; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 2
; RUN: llc -mtriple=amdgcn -mcpu=tahiti < %s | FileCheck --check-prefixes=GFX6 %s
; RUN: llc -mtriple=amdgcn -mcpu=tonga -mattr=-flat-for-global < %s | FileCheck --check-prefixes=GFX8 %s
; RUN: llc -mtriple=amdgcn -mcpu=gfx900 -mattr=-flat-for-global < %s | FileCheck --check-prefixes=GFX9 %s
; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 -mattr=-flat-for-global < %s | FileCheck --check-prefixes=GFX11,GFX11-TRUE16 %s
; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 -mattr=-flat-for-global < %s | FileCheck --check-prefixes=GFX11,GFX11-FAKE16 %s
; RUN: llc -mtriple=r600 -mcpu=redwood < %s | FileCheck --check-prefixes=R600 %s

define amdgpu_kernel void @round_f32(ptr addrspace(1) %out, float %x) #0 {
; GFX6-LABEL: round_f32:
; GFX6:       ; %bb.0:
; GFX6-NEXT:    s_load_dword s6, s[4:5], 0xb
; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
; GFX6-NEXT:    s_mov_b32 s3, 0xf000
; GFX6-NEXT:    s_mov_b32 s2, -1
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_trunc_f32_e32 v0, s6
; GFX6-NEXT:    v_sub_f32_e32 v1, s6, v0
; GFX6-NEXT:    v_cmp_ge_f32_e64 s[4:5], |v1|, 0.5
; GFX6-NEXT:    v_cndmask_b32_e64 v1, 0, 1.0, s[4:5]
; GFX6-NEXT:    s_brev_b32 s4, -2
; GFX6-NEXT:    v_mov_b32_e32 v2, s6
; GFX6-NEXT:    v_bfi_b32 v1, s4, v1, v2
; GFX6-NEXT:    v_add_f32_e32 v0, v0, v1
; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; GFX6-NEXT:    s_endpgm
;
; GFX8-LABEL: round_f32:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_load_dword s6, s[4:5], 0x2c
; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GFX8-NEXT:    s_mov_b32 s3, 0xf000
; GFX8-NEXT:    s_mov_b32 s2, -1
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_trunc_f32_e32 v0, s6
; GFX8-NEXT:    v_sub_f32_e32 v1, s6, v0
; GFX8-NEXT:    v_cmp_ge_f32_e64 s[4:5], |v1|, 0.5
; GFX8-NEXT:    v_cndmask_b32_e64 v1, 0, 1.0, s[4:5]
; GFX8-NEXT:    s_brev_b32 s4, -2
; GFX8-NEXT:    v_mov_b32_e32 v2, s6
; GFX8-NEXT:    v_bfi_b32 v1, s4, v1, v2
; GFX8-NEXT:    v_add_f32_e32 v0, v0, v1
; GFX8-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; GFX8-NEXT:    s_endpgm
;
; GFX9-LABEL: round_f32:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_load_dword s6, s[4:5], 0x2c
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GFX9-NEXT:    s_mov_b32 s3, 0xf000
; GFX9-NEXT:    s_mov_b32 s2, -1
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_trunc_f32_e32 v0, s6
; GFX9-NEXT:    v_sub_f32_e32 v1, s6, v0
; GFX9-NEXT:    v_cmp_ge_f32_e64 s[4:5], |v1|, 0.5
; GFX9-NEXT:    v_cndmask_b32_e64 v1, 0, 1.0, s[4:5]
; GFX9-NEXT:    s_brev_b32 s4, -2
; GFX9-NEXT:    v_mov_b32_e32 v2, s6
; GFX9-NEXT:    v_bfi_b32 v1, s4, v1, v2
; GFX9-NEXT:    v_add_f32_e32 v0, v0, v1
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; GFX9-NEXT:    s_endpgm
;
; GFX11-LABEL: round_f32:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x2c
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_trunc_f32_e32 v0, s2
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_sub_f32_e32 v1, s2, v0
; GFX11-NEXT:    v_cmp_ge_f32_e64 s3, |v1|, 0.5
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_cndmask_b32_e64 v1, 0, 1.0, s3
; GFX11-NEXT:    s_mov_b32 s3, 0x31016000
; GFX11-NEXT:    v_bfi_b32 v1, 0x7fffffff, v1, s2
; GFX11-NEXT:    s_mov_b32 s2, -1
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT:    v_add_f32_e32 v0, v0, v1
; GFX11-NEXT:    buffer_store_b32 v0, off, s[0:3], 0
; GFX11-NEXT:    s_endpgm
;
; R600-LABEL: round_f32:
; R600:       ; %bb.0:
; R600-NEXT:    ALU 7, @4, KC0[CB0:0-32], KC1[]
; R600-NEXT:    MEM_RAT_CACHELESS STORE_RAW T0.X, T1.X, 1
; R600-NEXT:    CF_END
; R600-NEXT:    PAD
; R600-NEXT:    ALU clause starting at 4:
; R600-NEXT:     TRUNC * T0.W, KC0[2].Z,
; R600-NEXT:     ADD * T1.W, KC0[2].Z, -PV.W,
; R600-NEXT:     SETGE * T1.W, |PV.W|, 0.5,
; R600-NEXT:     BFI_INT * T1.W, literal.x, PV.W, KC0[2].Z,
; R600-NEXT:    2147483647(nan), 0(0.000000e+00)
; R600-NEXT:     ADD T0.X, T0.W, PV.W,
; R600-NEXT:     LSHR * T1.X, KC0[2].Y, literal.x,
; R600-NEXT:    2(2.802597e-45), 0(0.000000e+00)
  %result = call float @llvm.round.f32(float %x) #1
  store float %result, ptr addrspace(1) %out
  ret void
}

; The vector tests are really difficult to verify, since it can be hard to
; predict how the scheduler will order the instructions.  We already have
; a test for the scalar case, so the vector tests just check that the
; compiler doesn't crash.
define amdgpu_kernel void @round_v2f32(ptr addrspace(1) %out, <2 x float> %in) #0 {
; GFX6-LABEL: round_v2f32:
; GFX6:       ; %bb.0:
; GFX6-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GFX6-NEXT:    s_brev_b32 s8, -2
; GFX6-NEXT:    s_mov_b32 s7, 0xf000
; GFX6-NEXT:    s_mov_b32 s6, -1
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_trunc_f32_e32 v0, s3
; GFX6-NEXT:    v_sub_f32_e32 v1, s3, v0
; GFX6-NEXT:    s_mov_b32 s4, s0
; GFX6-NEXT:    s_mov_b32 s5, s1
; GFX6-NEXT:    v_cmp_ge_f32_e64 s[0:1], |v1|, 0.5
; GFX6-NEXT:    v_cndmask_b32_e64 v1, 0, 1.0, s[0:1]
; GFX6-NEXT:    v_mov_b32_e32 v2, s3
; GFX6-NEXT:    v_bfi_b32 v1, s8, v1, v2
; GFX6-NEXT:    v_add_f32_e32 v1, v0, v1
; GFX6-NEXT:    v_trunc_f32_e32 v0, s2
; GFX6-NEXT:    v_sub_f32_e32 v2, s2, v0
; GFX6-NEXT:    v_cmp_ge_f32_e64 s[0:1], |v2|, 0.5
; GFX6-NEXT:    v_cndmask_b32_e64 v2, 0, 1.0, s[0:1]
; GFX6-NEXT:    v_mov_b32_e32 v3, s2
; GFX6-NEXT:    v_bfi_b32 v2, s8, v2, v3
; GFX6-NEXT:    v_add_f32_e32 v0, v0, v2
; GFX6-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
; GFX6-NEXT:    s_endpgm
;
; GFX8-LABEL: round_v2f32:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GFX8-NEXT:    s_brev_b32 s8, -2
; GFX8-NEXT:    s_mov_b32 s7, 0xf000
; GFX8-NEXT:    s_mov_b32 s6, -1
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_trunc_f32_e32 v0, s3
; GFX8-NEXT:    v_sub_f32_e32 v1, s3, v0
; GFX8-NEXT:    s_mov_b32 s4, s0
; GFX8-NEXT:    s_mov_b32 s5, s1
; GFX8-NEXT:    v_cmp_ge_f32_e64 s[0:1], |v1|, 0.5
; GFX8-NEXT:    v_cndmask_b32_e64 v1, 0, 1.0, s[0:1]
; GFX8-NEXT:    v_mov_b32_e32 v2, s3
; GFX8-NEXT:    v_bfi_b32 v1, s8, v1, v2
; GFX8-NEXT:    v_add_f32_e32 v1, v0, v1
; GFX8-NEXT:    v_trunc_f32_e32 v0, s2
; GFX8-NEXT:    v_sub_f32_e32 v2, s2, v0
; GFX8-NEXT:    v_cmp_ge_f32_e64 s[0:1], |v2|, 0.5
; GFX8-NEXT:    v_cndmask_b32_e64 v2, 0, 1.0, s[0:1]
; GFX8-NEXT:    v_mov_b32_e32 v3, s2
; GFX8-NEXT:    v_bfi_b32 v2, s8, v2, v3
; GFX8-NEXT:    v_add_f32_e32 v0, v0, v2
; GFX8-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
; GFX8-NEXT:    s_endpgm
;
; GFX9-LABEL: round_v2f32:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GFX9-NEXT:    s_brev_b32 s8, -2
; GFX9-NEXT:    s_mov_b32 s7, 0xf000
; GFX9-NEXT:    s_mov_b32 s6, -1
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_trunc_f32_e32 v0, s3
; GFX9-NEXT:    v_sub_f32_e32 v1, s3, v0
; GFX9-NEXT:    s_mov_b32 s4, s0
; GFX9-NEXT:    s_mov_b32 s5, s1
; GFX9-NEXT:    v_cmp_ge_f32_e64 s[0:1], |v1|, 0.5
; GFX9-NEXT:    v_cndmask_b32_e64 v1, 0, 1.0, s[0:1]
; GFX9-NEXT:    v_mov_b32_e32 v2, s3
; GFX9-NEXT:    v_bfi_b32 v1, s8, v1, v2
; GFX9-NEXT:    v_add_f32_e32 v1, v0, v1
; GFX9-NEXT:    v_trunc_f32_e32 v0, s2
; GFX9-NEXT:    v_sub_f32_e32 v2, s2, v0
; GFX9-NEXT:    v_cmp_ge_f32_e64 s[0:1], |v2|, 0.5
; GFX9-NEXT:    v_cndmask_b32_e64 v2, 0, 1.0, s[0:1]
; GFX9-NEXT:    v_mov_b32_e32 v3, s2
; GFX9-NEXT:    v_bfi_b32 v2, s8, v2, v3
; GFX9-NEXT:    v_add_f32_e32 v0, v0, v2
; GFX9-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
; GFX9-NEXT:    s_endpgm
;
; GFX11-LABEL: round_v2f32:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_trunc_f32_e32 v0, s3
; GFX11-NEXT:    v_trunc_f32_e32 v2, s2
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT:    v_sub_f32_e32 v1, s3, v0
; GFX11-NEXT:    v_sub_f32_e32 v3, s2, v2
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_cmp_ge_f32_e64 s4, |v1|, 0.5
; GFX11-NEXT:    v_cndmask_b32_e64 v1, 0, 1.0, s4
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT:    v_cmp_ge_f32_e64 s4, |v3|, 0.5
; GFX11-NEXT:    v_bfi_b32 v1, 0x7fffffff, v1, s3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-NEXT:    v_cndmask_b32_e64 v3, 0, 1.0, s4
; GFX11-NEXT:    s_mov_b32 s3, 0x31016000
; GFX11-NEXT:    v_add_f32_e32 v1, v0, v1
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_bfi_b32 v3, 0x7fffffff, v3, s2
; GFX11-NEXT:    s_mov_b32 s2, -1
; GFX11-NEXT:    v_add_f32_e32 v0, v2, v3
; GFX11-NEXT:    buffer_store_b64 v[0:1], off, s[0:3], 0
; GFX11-NEXT:    s_endpgm
;
; R600-LABEL: round_v2f32:
; R600:       ; %bb.0:
; R600-NEXT:    ALU 13, @4, KC0[CB0:0-32], KC1[]
; R600-NEXT:    MEM_RAT_CACHELESS STORE_RAW T0.XY, T1.X, 1
; R600-NEXT:    CF_END
; R600-NEXT:    PAD
; R600-NEXT:    ALU clause starting at 4:
; R600-NEXT:     TRUNC * T0.W, KC0[3].X,
; R600-NEXT:     ADD T1.W, KC0[3].X, -PV.W,
; R600-NEXT:     TRUNC * T2.W, KC0[2].W,
; R600-NEXT:     ADD T3.W, KC0[2].W, -PS,
; R600-NEXT:     SETGE * T1.W, |PV.W|, 0.5,
; R600-NEXT:     BFI_INT T1.W, literal.x, PS, KC0[3].X,
; R600-NEXT:     SETGE * T3.W, |PV.W|, 0.5,
; R600-NEXT:    2147483647(nan), 0(0.000000e+00)
; R600-NEXT:     ADD T0.Y, T0.W, PV.W,
; R600-NEXT:     BFI_INT * T0.W, literal.x, PS, KC0[2].W,
; R600-NEXT:    2147483647(nan), 0(0.000000e+00)
; R600-NEXT:     ADD T0.X, T2.W, PV.W,
; R600-NEXT:     LSHR * T1.X, KC0[2].Y, literal.x,
; R600-NEXT:    2(2.802597e-45), 0(0.000000e+00)
  %result = call <2 x float> @llvm.round.v2f32(<2 x float> %in) #1
  store <2 x float> %result, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @round_v4f32(ptr addrspace(1) %out, <4 x float> %in) #0 {
; GFX6-LABEL: round_v4f32:
; GFX6:       ; %bb.0:
; GFX6-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0xd
; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0x9
; GFX6-NEXT:    s_brev_b32 s10, -2
; GFX6-NEXT:    s_mov_b32 s7, 0xf000
; GFX6-NEXT:    s_mov_b32 s6, -1
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_trunc_f32_e32 v0, s3
; GFX6-NEXT:    v_sub_f32_e32 v1, s3, v0
; GFX6-NEXT:    v_cmp_ge_f32_e64 s[8:9], |v1|, 0.5
; GFX6-NEXT:    v_cndmask_b32_e64 v1, 0, 1.0, s[8:9]
; GFX6-NEXT:    v_mov_b32_e32 v2, s3
; GFX6-NEXT:    v_bfi_b32 v1, s10, v1, v2
; GFX6-NEXT:    v_add_f32_e32 v3, v0, v1
; GFX6-NEXT:    v_trunc_f32_e32 v0, s2
; GFX6-NEXT:    v_sub_f32_e32 v1, s2, v0
; GFX6-NEXT:    v_cmp_ge_f32_e64 s[8:9], |v1|, 0.5
; GFX6-NEXT:    v_cndmask_b32_e64 v1, 0, 1.0, s[8:9]
; GFX6-NEXT:    v_mov_b32_e32 v2, s2
; GFX6-NEXT:    v_bfi_b32 v1, s10, v1, v2
; GFX6-NEXT:    v_add_f32_e32 v2, v0, v1
; GFX6-NEXT:    v_trunc_f32_e32 v0, s1
; GFX6-NEXT:    v_sub_f32_e32 v1, s1, v0
; GFX6-NEXT:    v_cmp_ge_f32_e64 s[2:3], |v1|, 0.5
; GFX6-NEXT:    v_cndmask_b32_e64 v1, 0, 1.0, s[2:3]
; GFX6-NEXT:    v_mov_b32_e32 v4, s1
; GFX6-NEXT:    v_bfi_b32 v1, s10, v1, v4
; GFX6-NEXT:    v_add_f32_e32 v1, v0, v1
; GFX6-NEXT:    v_trunc_f32_e32 v0, s0
; GFX6-NEXT:    v_sub_f32_e32 v4, s0, v0
; GFX6-NEXT:    v_cmp_ge_f32_e64 s[2:3], |v4|, 0.5
; GFX6-NEXT:    v_cndmask_b32_e64 v4, 0, 1.0, s[2:3]
; GFX6-NEXT:    v_mov_b32_e32 v5, s0
; GFX6-NEXT:    v_bfi_b32 v4, s10, v4, v5
; GFX6-NEXT:    v_add_f32_e32 v0, v0, v4
; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], off, s[4:7], 0
; GFX6-NEXT:    s_endpgm
;
; GFX8-LABEL: round_v4f32:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x34
; GFX8-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0x24
; GFX8-NEXT:    s_brev_b32 s10, -2
; GFX8-NEXT:    s_mov_b32 s7, 0xf000
; GFX8-NEXT:    s_mov_b32 s6, -1
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_trunc_f32_e32 v0, s3
; GFX8-NEXT:    v_sub_f32_e32 v1, s3, v0
; GFX8-NEXT:    v_cmp_ge_f32_e64 s[8:9], |v1|, 0.5
; GFX8-NEXT:    v_cndmask_b32_e64 v1, 0, 1.0, s[8:9]
; GFX8-NEXT:    v_mov_b32_e32 v2, s3
; GFX8-NEXT:    v_bfi_b32 v1, s10, v1, v2
; GFX8-NEXT:    v_add_f32_e32 v3, v0, v1
; GFX8-NEXT:    v_trunc_f32_e32 v0, s2
; GFX8-NEXT:    v_sub_f32_e32 v1, s2, v0
; GFX8-NEXT:    v_cmp_ge_f32_e64 s[8:9], |v1|, 0.5
; GFX8-NEXT:    v_cndmask_b32_e64 v1, 0, 1.0, s[8:9]
; GFX8-NEXT:    v_mov_b32_e32 v2, s2
; GFX8-NEXT:    v_bfi_b32 v1, s10, v1, v2
; GFX8-NEXT:    v_add_f32_e32 v2, v0, v1
; GFX8-NEXT:    v_trunc_f32_e32 v0, s1
; GFX8-NEXT:    v_sub_f32_e32 v1, s1, v0
; GFX8-NEXT:    v_cmp_ge_f32_e64 s[2:3], |v1|, 0.5
; GFX8-NEXT:    v_cndmask_b32_e64 v1, 0, 1.0, s[2:3]
; GFX8-NEXT:    v_mov_b32_e32 v4, s1
; GFX8-NEXT:    v_bfi_b32 v1, s10, v1, v4
; GFX8-NEXT:    v_add_f32_e32 v1, v0, v1
; GFX8-NEXT:    v_trunc_f32_e32 v0, s0
; GFX8-NEXT:    v_sub_f32_e32 v4, s0, v0
; GFX8-NEXT:    v_cmp_ge_f32_e64 s[2:3], |v4|, 0.5
; GFX8-NEXT:    v_cndmask_b32_e64 v4, 0, 1.0, s[2:3]
; GFX8-NEXT:    v_mov_b32_e32 v5, s0
; GFX8-NEXT:    v_bfi_b32 v4, s10, v4, v5
; GFX8-NEXT:    v_add_f32_e32 v0, v0, v4
; GFX8-NEXT:    buffer_store_dwordx4 v[0:3], off, s[4:7], 0
; GFX8-NEXT:    s_endpgm
;
; GFX9-LABEL: round_v4f32:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x34
; GFX9-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x24
; GFX9-NEXT:    s_brev_b32 s6, -2
; GFX9-NEXT:    s_mov_b32 s11, 0xf000
; GFX9-NEXT:    s_mov_b32 s10, -1
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_trunc_f32_e32 v0, s3
; GFX9-NEXT:    v_sub_f32_e32 v1, s3, v0
; GFX9-NEXT:    v_cmp_ge_f32_e64 s[4:5], |v1|, 0.5
; GFX9-NEXT:    v_cndmask_b32_e64 v1, 0, 1.0, s[4:5]
; GFX9-NEXT:    v_mov_b32_e32 v2, s3
; GFX9-NEXT:    v_bfi_b32 v1, s6, v1, v2
; GFX9-NEXT:    v_add_f32_e32 v3, v0, v1
; GFX9-NEXT:    v_trunc_f32_e32 v0, s2
; GFX9-NEXT:    v_sub_f32_e32 v1, s2, v0
; GFX9-NEXT:    v_cmp_ge_f32_e64 s[4:5], |v1|, 0.5
; GFX9-NEXT:    v_cndmask_b32_e64 v1, 0, 1.0, s[4:5]
; GFX9-NEXT:    v_mov_b32_e32 v2, s2
; GFX9-NEXT:    v_bfi_b32 v1, s6, v1, v2
; GFX9-NEXT:    v_add_f32_e32 v2, v0, v1
; GFX9-NEXT:    v_trunc_f32_e32 v0, s1
; GFX9-NEXT:    v_sub_f32_e32 v1, s1, v0
; GFX9-NEXT:    v_cmp_ge_f32_e64 s[2:3], |v1|, 0.5
; GFX9-NEXT:    v_cndmask_b32_e64 v1, 0, 1.0, s[2:3]
; GFX9-NEXT:    v_mov_b32_e32 v4, s1
; GFX9-NEXT:    v_bfi_b32 v1, s6, v1, v4
; GFX9-NEXT:    v_add_f32_e32 v1, v0, v1
; GFX9-NEXT:    v_trunc_f32_e32 v0, s0
; GFX9-NEXT:    v_sub_f32_e32 v4, s0, v0
; GFX9-NEXT:    v_cmp_ge_f32_e64 s[2:3], |v4|, 0.5
; GFX9-NEXT:    v_cndmask_b32_e64 v4, 0, 1.0, s[2:3]
; GFX9-NEXT:    v_mov_b32_e32 v5, s0
; GFX9-NEXT:    v_bfi_b32 v4, s6, v4, v5
; GFX9-NEXT:    v_add_f32_e32 v0, v0, v4
; GFX9-NEXT:    buffer_store_dwordx4 v[0:3], off, s[8:11], 0
; GFX9-NEXT:    s_endpgm
;
; GFX11-LABEL: round_v4f32:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x34
; GFX11-NEXT:    s_load_b64 s[4:5], s[4:5], 0x24
; GFX11-NEXT:    s_mov_b32 s7, 0x31016000
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_trunc_f32_e32 v0, s3
; GFX11-NEXT:    v_trunc_f32_e32 v1, s2
; GFX11-NEXT:    v_trunc_f32_e32 v4, s1
; GFX11-NEXT:    v_trunc_f32_e32 v5, s0
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT:    v_dual_sub_f32 v2, s3, v0 :: v_dual_sub_f32 v3, s2, v1
; GFX11-NEXT:    v_dual_sub_f32 v6, s1, v4 :: v_dual_sub_f32 v7, s0, v5
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_cmp_ge_f32_e64 s6, |v2|, 0.5
; GFX11-NEXT:    v_cndmask_b32_e64 v2, 0, 1.0, s6
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT:    v_cmp_ge_f32_e64 s6, |v3|, 0.5
; GFX11-NEXT:    v_bfi_b32 v2, 0x7fffffff, v2, s3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-NEXT:    v_cndmask_b32_e64 v3, 0, 1.0, s6
; GFX11-NEXT:    v_cmp_ge_f32_e64 s6, |v6|, 0.5
; GFX11-NEXT:    v_bfi_b32 v8, 0x7fffffff, v3, s2
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_cndmask_b32_e64 v6, 0, 1.0, s6
; GFX11-NEXT:    v_cmp_ge_f32_e64 s6, |v7|, 0.5
; GFX11-NEXT:    v_dual_add_f32 v3, v0, v2 :: v_dual_add_f32 v2, v1, v8
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_bfi_b32 v6, 0x7fffffff, v6, s1
; GFX11-NEXT:    v_cndmask_b32_e64 v7, 0, 1.0, s6
; GFX11-NEXT:    s_mov_b32 s6, -1
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_bfi_b32 v7, 0x7fffffff, v7, s0
; GFX11-NEXT:    v_dual_add_f32 v1, v4, v6 :: v_dual_add_f32 v0, v5, v7
; GFX11-NEXT:    buffer_store_b128 v[0:3], off, s[4:7], 0
; GFX11-NEXT:    s_endpgm
;
; R600-LABEL: round_v4f32:
; R600:       ; %bb.0:
; R600-NEXT:    ALU 25, @4, KC0[CB0:0-32], KC1[]
; R600-NEXT:    MEM_RAT_CACHELESS STORE_RAW T4.XYZW, T0.X, 1
; R600-NEXT:    CF_END
; R600-NEXT:    PAD
; R600-NEXT:    ALU clause starting at 4:
; R600-NEXT:     TRUNC * T0.W, KC0[4].X,
; R600-NEXT:     ADD T1.W, KC0[4].X, -PV.W,
; R600-NEXT:     TRUNC * T2.W, KC0[3].W,
; R600-NEXT:     TRUNC T0.Z, KC0[3].Z,
; R600-NEXT:     ADD T3.W, KC0[3].W, -PS,
; R600-NEXT:     SETGE * T1.W, |PV.W|, 0.5,
; R600-NEXT:     BFI_INT T0.Y, literal.x, PS, KC0[4].X,
; R600-NEXT:     SETGE T1.Z, |PV.W|, 0.5,
; R600-NEXT:     ADD * T1.W, KC0[3].Z, -PV.Z,
; R600-NEXT:    2147483647(nan), 0(0.000000e+00)
; R600-NEXT:     TRUNC * T3.W, KC0[3].Y,
; R600-NEXT:     ADD T1.Y, KC0[3].Y, -PV.W,
; R600-NEXT:     SETGE T2.Z, |T1.W|, 0.5,
; R600-NEXT:     BFI_INT T1.W, literal.x, T1.Z, KC0[3].W,
; R600-NEXT:     ADD * T4.W, T0.W, T0.Y,
; R600-NEXT:    2147483647(nan), 0(0.000000e+00)
; R600-NEXT:     ADD T4.Z, T2.W, PV.W,
; R600-NEXT:     BFI_INT T0.W, literal.x, PV.Z, KC0[3].Z,
; R600-NEXT:     SETGE * T1.W, |PV.Y|, 0.5,
; R600-NEXT:    2147483647(nan), 0(0.000000e+00)
; R600-NEXT:     ADD T4.Y, T0.Z, PV.W,
; R600-NEXT:     BFI_INT * T0.W, literal.x, PS, KC0[3].Y,
; R600-NEXT:    2147483647(nan), 0(0.000000e+00)
; R600-NEXT:     ADD T4.X, T3.W, PV.W,
; R600-NEXT:     LSHR * T0.X, KC0[2].Y, literal.x,
; R600-NEXT:    2(2.802597e-45), 0(0.000000e+00)
  %result = call <4 x float> @llvm.round.v4f32(<4 x float> %in) #1
  store <4 x float> %result, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @round_v8f32(ptr addrspace(1) %out, <8 x float> %in) #0 {
; GFX6-LABEL: round_v8f32:
; GFX6:       ; %bb.0:
; GFX6-NEXT:    s_load_dwordx8 s[8:15], s[4:5], 0x11
; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
; GFX6-NEXT:    s_brev_b32 s6, -2
; GFX6-NEXT:    s_mov_b32 s3, 0xf000
; GFX6-NEXT:    s_mov_b32 s2, -1
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_trunc_f32_e32 v0, s11
; GFX6-NEXT:    v_sub_f32_e32 v1, s11, v0
; GFX6-NEXT:    v_cmp_ge_f32_e64 s[4:5], |v1|, 0.5
; GFX6-NEXT:    v_cndmask_b32_e64 v1, 0, 1.0, s[4:5]
; GFX6-NEXT:    v_mov_b32_e32 v2, s11
; GFX6-NEXT:    v_bfi_b32 v1, s6, v1, v2
; GFX6-NEXT:    v_add_f32_e32 v3, v0, v1
; GFX6-NEXT:    v_trunc_f32_e32 v0, s10
; GFX6-NEXT:    v_sub_f32_e32 v1, s10, v0
; GFX6-NEXT:    v_cmp_ge_f32_e64 s[4:5], |v1|, 0.5
; GFX6-NEXT:    v_cndmask_b32_e64 v1, 0, 1.0, s[4:5]
; GFX6-NEXT:    v_mov_b32_e32 v2, s10
; GFX6-NEXT:    v_bfi_b32 v1, s6, v1, v2
; GFX6-NEXT:    v_add_f32_e32 v2, v0, v1
; GFX6-NEXT:    v_trunc_f32_e32 v0, s9
; GFX6-NEXT:    v_sub_f32_e32 v1, s9, v0
; GFX6-NEXT:    v_cmp_ge_f32_e64 s[4:5], |v1|, 0.5
; GFX6-NEXT:    v_cndmask_b32_e64 v1, 0, 1.0, s[4:5]
; GFX6-NEXT:    v_mov_b32_e32 v4, s9
; GFX6-NEXT:    v_bfi_b32 v1, s6, v1, v4
; GFX6-NEXT:    v_add_f32_e32 v1, v0, v1
; GFX6-NEXT:    v_trunc_f32_e32 v0, s8
; GFX6-NEXT:    v_sub_f32_e32 v4, s8, v0
; GFX6-NEXT:    v_cmp_ge_f32_e64 s[4:5], |v4|, 0.5
; GFX6-NEXT:    v_cndmask_b32_e64 v4, 0, 1.0, s[4:5]
; GFX6-NEXT:    v_mov_b32_e32 v5, s8
; GFX6-NEXT:    v_bfi_b32 v4, s6, v4, v5
; GFX6-NEXT:    v_add_f32_e32 v0, v0, v4
; GFX6-NEXT:    v_trunc_f32_e32 v4, s15
; GFX6-NEXT:    v_sub_f32_e32 v5, s15, v4
; GFX6-NEXT:    v_cmp_ge_f32_e64 s[4:5], |v5|, 0.5
; GFX6-NEXT:    v_cndmask_b32_e64 v5, 0, 1.0, s[4:5]
; GFX6-NEXT:    v_mov_b32_e32 v6, s15
; GFX6-NEXT:    v_bfi_b32 v5, s6, v5, v6
; GFX6-NEXT:    v_add_f32_e32 v7, v4, v5
; GFX6-NEXT:    v_trunc_f32_e32 v4, s14
; GFX6-NEXT:    v_sub_f32_e32 v5, s14, v4
; GFX6-NEXT:    v_cmp_ge_f32_e64 s[4:5], |v5|, 0.5
; GFX6-NEXT:    v_cndmask_b32_e64 v5, 0, 1.0, s[4:5]
; GFX6-NEXT:    v_mov_b32_e32 v6, s14
; GFX6-NEXT:    v_bfi_b32 v5, s6, v5, v6
; GFX6-NEXT:    v_add_f32_e32 v6, v4, v5
; GFX6-NEXT:    v_trunc_f32_e32 v4, s13
; GFX6-NEXT:    v_sub_f32_e32 v5, s13, v4
; GFX6-NEXT:    v_cmp_ge_f32_e64 s[4:5], |v5|, 0.5
; GFX6-NEXT:    v_cndmask_b32_e64 v5, 0, 1.0, s[4:5]
; GFX6-NEXT:    v_mov_b32_e32 v8, s13
; GFX6-NEXT:    v_bfi_b32 v5, s6, v5, v8
; GFX6-NEXT:    v_add_f32_e32 v5, v4, v5
; GFX6-NEXT:    v_trunc_f32_e32 v4, s12
; GFX6-NEXT:    v_sub_f32_e32 v8, s12, v4
; GFX6-NEXT:    v_cmp_ge_f32_e64 s[4:5], |v8|, 0.5
; GFX6-NEXT:    v_cndmask_b32_e64 v8, 0, 1.0, s[4:5]
; GFX6-NEXT:    v_mov_b32_e32 v9, s12
; GFX6-NEXT:    v_bfi_b32 v8, s6, v8, v9
; GFX6-NEXT:    v_add_f32_e32 v4, v4, v8
; GFX6-NEXT:    buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:16
; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], off, s[0:3], 0
; GFX6-NEXT:    s_endpgm
;
; GFX8-LABEL: round_v8f32:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_load_dwordx8 s[8:15], s[4:5], 0x44
; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GFX8-NEXT:    s_brev_b32 s6, -2
; GFX8-NEXT:    s_mov_b32 s3, 0xf000
; GFX8-NEXT:    s_mov_b32 s2, -1
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_trunc_f32_e32 v0, s11
; GFX8-NEXT:    v_sub_f32_e32 v1, s11, v0
; GFX8-NEXT:    v_cmp_ge_f32_e64 s[4:5], |v1|, 0.5
; GFX8-NEXT:    v_cndmask_b32_e64 v1, 0, 1.0, s[4:5]
; GFX8-NEXT:    v_mov_b32_e32 v2, s11
; GFX8-NEXT:    v_bfi_b32 v1, s6, v1, v2
; GFX8-NEXT:    v_add_f32_e32 v3, v0, v1
; GFX8-NEXT:    v_trunc_f32_e32 v0, s10
; GFX8-NEXT:    v_sub_f32_e32 v1, s10, v0
; GFX8-NEXT:    v_cmp_ge_f32_e64 s[4:5], |v1|, 0.5
; GFX8-NEXT:    v_cndmask_b32_e64 v1, 0, 1.0, s[4:5]
; GFX8-NEXT:    v_mov_b32_e32 v2, s10
; GFX8-NEXT:    v_bfi_b32 v1, s6, v1, v2
; GFX8-NEXT:    v_add_f32_e32 v2, v0, v1
; GFX8-NEXT:    v_trunc_f32_e32 v0, s9
; GFX8-NEXT:    v_sub_f32_e32 v1, s9, v0
; GFX8-NEXT:    v_cmp_ge_f32_e64 s[4:5], |v1|, 0.5
; GFX8-NEXT:    v_cndmask_b32_e64 v1, 0, 1.0, s[4:5]
; GFX8-NEXT:    v_mov_b32_e32 v4, s9
; GFX8-NEXT:    v_bfi_b32 v1, s6, v1, v4
; GFX8-NEXT:    v_add_f32_e32 v1, v0, v1
; GFX8-NEXT:    v_trunc_f32_e32 v0, s8
; GFX8-NEXT:    v_sub_f32_e32 v4, s8, v0
; GFX8-NEXT:    v_cmp_ge_f32_e64 s[4:5], |v4|, 0.5
; GFX8-NEXT:    v_cndmask_b32_e64 v4, 0, 1.0, s[4:5]
; GFX8-NEXT:    v_mov_b32_e32 v5, s8
; GFX8-NEXT:    v_bfi_b32 v4, s6, v4, v5
; GFX8-NEXT:    v_add_f32_e32 v0, v0, v4
; GFX8-NEXT:    v_trunc_f32_e32 v4, s15
; GFX8-NEXT:    v_sub_f32_e32 v5, s15, v4
; GFX8-NEXT:    v_cmp_ge_f32_e64 s[4:5], |v5|, 0.5
; GFX8-NEXT:    v_cndmask_b32_e64 v5, 0, 1.0, s[4:5]
; GFX8-NEXT:    v_mov_b32_e32 v6, s15
; GFX8-NEXT:    v_bfi_b32 v5, s6, v5, v6
; GFX8-NEXT:    v_add_f32_e32 v7, v4, v5
; GFX8-NEXT:    v_trunc_f32_e32 v4, s14
; GFX8-NEXT:    v_sub_f32_e32 v5, s14, v4
; GFX8-NEXT:    v_cmp_ge_f32_e64 s[4:5], |v5|, 0.5
; GFX8-NEXT:    v_cndmask_b32_e64 v5, 0, 1.0, s[4:5]
; GFX8-NEXT:    v_mov_b32_e32 v6, s14
; GFX8-NEXT:    v_bfi_b32 v5, s6, v5, v6
; GFX8-NEXT:    v_add_f32_e32 v6, v4, v5
; GFX8-NEXT:    v_trunc_f32_e32 v4, s13
; GFX8-NEXT:    v_sub_f32_e32 v5, s13, v4
; GFX8-NEXT:    v_cmp_ge_f32_e64 s[4:5], |v5|, 0.5
; GFX8-NEXT:    v_cndmask_b32_e64 v5, 0, 1.0, s[4:5]
; GFX8-NEXT:    v_mov_b32_e32 v8, s13
; GFX8-NEXT:    v_bfi_b32 v5, s6, v5, v8
; GFX8-NEXT:    v_add_f32_e32 v5, v4, v5
; GFX8-NEXT:    v_trunc_f32_e32 v4, s12
; GFX8-NEXT:    v_sub_f32_e32 v8, s12, v4
; GFX8-NEXT:    v_cmp_ge_f32_e64 s[4:5], |v8|, 0.5
; GFX8-NEXT:    v_cndmask_b32_e64 v8, 0, 1.0, s[4:5]
; GFX8-NEXT:    v_mov_b32_e32 v9, s12
; GFX8-NEXT:    v_bfi_b32 v8, s6, v8, v9
; GFX8-NEXT:    v_add_f32_e32 v4, v4, v8
; GFX8-NEXT:    buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:16
; GFX8-NEXT:    buffer_store_dwordx4 v[0:3], off, s[0:3], 0
; GFX8-NEXT:    s_endpgm
;
; GFX9-LABEL: round_v8f32:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_load_dwordx8 s[8:15], s[4:5], 0x44
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GFX9-NEXT:    s_brev_b32 s6, -2
; GFX9-NEXT:    s_mov_b32 s3, 0xf000
; GFX9-NEXT:    s_mov_b32 s2, -1
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_trunc_f32_e32 v0, s11
; GFX9-NEXT:    v_sub_f32_e32 v1, s11, v0
; GFX9-NEXT:    v_cmp_ge_f32_e64 s[4:5], |v1|, 0.5
; GFX9-NEXT:    v_cndmask_b32_e64 v1, 0, 1.0, s[4:5]
; GFX9-NEXT:    v_mov_b32_e32 v2, s11
; GFX9-NEXT:    v_bfi_b32 v1, s6, v1, v2
; GFX9-NEXT:    v_add_f32_e32 v3, v0, v1
; GFX9-NEXT:    v_trunc_f32_e32 v0, s10
; GFX9-NEXT:    v_sub_f32_e32 v1, s10, v0
; GFX9-NEXT:    v_cmp_ge_f32_e64 s[4:5], |v1|, 0.5
; GFX9-NEXT:    v_cndmask_b32_e64 v1, 0, 1.0, s[4:5]
; GFX9-NEXT:    v_mov_b32_e32 v2, s10
; GFX9-NEXT:    v_bfi_b32 v1, s6, v1, v2
; GFX9-NEXT:    v_add_f32_e32 v2, v0, v1
; GFX9-NEXT:    v_trunc_f32_e32 v0, s9
; GFX9-NEXT:    v_sub_f32_e32 v1, s9, v0
; GFX9-NEXT:    v_cmp_ge_f32_e64 s[4:5], |v1|, 0.5
; GFX9-NEXT:    v_cndmask_b32_e64 v1, 0, 1.0, s[4:5]
; GFX9-NEXT:    v_mov_b32_e32 v4, s9
; GFX9-NEXT:    v_bfi_b32 v1, s6, v1, v4
; GFX9-NEXT:    v_add_f32_e32 v1, v0, v1
; GFX9-NEXT:    v_trunc_f32_e32 v0, s8
; GFX9-NEXT:    v_sub_f32_e32 v4, s8, v0
; GFX9-NEXT:    v_cmp_ge_f32_e64 s[4:5], |v4|, 0.5
; GFX9-NEXT:    v_cndmask_b32_e64 v4, 0, 1.0, s[4:5]
; GFX9-NEXT:    v_mov_b32_e32 v5, s8
; GFX9-NEXT:    v_bfi_b32 v4, s6, v4, v5
; GFX9-NEXT:    v_add_f32_e32 v0, v0, v4
; GFX9-NEXT:    v_trunc_f32_e32 v4, s15
; GFX9-NEXT:    v_sub_f32_e32 v5, s15, v4
; GFX9-NEXT:    v_cmp_ge_f32_e64 s[4:5], |v5|, 0.5
; GFX9-NEXT:    v_cndmask_b32_e64 v5, 0, 1.0, s[4:5]
; GFX9-NEXT:    v_mov_b32_e32 v6, s15
; GFX9-NEXT:    v_bfi_b32 v5, s6, v5, v6
; GFX9-NEXT:    v_add_f32_e32 v7, v4, v5
; GFX9-NEXT:    v_trunc_f32_e32 v4, s14
; GFX9-NEXT:    v_sub_f32_e32 v5, s14, v4
; GFX9-NEXT:    v_cmp_ge_f32_e64 s[4:5], |v5|, 0.5
; GFX9-NEXT:    v_cndmask_b32_e64 v5, 0, 1.0, s[4:5]
; GFX9-NEXT:    v_mov_b32_e32 v6, s14
; GFX9-NEXT:    v_bfi_b32 v5, s6, v5, v6
; GFX9-NEXT:    v_add_f32_e32 v6, v4, v5
; GFX9-NEXT:    v_trunc_f32_e32 v4, s13
; GFX9-NEXT:    v_sub_f32_e32 v5, s13, v4
; GFX9-NEXT:    v_cmp_ge_f32_e64 s[4:5], |v5|, 0.5
; GFX9-NEXT:    v_cndmask_b32_e64 v5, 0, 1.0, s[4:5]
; GFX9-NEXT:    v_mov_b32_e32 v8, s13
; GFX9-NEXT:    v_bfi_b32 v5, s6, v5, v8
; GFX9-NEXT:    v_add_f32_e32 v5, v4, v5
; GFX9-NEXT:    v_trunc_f32_e32 v4, s12
; GFX9-NEXT:    v_sub_f32_e32 v8, s12, v4
; GFX9-NEXT:    v_cmp_ge_f32_e64 s[4:5], |v8|, 0.5
; GFX9-NEXT:    v_cndmask_b32_e64 v8, 0, 1.0, s[4:5]
; GFX9-NEXT:    v_mov_b32_e32 v9, s12
; GFX9-NEXT:    v_bfi_b32 v8, s6, v8, v9
; GFX9-NEXT:    v_add_f32_e32 v4, v4, v8
; GFX9-NEXT:    buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:16
; GFX9-NEXT:    buffer_store_dwordx4 v[0:3], off, s[0:3], 0
; GFX9-NEXT:    s_endpgm
;
; GFX11-LABEL: round_v8f32:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b256 s[8:15], s[4:5], 0x44
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX11-NEXT:    s_mov_b32 s3, 0x31016000
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_trunc_f32_e32 v0, s11
; GFX11-NEXT:    v_trunc_f32_e32 v1, s10
; GFX11-NEXT:    v_trunc_f32_e32 v4, s9
; GFX11-NEXT:    v_trunc_f32_e32 v8, s8
; GFX11-NEXT:    v_trunc_f32_e32 v5, s15
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT:    v_dual_sub_f32 v2, s11, v0 :: v_dual_sub_f32 v3, s10, v1
; GFX11-NEXT:    v_sub_f32_e32 v7, s9, v4
; GFX11-NEXT:    v_trunc_f32_e32 v9, s13
; GFX11-NEXT:    v_sub_f32_e32 v11, s8, v8
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-NEXT:    v_cmp_ge_f32_e64 s2, |v2|, 0.5
; GFX11-NEXT:    v_sub_f32_e32 v12, s15, v5
; GFX11-NEXT:    v_trunc_f32_e32 v6, s14
; GFX11-NEXT:    v_sub_f32_e32 v14, s13, v9
; GFX11-NEXT:    v_trunc_f32_e32 v10, s12
; GFX11-NEXT:    v_cndmask_b32_e64 v2, 0, 1.0, s2
; GFX11-NEXT:    v_cmp_ge_f32_e64 s2, |v3|, 0.5
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT:    v_bfi_b32 v2, 0x7fffffff, v2, s11
; GFX11-NEXT:    v_cndmask_b32_e64 v3, 0, 1.0, s2
; GFX11-NEXT:    v_cmp_ge_f32_e64 s2, |v7|, 0.5
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT:    v_bfi_b32 v16, 0x7fffffff, v3, s10
; GFX11-NEXT:    v_cndmask_b32_e64 v7, 0, 1.0, s2
; GFX11-NEXT:    v_cmp_ge_f32_e64 s2, |v11|, 0.5
; GFX11-NEXT:    v_sub_f32_e32 v13, s14, v6
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT:    v_dual_add_f32 v3, v0, v2 :: v_dual_add_f32 v2, v1, v16
; GFX11-NEXT:    v_bfi_b32 v7, 0x7fffffff, v7, s9
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_cndmask_b32_e64 v11, 0, 1.0, s2
; GFX11-NEXT:    v_cmp_ge_f32_e64 s2, |v12|, 0.5
; GFX11-NEXT:    v_add_f32_e32 v1, v4, v7
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_bfi_b32 v11, 0x7fffffff, v11, s8
; GFX11-NEXT:    v_cndmask_b32_e64 v12, 0, 1.0, s2
; GFX11-NEXT:    v_cmp_ge_f32_e64 s2, |v13|, 0.5
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT:    v_bfi_b32 v12, 0x7fffffff, v12, s15
; GFX11-NEXT:    v_cndmask_b32_e64 v13, 0, 1.0, s2
; GFX11-NEXT:    v_cmp_ge_f32_e64 s2, |v14|, 0.5
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_add_f32_e32 v7, v5, v12
; GFX11-NEXT:    v_bfi_b32 v13, 0x7fffffff, v13, s14
; GFX11-NEXT:    v_sub_f32_e32 v15, s12, v10
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_cndmask_b32_e64 v14, 0, 1.0, s2
; GFX11-NEXT:    v_add_f32_e32 v6, v6, v13
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_cmp_ge_f32_e64 s2, |v15|, 0.5
; GFX11-NEXT:    v_bfi_b32 v0, 0x7fffffff, v14, s13
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT:    v_cndmask_b32_e64 v15, 0, 1.0, s2
; GFX11-NEXT:    v_dual_add_f32 v5, v9, v0 :: v_dual_add_f32 v0, v8, v11
; GFX11-NEXT:    s_mov_b32 s2, -1
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_bfi_b32 v4, 0x7fffffff, v15, s12
; GFX11-NEXT:    v_add_f32_e32 v4, v10, v4
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    buffer_store_b128 v[4:7], off, s[0:3], 0 offset:16
; GFX11-NEXT:    buffer_store_b128 v[0:3], off, s[0:3], 0
; GFX11-NEXT:    s_endpgm
;
; R600-LABEL: round_v8f32:
; R600:       ; %bb.0:
; R600-NEXT:    ALU 50, @4, KC0[CB0:0-32], KC1[]
; R600-NEXT:    MEM_RAT_CACHELESS STORE_RAW T0.XYZW, T2.X, 0
; R600-NEXT:    MEM_RAT_CACHELESS STORE_RAW T3.XYZW, T1.X, 1
; R600-NEXT:    CF_END
; R600-NEXT:    ALU clause starting at 4:
; R600-NEXT:     TRUNC * T0.W, KC0[6].X,
; R600-NEXT:     ADD T0.Z, KC0[6].X, -PV.W,
; R600-NEXT:     TRUNC * T1.W, KC0[5].X,
; R600-NEXT:     TRUNC * T2.W, KC0[4].W,
; R600-NEXT:     ADD T1.Z, KC0[4].W, -PV.W,
; R600-NEXT:     ADD T3.W, KC0[5].X, -T1.W,
; R600-NEXT:     SETGE * T4.W, |T0.Z|, 0.5,
; R600-NEXT:     BFI_INT T0.Y, literal.x, PS, KC0[6].X,
; R600-NEXT:     SETGE T0.Z, |PV.W|, 0.5,
; R600-NEXT:     SETGE T3.W, |PV.Z|, 0.5,
; R600-NEXT:     TRUNC * T4.W, KC0[5].Y,
; R600-NEXT:    2147483647(nan), 0(0.000000e+00)
; R600-NEXT:     ADD T1.Y, KC0[5].Y, -PS,
; R600-NEXT:     BFI_INT T1.Z, literal.x, PV.W, KC0[4].W,
; R600-NEXT:     BFI_INT T3.W, literal.x, PV.Z, KC0[5].X,
; R600-NEXT:     TRUNC * T5.W, KC0[4].Z,
; R600-NEXT:    2147483647(nan), 0(0.000000e+00)
; R600-NEXT:     TRUNC T0.Z, KC0[4].Y,
; R600-NEXT:     TRUNC * T6.W, KC0[5].W,
; R600-NEXT:     ADD * T7.W, KC0[4].Z, -T5.W,
; R600-NEXT:     TRUNC T0.X, KC0[5].Z,
; R600-NEXT:     SETGE T2.Y, |PV.W|, 0.5,
; R600-NEXT:     ADD T2.Z, KC0[5].W, -T6.W, BS:VEC_102/SCL_221
; R600-NEXT:     ADD T7.W, KC0[4].Y, -T0.Z,
; R600-NEXT:     ADD * T3.W, T1.W, T3.W,
; R600-NEXT:     SETGE T1.X, |PV.W|, 0.5,
; R600-NEXT:     SETGE T4.Y, |PV.Z|, 0.5,
; R600-NEXT:     ADD T3.Z, T2.W, T1.Z,
; R600-NEXT:     BFI_INT T1.W, literal.x, PV.Y, KC0[4].Z,
; R600-NEXT:     ADD * T2.W, KC0[5].Z, -PV.X,
; R600-NEXT:    2147483647(nan), 0(0.000000e+00)
; R600-NEXT:     SETGE T2.X, |PS|, 0.5,
; R600-NEXT:     ADD T3.Y, T5.W, PV.W,
; R600-NEXT:     BFI_INT T1.Z, literal.x, PV.Y, KC0[5].W,
; R600-NEXT:     BFI_INT T1.W, literal.x, PV.X, KC0[4].Y,
; R600-NEXT:     ADD * T0.W, T0.W, T0.Y,
; R600-NEXT:    2147483647(nan), 0(0.000000e+00)
; R600-NEXT:     ADD T3.X, T0.Z, PV.W,
; R600-NEXT:     ADD T0.Z, T6.W, PV.Z,
; R600-NEXT:     BFI_INT T1.W, literal.x, PV.X, KC0[5].Z,
; R600-NEXT:     SETGE * T2.W, |T1.Y|, 0.5,
; R600-NEXT:    2147483647(nan), 0(0.000000e+00)
; R600-NEXT:     LSHR T1.X, KC0[2].Y, literal.x,
; R600-NEXT:     ADD T0.Y, T0.X, PV.W,
; R600-NEXT:     BFI_INT * T1.W, literal.y, PS, KC0[5].Y,
; R600-NEXT:    2(2.802597e-45), 2147483647(nan)
; R600-NEXT:     ADD T0.X, T4.W, PV.W,
; R600-NEXT:     ADD_INT * T1.W, KC0[2].Y, literal.x,
; R600-NEXT:    16(2.242078e-44), 0(0.000000e+00)
; R600-NEXT:     LSHR * T2.X, PV.W, literal.x,
; R600-NEXT:    2(2.802597e-45), 0(0.000000e+00)
  %result = call <8 x float> @llvm.round.v8f32(<8 x float> %in) #1
  store <8 x float> %result, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @round_f16(ptr addrspace(1) %out, i32 %x.arg) #0 {
; GFX6-LABEL: round_f16:
; GFX6:       ; %bb.0:
; GFX6-NEXT:    s_load_dword s0, s[4:5], 0xb
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_cvt_f32_f16_e32 v0, s0
; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
; GFX6-NEXT:    v_trunc_f32_e32 v1, v0
; GFX6-NEXT:    v_sub_f32_e32 v2, v0, v1
; GFX6-NEXT:    v_cmp_ge_f32_e64 s[2:3], |v2|, 0.5
; GFX6-NEXT:    v_cndmask_b32_e64 v2, 0, 1.0, s[2:3]
; GFX6-NEXT:    s_brev_b32 s2, -2
; GFX6-NEXT:    v_bfi_b32 v0, s2, v2, v0
; GFX6-NEXT:    v_add_f32_e32 v0, v1, v0
; GFX6-NEXT:    v_cvt_f16_f32_e32 v0, v0
; GFX6-NEXT:    s_mov_b32 s3, 0xf000
; GFX6-NEXT:    s_mov_b32 s2, -1
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    buffer_store_short v0, off, s[0:3], 0
; GFX6-NEXT:    s_endpgm
;
; GFX8-LABEL: round_f16:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_load_dword s6, s[4:5], 0x2c
; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GFX8-NEXT:    v_mov_b32_e32 v0, 0x3c00
; GFX8-NEXT:    s_movk_i32 s4, 0x7fff
; GFX8-NEXT:    s_mov_b32 s3, 0xf000
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_trunc_f16_e32 v1, s6
; GFX8-NEXT:    v_sub_f16_e32 v2, s6, v1
; GFX8-NEXT:    v_cmp_ge_f16_e64 vcc, |v2|, 0.5
; GFX8-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
; GFX8-NEXT:    v_mov_b32_e32 v2, s6
; GFX8-NEXT:    v_bfi_b32 v0, s4, v0, v2
; GFX8-NEXT:    s_mov_b32 s2, -1
; GFX8-NEXT:    v_add_f16_e32 v0, v1, v0
; GFX8-NEXT:    buffer_store_short v0, off, s[0:3], 0
; GFX8-NEXT:    s_endpgm
;
; GFX9-LABEL: round_f16:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_load_dword s6, s[4:5], 0x2c
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0x3c00
; GFX9-NEXT:    s_movk_i32 s4, 0x7fff
; GFX9-NEXT:    s_mov_b32 s3, 0xf000
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_trunc_f16_e32 v1, s6
; GFX9-NEXT:    v_sub_f16_e32 v2, s6, v1
; GFX9-NEXT:    v_cmp_ge_f16_e64 vcc, |v2|, 0.5
; GFX9-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
; GFX9-NEXT:    v_mov_b32_e32 v2, s6
; GFX9-NEXT:    v_bfi_b32 v0, s4, v0, v2
; GFX9-NEXT:    s_mov_b32 s2, -1
; GFX9-NEXT:    v_add_f16_e32 v0, v1, v0
; GFX9-NEXT:    buffer_store_short v0, off, s[0:3], 0
; GFX9-NEXT:    s_endpgm
;
; GFX11-TRUE16-LABEL: round_f16:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_load_b32 s2, s[4:5], 0x2c
; GFX11-TRUE16-NEXT:    s_mov_b32 s3, 0x31016000
; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT:    v_trunc_f16_e32 v0.l, s2
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_sub_f16_e32 v0.h, s2, v0.l
; GFX11-TRUE16-NEXT:    v_cmp_ge_f16_e64 s0, |v0.h|, 0.5
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, 0, 0x3c00, s0
; GFX11-TRUE16-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX11-TRUE16-NEXT:    v_bfi_b32 v1, 0x7fff, v1, s2
; GFX11-TRUE16-NEXT:    s_mov_b32 s2, -1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add_f16_e32 v0.l, v0.l, v1.l
; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT:    buffer_store_b16 v0, off, s[0:3], 0
; GFX11-TRUE16-NEXT:    s_endpgm
;
; GFX11-FAKE16-LABEL: round_f16:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_clause 0x1
; GFX11-FAKE16-NEXT:    s_load_b32 s2, s[4:5], 0x2c
; GFX11-FAKE16-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-FAKE16-NEXT:    v_trunc_f16_e32 v0, s2
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_sub_f16_e32 v1, s2, v0
; GFX11-FAKE16-NEXT:    v_cmp_ge_f16_e64 s3, |v1|, 0.5
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e64 v1, 0, 0x3c00, s3
; GFX11-FAKE16-NEXT:    s_mov_b32 s3, 0x31016000
; GFX11-FAKE16-NEXT:    v_bfi_b32 v1, 0x7fff, v1, s2
; GFX11-FAKE16-NEXT:    s_mov_b32 s2, -1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add_f16_e32 v0, v0, v1
; GFX11-FAKE16-NEXT:    buffer_store_b16 v0, off, s[0:3], 0
; GFX11-FAKE16-NEXT:    s_endpgm
;
; R600-LABEL: round_f16:
; R600:       ; %bb.0:
; R600-NEXT:    ALU 17, @4, KC0[CB0:0-32], KC1[]
; R600-NEXT:    MEM_RAT MSKOR T0.XW, T1.X
; R600-NEXT:    CF_END
; R600-NEXT:    PAD
; R600-NEXT:    ALU clause starting at 4:
; R600-NEXT:     FLT16_TO_FLT32 * T0.W, KC0[2].Z,
; R600-NEXT:     TRUNC * T1.W, PV.W,
; R600-NEXT:     ADD * T2.W, T0.W, -PV.W,
; R600-NEXT:     SETGE * T2.W, |PV.W|, 0.5,
; R600-NEXT:     BFI_INT T0.W, literal.x, PV.W, T0.W,
; R600-NEXT:     AND_INT * T2.W, KC0[2].Y, literal.y,
; R600-NEXT:    2147483647(nan), 3(4.203895e-45)
; R600-NEXT:     ADD * T0.W, T1.W, PV.W,
; R600-NEXT:     FLT32_TO_FLT16 T0.W, PV.W,
; R600-NEXT:     LSHL * T1.W, T2.W, literal.x,
; R600-NEXT:    3(4.203895e-45), 0(0.000000e+00)
; R600-NEXT:     LSHL T0.X, PV.W, PS,
; R600-NEXT:     LSHL * T0.W, literal.x, PS,
; R600-NEXT:    65535(9.183409e-41), 0(0.000000e+00)
; R600-NEXT:     MOV T0.Y, 0.0,
; R600-NEXT:     MOV * T0.Z, 0.0,
; R600-NEXT:     LSHR * T1.X, KC0[2].Y, literal.x,
; R600-NEXT:    2(2.802597e-45), 0(0.000000e+00)
  %x.arg.trunc = trunc i32 %x.arg to i16
  %x = bitcast i16 %x.arg.trunc to half
  %result = call half @llvm.round.f16(half %x) #1
  store half %result, ptr addrspace(1) %out
  ret void
}

; Should be scalarized
define amdgpu_kernel void @round_v2f16(ptr addrspace(1) %out, i32 %in.arg) #0 {
; GFX6-LABEL: round_v2f16:
; GFX6:       ; %bb.0:
; GFX6-NEXT:    s_load_dword s0, s[4:5], 0xb
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    s_lshr_b32 s1, s0, 16
; GFX6-NEXT:    v_cvt_f32_f16_e32 v1, s1
; GFX6-NEXT:    v_cvt_f32_f16_e32 v0, s0
; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
; GFX6-NEXT:    s_brev_b32 s4, -2
; GFX6-NEXT:    v_trunc_f32_e32 v3, v1
; GFX6-NEXT:    v_sub_f32_e32 v5, v1, v3
; GFX6-NEXT:    v_trunc_f32_e32 v2, v0
; GFX6-NEXT:    v_cmp_ge_f32_e64 s[2:3], |v5|, 0.5
; GFX6-NEXT:    v_sub_f32_e32 v4, v0, v2
; GFX6-NEXT:    v_cndmask_b32_e64 v5, 0, 1.0, s[2:3]
; GFX6-NEXT:    v_bfi_b32 v1, s4, v5, v1
; GFX6-NEXT:    v_cmp_ge_f32_e64 s[2:3], |v4|, 0.5
; GFX6-NEXT:    v_add_f32_e32 v1, v3, v1
; GFX6-NEXT:    v_cndmask_b32_e64 v3, 0, 1.0, s[2:3]
; GFX6-NEXT:    v_bfi_b32 v0, s4, v3, v0
; GFX6-NEXT:    v_cvt_f16_f32_e32 v1, v1
; GFX6-NEXT:    v_add_f32_e32 v0, v2, v0
; GFX6-NEXT:    v_cvt_f16_f32_e32 v0, v0
; GFX6-NEXT:    s_mov_b32 s3, 0xf000
; GFX6-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX6-NEXT:    s_mov_b32 s2, -1
; GFX6-NEXT:    v_or_b32_e32 v0, v0, v1
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; GFX6-NEXT:    s_endpgm
;
; GFX8-LABEL: round_v2f16:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_load_dword s6, s[4:5], 0x2c
; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GFX8-NEXT:    v_mov_b32_e32 v0, 0x3c00
; GFX8-NEXT:    s_movk_i32 s5, 0x7fff
; GFX8-NEXT:    s_mov_b32 s3, 0xf000
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    s_lshr_b32 s4, s6, 16
; GFX8-NEXT:    v_trunc_f16_e32 v1, s4
; GFX8-NEXT:    v_sub_f16_e32 v2, s4, v1
; GFX8-NEXT:    v_cmp_ge_f16_e64 vcc, |v2|, 0.5
; GFX8-NEXT:    v_cndmask_b32_e32 v2, 0, v0, vcc
; GFX8-NEXT:    v_mov_b32_e32 v3, s4
; GFX8-NEXT:    v_bfi_b32 v2, s5, v2, v3
; GFX8-NEXT:    v_add_f16_sdwa v1, v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX8-NEXT:    v_trunc_f16_e32 v2, s6
; GFX8-NEXT:    v_sub_f16_e32 v3, s6, v2
; GFX8-NEXT:    v_cmp_ge_f16_e64 vcc, |v3|, 0.5
; GFX8-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
; GFX8-NEXT:    v_mov_b32_e32 v3, s6
; GFX8-NEXT:    v_bfi_b32 v0, s5, v0, v3
; GFX8-NEXT:    v_add_f16_e32 v0, v2, v0
; GFX8-NEXT:    s_mov_b32 s2, -1
; GFX8-NEXT:    v_or_b32_e32 v0, v0, v1
; GFX8-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; GFX8-NEXT:    s_endpgm
;
; GFX9-LABEL: round_v2f16:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_load_dword s6, s[4:5], 0x2c
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0x3c00
; GFX9-NEXT:    s_movk_i32 s5, 0x7fff
; GFX9-NEXT:    s_mov_b32 s3, 0xf000
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshr_b32 s4, s6, 16
; GFX9-NEXT:    v_trunc_f16_e32 v1, s4
; GFX9-NEXT:    v_sub_f16_e32 v2, s4, v1
; GFX9-NEXT:    v_cmp_ge_f16_e64 vcc, |v2|, 0.5
; GFX9-NEXT:    v_cndmask_b32_e32 v2, 0, v0, vcc
; GFX9-NEXT:    v_mov_b32_e32 v3, s4
; GFX9-NEXT:    v_bfi_b32 v2, s5, v2, v3
; GFX9-NEXT:    v_add_f16_e32 v1, v1, v2
; GFX9-NEXT:    v_trunc_f16_e32 v2, s6
; GFX9-NEXT:    v_sub_f16_e32 v3, s6, v2
; GFX9-NEXT:    v_cmp_ge_f16_e64 vcc, |v3|, 0.5
; GFX9-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
; GFX9-NEXT:    v_mov_b32_e32 v3, s6
; GFX9-NEXT:    v_bfi_b32 v0, s5, v0, v3
; GFX9-NEXT:    v_add_f16_e32 v0, v2, v0
; GFX9-NEXT:    s_mov_b32 s2, -1
; GFX9-NEXT:    v_pack_b32_f16 v0, v0, v1
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; GFX9-NEXT:    s_endpgm
;
; GFX11-TRUE16-LABEL: round_v2f16:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_load_b32 s2, s[4:5], 0x2c
; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT:    s_lshr_b32 s3, s2, 16
; GFX11-TRUE16-NEXT:    v_trunc_f16_e32 v0.h, s2
; GFX11-TRUE16-NEXT:    v_trunc_f16_e32 v0.l, s3
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_sub_f16_e32 v1.h, s2, v0.h
; GFX11-TRUE16-NEXT:    v_sub_f16_e32 v1.l, s3, v0.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_cmp_ge_f16_e64 s1, |v1.h|, 0.5
; GFX11-TRUE16-NEXT:    v_cmp_ge_f16_e64 s0, |v1.l|, 0.5
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v2.l, 0, 0x3c00, s1
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, 0, 0x3c00, s0
; GFX11-TRUE16-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_bfi_b32 v2, 0x7fff, v2, s2
; GFX11-TRUE16-NEXT:    v_bfi_b32 v1, 0x7fff, v1, s3
; GFX11-TRUE16-NEXT:    s_mov_b32 s3, 0x31016000
; GFX11-TRUE16-NEXT:    s_mov_b32 s2, -1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_add_f16_e32 v0.h, v0.h, v2.l
; GFX11-TRUE16-NEXT:    v_add_f16_e32 v0.l, v0.l, v1.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_pack_b32_f16 v0, v0.h, v0.l
; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT:    buffer_store_b32 v0, off, s[0:3], 0
; GFX11-TRUE16-NEXT:    s_endpgm
;
; GFX11-FAKE16-LABEL: round_v2f16:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_clause 0x1
; GFX11-FAKE16-NEXT:    s_load_b32 s2, s[4:5], 0x2c
; GFX11-FAKE16-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_lshr_b32 s3, s2, 16
; GFX11-FAKE16-NEXT:    v_trunc_f16_e32 v1, s2
; GFX11-FAKE16-NEXT:    v_trunc_f16_e32 v0, s3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_sub_f16_e32 v3, s2, v1
; GFX11-FAKE16-NEXT:    v_sub_f16_e32 v2, s3, v0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_cmp_ge_f16_e64 s4, |v2|, 0.5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e64 v2, 0, 0x3c00, s4
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cmp_ge_f16_e64 s4, |v3|, 0.5
; GFX11-FAKE16-NEXT:    v_bfi_b32 v2, 0x7fff, v2, s3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e64 v3, 0, 0x3c00, s4
; GFX11-FAKE16-NEXT:    s_mov_b32 s3, 0x31016000
; GFX11-FAKE16-NEXT:    v_add_f16_e32 v0, v0, v2
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_bfi_b32 v3, 0x7fff, v3, s2
; GFX11-FAKE16-NEXT:    s_mov_b32 s2, -1
; GFX11-FAKE16-NEXT:    v_add_f16_e32 v1, v1, v3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_pack_b32_f16 v0, v1, v0
; GFX11-FAKE16-NEXT:    buffer_store_b32 v0, off, s[0:3], 0
; GFX11-FAKE16-NEXT:    s_endpgm
;
; R600-LABEL: round_v2f16:
; R600:       ; %bb.0:
; R600-NEXT:    ALU 22, @4, KC0[CB0:0-32], KC1[]
; R600-NEXT:    MEM_RAT_CACHELESS STORE_RAW T0.X, T1.X, 1
; R600-NEXT:    CF_END
; R600-NEXT:    PAD
; R600-NEXT:    ALU clause starting at 4:
; R600-NEXT:     LSHR * T0.W, KC0[2].Z, literal.x,
; R600-NEXT:    16(2.242078e-44), 0(0.000000e+00)
; R600-NEXT:     FLT16_TO_FLT32 * T0.W, PV.W,
; R600-NEXT:     FLT16_TO_FLT32 T1.W, KC0[2].Z,
; R600-NEXT:     TRUNC * T2.W, PV.W,
; R600-NEXT:     ADD T3.W, T0.W, -PS,
; R600-NEXT:     TRUNC * T4.W, PV.W,
; R600-NEXT:     ADD T5.W, T1.W, -PS,
; R600-NEXT:     SETGE * T3.W, |PV.W|, 0.5,
; R600-NEXT:     BFI_INT T0.W, literal.x, PS, T0.W,
; R600-NEXT:     SETGE * T3.W, |PV.W|, 0.5,
; R600-NEXT:    2147483647(nan), 0(0.000000e+00)
; R600-NEXT:     BFI_INT T1.W, literal.x, PS, T1.W, BS:VEC_021/SCL_122
; R600-NEXT:     ADD * T0.W, T2.W, PV.W,
; R600-NEXT:    2147483647(nan), 0(0.000000e+00)
; R600-NEXT:     FLT32_TO_FLT16 T0.W, PS,
; R600-NEXT:     ADD * T1.W, T4.W, PV.W,
; R600-NEXT:     FLT32_TO_FLT16 T1.W, PS,
; R600-NEXT:     LSHL * T0.W, PV.W, literal.x,
; R600-NEXT:    16(2.242078e-44), 0(0.000000e+00)
; R600-NEXT:     OR_INT T0.X, PV.W, PS,
; R600-NEXT:     LSHR * T1.X, KC0[2].Y, literal.x,
; R600-NEXT:    2(2.802597e-45), 0(0.000000e+00)
  %in = bitcast i32 %in.arg to <2 x half>
  %result = call <2 x half> @llvm.round.v2f16(<2 x half> %in)
  store <2 x half> %result, ptr addrspace(1) %out
  ret void
}

declare float @llvm.round.f32(float) #1
declare <2 x float> @llvm.round.v2f32(<2 x float>) #1
declare <4 x float> @llvm.round.v4f32(<4 x float>) #1
declare <8 x float> @llvm.round.v8f32(<8 x float>) #1

declare half @llvm.round.f16(half) #1
declare <2 x half> @llvm.round.v2f16(<2 x half>) #1
declare <4 x half> @llvm.round.v4f16(<4 x half>) #1
declare <8 x half> @llvm.round.v8f16(<8 x half>) #1

attributes #0 = { nounwind }
attributes #1 = { nounwind readnone }