i64 1 monotonic
  %ret.cast = bitcast i64 %ret to <2 x float>
  ret <2 x float> %ret.cast
}

!0 = !{i32 0, i32 1024}