; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc -global-isel -amdgpu-codegenprepare-disable-idiv-expansion=1 -amdgpu-bypass-slow-div=0 -mtriple=amdgcn-amd-amdhsa -mcpu=tonga -verify-machineinstrs < %s | FileCheck --check-prefix=GFX8 %s
; RUN: llc -global-isel -amdgpu-codegenprepare-disable-idiv-expansion=1 -amdgpu-bypass-slow-div=0 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefix=GFX9 %s
; RUN: llc -global-isel -amdgpu-codegenprepare-disable-idiv-expansion=1 -amdgpu-bypass-slow-div=0 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck -check-prefix=GFX10 %s
define amdgpu_kernel void @udivrem_i32(i32 addrspace(1)* %out0, i32 addrspace(1)* %out1, i32 %x, i32 %y) {
; GFX8-LABEL: udivrem_i32:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x10
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
; GFX8-NEXT: v_cvt_f32_u32_e32 v0, s7
; GFX8-NEXT: s_sub_i32 s0, 0, s7
; GFX8-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX8-NEXT: v_mul_f32_e32 v0, 0x4f7ffffe, v0
; GFX8-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX8-NEXT: v_mul_lo_u32 v1, s0, v0
; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x0
; GFX8-NEXT: v_mul_hi_u32 v1, v0, v1
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v0, v1
; GFX8-NEXT: v_mul_hi_u32 v2, s6, v0
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
; GFX8-NEXT: v_mov_b32_e32 v0, s0
; GFX8-NEXT: v_mov_b32_e32 v1, s1
; GFX8-NEXT: v_mul_lo_u32 v3, v2, s7
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 1, v2
; GFX8-NEXT: v_sub_u32_e32 v3, vcc, s6, v3
; GFX8-NEXT: v_cmp_le_u32_e32 vcc, s7, v3
; GFX8-NEXT: v_cndmask_b32_e32 v2, v2, v4, vcc
; GFX8-NEXT: v_subrev_u32_e64 v4, s[0:1], s7, v3
; GFX8-NEXT: v_cndmask_b32_e32 v3, v3, v4, vcc
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 1, v2
; GFX8-NEXT: v_cmp_le_u32_e32 vcc, s7, v3
; GFX8-NEXT: v_cndmask_b32_e32 v2, v2, v4, vcc
; GFX8-NEXT: v_subrev_u32_e64 v4, s[0:1], s7, v3
; GFX8-NEXT: flat_store_dword v[0:1], v2
; GFX8-NEXT: v_mov_b32_e32 v0, s2
; GFX8-NEXT: v_cndmask_b32_e32 v3, v3, v4, vcc
; GFX8-NEXT: v_mov_b32_e32 v1, s3
; GFX8-NEXT: flat_store_dword v[0:1], v3
; GFX8-NEXT: s_endpgm
;
; GFX9-LABEL: udivrem_i32:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x10
; GFX9-NEXT: v_mov_b32_e32 v2, 0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: v_cvt_f32_u32_e32 v0, s7
; GFX9-NEXT: s_sub_i32 s0, 0, s7
; GFX9-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX9-NEXT: v_mul_f32_e32 v0, 0x4f7ffffe, v0
; GFX9-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX9-NEXT: v_mul_lo_u32 v1, s0, v0
; GFX9-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x0
; GFX9-NEXT: v_mul_hi_u32 v1, v0, v1
; GFX9-NEXT: v_add_u32_e32 v0, v0, v1
; GFX9-NEXT: v_mul_hi_u32 v0, s6, v0
; GFX9-NEXT: v_mul_lo_u32 v1, v0, s7
; GFX9-NEXT: v_add_u32_e32 v3, 1, v0
; GFX9-NEXT: v_sub_u32_e32 v1, s6, v1
; GFX9-NEXT: v_cmp_le_u32_e32 vcc, s7, v1
; GFX9-NEXT: v_cndmask_b32_e32 v0, v0, v3, vcc
; GFX9-NEXT: v_subrev_u32_e32 v3, s7, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX9-NEXT: v_add_u32_e32 v3, 1, v0
; GFX9-NEXT: v_cmp_le_u32_e32 vcc, s7, v1
; GFX9-NEXT: v_cndmask_b32_e32 v0, v0, v3, vcc
; GFX9-NEXT: v_subrev_u32_e32 v3, s7, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: global_store_dword v2, v0, s[0:1]
; GFX9-NEXT: global_store_dword v2, v1, s[2:3]
; GFX9-NEXT: s_endpgm
;
; GFX10-LABEL: udivrem_i32:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x10
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: v_cvt_f32_u32_e32 v0, s7
; GFX10-NEXT: s_sub_i32 s0, 0, s7
; GFX10-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX10-NEXT: v_mul_f32_e32 v0, 0x4f7ffffe, v0
; GFX10-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX10-NEXT: v_mul_lo_u32 v1, s0, v0
; GFX10-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x0
; GFX10-NEXT: v_mul_hi_u32 v1, v0, v1
; GFX10-NEXT: v_add_nc_u32_e32 v0, v0, v1
; GFX10-NEXT: v_mul_hi_u32 v0, s6, v0
; GFX10-NEXT: v_mul_lo_u32 v1, v0, s7
; GFX10-NEXT: v_add_nc_u32_e32 v2, 1, v0
; GFX10-NEXT: v_sub_nc_u32_e32 v1, s6, v1
; GFX10-NEXT: v_subrev_nc_u32_e32 v3, s7, v1
; GFX10-NEXT: v_cmp_le_u32_e32 vcc_lo, s7, v1
; GFX10-NEXT: v_cndmask_b32_e32 v0, v0, v2, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX10-NEXT: v_add_nc_u32_e32 v2, 1, v0
; GFX10-NEXT: v_cmp_le_u32_e32 vcc_lo, s7, v1
; GFX10-NEXT: v_subrev_nc_u32_e32 v3, s7, v1
; GFX10-NEXT: v_cndmask_b32_e32 v0, v0, v2, vcc_lo
; GFX10-NEXT: v_mov_b32_e32 v2, 0
; GFX10-NEXT: v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: global_store_dword v2, v0, s[0:1]
; GFX10-NEXT: global_store_dword v2, v1, s[2:3]
; GFX10-NEXT: s_endpgm
%div = udiv i32 %x, %y
store i32 %div, i32 addrspace(1)* %out0
%rem = urem i32 %x, %y
store i32 %rem, i32 addrspace(1)* %out1
ret void
}
define amdgpu_kernel void @udivrem_i64(i64 addrspace(1)* %out0, i64 addrspace(1)* %out1, i64 %x, i64 %y) {
; GFX8-LABEL: udivrem_i64:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_load_dwordx8 s[4:11], s[4:5], 0x0
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
; GFX8-NEXT: v_cvt_f32_u32_e32 v0, s11
; GFX8-NEXT: v_cvt_f32_u32_e32 v1, s10
; GFX8-NEXT: s_sub_u32 s2, 0, s10
; GFX8-NEXT: s_subb_u32 s3, 0, s11
; GFX8-NEXT: v_mul_f32_e32 v0, 0x4f800000, v0
; GFX8-NEXT: v_add_f32_e32 v0, v0, v1
; GFX8-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX8-NEXT: v_mul_f32_e32 v0, 0x5f7ffffc, v0
; GFX8-NEXT: v_mul_f32_e32 v1, 0x2f800000, v0
; GFX8-NEXT: v_trunc_f32_e32 v2, v1
; GFX8-NEXT: v_mul_f32_e32 v1, 0xcf800000, v2
; GFX8-NEXT: v_add_f32_e32 v0, v1, v0
; GFX8-NEXT: v_cvt_u32_f32_e32 v3, v0
; GFX8-NEXT: v_cvt_u32_f32_e32 v4, v2
; GFX8-NEXT: v_mad_u64_u32 v[0:1], s[0:1], s2, v3, 0
; GFX8-NEXT: v_mad_u64_u32 v[1:2], s[0:1], s2, v4, v[1:2]
; GFX8-NEXT: v_mul_hi_u32 v5, v3, v0
; GFX8-NEXT: v_mad_u64_u32 v[1:2], s[0:1], s3, v3, v[1:2]
; GFX8-NEXT: v_mul_lo_u32 v2, v4, v0
; GFX8-NEXT: v_mul_hi_u32 v0, v4, v0
; GFX8-NEXT: v_mul_lo_u32 v6, v3, v1
; GFX8-NEXT: v_mul_lo_u32 v7, v4, v1
; GFX8-NEXT: v_mul_hi_u32 v8, v3, v1
; GFX8-NEXT: v_mul_hi_u32 v1, v4, v1
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v6
; GFX8-NEXT: v_cndmask_b32_e64 v6, 0, 1, vcc
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v7, v0
; GFX8-NEXT: v_cndmask_b32_e64 v7, 0, 1, vcc
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v5
; GFX8-NEXT: v_cndmask_b32_e64 v2, 0, 1, vcc
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v0, v8
; GFX8-NEXT: v_cndmask_b32_e64 v5, 0, 1, vcc
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v6, v2
; GFX8-NEXT: v_add_u32_e32 v5, vcc, v7, v5
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v0, v2
; GFX8-NEXT: v_cndmask_b32_e64 v2, 0, 1, vcc
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v5, v2
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v2
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v0
; GFX8-NEXT: v_addc_u32_e32 v4, vcc, v4, v1, vcc
; GFX8-NEXT: v_mad_u64_u32 v[0:1], s[0:1], s2, v3, 0
; GFX8-NEXT: v_mad_u64_u32 v[1:2], s[0:1], s2, v4, v[1:2]
; GFX8-NEXT: v_mul_hi_u32 v6, v3, v0
; GFX8-NEXT: v_mad_u64_u32 v[1:2], s[0:1], s3, v3, v[1:2]
; GFX8-NEXT: v_mul_lo_u32 v2, v4, v0
; GFX8-NEXT: v_mul_hi_u32 v0, v4, v0
; GFX8-NEXT: v_mul_lo_u32 v5, v3, v1
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v5
; GFX8-NEXT: v_cndmask_b32_e64 v5, 0, 1, vcc
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v6
; GFX8-NEXT: v_cndmask_b32_e64 v2, 0, 1, vcc
; GFX8-NEXT: v_mul_lo_u32 v6, v4, v1
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v5, v2
; GFX8-NEXT: v_mul_hi_u32 v5, v3, v1
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v6, v0
; GFX8-NEXT: v_cndmask_b32_e64 v6, 0, 1, vcc
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v0, v5
; GFX8-NEXT: v_cndmask_b32_e64 v5, 0, 1, vcc
; GFX8-NEXT: v_add_u32_e32 v5, vcc, v6, v5
; GFX8-NEXT: v_mul_hi_u32 v1, v4, v1
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v0, v2
; GFX8-NEXT: v_cndmask_b32_e64 v2, 0, 1, vcc
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v5, v2
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v2
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v3, v0
; GFX8-NEXT: v_addc_u32_e32 v1, vcc, v4, v1, vcc
; GFX8-NEXT: v_mul_lo_u32 v2, s9, v0
; GFX8-NEXT: v_mul_lo_u32 v3, s8, v1
; GFX8-NEXT: v_mul_hi_u32 v4, s8, v0
; GFX8-NEXT: v_mul_hi_u32 v0, s9, v0
; GFX8-NEXT: v_mul_hi_u32 v5, s9, v1
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v3
; GFX8-NEXT: v_cndmask_b32_e64 v3, 0, 1, vcc
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v4
; GFX8-NEXT: v_cndmask_b32_e64 v2, 0, 1, vcc
; GFX8-NEXT: v_mul_lo_u32 v4, s9, v1
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v3, v2
; GFX8-NEXT: v_mul_hi_u32 v3, s8, v1
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v4, v0
; GFX8-NEXT: v_cndmask_b32_e64 v4, 0, 1, vcc
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v0, v3
; GFX8-NEXT: v_cndmask_b32_e64 v3, 0, 1, vcc
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v4, v3
; GFX8-NEXT: v_add_u32_e32 v4, vcc, v0, v2
; GFX8-NEXT: v_mad_u64_u32 v[0:1], s[0:1], s10, v4, 0
; GFX8-NEXT: v_cndmask_b32_e64 v2, 0, 1, vcc
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v3, v2
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v5, v2
; GFX8-NEXT: v_mad_u64_u32 v[1:2], s[0:1], s10, v3, v[1:2]
; GFX8-NEXT: v_mov_b32_e32 v6, s9
; GFX8-NEXT: v_mov_b32_e32 v5, s11
; GFX8-NEXT: v_mad_u64_u32 v[1:2], s[0:1], s11, v4, v[1:2]
; GFX8-NEXT: v_sub_u32_e32 v2, vcc, s8, v0
; GFX8-NEXT: v_subb_u32_e64 v6, s[0:1], v6, v1, vcc
; GFX8-NEXT: v_sub_u32_e64 v0, s[0:1], s9, v1
; GFX8-NEXT: v_cmp_le_u32_e64 s[0:1], s11, v6
; GFX8-NEXT: v_cndmask_b32_e64 v1, 0, -1, s[0:1]
; GFX8-NEXT: v_cmp_le_u32_e64 s[0:1], s10, v2
; GFX8-NEXT: v_cndmask_b32_e64 v7, 0, -1, s[0:1]
; GFX8-NEXT: v_cmp_eq_u32_e64 s[0:1], s11, v6
; GFX8-NEXT: v_subb_u32_e32 v0, vcc, v0, v5, vcc
; GFX8-NEXT: v_cndmask_b32_e64 v1, v1, v7, s[0:1]
; GFX8-NEXT: v_subrev_u32_e32 v7, vcc, s10, v2
; GFX8-NEXT: v_subbrev_u32_e64 v8, s[0:1], 0, v0, vcc
; GFX8-NEXT: v_add_u32_e64 v9, s[0:1], 1, v4
; GFX8-NEXT: v_addc_u32_e64 v10, s[0:1], 0, v3, s[0:1]
; GFX8-NEXT: v_cmp_le_u32_e64 s[0:1], s11, v8
; GFX8-NEXT: v_cndmask_b32_e64 v11, 0, -1, s[0:1]
; GFX8-NEXT: v_cmp_le_u32_e64 s[0:1], s10, v7
; GFX8-NEXT: v_subb_u32_e32 v0, vcc, v0, v5, vcc
; GFX8-NEXT: v_cndmask_b32_e64 v12, 0, -1, s[0:1]
; GFX8-NEXT: v_cmp_eq_u32_e64 s[0:1], s11, v8
; GFX8-NEXT: v_subrev_u32_e32 v5, vcc, s10, v7
; GFX8-NEXT: v_cndmask_b32_e64 v11, v11, v12, s[0:1]
; GFX8-NEXT: v_add_u32_e64 v12, s[0:1], 1, v9
; GFX8-NEXT: v_subbrev_u32_e32 v14, vcc, 0, v0, vcc
; GFX8-NEXT: v_addc_u32_e64 v13, s[0:1], 0, v10, s[0:1]
; GFX8-NEXT: v_cmp_ne_u32_e32 vcc, 0, v11
; GFX8-NEXT: v_cndmask_b32_e32 v0, v9, v12, vcc
; GFX8-NEXT: v_cndmask_b32_e32 v9, v10, v13, vcc
; GFX8-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
; GFX8-NEXT: v_cmp_ne_u32_e64 s[0:1], 0, v11
; GFX8-NEXT: v_cndmask_b32_e32 v0, v4, v0, vcc
; GFX8-NEXT: v_cndmask_b32_e32 v1, v3, v9, vcc
; GFX8-NEXT: v_cndmask_b32_e64 v3, v7, v5, s[0:1]
; GFX8-NEXT: v_cndmask_b32_e64 v4, v8, v14, s[0:1]
; GFX8-NEXT: v_cndmask_b32_e32 v2, v2, v3, vcc
; GFX8-NEXT: v_cndmask_b32_e32 v3, v6, v4, vcc
; GFX8-NEXT: v_mov_b32_e32 v4, s4
; GFX8-NEXT: v_mov_b32_e32 v5, s5
; GFX8-NEXT: flat_store_dwordx2 v[4:5], v[0:1]
; GFX8-NEXT: v_mov_b32_e32 v0, s6
; GFX8-NEXT: v_mov_b32_e32 v1, s7
; GFX8-NEXT: flat_store_dwordx2 v[0:1], v[2:3]
; GFX8-NEXT: s_endpgm
;
; GFX9-LABEL: udivrem_i64:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_load_dwordx8 s[4:11], s[4:5], 0x0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: v_cvt_f32_u32_e32 v0, s11
; GFX9-NEXT: v_cvt_f32_u32_e32 v1, s10
; GFX9-NEXT: s_sub_u32 s2, 0, s10
; GFX9-NEXT: s_subb_u32 s3, 0, s11
; GFX9-NEXT: v_mul_f32_e32 v0, 0x4f800000, v0
; GFX9-NEXT: v_add_f32_e32 v0, v0, v1
; GFX9-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX9-NEXT: v_mul_f32_e32 v0, 0x5f7ffffc, v0
; GFX9-NEXT: v_mul_f32_e32 v1, 0x2f800000, v0
; GFX9-NEXT: v_trunc_f32_e32 v2, v1
; GFX9-NEXT: v_mul_f32_e32 v1, 0xcf800000, v2
; GFX9-NEXT: v_add_f32_e32 v0, v1, v0
; GFX9-NEXT: v_cvt_u32_f32_e32 v3, v0
; GFX9-NEXT: v_cvt_u32_f32_e32 v4, v2
; GFX9-NEXT: v_mad_u64_u32 v[0:1], s[0:1], s2, v3, 0
; GFX9-NEXT: v_mad_u64_u32 v[1:2], s[0:1], s2, v4, v[1:2]
; GFX9-NEXT: v_mul_hi_u32 v5, v3, v0
; GFX9-NEXT: v_mad_u64_u32 v[1:2], s[0:1], s3, v3, v[1:2]
; GFX9-NEXT: v_mul_lo_u32 v2, v4, v0
; GFX9-NEXT: v_mul_hi_u32 v0, v4, v0
; GFX9-NEXT: v_mul_lo_u32 v6, v3, v1
; GFX9-NEXT: v_mul_lo_u32 v7, v4, v1
; GFX9-NEXT: v_mul_hi_u32 v8, v3, v1
; GFX9-NEXT: v_mul_hi_u32 v1, v4, v1
; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, v2, v6
; GFX9-NEXT: v_cndmask_b32_e64 v6, 0, 1, vcc
; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, v7, v0
; GFX9-NEXT: v_cndmask_b32_e64 v7, 0, 1, vcc
; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, v2, v5
; GFX9-NEXT: v_cndmask_b32_e64 v2, 0, 1, vcc
; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, v0, v8
; GFX9-NEXT: v_add_u32_e32 v2, v6, v2
; GFX9-NEXT: v_cndmask_b32_e64 v5, 0, 1, vcc
; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, v0, v2
; GFX9-NEXT: v_add_u32_e32 v5, v7, v5
; GFX9-NEXT: v_cndmask_b32_e64 v2, 0, 1, vcc
; GFX9-NEXT: v_add3_u32 v1, v5, v2, v1
; GFX9-NEXT: v_add_co_u32_e32 v3, vcc, v3, v0
; GFX9-NEXT: v_addc_co_u32_e32 v4, vcc, v4, v1, vcc
; GFX9-NEXT: v_mad_u64_u32 v[0:1], s[0:1], s2, v3, 0
; GFX9-NEXT: v_mov_b32_e32 v7, s11
; GFX9-NEXT: v_mad_u64_u32 v[1:2], s[0:1], s2, v4, v[1:2]
; GFX9-NEXT: v_mul_hi_u32 v6, v3, v0
; GFX9-NEXT: v_mad_u64_u32 v[1:2], s[0:1], s3, v3, v[1:2]
; GFX9-NEXT: v_mul_lo_u32 v2, v4, v0
; GFX9-NEXT: v_mul_hi_u32 v0, v4, v0
; GFX9-NEXT: v_mul_lo_u32 v5, v3, v1
; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, v2, v5
; GFX9-NEXT: v_cndmask_b32_e64 v5, 0, 1, vcc
; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, v2, v6
; GFX9-NEXT: v_cndmask_b32_e64 v2, 0, 1, vcc
; GFX9-NEXT: v_mul_lo_u32 v6, v4, v1
; GFX9-NEXT: v_add_u32_e32 v2, v5, v2
; GFX9-NEXT: v_mul_hi_u32 v5, v3, v1
; GFX9-NEXT: v_mul_hi_u32 v1, v4, v1
; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, v6, v0
; GFX9-NEXT: v_cndmask_b32_e64 v6, 0, 1, vcc
; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, v0, v5
; GFX9-NEXT: v_cndmask_b32_e64 v5, 0, 1, vcc
; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, v0, v2
; GFX9-NEXT: v_add_u32_e32 v5, v6, v5
; GFX9-NEXT: v_cndmask_b32_e64 v2, 0, 1, vcc
; GFX9-NEXT: v_add3_u32 v1, v5, v2, v1
; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, v3, v0
; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, v4, v1, vcc
; GFX9-NEXT: v_mul_lo_u32 v2, s9, v0
; GFX9-NEXT: v_mul_lo_u32 v3, s8, v1
; GFX9-NEXT: v_mul_hi_u32 v4, s8, v0
; GFX9-NEXT: v_mul_hi_u32 v0, s9, v0
; GFX9-NEXT: v_mul_hi_u32 v6, s9, v1
; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, v2, v3
; GFX9-NEXT: v_cndmask_b32_e64 v3, 0, 1, vcc
; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, v2, v4
; GFX9-NEXT: v_cndmask_b32_e64 v2, 0, 1, vcc
; GFX9-NEXT: v_mul_lo_u32 v4, s9, v1
; GFX9-NEXT: v_add_u32_e32 v2, v3, v2
; GFX9-NEXT: v_mul_hi_u32 v3, s8, v1
; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, v4, v0
; GFX9-NEXT: v_cndmask_b32_e64 v4, 0, 1, vcc
; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, v0, v3
; GFX9-NEXT: v_cndmask_b32_e64 v3, 0, 1, vcc
; GFX9-NEXT: v_add_co_u32_e32 v5, vcc, v0, v2
; GFX9-NEXT: v_mad_u64_u32 v[0:1], s[0:1], s10, v5, 0
; GFX9-NEXT: v_cndmask_b32_e64 v2, 0, 1, vcc
; GFX9-NEXT: v_add_u32_e32 v3, v4, v3
; GFX9-NEXT: v_add3_u32 v3, v3, v2, v6
; GFX9-NEXT: v_mad_u64_u32 v[1:2], s[0:1], s10, v3, v[1:2]
; GFX9-NEXT: v_mov_b32_e32 v6, s9
; GFX9-NEXT: v_mov_b32_e32 v4, 0
; GFX9-NEXT: v_mad_u64_u32 v[1:2], s[0:1], s11, v5, v[1:2]
; GFX9-NEXT: v_sub_co_u32_e32 v2, vcc, s8, v0
; GFX9-NEXT: v_subb_co_u32_e64 v6, s[0:1], v6, v1, vcc
; GFX9-NEXT: v_cmp_le_u32_e64 s[0:1], s11, v6
; GFX9-NEXT: v_sub_u32_e32 v0, s9, v1
; GFX9-NEXT: v_cndmask_b32_e64 v1, 0, -1, s[0:1]
; GFX9-NEXT: v_cmp_le_u32_e64 s[0:1], s10, v2
; GFX9-NEXT: v_cndmask_b32_e64 v8, 0, -1, s[0:1]
; GFX9-NEXT: v_cmp_eq_u32_e64 s[0:1], s11, v6
; GFX9-NEXT: v_subb_co_u32_e32 v0, vcc, v0, v7, vcc
; GFX9-NEXT: v_cndmask_b32_e64 v1, v1, v8, s[0:1]
; GFX9-NEXT: v_subrev_co_u32_e32 v8, vcc, s10, v2
; GFX9-NEXT: v_subbrev_co_u32_e64 v9, s[0:1], 0, v0, vcc
; GFX9-NEXT: v_add_co_u32_e64 v10, s[0:1], 1, v5
; GFX9-NEXT: v_addc_co_u32_e64 v11, s[0:1], 0, v3, s[0:1]
; GFX9-NEXT: v_cmp_le_u32_e64 s[0:1], s11, v9
; GFX9-NEXT: v_cndmask_b32_e64 v12, 0, -1, s[0:1]
; GFX9-NEXT: v_cmp_le_u32_e64 s[0:1], s10, v8
; GFX9-NEXT: v_subb_co_u32_e32 v0, vcc, v0, v7, vcc
; GFX9-NEXT: v_cndmask_b32_e64 v13, 0, -1, s[0:1]
; GFX9-NEXT: v_cmp_eq_u32_e64 s[0:1], s11, v9
; GFX9-NEXT: v_subrev_co_u32_e32 v7, vcc, s10, v8
; GFX9-NEXT: v_cndmask_b32_e64 v12, v12, v13, s[0:1]
; GFX9-NEXT: v_add_co_u32_e64 v13, s[0:1], 1, v10
; GFX9-NEXT: v_subbrev_co_u32_e32 v15, vcc, 0, v0, vcc
; GFX9-NEXT: v_addc_co_u32_e64 v14, s[0:1], 0, v11, s[0:1]
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v12
; GFX9-NEXT: v_cndmask_b32_e32 v0, v10, v13, vcc
; GFX9-NEXT: v_cndmask_b32_e32 v10, v11, v14, vcc
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1
; GFX9-NEXT: v_cmp_ne_u32_e64 s[0:1], 0, v12
; GFX9-NEXT: v_cndmask_b32_e32 v0, v5, v0, vcc
; GFX9-NEXT: v_cndmask_b32_e32 v1, v3, v10, vcc
; GFX9-NEXT: v_cndmask_b32_e64 v3, v8, v7, s[0:1]
; GFX9-NEXT: v_cndmask_b32_e64 v5, v9, v15, s[0:1]
; GFX9-NEXT: v_cndmask_b32_e32 v2, v2, v3, vcc
; GFX9-NEXT: v_cndmask_b32_e32 v3, v6, v5, vcc
; GFX9-NEXT: global_store_dwordx2 v4, v[0:1], s[4:5]
; GFX9-NEXT: global_store_dwordx2 v4, v[2:3], s[6:7]
; GFX9-NEXT: s_endpgm
;
; GFX10-LABEL: udivrem_i64:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_load_dwordx8 s[4:11], s[4:5], 0x0
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: v_cvt_f32_u32_e32 v0, s11
; GFX10-NEXT: v_cvt_f32_u32_e32 v1, s10
; GFX10-NEXT: s_sub_u32 s0, 0, s10
; GFX10-NEXT: s_subb_u32 s1, 0, s11
; GFX10-NEXT: v_mul_f32_e32 v0, 0x4f800000, v0
; GFX10-NEXT: v_add_f32_e32 v0, v0, v1
; GFX10-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX10-NEXT: v_mul_f32_e32 v0, 0x5f7ffffc, v0
; GFX10-NEXT: v_mul_f32_e32 v1, 0x2f800000, v0
; GFX10-NEXT: v_trunc_f32_e32 v1, v1
; GFX10-NEXT: v_mul_f32_e32 v2, 0xcf800000, v1
; GFX10-NEXT: v_add_f32_e32 v0, v2, v0
; GFX10-NEXT: v_cvt_u32_f32_e32 v2, v1
; GFX10-NEXT: v_cvt_u32_f32_e32 v3, v0
; GFX10-NEXT: v_mul_lo_u32 v4, s0, v2
; GFX10-NEXT: v_mad_u64_u32 v[0:1], s2, s0, v3, 0
; GFX10-NEXT: v_mul_lo_u32 v5, s1, v3
; GFX10-NEXT: v_mul_hi_u32 v6, v2, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v4, v5
; GFX10-NEXT: v_mul_lo_u32 v4, v2, v0
; GFX10-NEXT: v_mul_hi_u32 v0, v3, v0
; GFX10-NEXT: v_mul_lo_u32 v5, v3, v1
; GFX10-NEXT: v_mul_lo_u32 v7, v2, v1
; GFX10-NEXT: v_mul_hi_u32 v8, v3, v1
; GFX10-NEXT: v_mul_hi_u32 v1, v2, v1
; GFX10-NEXT: v_add_co_u32 v4, s2, v4, v5
; GFX10-NEXT: v_cndmask_b32_e64 v5, 0, 1, s2
; GFX10-NEXT: v_add_co_u32 v6, s2, v7, v6
; GFX10-NEXT: v_cndmask_b32_e64 v7, 0, 1, s2
; GFX10-NEXT: v_add_co_u32 v0, s2, v4, v0
; GFX10-NEXT: v_cndmask_b32_e64 v0, 0, 1, s2
; GFX10-NEXT: v_add_co_u32 v4, s2, v6, v8
; GFX10-NEXT: v_cndmask_b32_e64 v6, 0, 1, s2
; GFX10-NEXT: v_add_nc_u32_e32 v0, v5, v0
; GFX10-NEXT: v_add_nc_u32_e32 v5, v7, v6
; GFX10-NEXT: v_add_co_u32 v0, s2, v4, v0
; GFX10-NEXT: v_cndmask_b32_e64 v4, 0, 1, s2
; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, v3, v0
; GFX10-NEXT: v_add3_u32 v1, v5, v4, v1
; GFX10-NEXT: v_mul_lo_u32 v4, s1, v3
; GFX10-NEXT: v_add_co_ci_u32_e32 v2, vcc_lo, v2, v1, vcc_lo
; GFX10-NEXT: v_mad_u64_u32 v[0:1], s2, s0, v3, 0
; GFX10-NEXT: v_mul_lo_u32 v5, s0, v2
; GFX10-NEXT: v_mul_hi_u32 v6, v2, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v5, v4
; GFX10-NEXT: v_mul_lo_u32 v4, v2, v0
; GFX10-NEXT: v_mul_hi_u32 v0, v3, v0
; GFX10-NEXT: v_mul_lo_u32 v5, v3, v1
; GFX10-NEXT: v_mul_lo_u32 v7, v2, v1
; GFX10-NEXT: v_mul_hi_u32 v8, v3, v1
; GFX10-NEXT: v_mul_hi_u32 v1, v2, v1
; GFX10-NEXT: v_add_co_u32 v4, s0, v4, v5
; GFX10-NEXT: v_cndmask_b32_e64 v5, 0, 1, s0
; GFX10-NEXT: v_add_co_u32 v6, s0, v7, v6
; GFX10-NEXT: v_cndmask_b32_e64 v7, 0, 1, s0
; GFX10-NEXT: v_add_co_u32 v0, s0, v4, v0
; GFX10-NEXT: v_cndmask_b32_e64 v0, 0, 1, s0
; GFX10-NEXT: v_add_co_u32 v4, s0, v6, v8
; GFX10-NEXT: v_cndmask_b32_e64 v6, 0, 1, s0
; GFX10-NEXT: v_add_nc_u32_e32 v0, v5, v0
; GFX10-NEXT: v_add_nc_u32_e32 v5, v7, v6
; GFX10-NEXT: v_add_co_u32 v0, s0, v4, v0
; GFX10-NEXT: v_cndmask_b32_e64 v4, 0, 1, s0
; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, v3, v0
; GFX10-NEXT: v_add3_u32 v1, v5, v4, v1
; GFX10-NEXT: v_mul_hi_u32 v4, s9, v0
; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, v2, v1, vcc_lo
; GFX10-NEXT: v_mul_lo_u32 v2, s9, v0
; GFX10-NEXT: v_mul_hi_u32 v0, s8, v0
; GFX10-NEXT: v_mul_lo_u32 v3, s8, v1
; GFX10-NEXT: v_mul_lo_u32 v5, s9, v1
; GFX10-NEXT: v_mul_hi_u32 v6, s8, v1
; GFX10-NEXT: v_mul_hi_u32 v1, s9, v1
; GFX10-NEXT: v_add_co_u32 v2, s0, v2, v3
; GFX10-NEXT: v_cndmask_b32_e64 v3, 0, 1, s0
; GFX10-NEXT: v_add_co_u32 v4, s0, v5, v4
; GFX10-NEXT: v_cndmask_b32_e64 v5, 0, 1, s0
; GFX10-NEXT: v_add_co_u32 v0, s0, v2, v0
; GFX10-NEXT: v_cndmask_b32_e64 v0, 0, 1, s0
; GFX10-NEXT: v_add_co_u32 v2, s0, v4, v6
; GFX10-NEXT: v_cndmask_b32_e64 v4, 0, 1, s0
; GFX10-NEXT: v_add_nc_u32_e32 v0, v3, v0
; GFX10-NEXT: v_add_nc_u32_e32 v3, v5, v4
; GFX10-NEXT: v_add_co_u32 v2, s0, v2, v0
; GFX10-NEXT: v_cndmask_b32_e64 v0, 0, 1, s0
; GFX10-NEXT: v_mul_lo_u32 v4, s11, v2
; GFX10-NEXT: v_add_co_u32 v6, vcc_lo, v2, 1
; GFX10-NEXT: v_add3_u32 v3, v3, v0, v1
; GFX10-NEXT: v_mad_u64_u32 v[0:1], s0, s10, v2, 0
; GFX10-NEXT: v_mul_lo_u32 v5, s10, v3
; GFX10-NEXT: v_add_co_ci_u32_e32 v7, vcc_lo, 0, v3, vcc_lo
; GFX10-NEXT: v_add3_u32 v1, v1, v5, v4
; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, v6, 1
; GFX10-NEXT: v_add_co_ci_u32_e32 v5, vcc_lo, 0, v7, vcc_lo
; GFX10-NEXT: v_sub_nc_u32_e32 v8, s9, v1
; GFX10-NEXT: v_sub_co_u32 v9, vcc_lo, s8, v0
; GFX10-NEXT: v_sub_co_ci_u32_e64 v10, s0, s9, v1, vcc_lo
; GFX10-NEXT: v_subrev_co_ci_u32_e32 v0, vcc_lo, s11, v8, vcc_lo
; GFX10-NEXT: v_cmp_le_u32_e32 vcc_lo, s10, v9
; GFX10-NEXT: v_cndmask_b32_e64 v1, 0, -1, vcc_lo
; GFX10-NEXT: v_sub_co_u32 v8, vcc_lo, v9, s10
; GFX10-NEXT: v_subrev_co_ci_u32_e64 v11, s0, 0, v0, vcc_lo
; GFX10-NEXT: v_cmp_le_u32_e64 s0, s11, v10
; GFX10-NEXT: v_subrev_co_ci_u32_e32 v0, vcc_lo, s11, v0, vcc_lo
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, s11, v11
; GFX10-NEXT: v_cndmask_b32_e64 v12, 0, -1, s0
; GFX10-NEXT: v_cmp_le_u32_e64 s0, s10, v8
; GFX10-NEXT: v_cndmask_b32_e64 v13, 0, -1, s0
; GFX10-NEXT: v_cmp_le_u32_e64 s0, s11, v11
; GFX10-NEXT: v_cndmask_b32_e64 v14, 0, -1, s0
; GFX10-NEXT: v_cmp_eq_u32_e64 s0, s11, v10
; GFX10-NEXT: v_cndmask_b32_e64 v1, v12, v1, s0
; GFX10-NEXT: v_cndmask_b32_e32 v12, v14, v13, vcc_lo
; GFX10-NEXT: v_sub_co_u32 v13, vcc_lo, v8, s10
; GFX10-NEXT: v_subrev_co_ci_u32_e32 v0, vcc_lo, 0, v0, vcc_lo
; GFX10-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v12
; GFX10-NEXT: v_cmp_ne_u32_e64 s0, 0, v12
; GFX10-NEXT: v_cmp_ne_u32_e64 s1, 0, v1
; GFX10-NEXT: v_cndmask_b32_e32 v4, v6, v4, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v1, v7, v5, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e64 v6, v8, v13, s0
; GFX10-NEXT: v_cndmask_b32_e64 v5, v11, v0, s0
; GFX10-NEXT: v_mov_b32_e32 v7, 0
; GFX10-NEXT: v_cndmask_b32_e64 v0, v2, v4, s1
; GFX10-NEXT: v_cndmask_b32_e64 v1, v3, v1, s1
; GFX10-NEXT: v_cndmask_b32_e64 v2, v9, v6, s1
; GFX10-NEXT: v_cndmask_b32_e64 v3, v10, v5, s1
; GFX10-NEXT: global_store_dwordx2 v7, v[0:1], s[4:5]
; GFX10-NEXT: global_store_dwordx2 v7, v[2:3], s[6:7]
; GFX10-NEXT: s_endpgm
%div = udiv i64 %x, %y
store i64 %div, i64 addrspace(1)* %out0
%rem = urem i64 %x, %y
store i64 %rem, i64 addrspace(1)* %out1
ret void
}
define amdgpu_kernel void @udivrem_v2i32(<2 x i32> addrspace(1)* %out0, <2 x i32> addrspace(1)* %out1, <2 x i32> %x, <2 x i32> %y) {
; GFX8-LABEL: udivrem_v2i32:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_load_dwordx4 s[8:11], s[4:5], 0x10
; GFX8-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x0
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
; GFX8-NEXT: v_cvt_f32_u32_e32 v0, s10
; GFX8-NEXT: v_cvt_f32_u32_e32 v1, s11
; GFX8-NEXT: s_sub_i32 s0, 0, s10
; GFX8-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX8-NEXT: v_rcp_iflag_f32_e32 v1, v1
; GFX8-NEXT: v_mul_f32_e32 v0, 0x4f7ffffe, v0
; GFX8-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX8-NEXT: v_mul_f32_e32 v1, 0x4f7ffffe, v1
; GFX8-NEXT: v_cvt_u32_f32_e32 v1, v1
; GFX8-NEXT: v_mul_lo_u32 v2, s0, v0
; GFX8-NEXT: s_sub_i32 s0, 0, s11
; GFX8-NEXT: v_mul_lo_u32 v3, s0, v1
; GFX8-NEXT: v_mul_hi_u32 v2, v0, v2
; GFX8-NEXT: v_mul_hi_u32 v3, v1, v3
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v0, v2
; GFX8-NEXT: v_mul_hi_u32 v0, s8, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v3
; GFX8-NEXT: v_mul_hi_u32 v1, s9, v1
; GFX8-NEXT: v_mul_lo_u32 v2, v0, s10
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 1, v0
; GFX8-NEXT: v_mul_lo_u32 v4, v1, s11
; GFX8-NEXT: v_sub_u32_e32 v2, vcc, s8, v2
; GFX8-NEXT: v_cmp_le_u32_e32 vcc, s10, v2
; GFX8-NEXT: v_cndmask_b32_e32 v0, v0, v3, vcc
; GFX8-NEXT: v_subrev_u32_e64 v3, s[0:1], s10, v2
; GFX8-NEXT: v_cndmask_b32_e32 v2, v2, v3, vcc
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 1, v0
; GFX8-NEXT: v_cmp_le_u32_e32 vcc, s10, v2
; GFX8-NEXT: v_cndmask_b32_e32 v0, v0, v3, vcc
; GFX8-NEXT: v_subrev_u32_e64 v3, s[0:1], s10, v2
; GFX8-NEXT: v_cndmask_b32_e32 v2, v2, v3, vcc
; GFX8-NEXT: v_sub_u32_e32 v3, vcc, s9, v4
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 1, v1
; GFX8-NEXT: v_cmp_le_u32_e32 vcc, s11, v3
; GFX8-NEXT: v_cndmask_b32_e32 v1, v1, v4, vcc
; GFX8-NEXT: v_subrev_u32_e64 v4, s[0:1], s11, v3
; GFX8-NEXT: v_cndmask_b32_e32 v3, v3, v4, vcc
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 1, v1
; GFX8-NEXT: v_cmp_le_u32_e32 vcc, s11, v3
; GFX8-NEXT: v_cndmask_b32_e32 v1, v1, v4, vcc
; GFX8-NEXT: v_subrev_u32_e64 v4, s[0:1], s11, v3
; GFX8-NEXT: v_cndmask_b32_e32 v3, v3, v4, vcc
; GFX8-NEXT: v_mov_b32_e32 v4, s4
; GFX8-NEXT: v_mov_b32_e32 v5, s5
; GFX8-NEXT: flat_store_dwordx2 v[4:5], v[0:1]
; GFX8-NEXT: v_mov_b32_e32 v0, s6
; GFX8-NEXT: v_mov_b32_e32 v1, s7
; GFX8-NEXT: flat_store_dwordx2 v[0:1], v[2:3]
; GFX8-NEXT: s_endpgm
;
; GFX9-LABEL: udivrem_v2i32:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x10
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: v_cvt_f32_u32_e32 v0, s2
; GFX9-NEXT: v_cvt_f32_u32_e32 v1, s3
; GFX9-NEXT: s_sub_i32 s6, 0, s2
; GFX9-NEXT: s_sub_i32 s7, 0, s3
; GFX9-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX9-NEXT: v_rcp_iflag_f32_e32 v1, v1
; GFX9-NEXT: v_mul_f32_e32 v0, 0x4f7ffffe, v0
; GFX9-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX9-NEXT: v_mul_f32_e32 v1, 0x4f7ffffe, v1
; GFX9-NEXT: v_cvt_u32_f32_e32 v1, v1
; GFX9-NEXT: v_mul_lo_u32 v2, s6, v0
; GFX9-NEXT: v_mul_lo_u32 v3, s7, v1
; GFX9-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x0
; GFX9-NEXT: v_mul_hi_u32 v2, v0, v2
; GFX9-NEXT: v_mul_hi_u32 v3, v1, v3
; GFX9-NEXT: v_add_u32_e32 v0, v0, v2
; GFX9-NEXT: v_mul_hi_u32 v0, s0, v0
; GFX9-NEXT: v_add_u32_e32 v1, v1, v3
; GFX9-NEXT: v_mul_hi_u32 v1, s1, v1
; GFX9-NEXT: v_mul_lo_u32 v2, v0, s2
; GFX9-NEXT: v_add_u32_e32 v4, 1, v0
; GFX9-NEXT: v_mul_lo_u32 v3, v1, s3
; GFX9-NEXT: v_add_u32_e32 v5, 1, v1
; GFX9-NEXT: v_sub_u32_e32 v2, s0, v2
; GFX9-NEXT: v_cmp_le_u32_e32 vcc, s2, v2
; GFX9-NEXT: v_sub_u32_e32 v3, s1, v3
; GFX9-NEXT: v_cndmask_b32_e32 v0, v0, v4, vcc
; GFX9-NEXT: v_subrev_u32_e32 v4, s2, v2
; GFX9-NEXT: v_cmp_le_u32_e64 s[0:1], s3, v3
; GFX9-NEXT: v_cndmask_b32_e32 v2, v2, v4, vcc
; GFX9-NEXT: v_cndmask_b32_e64 v1, v1, v5, s[0:1]
; GFX9-NEXT: v_subrev_u32_e32 v5, s3, v3
; GFX9-NEXT: v_add_u32_e32 v4, 1, v0
; GFX9-NEXT: v_cmp_le_u32_e32 vcc, s2, v2
; GFX9-NEXT: v_cndmask_b32_e64 v3, v3, v5, s[0:1]
; GFX9-NEXT: v_cndmask_b32_e32 v0, v0, v4, vcc
; GFX9-NEXT: v_subrev_u32_e32 v4, s2, v2
; GFX9-NEXT: v_add_u32_e32 v5, 1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v2, v2, v4, vcc
; GFX9-NEXT: v_cmp_le_u32_e32 vcc, s3, v3
; GFX9-NEXT: v_subrev_u32_e32 v4, s3, v3
; GFX9-NEXT: v_cndmask_b32_e32 v1, v1, v5, vcc
; GFX9-NEXT: v_cndmask_b32_e32 v3, v3, v4, vcc
; GFX9-NEXT: v_mov_b32_e32 v4, 0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: global_store_dwordx2 v4, v[0:1], s[4:5]
; GFX9-NEXT: global_store_dwordx2 v4, v[2:3], s[6:7]
; GFX9-NEXT: s_endpgm
;
; GFX10-LABEL: udivrem_v2i32:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x10
; GFX10-NEXT: v_mov_b32_e32 v8, 0
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: v_cvt_f32_u32_e32 v0, s2
; GFX10-NEXT: v_cvt_f32_u32_e32 v1, s3
; GFX10-NEXT: s_sub_i32 s6, 0, s2
; GFX10-NEXT: s_sub_i32 s7, 0, s3
; GFX10-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX10-NEXT: v_rcp_iflag_f32_e32 v1, v1
; GFX10-NEXT: v_mul_f32_e32 v0, 0x4f7ffffe, v0
; GFX10-NEXT: v_mul_f32_e32 v1, 0x4f7ffffe, v1
; GFX10-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX10-NEXT: v_cvt_u32_f32_e32 v1, v1
; GFX10-NEXT: v_mul_lo_u32 v2, s6, v0
; GFX10-NEXT: v_mul_lo_u32 v3, s7, v1
; GFX10-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x0
; GFX10-NEXT: v_mul_hi_u32 v2, v0, v2
; GFX10-NEXT: v_mul_hi_u32 v3, v1, v3
; GFX10-NEXT: v_add_nc_u32_e32 v0, v0, v2
; GFX10-NEXT: v_add_nc_u32_e32 v1, v1, v3
; GFX10-NEXT: v_mul_hi_u32 v0, s0, v0
; GFX10-NEXT: v_mul_hi_u32 v1, s1, v1
; GFX10-NEXT: v_mul_lo_u32 v2, v0, s2
; GFX10-NEXT: v_mul_lo_u32 v3, v1, s3
; GFX10-NEXT: v_add_nc_u32_e32 v4, 1, v0
; GFX10-NEXT: v_add_nc_u32_e32 v5, 1, v1
; GFX10-NEXT: v_sub_nc_u32_e32 v2, s0, v2
; GFX10-NEXT: v_sub_nc_u32_e32 v3, s1, v3
; GFX10-NEXT: v_subrev_nc_u32_e32 v6, s2, v2
; GFX10-NEXT: v_cmp_le_u32_e32 vcc_lo, s3, v3
; GFX10-NEXT: v_cmp_le_u32_e64 s0, s2, v2
; GFX10-NEXT: v_subrev_nc_u32_e32 v7, s3, v3
; GFX10-NEXT: v_cndmask_b32_e32 v1, v1, v5, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e64 v0, v0, v4, s0
; GFX10-NEXT: v_cndmask_b32_e64 v2, v2, v6, s0
; GFX10-NEXT: v_cndmask_b32_e32 v3, v3, v7, vcc_lo
; GFX10-NEXT: v_add_nc_u32_e32 v5, 1, v1
; GFX10-NEXT: v_add_nc_u32_e32 v4, 1, v0
; GFX10-NEXT: v_cmp_le_u32_e64 s0, s2, v2
; GFX10-NEXT: v_cmp_le_u32_e32 vcc_lo, s3, v3
; GFX10-NEXT: v_subrev_nc_u32_e32 v6, s2, v2
; GFX10-NEXT: v_subrev_nc_u32_e32 v7, s3, v3
; GFX10-NEXT: v_cndmask_b32_e64 v0, v0, v4, s0
; GFX10-NEXT: v_cndmask_b32_e32 v1, v1, v5, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e64 v2, v2, v6, s0
; GFX10-NEXT: v_cndmask_b32_e32 v3, v3, v7, vcc_lo
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: global_store_dwordx2 v8, v[0:1], s[4:5]
; GFX10-NEXT: global_store_dwordx2 v8, v[2:3], s[6:7]
; GFX10-NEXT: s_endpgm
%div = udiv <2 x i32> %x, %y
store <2 x i32> %div, <2 x i32> addrspace(1)* %out0
%rem = urem <2 x i32> %x, %y
store <2 x i32> %rem, <2 x i32> addrspace(1)* %out1
ret void
}
define amdgpu_kernel void @udivrem_v4i32(<4 x i32> addrspace(1)* %out0, <4 x i32> addrspace(1)* %out1, <4 x i32> %x, <4 x i32> %y) {
; GFX8-LABEL: udivrem_v4i32:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_load_dwordx8 s[8:15], s[4:5], 0x10
; GFX8-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x0
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
; GFX8-NEXT: v_cvt_f32_u32_e32 v0, s12
; GFX8-NEXT: v_cvt_f32_u32_e32 v1, s13
; GFX8-NEXT: s_sub_i32 s0, 0, s12
; GFX8-NEXT: v_cvt_f32_u32_e32 v6, s14
; GFX8-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX8-NEXT: v_rcp_iflag_f32_e32 v1, v1
; GFX8-NEXT: v_mul_f32_e32 v0, 0x4f7ffffe, v0
; GFX8-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX8-NEXT: v_mul_f32_e32 v1, 0x4f7ffffe, v1
; GFX8-NEXT: v_cvt_u32_f32_e32 v1, v1
; GFX8-NEXT: v_mul_lo_u32 v2, s0, v0
; GFX8-NEXT: s_sub_i32 s0, 0, s13
; GFX8-NEXT: v_mul_lo_u32 v3, s0, v1
; GFX8-NEXT: v_mul_hi_u32 v2, v0, v2
; GFX8-NEXT: v_mul_hi_u32 v3, v1, v3
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v0, v2
; GFX8-NEXT: v_mul_hi_u32 v0, s8, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v3
; GFX8-NEXT: v_mul_hi_u32 v1, s9, v1
; GFX8-NEXT: v_mul_lo_u32 v2, v0, s12
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 1, v0
; GFX8-NEXT: v_mul_lo_u32 v5, v1, s13
; GFX8-NEXT: v_sub_u32_e32 v2, vcc, s8, v2
; GFX8-NEXT: v_cmp_le_u32_e32 vcc, s12, v2
; GFX8-NEXT: v_cndmask_b32_e32 v0, v0, v3, vcc
; GFX8-NEXT: v_subrev_u32_e64 v3, s[0:1], s12, v2
; GFX8-NEXT: v_cndmask_b32_e32 v2, v2, v3, vcc
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 1, v0
; GFX8-NEXT: v_cmp_le_u32_e32 vcc, s12, v2
; GFX8-NEXT: v_cndmask_b32_e32 v0, v0, v3, vcc
; GFX8-NEXT: v_subrev_u32_e64 v3, s[0:1], s12, v2
; GFX8-NEXT: v_cndmask_b32_e32 v4, v2, v3, vcc
; GFX8-NEXT: v_rcp_iflag_f32_e32 v3, v6
; GFX8-NEXT: v_sub_u32_e32 v2, vcc, s9, v5
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 1, v1
; GFX8-NEXT: v_mul_f32_e32 v3, 0x4f7ffffe, v3
; GFX8-NEXT: v_cvt_u32_f32_e32 v3, v3
; GFX8-NEXT: v_cmp_le_u32_e32 vcc, s13, v2
; GFX8-NEXT: v_cndmask_b32_e32 v1, v1, v5, vcc
; GFX8-NEXT: v_subrev_u32_e64 v5, s[0:1], s13, v2
; GFX8-NEXT: s_sub_i32 s0, 0, s14
; GFX8-NEXT: v_mul_lo_u32 v6, s0, v3
; GFX8-NEXT: v_cndmask_b32_e32 v2, v2, v5, vcc
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 1, v1
; GFX8-NEXT: v_cmp_le_u32_e32 vcc, s13, v2
; GFX8-NEXT: v_mul_hi_u32 v6, v3, v6
; GFX8-NEXT: v_cndmask_b32_e32 v1, v1, v5, vcc
; GFX8-NEXT: v_cvt_f32_u32_e32 v5, s15
; GFX8-NEXT: v_add_u32_e64 v3, s[0:1], v3, v6
; GFX8-NEXT: v_rcp_iflag_f32_e32 v6, v5
; GFX8-NEXT: v_mul_hi_u32 v3, s10, v3
; GFX8-NEXT: v_subrev_u32_e64 v5, s[0:1], s13, v2
; GFX8-NEXT: v_mul_f32_e32 v6, 0x4f7ffffe, v6
; GFX8-NEXT: v_cvt_u32_f32_e32 v6, v6
; GFX8-NEXT: s_sub_i32 s0, 0, s15
; GFX8-NEXT: v_cndmask_b32_e32 v5, v2, v5, vcc
; GFX8-NEXT: v_mul_lo_u32 v2, v3, s14
; GFX8-NEXT: v_mul_lo_u32 v7, s0, v6
; GFX8-NEXT: v_add_u32_e32 v8, vcc, 1, v3
; GFX8-NEXT: v_sub_u32_e32 v2, vcc, s10, v2
; GFX8-NEXT: v_mul_hi_u32 v7, v6, v7
; GFX8-NEXT: v_cmp_le_u32_e32 vcc, s14, v2
; GFX8-NEXT: v_cndmask_b32_e32 v3, v3, v8, vcc
; GFX8-NEXT: v_subrev_u32_e64 v8, s[0:1], s14, v2
; GFX8-NEXT: v_cndmask_b32_e32 v8, v2, v8, vcc
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v6, v7
; GFX8-NEXT: v_mul_hi_u32 v7, s11, v2
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 1, v3
; GFX8-NEXT: v_cmp_le_u32_e32 vcc, s14, v8
; GFX8-NEXT: v_cndmask_b32_e32 v2, v3, v2, vcc
; GFX8-NEXT: v_mul_lo_u32 v3, v7, s15
; GFX8-NEXT: v_subrev_u32_e64 v6, s[0:1], s14, v8
; GFX8-NEXT: v_cndmask_b32_e32 v6, v8, v6, vcc
; GFX8-NEXT: v_sub_u32_e32 v3, vcc, s11, v3
; GFX8-NEXT: v_add_u32_e32 v8, vcc, 1, v7
; GFX8-NEXT: v_cmp_le_u32_e32 vcc, s15, v3
; GFX8-NEXT: v_cndmask_b32_e32 v7, v7, v8, vcc
; GFX8-NEXT: v_subrev_u32_e64 v8, s[0:1], s15, v3
; GFX8-NEXT: v_cndmask_b32_e32 v8, v3, v8, vcc
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 1, v7
; GFX8-NEXT: v_cmp_le_u32_e32 vcc, s15, v8
; GFX8-NEXT: v_cndmask_b32_e32 v3, v7, v3, vcc
; GFX8-NEXT: v_subrev_u32_e64 v7, s[0:1], s15, v8
; GFX8-NEXT: v_cndmask_b32_e32 v7, v8, v7, vcc
; GFX8-NEXT: v_mov_b32_e32 v9, s5
; GFX8-NEXT: v_mov_b32_e32 v8, s4
; GFX8-NEXT: flat_store_dwordx4 v[8:9], v[0:3]
; GFX8-NEXT: s_nop 0
; GFX8-NEXT: v_mov_b32_e32 v0, s6
; GFX8-NEXT: v_mov_b32_e32 v1, s7
; GFX8-NEXT: flat_store_dwordx4 v[0:1], v[4:7]
; GFX8-NEXT: s_endpgm
;
; GFX9-LABEL: udivrem_v4i32:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_load_dwordx8 s[8:15], s[4:5], 0x10
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: v_cvt_f32_u32_e32 v0, s12
; GFX9-NEXT: v_cvt_f32_u32_e32 v1, s13
; GFX9-NEXT: s_sub_i32 s0, 0, s12
; GFX9-NEXT: s_sub_i32 s1, 0, s13
; GFX9-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX9-NEXT: v_rcp_iflag_f32_e32 v1, v1
; GFX9-NEXT: v_cvt_f32_u32_e32 v4, s14
; GFX9-NEXT: v_mul_f32_e32 v0, 0x4f7ffffe, v0
; GFX9-NEXT: v_mul_f32_e32 v1, 0x4f7ffffe, v1
; GFX9-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX9-NEXT: v_cvt_u32_f32_e32 v1, v1
; GFX9-NEXT: v_rcp_iflag_f32_e32 v4, v4
; GFX9-NEXT: v_mul_lo_u32 v2, s0, v0
; GFX9-NEXT: v_mul_lo_u32 v3, s1, v1
; GFX9-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x0
; GFX9-NEXT: s_sub_i32 s4, 0, s14
; GFX9-NEXT: v_mul_hi_u32 v2, v0, v2
; GFX9-NEXT: v_mul_hi_u32 v3, v1, v3
; GFX9-NEXT: v_add_u32_e32 v0, v0, v2
; GFX9-NEXT: v_add_u32_e32 v1, v1, v3
; GFX9-NEXT: v_mul_hi_u32 v0, s8, v0
; GFX9-NEXT: v_mul_hi_u32 v1, s9, v1
; GFX9-NEXT: v_mul_f32_e32 v2, 0x4f7ffffe, v4
; GFX9-NEXT: v_cvt_u32_f32_e32 v2, v2
; GFX9-NEXT: v_mul_lo_u32 v3, v0, s12
; GFX9-NEXT: v_mul_lo_u32 v4, v1, s13
; GFX9-NEXT: v_add_u32_e32 v5, 1, v0
; GFX9-NEXT: v_add_u32_e32 v6, 1, v1
; GFX9-NEXT: v_sub_u32_e32 v3, s8, v3
; GFX9-NEXT: v_sub_u32_e32 v7, s9, v4
; GFX9-NEXT: v_cmp_le_u32_e32 vcc, s12, v3
; GFX9-NEXT: v_subrev_u32_e32 v4, s12, v3
; GFX9-NEXT: v_cndmask_b32_e32 v0, v0, v5, vcc
; GFX9-NEXT: v_cndmask_b32_e32 v3, v3, v4, vcc
; GFX9-NEXT: v_add_u32_e32 v4, 1, v0
; GFX9-NEXT: v_cmp_le_u32_e32 vcc, s12, v3
; GFX9-NEXT: v_cndmask_b32_e32 v0, v0, v4, vcc
; GFX9-NEXT: v_subrev_u32_e32 v4, s12, v3
; GFX9-NEXT: v_cndmask_b32_e32 v4, v3, v4, vcc
; GFX9-NEXT: v_mul_lo_u32 v3, s4, v2
; GFX9-NEXT: v_cvt_f32_u32_e32 v5, s15
; GFX9-NEXT: v_cmp_le_u32_e32 vcc, s13, v7
; GFX9-NEXT: v_cndmask_b32_e32 v1, v1, v6, vcc
; GFX9-NEXT: v_mul_hi_u32 v3, v2, v3
; GFX9-NEXT: v_rcp_iflag_f32_e32 v5, v5
; GFX9-NEXT: v_subrev_u32_e32 v6, s13, v7
; GFX9-NEXT: v_cndmask_b32_e32 v6, v7, v6, vcc
; GFX9-NEXT: v_add_u32_e32 v2, v2, v3
; GFX9-NEXT: v_mul_f32_e32 v3, 0x4f7ffffe, v5
; GFX9-NEXT: v_mul_hi_u32 v2, s10, v2
; GFX9-NEXT: v_cvt_u32_f32_e32 v3, v3
; GFX9-NEXT: v_add_u32_e32 v7, 1, v1
; GFX9-NEXT: v_cmp_le_u32_e32 vcc, s13, v6
; GFX9-NEXT: s_sub_i32 s4, 0, s15
; GFX9-NEXT: v_cndmask_b32_e32 v1, v1, v7, vcc
; GFX9-NEXT: v_mul_lo_u32 v7, v2, s14
; GFX9-NEXT: v_mul_lo_u32 v8, s4, v3
; GFX9-NEXT: v_subrev_u32_e32 v5, s13, v6
; GFX9-NEXT: v_cndmask_b32_e32 v5, v6, v5, vcc
; GFX9-NEXT: v_sub_u32_e32 v6, s10, v7
; GFX9-NEXT: v_mul_hi_u32 v7, v3, v8
; GFX9-NEXT: v_add_u32_e32 v8, 1, v2
; GFX9-NEXT: v_cmp_le_u32_e32 vcc, s14, v6
; GFX9-NEXT: v_cndmask_b32_e32 v2, v2, v8, vcc
; GFX9-NEXT: v_add_u32_e32 v3, v3, v7
; GFX9-NEXT: v_mul_hi_u32 v3, s11, v3
; GFX9-NEXT: v_subrev_u32_e32 v7, s14, v6
; GFX9-NEXT: v_cndmask_b32_e32 v6, v6, v7, vcc
; GFX9-NEXT: v_add_u32_e32 v7, 1, v2
; GFX9-NEXT: v_mul_lo_u32 v8, v3, s15
; GFX9-NEXT: v_cmp_le_u32_e32 vcc, s14, v6
; GFX9-NEXT: v_cndmask_b32_e32 v2, v2, v7, vcc
; GFX9-NEXT: v_subrev_u32_e32 v7, s14, v6
; GFX9-NEXT: v_cndmask_b32_e32 v6, v6, v7, vcc
; GFX9-NEXT: v_sub_u32_e32 v7, s11, v8
; GFX9-NEXT: v_add_u32_e32 v8, 1, v3
; GFX9-NEXT: v_cmp_le_u32_e32 vcc, s15, v7
; GFX9-NEXT: v_cndmask_b32_e32 v3, v3, v8, vcc
; GFX9-NEXT: v_subrev_u32_e32 v8, s15, v7
; GFX9-NEXT: v_cndmask_b32_e32 v7, v7, v8, vcc
; GFX9-NEXT: v_add_u32_e32 v8, 1, v3
; GFX9-NEXT: v_cmp_le_u32_e32 vcc, s15, v7
; GFX9-NEXT: v_cndmask_b32_e32 v3, v3, v8, vcc
; GFX9-NEXT: v_subrev_u32_e32 v8, s15, v7
; GFX9-NEXT: v_cndmask_b32_e32 v7, v7, v8, vcc
; GFX9-NEXT: v_mov_b32_e32 v8, 0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1]
; GFX9-NEXT: global_store_dwordx4 v8, v[4:7], s[2:3]
; GFX9-NEXT: s_endpgm
;
; GFX10-LABEL: udivrem_v4i32:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_load_dwordx8 s[8:15], s[4:5], 0x10
; GFX10-NEXT: v_mov_b32_e32 v8, 0
; GFX10-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x0
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: v_cvt_f32_u32_e32 v0, s12
; GFX10-NEXT: v_cvt_f32_u32_e32 v1, s13
; GFX10-NEXT: v_cvt_f32_u32_e32 v2, s14
; GFX10-NEXT: v_cvt_f32_u32_e32 v3, s15
; GFX10-NEXT: s_sub_i32 s0, 0, s12
; GFX10-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX10-NEXT: v_rcp_iflag_f32_e32 v1, v1
; GFX10-NEXT: v_rcp_iflag_f32_e32 v2, v2
; GFX10-NEXT: v_rcp_iflag_f32_e32 v3, v3
; GFX10-NEXT: s_sub_i32 s1, 0, s13
; GFX10-NEXT: s_sub_i32 s2, 0, s14
; GFX10-NEXT: v_mul_f32_e32 v0, 0x4f7ffffe, v0
; GFX10-NEXT: v_mul_f32_e32 v1, 0x4f7ffffe, v1
; GFX10-NEXT: v_mul_f32_e32 v2, 0x4f7ffffe, v2
; GFX10-NEXT: v_mul_f32_e32 v3, 0x4f7ffffe, v3
; GFX10-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX10-NEXT: v_cvt_u32_f32_e32 v1, v1
; GFX10-NEXT: v_cvt_u32_f32_e32 v2, v2
; GFX10-NEXT: v_cvt_u32_f32_e32 v3, v3
; GFX10-NEXT: v_mul_lo_u32 v4, s0, v0
; GFX10-NEXT: v_mul_lo_u32 v5, s1, v1
; GFX10-NEXT: v_mul_lo_u32 v6, s2, v2
; GFX10-NEXT: s_sub_i32 s0, 0, s15
; GFX10-NEXT: v_mul_lo_u32 v7, s0, v3
; GFX10-NEXT: v_mul_hi_u32 v4, v0, v4
; GFX10-NEXT: v_mul_hi_u32 v5, v1, v5
; GFX10-NEXT: v_mul_hi_u32 v6, v2, v6
; GFX10-NEXT: v_mul_hi_u32 v7, v3, v7
; GFX10-NEXT: v_add_nc_u32_e32 v0, v0, v4
; GFX10-NEXT: v_add_nc_u32_e32 v1, v1, v5
; GFX10-NEXT: v_add_nc_u32_e32 v2, v2, v6
; GFX10-NEXT: v_add_nc_u32_e32 v3, v3, v7
; GFX10-NEXT: v_mul_hi_u32 v0, s8, v0
; GFX10-NEXT: v_mul_hi_u32 v1, s9, v1
; GFX10-NEXT: v_mul_hi_u32 v2, s10, v2
; GFX10-NEXT: v_mul_hi_u32 v3, s11, v3
; GFX10-NEXT: v_mul_lo_u32 v4, v0, s12
; GFX10-NEXT: v_mul_lo_u32 v5, v1, s13
; GFX10-NEXT: v_mul_lo_u32 v6, v2, s14
; GFX10-NEXT: v_add_nc_u32_e32 v9, 1, v0
; GFX10-NEXT: v_mul_lo_u32 v7, v3, s15
; GFX10-NEXT: v_add_nc_u32_e32 v10, 1, v1
; GFX10-NEXT: v_add_nc_u32_e32 v11, 1, v2
; GFX10-NEXT: v_add_nc_u32_e32 v12, 1, v3
; GFX10-NEXT: v_sub_nc_u32_e32 v4, s8, v4
; GFX10-NEXT: v_sub_nc_u32_e32 v5, s9, v5
; GFX10-NEXT: v_sub_nc_u32_e32 v6, s10, v6
; GFX10-NEXT: v_sub_nc_u32_e32 v7, s11, v7
; GFX10-NEXT: v_cmp_le_u32_e32 vcc_lo, s12, v4
; GFX10-NEXT: v_cmp_le_u32_e64 s0, s13, v5
; GFX10-NEXT: v_cmp_le_u32_e64 s1, s14, v6
; GFX10-NEXT: v_cmp_le_u32_e64 s2, s15, v7
; GFX10-NEXT: v_cndmask_b32_e32 v0, v0, v9, vcc_lo
; GFX10-NEXT: v_subrev_nc_u32_e32 v9, s12, v4
; GFX10-NEXT: v_cndmask_b32_e64 v1, v1, v10, s0
; GFX10-NEXT: v_subrev_nc_u32_e32 v10, s13, v5
; GFX10-NEXT: v_cndmask_b32_e64 v2, v2, v11, s1
; GFX10-NEXT: v_subrev_nc_u32_e32 v11, s14, v6
; GFX10-NEXT: v_cndmask_b32_e64 v3, v3, v12, s2
; GFX10-NEXT: v_subrev_nc_u32_e32 v12, s15, v7
; GFX10-NEXT: v_cndmask_b32_e32 v4, v4, v9, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e64 v5, v5, v10, s0
; GFX10-NEXT: v_cndmask_b32_e64 v6, v6, v11, s1
; GFX10-NEXT: v_add_nc_u32_e32 v9, 1, v0
; GFX10-NEXT: v_cndmask_b32_e64 v7, v7, v12, s2
; GFX10-NEXT: v_add_nc_u32_e32 v10, 1, v1
; GFX10-NEXT: v_add_nc_u32_e32 v11, 1, v2
; GFX10-NEXT: v_cmp_le_u32_e32 vcc_lo, s12, v4
; GFX10-NEXT: v_cmp_le_u32_e64 s0, s13, v5
; GFX10-NEXT: v_cmp_le_u32_e64 s1, s14, v6
; GFX10-NEXT: v_add_nc_u32_e32 v12, 1, v3
; GFX10-NEXT: v_cmp_le_u32_e64 s2, s15, v7
; GFX10-NEXT: v_cndmask_b32_e32 v0, v0, v9, vcc_lo
; GFX10-NEXT: v_subrev_nc_u32_e32 v9, s12, v4
; GFX10-NEXT: v_cndmask_b32_e64 v1, v1, v10, s0
; GFX10-NEXT: v_subrev_nc_u32_e32 v10, s13, v5
; GFX10-NEXT: v_cndmask_b32_e64 v2, v2, v11, s1
; GFX10-NEXT: v_subrev_nc_u32_e32 v11, s14, v6
; GFX10-NEXT: v_subrev_nc_u32_e32 v13, s15, v7
; GFX10-NEXT: v_cndmask_b32_e64 v3, v3, v12, s2
; GFX10-NEXT: v_cndmask_b32_e32 v4, v4, v9, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e64 v5, v5, v10, s0
; GFX10-NEXT: v_cndmask_b32_e64 v6, v6, v11, s1
; GFX10-NEXT: v_cndmask_b32_e64 v7, v7, v13, s2
; GFX10-NEXT: global_store_dwordx4 v8, v[0:3], s[4:5]
; GFX10-NEXT: global_store_dwordx4 v8, v[4:7], s[6:7]
; GFX10-NEXT: s_endpgm
%div = udiv <4 x i32> %x, %y
store <4 x i32> %div, <4 x i32> addrspace(1)* %out0
%rem = urem <4 x i32> %x, %y
store <4 x i32> %rem, <4 x i32> addrspace(1)* %out1
ret void
}
define amdgpu_kernel void @udivrem_v2i64(<2 x i64> addrspace(1)* %out0, <2 x i64> addrspace(1)* %out1, <2 x i64> %x, <2 x i64> %y) {
; GFX8-LABEL: udivrem_v2i64:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_load_dwordx8 s[8:15], s[4:5], 0x10
; GFX8-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x0
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
; GFX8-NEXT: v_cvt_f32_u32_e32 v0, s13
; GFX8-NEXT: v_cvt_f32_u32_e32 v1, s12
; GFX8-NEXT: s_sub_u32 s2, 0, s12
; GFX8-NEXT: s_subb_u32 s3, 0, s13
; GFX8-NEXT: v_mul_f32_e32 v0, 0x4f800000, v0
; GFX8-NEXT: v_add_f32_e32 v0, v0, v1
; GFX8-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX8-NEXT: v_mul_f32_e32 v0, 0x5f7ffffc, v0
; GFX8-NEXT: v_mul_f32_e32 v1, 0x2f800000, v0
; GFX8-NEXT: v_trunc_f32_e32 v2, v1
; GFX8-NEXT: v_mul_f32_e32 v1, 0xcf800000, v2
; GFX8-NEXT: v_add_f32_e32 v0, v1, v0
; GFX8-NEXT: v_cvt_u32_f32_e32 v3, v0
; GFX8-NEXT: v_cvt_u32_f32_e32 v4, v2
; GFX8-NEXT: v_mad_u64_u32 v[0:1], s[0:1], s2, v3, 0
; GFX8-NEXT: v_mad_u64_u32 v[1:2], s[0:1], s2, v4, v[1:2]
; GFX8-NEXT: v_mul_hi_u32 v5, v3, v0
; GFX8-NEXT: v_mad_u64_u32 v[1:2], s[0:1], s3, v3, v[1:2]
; GFX8-NEXT: v_mul_lo_u32 v2, v4, v0
; GFX8-NEXT: v_mul_hi_u32 v0, v4, v0
; GFX8-NEXT: v_mul_lo_u32 v6, v3, v1
; GFX8-NEXT: v_mul_lo_u32 v7, v4, v1
; GFX8-NEXT: v_mul_hi_u32 v8, v3, v1
; GFX8-NEXT: v_mul_hi_u32 v1, v4, v1
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v6
; GFX8-NEXT: v_cndmask_b32_e64 v6, 0, 1, vcc
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v7, v0
; GFX8-NEXT: v_cndmask_b32_e64 v7, 0, 1, vcc
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v5
; GFX8-NEXT: v_cndmask_b32_e64 v2, 0, 1, vcc
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v0, v8
; GFX8-NEXT: v_cndmask_b32_e64 v5, 0, 1, vcc
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v6, v2
; GFX8-NEXT: v_add_u32_e32 v5, vcc, v7, v5
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v0, v2
; GFX8-NEXT: v_cndmask_b32_e64 v2, 0, 1, vcc
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v5, v2
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v2
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v0
; GFX8-NEXT: v_addc_u32_e32 v4, vcc, v4, v1, vcc
; GFX8-NEXT: v_mad_u64_u32 v[0:1], s[0:1], s2, v3, 0
; GFX8-NEXT: v_mad_u64_u32 v[1:2], s[0:1], s2, v4, v[1:2]
; GFX8-NEXT: v_mul_hi_u32 v6, v3, v0
; GFX8-NEXT: s_sub_u32 s2, 0, s14
; GFX8-NEXT: v_mad_u64_u32 v[1:2], s[0:1], s3, v3, v[1:2]
; GFX8-NEXT: v_mul_lo_u32 v2, v4, v0
; GFX8-NEXT: v_mul_hi_u32 v0, v4, v0
; GFX8-NEXT: v_mul_lo_u32 v5, v3, v1
; GFX8-NEXT: s_subb_u32 s3, 0, s15
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v5
; GFX8-NEXT: v_cndmask_b32_e64 v5, 0, 1, vcc
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v6
; GFX8-NEXT: v_cndmask_b32_e64 v2, 0, 1, vcc
; GFX8-NEXT: v_mul_lo_u32 v6, v4, v1
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v5, v2
; GFX8-NEXT: v_mul_hi_u32 v5, v3, v1
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v6, v0
; GFX8-NEXT: v_cndmask_b32_e64 v6, 0, 1, vcc
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v0, v5
; GFX8-NEXT: v_cndmask_b32_e64 v5, 0, 1, vcc
; GFX8-NEXT: v_add_u32_e32 v5, vcc, v6, v5
; GFX8-NEXT: v_mul_hi_u32 v1, v4, v1
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v0, v2
; GFX8-NEXT: v_cndmask_b32_e64 v2, 0, 1, vcc
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v5, v2
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v2
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v3, v0
; GFX8-NEXT: v_addc_u32_e32 v1, vcc, v4, v1, vcc
; GFX8-NEXT: v_mul_lo_u32 v2, s9, v0
; GFX8-NEXT: v_mul_lo_u32 v3, s8, v1
; GFX8-NEXT: v_mul_hi_u32 v4, s8, v0
; GFX8-NEXT: v_mul_hi_u32 v0, s9, v0
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v3
; GFX8-NEXT: v_cndmask_b32_e64 v3, 0, 1, vcc
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v4
; GFX8-NEXT: v_cndmask_b32_e64 v2, 0, 1, vcc
; GFX8-NEXT: v_mul_lo_u32 v4, s9, v1
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v3, v2
; GFX8-NEXT: v_mul_hi_u32 v3, s8, v1
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v4, v0
; GFX8-NEXT: v_cndmask_b32_e64 v4, 0, 1, vcc
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v0, v3
; GFX8-NEXT: v_cndmask_b32_e64 v3, 0, 1, vcc
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v4, v3
; GFX8-NEXT: v_add_u32_e32 v5, vcc, v0, v2
; GFX8-NEXT: v_mul_hi_u32 v4, s9, v1
; GFX8-NEXT: v_mad_u64_u32 v[0:1], s[0:1], s12, v5, 0
; GFX8-NEXT: v_cndmask_b32_e64 v2, 0, 1, vcc
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v3, v2
; GFX8-NEXT: v_add_u32_e32 v6, vcc, v4, v2
; GFX8-NEXT: v_mad_u64_u32 v[1:2], s[0:1], s12, v6, v[1:2]
; GFX8-NEXT: v_mov_b32_e32 v4, s9
; GFX8-NEXT: v_sub_u32_e32 v7, vcc, s8, v0
; GFX8-NEXT: v_mad_u64_u32 v[1:2], s[0:1], s13, v5, v[1:2]
; GFX8-NEXT: v_mov_b32_e32 v3, s13
; GFX8-NEXT: v_subb_u32_e64 v8, s[0:1], v4, v1, vcc
; GFX8-NEXT: v_sub_u32_e64 v0, s[0:1], s9, v1
; GFX8-NEXT: v_cmp_le_u32_e64 s[0:1], s13, v8
; GFX8-NEXT: v_cndmask_b32_e64 v1, 0, -1, s[0:1]
; GFX8-NEXT: v_cmp_le_u32_e64 s[0:1], s12, v7
; GFX8-NEXT: v_cndmask_b32_e64 v2, 0, -1, s[0:1]
; GFX8-NEXT: v_cmp_eq_u32_e64 s[0:1], s13, v8
; GFX8-NEXT: v_cndmask_b32_e64 v9, v1, v2, s[0:1]
; GFX8-NEXT: v_cvt_f32_u32_e32 v1, s15
; GFX8-NEXT: v_cvt_f32_u32_e32 v2, s14
; GFX8-NEXT: v_subb_u32_e32 v4, vcc, v0, v3, vcc
; GFX8-NEXT: v_mul_f32_e32 v0, 0x4f800000, v1
; GFX8-NEXT: v_add_f32_e32 v0, v0, v2
; GFX8-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX8-NEXT: v_subrev_u32_e32 v10, vcc, s12, v7
; GFX8-NEXT: v_subbrev_u32_e64 v11, s[0:1], 0, v4, vcc
; GFX8-NEXT: v_mul_f32_e32 v0, 0x5f7ffffc, v0
; GFX8-NEXT: v_mul_f32_e32 v1, 0x2f800000, v0
; GFX8-NEXT: v_trunc_f32_e32 v2, v1
; GFX8-NEXT: v_mul_f32_e32 v1, 0xcf800000, v2
; GFX8-NEXT: v_add_f32_e32 v0, v1, v0
; GFX8-NEXT: v_cvt_u32_f32_e32 v12, v0
; GFX8-NEXT: v_add_u32_e64 v13, s[0:1], 1, v5
; GFX8-NEXT: v_addc_u32_e64 v14, s[0:1], 0, v6, s[0:1]
; GFX8-NEXT: v_mad_u64_u32 v[0:1], s[0:1], s2, v12, 0
; GFX8-NEXT: v_cvt_u32_f32_e32 v15, v2
; GFX8-NEXT: v_cmp_le_u32_e64 s[0:1], s13, v11
; GFX8-NEXT: v_cndmask_b32_e64 v16, 0, -1, s[0:1]
; GFX8-NEXT: v_subb_u32_e32 v3, vcc, v4, v3, vcc
; GFX8-NEXT: v_mad_u64_u32 v[1:2], s[0:1], s2, v15, v[1:2]
; GFX8-NEXT: v_cmp_le_u32_e64 s[0:1], s12, v10
; GFX8-NEXT: v_cndmask_b32_e64 v17, 0, -1, s[0:1]
; GFX8-NEXT: v_mad_u64_u32 v[1:2], s[0:1], s3, v12, v[1:2]
; GFX8-NEXT: v_cmp_eq_u32_e64 s[0:1], s13, v11
; GFX8-NEXT: v_cndmask_b32_e64 v16, v16, v17, s[0:1]
; GFX8-NEXT: v_mul_lo_u32 v2, v15, v0
; GFX8-NEXT: v_mul_lo_u32 v17, v12, v1
; GFX8-NEXT: v_mul_hi_u32 v4, v12, v0
; GFX8-NEXT: v_mul_hi_u32 v0, v15, v0
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v17
; GFX8-NEXT: v_cndmask_b32_e64 v17, 0, 1, vcc
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v4
; GFX8-NEXT: v_cndmask_b32_e64 v2, 0, 1, vcc
; GFX8-NEXT: v_mul_lo_u32 v4, v15, v1
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v17, v2
; GFX8-NEXT: v_mul_hi_u32 v17, v12, v1
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v4, v0
; GFX8-NEXT: v_cndmask_b32_e64 v4, 0, 1, vcc
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v0, v17
; GFX8-NEXT: v_cndmask_b32_e64 v17, 0, 1, vcc
; GFX8-NEXT: v_add_u32_e32 v4, vcc, v4, v17
; GFX8-NEXT: v_add_u32_e32 v17, vcc, 1, v13
; GFX8-NEXT: v_addc_u32_e32 v18, vcc, 0, v14, vcc
; GFX8-NEXT: v_subrev_u32_e32 v19, vcc, s12, v10
; GFX8-NEXT: v_mul_hi_u32 v1, v15, v1
; GFX8-NEXT: v_subbrev_u32_e32 v20, vcc, 0, v3, vcc
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v0, v2
; GFX8-NEXT: v_cndmask_b32_e64 v2, 0, 1, vcc
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v4, v2
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v2
; GFX8-NEXT: v_add_u32_e32 v12, vcc, v12, v0
; GFX8-NEXT: v_mad_u64_u32 v[2:3], s[0:1], s2, v12, 0
; GFX8-NEXT: v_addc_u32_e32 v15, vcc, v15, v1, vcc
; GFX8-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
; GFX8-NEXT: v_cndmask_b32_e32 v1, v13, v17, vcc
; GFX8-NEXT: v_mov_b32_e32 v0, v3
; GFX8-NEXT: v_mad_u64_u32 v[3:4], s[0:1], s2, v15, v[0:1]
; GFX8-NEXT: v_cndmask_b32_e32 v13, v14, v18, vcc
; GFX8-NEXT: v_cmp_ne_u32_e32 vcc, 0, v9
; GFX8-NEXT: v_mad_u64_u32 v[3:4], s[0:1], s3, v12, v[3:4]
; GFX8-NEXT: v_cmp_ne_u32_e64 s[0:1], 0, v16
; GFX8-NEXT: v_cndmask_b32_e32 v0, v5, v1, vcc
; GFX8-NEXT: v_cndmask_b32_e32 v1, v6, v13, vcc
; GFX8-NEXT: v_cndmask_b32_e64 v4, v10, v19, s[0:1]
; GFX8-NEXT: v_mul_lo_u32 v6, v15, v2
; GFX8-NEXT: v_mul_lo_u32 v9, v12, v3
; GFX8-NEXT: v_cndmask_b32_e32 v4, v7, v4, vcc
; GFX8-NEXT: v_mul_hi_u32 v7, v12, v2
; GFX8-NEXT: v_cndmask_b32_e64 v5, v11, v20, s[0:1]
; GFX8-NEXT: v_add_u32_e64 v6, s[0:1], v6, v9
; GFX8-NEXT: v_cndmask_b32_e64 v9, 0, 1, s[0:1]
; GFX8-NEXT: v_add_u32_e64 v6, s[0:1], v6, v7
; GFX8-NEXT: v_cndmask_b32_e64 v6, 0, 1, s[0:1]
; GFX8-NEXT: v_mul_lo_u32 v7, v15, v3
; GFX8-NEXT: v_mul_hi_u32 v2, v15, v2
; GFX8-NEXT: v_add_u32_e64 v6, s[0:1], v9, v6
; GFX8-NEXT: v_mul_hi_u32 v9, v12, v3
; GFX8-NEXT: v_add_u32_e64 v2, s[0:1], v7, v2
; GFX8-NEXT: v_cndmask_b32_e64 v7, 0, 1, s[0:1]
; GFX8-NEXT: v_add_u32_e64 v2, s[0:1], v2, v9
; GFX8-NEXT: v_cndmask_b32_e64 v9, 0, 1, s[0:1]
; GFX8-NEXT: v_add_u32_e64 v7, s[0:1], v7, v9
; GFX8-NEXT: v_mul_hi_u32 v3, v15, v3
; GFX8-NEXT: v_add_u32_e64 v2, s[0:1], v2, v6
; GFX8-NEXT: v_cndmask_b32_e64 v6, 0, 1, s[0:1]
; GFX8-NEXT: v_add_u32_e64 v6, s[0:1], v7, v6
; GFX8-NEXT: v_add_u32_e64 v3, s[0:1], v3, v6
; GFX8-NEXT: v_add_u32_e64 v2, s[0:1], v12, v2
; GFX8-NEXT: v_addc_u32_e64 v3, s[0:1], v15, v3, s[0:1]
; GFX8-NEXT: v_mul_lo_u32 v6, s11, v2
; GFX8-NEXT: v_mul_lo_u32 v7, s10, v3
; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v5, vcc
; GFX8-NEXT: v_mul_hi_u32 v8, s10, v2
; GFX8-NEXT: v_mul_hi_u32 v2, s11, v2
; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v7
; GFX8-NEXT: v_cndmask_b32_e64 v7, 0, 1, vcc
; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v8
; GFX8-NEXT: v_cndmask_b32_e64 v6, 0, 1, vcc
; GFX8-NEXT: v_mul_lo_u32 v8, s11, v3
; GFX8-NEXT: v_add_u32_e32 v6, vcc, v7, v6
; GFX8-NEXT: v_mul_hi_u32 v7, s10, v3
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v8, v2
; GFX8-NEXT: v_cndmask_b32_e64 v8, 0, 1, vcc
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v7
; GFX8-NEXT: v_cndmask_b32_e64 v7, 0, 1, vcc
; GFX8-NEXT: v_add_u32_e32 v7, vcc, v8, v7
; GFX8-NEXT: v_add_u32_e32 v8, vcc, v2, v6
; GFX8-NEXT: v_mul_hi_u32 v9, s11, v3
; GFX8-NEXT: v_mad_u64_u32 v[2:3], s[0:1], s14, v8, 0
; GFX8-NEXT: v_cndmask_b32_e64 v6, 0, 1, vcc
; GFX8-NEXT: v_add_u32_e32 v6, vcc, v7, v6
; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
; GFX8-NEXT: v_mad_u64_u32 v[6:7], s[0:1], s14, v9, v[3:4]
; GFX8-NEXT: v_mov_b32_e32 v10, s11
; GFX8-NEXT: v_mov_b32_e32 v3, s15
; GFX8-NEXT: v_mad_u64_u32 v[6:7], s[0:1], s15, v8, v[6:7]
; GFX8-NEXT: v_sub_u32_e32 v7, vcc, s10, v2
; GFX8-NEXT: v_subb_u32_e64 v10, s[0:1], v10, v6, vcc
; GFX8-NEXT: v_sub_u32_e64 v2, s[0:1], s11, v6
; GFX8-NEXT: v_cmp_le_u32_e64 s[0:1], s15, v10
; GFX8-NEXT: v_cndmask_b32_e64 v6, 0, -1, s[0:1]
; GFX8-NEXT: v_cmp_le_u32_e64 s[0:1], s14, v7
; GFX8-NEXT: v_cndmask_b32_e64 v11, 0, -1, s[0:1]
; GFX8-NEXT: v_cmp_eq_u32_e64 s[0:1], s15, v10
; GFX8-NEXT: v_subb_u32_e32 v2, vcc, v2, v3, vcc
; GFX8-NEXT: v_cndmask_b32_e64 v6, v6, v11, s[0:1]
; GFX8-NEXT: v_subrev_u32_e32 v11, vcc, s14, v7
; GFX8-NEXT: v_subbrev_u32_e64 v12, s[0:1], 0, v2, vcc
; GFX8-NEXT: v_cmp_le_u32_e64 s[0:1], s15, v12
; GFX8-NEXT: v_cndmask_b32_e64 v13, 0, -1, s[0:1]
; GFX8-NEXT: v_cmp_le_u32_e64 s[0:1], s14, v11
; GFX8-NEXT: v_cndmask_b32_e64 v14, 0, -1, s[0:1]
; GFX8-NEXT: v_cmp_eq_u32_e64 s[0:1], s15, v12
; GFX8-NEXT: v_cndmask_b32_e64 v13, v13, v14, s[0:1]
; GFX8-NEXT: v_add_u32_e64 v14, s[0:1], 1, v8
; GFX8-NEXT: v_subb_u32_e32 v2, vcc, v2, v3, vcc
; GFX8-NEXT: v_addc_u32_e64 v15, s[0:1], 0, v9, s[0:1]
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 1, v14
; GFX8-NEXT: v_addc_u32_e32 v16, vcc, 0, v15, vcc
; GFX8-NEXT: v_cmp_ne_u32_e32 vcc, 0, v13
; GFX8-NEXT: v_subrev_u32_e64 v13, s[0:1], s14, v11
; GFX8-NEXT: v_cndmask_b32_e32 v3, v14, v3, vcc
; GFX8-NEXT: v_subbrev_u32_e64 v14, s[0:1], 0, v2, s[0:1]
; GFX8-NEXT: v_cmp_ne_u32_e64 s[0:1], 0, v6
; GFX8-NEXT: v_cndmask_b32_e32 v15, v15, v16, vcc
; GFX8-NEXT: v_cndmask_b32_e64 v2, v8, v3, s[0:1]
; GFX8-NEXT: v_cndmask_b32_e32 v6, v11, v13, vcc
; GFX8-NEXT: v_cndmask_b32_e32 v8, v12, v14, vcc
; GFX8-NEXT: v_cndmask_b32_e64 v3, v9, v15, s[0:1]
; GFX8-NEXT: v_cndmask_b32_e64 v6, v7, v6, s[0:1]
; GFX8-NEXT: v_cndmask_b32_e64 v7, v10, v8, s[0:1]
; GFX8-NEXT: v_mov_b32_e32 v9, s5
; GFX8-NEXT: v_mov_b32_e32 v8, s4
; GFX8-NEXT: flat_store_dwordx4 v[8:9], v[0:3]
; GFX8-NEXT: s_nop 0
; GFX8-NEXT: v_mov_b32_e32 v0, s6
; GFX8-NEXT: v_mov_b32_e32 v1, s7
; GFX8-NEXT: flat_store_dwordx4 v[0:1], v[4:7]
; GFX8-NEXT: s_endpgm
;
; GFX9-LABEL: udivrem_v2i64:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_load_dwordx8 s[8:15], s[4:5], 0x10
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: v_cvt_f32_u32_e32 v0, s13
; GFX9-NEXT: v_cvt_f32_u32_e32 v1, s12
; GFX9-NEXT: s_sub_u32 s2, 0, s12
; GFX9-NEXT: s_subb_u32 s3, 0, s13
; GFX9-NEXT: v_mul_f32_e32 v0, 0x4f800000, v0
; GFX9-NEXT: v_add_f32_e32 v0, v0, v1
; GFX9-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX9-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x0
; GFX9-NEXT: v_mul_f32_e32 v0, 0x5f7ffffc, v0
; GFX9-NEXT: v_mul_f32_e32 v1, 0x2f800000, v0
; GFX9-NEXT: v_trunc_f32_e32 v2, v1
; GFX9-NEXT: v_mul_f32_e32 v1, 0xcf800000, v2
; GFX9-NEXT: v_add_f32_e32 v0, v1, v0
; GFX9-NEXT: v_cvt_u32_f32_e32 v3, v0
; GFX9-NEXT: v_cvt_u32_f32_e32 v4, v2
; GFX9-NEXT: v_mad_u64_u32 v[0:1], s[0:1], s2, v3, 0
; GFX9-NEXT: v_mad_u64_u32 v[1:2], s[0:1], s2, v4, v[1:2]
; GFX9-NEXT: v_mul_hi_u32 v5, v3, v0
; GFX9-NEXT: v_mad_u64_u32 v[1:2], s[0:1], s3, v3, v[1:2]
; GFX9-NEXT: v_mul_lo_u32 v2, v4, v0
; GFX9-NEXT: v_mul_hi_u32 v0, v4, v0
; GFX9-NEXT: v_mul_lo_u32 v6, v3, v1
; GFX9-NEXT: v_mul_lo_u32 v7, v4, v1
; GFX9-NEXT: v_mul_hi_u32 v8, v3, v1
; GFX9-NEXT: v_mul_hi_u32 v1, v4, v1
; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, v2, v6
; GFX9-NEXT: v_cndmask_b32_e64 v6, 0, 1, vcc
; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, v7, v0
; GFX9-NEXT: v_cndmask_b32_e64 v7, 0, 1, vcc
; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, v2, v5
; GFX9-NEXT: v_cndmask_b32_e64 v2, 0, 1, vcc
; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, v0, v8
; GFX9-NEXT: v_add_u32_e32 v2, v6, v2
; GFX9-NEXT: v_cndmask_b32_e64 v5, 0, 1, vcc
; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, v0, v2
; GFX9-NEXT: v_add_u32_e32 v5, v7, v5
; GFX9-NEXT: v_cndmask_b32_e64 v2, 0, 1, vcc
; GFX9-NEXT: v_add3_u32 v1, v5, v2, v1
; GFX9-NEXT: v_add_co_u32_e32 v3, vcc, v3, v0
; GFX9-NEXT: v_addc_co_u32_e32 v4, vcc, v4, v1, vcc
; GFX9-NEXT: v_mad_u64_u32 v[0:1], s[0:1], s2, v3, 0
; GFX9-NEXT: v_mad_u64_u32 v[1:2], s[0:1], s2, v4, v[1:2]
; GFX9-NEXT: v_mul_hi_u32 v6, v3, v0
; GFX9-NEXT: s_sub_u32 s2, 0, s14
; GFX9-NEXT: v_mad_u64_u32 v[1:2], s[0:1], s3, v3, v[1:2]
; GFX9-NEXT: v_mul_lo_u32 v2, v4, v0
; GFX9-NEXT: v_mul_hi_u32 v0, v4, v0
; GFX9-NEXT: v_mul_lo_u32 v5, v3, v1
; GFX9-NEXT: s_subb_u32 s3, 0, s15
; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, v2, v5
; GFX9-NEXT: v_cndmask_b32_e64 v5, 0, 1, vcc
; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, v2, v6
; GFX9-NEXT: v_cndmask_b32_e64 v2, 0, 1, vcc
; GFX9-NEXT: v_mul_lo_u32 v6, v4, v1
; GFX9-NEXT: v_add_u32_e32 v2, v5, v2
; GFX9-NEXT: v_mul_hi_u32 v5, v3, v1
; GFX9-NEXT: v_mul_hi_u32 v1, v4, v1
; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, v6, v0
; GFX9-NEXT: v_cndmask_b32_e64 v6, 0, 1, vcc
; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, v0, v5
; GFX9-NEXT: v_cndmask_b32_e64 v5, 0, 1, vcc
; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, v0, v2
; GFX9-NEXT: v_add_u32_e32 v5, v6, v5
; GFX9-NEXT: v_cndmask_b32_e64 v2, 0, 1, vcc
; GFX9-NEXT: v_add3_u32 v1, v5, v2, v1
; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, v3, v0
; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, v4, v1, vcc
; GFX9-NEXT: v_mul_lo_u32 v2, s9, v0
; GFX9-NEXT: v_mul_lo_u32 v3, s8, v1
; GFX9-NEXT: v_mul_hi_u32 v4, s8, v0
; GFX9-NEXT: v_mul_hi_u32 v0, s9, v0
; GFX9-NEXT: v_mul_hi_u32 v6, s9, v1
; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, v2, v3
; GFX9-NEXT: v_cndmask_b32_e64 v3, 0, 1, vcc
; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, v2, v4
; GFX9-NEXT: v_cndmask_b32_e64 v2, 0, 1, vcc
; GFX9-NEXT: v_mul_lo_u32 v4, s9, v1
; GFX9-NEXT: v_add_u32_e32 v2, v3, v2
; GFX9-NEXT: v_mul_hi_u32 v3, s8, v1
; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, v4, v0
; GFX9-NEXT: v_cndmask_b32_e64 v4, 0, 1, vcc
; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, v0, v3
; GFX9-NEXT: v_cndmask_b32_e64 v3, 0, 1, vcc
; GFX9-NEXT: v_add_co_u32_e32 v5, vcc, v0, v2
; GFX9-NEXT: v_mad_u64_u32 v[0:1], s[0:1], s12, v5, 0
; GFX9-NEXT: v_cndmask_b32_e64 v2, 0, 1, vcc
; GFX9-NEXT: v_add_u32_e32 v3, v4, v3
; GFX9-NEXT: v_add3_u32 v6, v3, v2, v6
; GFX9-NEXT: v_mad_u64_u32 v[1:2], s[0:1], s12, v6, v[1:2]
; GFX9-NEXT: v_mov_b32_e32 v4, s9
; GFX9-NEXT: v_sub_co_u32_e32 v7, vcc, s8, v0
; GFX9-NEXT: v_mad_u64_u32 v[1:2], s[0:1], s13, v5, v[1:2]
; GFX9-NEXT: v_mov_b32_e32 v3, s13
; GFX9-NEXT: v_subb_co_u32_e64 v8, s[0:1], v4, v1, vcc
; GFX9-NEXT: v_cmp_le_u32_e64 s[0:1], s13, v8
; GFX9-NEXT: v_sub_u32_e32 v0, s9, v1
; GFX9-NEXT: v_cndmask_b32_e64 v1, 0, -1, s[0:1]
; GFX9-NEXT: v_cmp_le_u32_e64 s[0:1], s12, v7
; GFX9-NEXT: v_cndmask_b32_e64 v2, 0, -1, s[0:1]
; GFX9-NEXT: v_cmp_eq_u32_e64 s[0:1], s13, v8
; GFX9-NEXT: v_cndmask_b32_e64 v9, v1, v2, s[0:1]
; GFX9-NEXT: v_cvt_f32_u32_e32 v1, s15
; GFX9-NEXT: v_cvt_f32_u32_e32 v2, s14
; GFX9-NEXT: v_subb_co_u32_e32 v4, vcc, v0, v3, vcc
; GFX9-NEXT: v_mul_f32_e32 v0, 0x4f800000, v1
; GFX9-NEXT: v_add_f32_e32 v0, v0, v2
; GFX9-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX9-NEXT: v_subrev_co_u32_e32 v10, vcc, s12, v7
; GFX9-NEXT: v_subbrev_co_u32_e64 v11, s[0:1], 0, v4, vcc
; GFX9-NEXT: v_mul_f32_e32 v0, 0x5f7ffffc, v0
; GFX9-NEXT: v_mul_f32_e32 v1, 0x2f800000, v0
; GFX9-NEXT: v_trunc_f32_e32 v2, v1
; GFX9-NEXT: v_mul_f32_e32 v1, 0xcf800000, v2
; GFX9-NEXT: v_add_f32_e32 v0, v1, v0
; GFX9-NEXT: v_cvt_u32_f32_e32 v12, v0
; GFX9-NEXT: v_add_co_u32_e64 v13, s[0:1], 1, v5
; GFX9-NEXT: v_addc_co_u32_e64 v14, s[0:1], 0, v6, s[0:1]
; GFX9-NEXT: v_mad_u64_u32 v[0:1], s[0:1], s2, v12, 0
; GFX9-NEXT: v_cvt_u32_f32_e32 v15, v2
; GFX9-NEXT: v_cmp_le_u32_e64 s[0:1], s13, v11
; GFX9-NEXT: v_cndmask_b32_e64 v16, 0, -1, s[0:1]
; GFX9-NEXT: v_subb_co_u32_e32 v3, vcc, v4, v3, vcc
; GFX9-NEXT: v_mad_u64_u32 v[1:2], s[0:1], s2, v15, v[1:2]
; GFX9-NEXT: v_cmp_le_u32_e64 s[0:1], s12, v10
; GFX9-NEXT: v_cndmask_b32_e64 v17, 0, -1, s[0:1]
; GFX9-NEXT: v_mad_u64_u32 v[1:2], s[0:1], s3, v12, v[1:2]
; GFX9-NEXT: v_cmp_eq_u32_e64 s[0:1], s13, v11
; GFX9-NEXT: v_cndmask_b32_e64 v16, v16, v17, s[0:1]
; GFX9-NEXT: v_mul_lo_u32 v2, v15, v0
; GFX9-NEXT: v_mul_lo_u32 v17, v12, v1
; GFX9-NEXT: v_mul_hi_u32 v4, v12, v0
; GFX9-NEXT: v_mul_hi_u32 v0, v15, v0
; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, v2, v17
; GFX9-NEXT: v_cndmask_b32_e64 v17, 0, 1, vcc
; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, v2, v4
; GFX9-NEXT: v_cndmask_b32_e64 v2, 0, 1, vcc
; GFX9-NEXT: v_mul_lo_u32 v4, v15, v1
; GFX9-NEXT: v_add_u32_e32 v2, v17, v2
; GFX9-NEXT: v_mul_hi_u32 v17, v12, v1
; GFX9-NEXT: v_mul_hi_u32 v1, v15, v1
; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, v4, v0
; GFX9-NEXT: v_cndmask_b32_e64 v4, 0, 1, vcc
; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, v0, v17
; GFX9-NEXT: v_cndmask_b32_e64 v17, 0, 1, vcc
; GFX9-NEXT: v_add_u32_e32 v4, v4, v17
; GFX9-NEXT: v_add_co_u32_e32 v17, vcc, 1, v13
; GFX9-NEXT: v_addc_co_u32_e32 v18, vcc, 0, v14, vcc
; GFX9-NEXT: v_subrev_co_u32_e32 v19, vcc, s12, v10
; GFX9-NEXT: v_subbrev_co_u32_e32 v20, vcc, 0, v3, vcc
; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, v0, v2
; GFX9-NEXT: v_cndmask_b32_e64 v2, 0, 1, vcc
; GFX9-NEXT: v_add_co_u32_e32 v12, vcc, v12, v0
; GFX9-NEXT: v_add3_u32 v1, v4, v2, v1
; GFX9-NEXT: v_mad_u64_u32 v[2:3], s[0:1], s2, v12, 0
; GFX9-NEXT: v_addc_co_u32_e32 v15, vcc, v15, v1, vcc
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
; GFX9-NEXT: v_cndmask_b32_e32 v1, v13, v17, vcc
; GFX9-NEXT: v_mov_b32_e32 v0, v3
; GFX9-NEXT: v_mad_u64_u32 v[3:4], s[0:1], s2, v15, v[0:1]
; GFX9-NEXT: v_cndmask_b32_e32 v13, v14, v18, vcc
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v9
; GFX9-NEXT: v_mad_u64_u32 v[3:4], s[0:1], s3, v12, v[3:4]
; GFX9-NEXT: v_cndmask_b32_e32 v0, v5, v1, vcc
; GFX9-NEXT: v_cndmask_b32_e32 v1, v6, v13, vcc
; GFX9-NEXT: v_cmp_ne_u32_e64 s[0:1], 0, v16
; GFX9-NEXT: v_mul_lo_u32 v5, v15, v2
; GFX9-NEXT: v_mul_lo_u32 v6, v12, v3
; GFX9-NEXT: v_cndmask_b32_e64 v4, v10, v19, s[0:1]
; GFX9-NEXT: v_mul_hi_u32 v10, v12, v2
; GFX9-NEXT: v_cndmask_b32_e64 v9, v11, v20, s[0:1]
; GFX9-NEXT: v_add_co_u32_e64 v5, s[0:1], v5, v6
; GFX9-NEXT: v_cndmask_b32_e64 v6, 0, 1, s[0:1]
; GFX9-NEXT: v_add_co_u32_e64 v5, s[0:1], v5, v10
; GFX9-NEXT: v_cndmask_b32_e64 v5, 0, 1, s[0:1]
; GFX9-NEXT: v_mul_lo_u32 v10, v15, v3
; GFX9-NEXT: v_mul_hi_u32 v2, v15, v2
; GFX9-NEXT: v_add_u32_e32 v5, v6, v5
; GFX9-NEXT: v_mul_hi_u32 v6, v12, v3
; GFX9-NEXT: v_mul_hi_u32 v3, v15, v3
; GFX9-NEXT: v_add_co_u32_e64 v2, s[0:1], v10, v2
; GFX9-NEXT: v_cndmask_b32_e64 v10, 0, 1, s[0:1]
; GFX9-NEXT: v_add_co_u32_e64 v2, s[0:1], v2, v6
; GFX9-NEXT: v_cndmask_b32_e64 v6, 0, 1, s[0:1]
; GFX9-NEXT: v_add_co_u32_e64 v2, s[0:1], v2, v5
; GFX9-NEXT: v_add_u32_e32 v6, v10, v6
; GFX9-NEXT: v_cndmask_b32_e64 v5, 0, 1, s[0:1]
; GFX9-NEXT: v_add3_u32 v3, v6, v5, v3
; GFX9-NEXT: v_add_co_u32_e64 v2, s[0:1], v12, v2
; GFX9-NEXT: v_addc_co_u32_e64 v3, s[0:1], v15, v3, s[0:1]
; GFX9-NEXT: v_mul_lo_u32 v5, s11, v2
; GFX9-NEXT: v_mul_lo_u32 v6, s10, v3
; GFX9-NEXT: v_cndmask_b32_e32 v4, v7, v4, vcc
; GFX9-NEXT: v_mul_hi_u32 v7, s10, v2
; GFX9-NEXT: v_mul_hi_u32 v2, s11, v2
; GFX9-NEXT: v_add_co_u32_e64 v5, s[0:1], v5, v6
; GFX9-NEXT: v_cndmask_b32_e64 v6, 0, 1, s[0:1]
; GFX9-NEXT: v_add_co_u32_e64 v5, s[0:1], v5, v7
; GFX9-NEXT: v_cndmask_b32_e64 v5, 0, 1, s[0:1]
; GFX9-NEXT: v_mul_lo_u32 v7, s11, v3
; GFX9-NEXT: v_add_u32_e32 v5, v6, v5
; GFX9-NEXT: v_mul_hi_u32 v6, s10, v3
; GFX9-NEXT: v_mul_hi_u32 v12, s11, v3
; GFX9-NEXT: v_add_co_u32_e64 v2, s[0:1], v7, v2
; GFX9-NEXT: v_cndmask_b32_e64 v7, 0, 1, s[0:1]
; GFX9-NEXT: v_add_co_u32_e64 v2, s[0:1], v2, v6
; GFX9-NEXT: v_cndmask_b32_e64 v6, 0, 1, s[0:1]
; GFX9-NEXT: v_add_co_u32_e64 v10, s[0:1], v2, v5
; GFX9-NEXT: v_cndmask_b32_e64 v11, 0, 1, s[0:1]
; GFX9-NEXT: v_mad_u64_u32 v[2:3], s[0:1], s14, v10, 0
; GFX9-NEXT: v_add_u32_e32 v6, v7, v6
; GFX9-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
; GFX9-NEXT: v_add3_u32 v8, v6, v11, v12
; GFX9-NEXT: v_mad_u64_u32 v[6:7], s[0:1], s14, v8, v[3:4]
; GFX9-NEXT: v_mov_b32_e32 v9, s11
; GFX9-NEXT: v_mov_b32_e32 v3, s15
; GFX9-NEXT: v_mad_u64_u32 v[6:7], s[0:1], s15, v10, v[6:7]
; GFX9-NEXT: v_sub_co_u32_e32 v7, vcc, s10, v2
; GFX9-NEXT: v_subb_co_u32_e64 v9, s[0:1], v9, v6, vcc
; GFX9-NEXT: v_cmp_le_u32_e64 s[0:1], s15, v9
; GFX9-NEXT: v_sub_u32_e32 v2, s11, v6
; GFX9-NEXT: v_cndmask_b32_e64 v6, 0, -1, s[0:1]
; GFX9-NEXT: v_cmp_le_u32_e64 s[0:1], s14, v7
; GFX9-NEXT: v_cndmask_b32_e64 v11, 0, -1, s[0:1]
; GFX9-NEXT: v_cmp_eq_u32_e64 s[0:1], s15, v9
; GFX9-NEXT: v_subb_co_u32_e32 v2, vcc, v2, v3, vcc
; GFX9-NEXT: v_cndmask_b32_e64 v6, v6, v11, s[0:1]
; GFX9-NEXT: v_subrev_co_u32_e32 v11, vcc, s14, v7
; GFX9-NEXT: v_subbrev_co_u32_e64 v12, s[0:1], 0, v2, vcc
; GFX9-NEXT: v_cmp_le_u32_e64 s[0:1], s15, v12
; GFX9-NEXT: v_cndmask_b32_e64 v13, 0, -1, s[0:1]
; GFX9-NEXT: v_cmp_le_u32_e64 s[0:1], s14, v11
; GFX9-NEXT: v_cndmask_b32_e64 v14, 0, -1, s[0:1]
; GFX9-NEXT: v_cmp_eq_u32_e64 s[0:1], s15, v12
; GFX9-NEXT: v_cndmask_b32_e64 v13, v13, v14, s[0:1]
; GFX9-NEXT: v_add_co_u32_e64 v14, s[0:1], 1, v10
; GFX9-NEXT: v_subb_co_u32_e32 v2, vcc, v2, v3, vcc
; GFX9-NEXT: v_addc_co_u32_e64 v15, s[0:1], 0, v8, s[0:1]
; GFX9-NEXT: v_add_co_u32_e32 v3, vcc, 1, v14
; GFX9-NEXT: v_addc_co_u32_e32 v16, vcc, 0, v15, vcc
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v13
; GFX9-NEXT: v_cndmask_b32_e32 v3, v14, v3, vcc
; GFX9-NEXT: v_cndmask_b32_e32 v14, v15, v16, vcc
; GFX9-NEXT: v_subrev_co_u32_e64 v15, s[0:1], s14, v11
; GFX9-NEXT: v_subbrev_co_u32_e64 v16, s[0:1], 0, v2, s[0:1]
; GFX9-NEXT: v_cmp_ne_u32_e64 s[0:1], 0, v6
; GFX9-NEXT: v_mov_b32_e32 v13, 0
; GFX9-NEXT: v_cndmask_b32_e64 v2, v10, v3, s[0:1]
; GFX9-NEXT: v_cndmask_b32_e64 v3, v8, v14, s[0:1]
; GFX9-NEXT: v_cndmask_b32_e32 v6, v11, v15, vcc
; GFX9-NEXT: v_cndmask_b32_e32 v8, v12, v16, vcc
; GFX9-NEXT: v_cndmask_b32_e64 v6, v7, v6, s[0:1]
; GFX9-NEXT: v_cndmask_b32_e64 v7, v9, v8, s[0:1]
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: global_store_dwordx4 v13, v[0:3], s[4:5]
; GFX9-NEXT: global_store_dwordx4 v13, v[4:7], s[6:7]
; GFX9-NEXT: s_endpgm
;
; GFX10-LABEL: udivrem_v2i64:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_load_dwordx8 s[8:15], s[4:5], 0x10
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: v_cvt_f32_u32_e32 v0, s13
; GFX10-NEXT: v_cvt_f32_u32_e32 v1, s15
; GFX10-NEXT: v_cvt_f32_u32_e32 v2, s12
; GFX10-NEXT: v_cvt_f32_u32_e32 v3, s14
; GFX10-NEXT: s_sub_u32 s0, 0, s12
; GFX10-NEXT: v_mul_f32_e32 v0, 0x4f800000, v0
; GFX10-NEXT: v_mul_f32_e32 v1, 0x4f800000, v1
; GFX10-NEXT: s_subb_u32 s1, 0, s13
; GFX10-NEXT: s_sub_u32 s2, 0, s14
; GFX10-NEXT: s_subb_u32 s3, 0, s15
; GFX10-NEXT: v_add_f32_e32 v0, v0, v2
; GFX10-NEXT: v_add_f32_e32 v1, v1, v3
; GFX10-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX10-NEXT: v_rcp_iflag_f32_e32 v1, v1
; GFX10-NEXT: v_mul_f32_e32 v0, 0x5f7ffffc, v0
; GFX10-NEXT: v_mul_f32_e32 v1, 0x5f7ffffc, v1
; GFX10-NEXT: v_mul_f32_e32 v2, 0x2f800000, v0
; GFX10-NEXT: v_mul_f32_e32 v3, 0x2f800000, v1
; GFX10-NEXT: v_trunc_f32_e32 v2, v2
; GFX10-NEXT: v_trunc_f32_e32 v3, v3
; GFX10-NEXT: v_mul_f32_e32 v4, 0xcf800000, v2
; GFX10-NEXT: v_mul_f32_e32 v5, 0xcf800000, v3
; GFX10-NEXT: v_cvt_u32_f32_e32 v6, v3
; GFX10-NEXT: v_add_f32_e32 v0, v4, v0
; GFX10-NEXT: v_add_f32_e32 v1, v5, v1
; GFX10-NEXT: v_cvt_u32_f32_e32 v4, v2
; GFX10-NEXT: v_mul_lo_u32 v10, s2, v6
; GFX10-NEXT: v_cvt_u32_f32_e32 v5, v0
; GFX10-NEXT: v_cvt_u32_f32_e32 v8, v1
; GFX10-NEXT: v_mul_lo_u32 v7, s0, v4
; GFX10-NEXT: v_mad_u64_u32 v[0:1], s6, s0, v5, 0
; GFX10-NEXT: v_mul_lo_u32 v9, s1, v5
; GFX10-NEXT: v_mad_u64_u32 v[2:3], s6, s2, v8, 0
; GFX10-NEXT: v_mul_lo_u32 v11, s3, v8
; GFX10-NEXT: v_add3_u32 v1, v1, v7, v9
; GFX10-NEXT: v_mul_lo_u32 v7, v4, v0
; GFX10-NEXT: v_mul_hi_u32 v9, v5, v0
; GFX10-NEXT: v_add3_u32 v3, v3, v10, v11
; GFX10-NEXT: v_mul_hi_u32 v0, v4, v0
; GFX10-NEXT: v_mul_lo_u32 v12, v5, v1
; GFX10-NEXT: v_mul_lo_u32 v13, v4, v1
; GFX10-NEXT: v_mul_lo_u32 v10, v6, v2
; GFX10-NEXT: v_mul_lo_u32 v15, v8, v3
; GFX10-NEXT: v_mul_hi_u32 v11, v8, v2
; GFX10-NEXT: v_mul_hi_u32 v2, v6, v2
; GFX10-NEXT: v_mul_lo_u32 v16, v6, v3
; GFX10-NEXT: v_mul_hi_u32 v14, v5, v1
; GFX10-NEXT: v_add_co_u32 v7, s6, v7, v12
; GFX10-NEXT: v_cndmask_b32_e64 v12, 0, 1, s6
; GFX10-NEXT: v_add_co_u32 v0, s6, v13, v0
; GFX10-NEXT: v_cndmask_b32_e64 v13, 0, 1, s6
; GFX10-NEXT: v_add_co_u32 v10, s6, v10, v15
; GFX10-NEXT: v_cndmask_b32_e64 v15, 0, 1, s6
; GFX10-NEXT: v_add_co_u32 v2, s6, v16, v2
; GFX10-NEXT: v_cndmask_b32_e64 v16, 0, 1, s6
; GFX10-NEXT: v_add_co_u32 v7, s6, v7, v9
; GFX10-NEXT: v_mul_hi_u32 v17, v8, v3
; GFX10-NEXT: v_cndmask_b32_e64 v7, 0, 1, s6
; GFX10-NEXT: v_add_co_u32 v0, s6, v0, v14
; GFX10-NEXT: v_cndmask_b32_e64 v9, 0, 1, s6
; GFX10-NEXT: v_add_co_u32 v10, s6, v10, v11
; GFX10-NEXT: v_cndmask_b32_e64 v10, 0, 1, s6
; GFX10-NEXT: v_add_nc_u32_e32 v7, v12, v7
; GFX10-NEXT: v_add_co_u32 v2, s6, v2, v17
; GFX10-NEXT: v_mul_hi_u32 v1, v4, v1
; GFX10-NEXT: v_add_nc_u32_e32 v10, v15, v10
; GFX10-NEXT: v_cndmask_b32_e64 v11, 0, 1, s6
; GFX10-NEXT: v_add_co_u32 v0, s6, v0, v7
; GFX10-NEXT: v_add_nc_u32_e32 v9, v13, v9
; GFX10-NEXT: v_cndmask_b32_e64 v7, 0, 1, s6
; GFX10-NEXT: v_mul_hi_u32 v3, v6, v3
; GFX10-NEXT: v_add_co_u32 v2, s6, v2, v10
; GFX10-NEXT: v_add_nc_u32_e32 v11, v16, v11
; GFX10-NEXT: v_cndmask_b32_e64 v10, 0, 1, s6
; GFX10-NEXT: v_add3_u32 v1, v9, v7, v1
; GFX10-NEXT: v_add_co_u32 v5, vcc_lo, v5, v0
; GFX10-NEXT: v_add3_u32 v3, v11, v10, v3
; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, v4, v1, vcc_lo
; GFX10-NEXT: v_add_co_u32 v8, vcc_lo, v8, v2
; GFX10-NEXT: v_add_co_ci_u32_e32 v6, vcc_lo, v6, v3, vcc_lo
; GFX10-NEXT: v_mad_u64_u32 v[0:1], s6, s0, v5, 0
; GFX10-NEXT: v_mul_lo_u32 v7, s1, v5
; GFX10-NEXT: v_mul_lo_u32 v9, s0, v4
; GFX10-NEXT: v_mad_u64_u32 v[2:3], s0, s2, v8, 0
; GFX10-NEXT: v_mul_lo_u32 v10, s3, v8
; GFX10-NEXT: v_mul_lo_u32 v11, s2, v6
; GFX10-NEXT: v_mul_lo_u32 v12, v4, v0
; GFX10-NEXT: v_mul_hi_u32 v13, v5, v0
; GFX10-NEXT: v_mul_hi_u32 v0, v4, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v9, v7
; GFX10-NEXT: v_mul_lo_u32 v7, v6, v2
; GFX10-NEXT: v_mul_hi_u32 v9, v8, v2
; GFX10-NEXT: v_mul_hi_u32 v2, v6, v2
; GFX10-NEXT: v_add3_u32 v3, v3, v11, v10
; GFX10-NEXT: v_mul_lo_u32 v10, v5, v1
; GFX10-NEXT: v_mul_lo_u32 v11, v4, v1
; GFX10-NEXT: v_mul_hi_u32 v14, v5, v1
; GFX10-NEXT: v_mul_hi_u32 v1, v4, v1
; GFX10-NEXT: v_mul_lo_u32 v15, v8, v3
; GFX10-NEXT: v_mul_lo_u32 v16, v6, v3
; GFX10-NEXT: v_mul_hi_u32 v17, v8, v3
; GFX10-NEXT: v_mul_hi_u32 v3, v6, v3
; GFX10-NEXT: v_add_co_u32 v10, s0, v12, v10
; GFX10-NEXT: v_cndmask_b32_e64 v12, 0, 1, s0
; GFX10-NEXT: v_add_co_u32 v0, s0, v11, v0
; GFX10-NEXT: v_cndmask_b32_e64 v11, 0, 1, s0
; GFX10-NEXT: v_add_co_u32 v7, s0, v7, v15
; GFX10-NEXT: v_cndmask_b32_e64 v15, 0, 1, s0
; GFX10-NEXT: v_add_co_u32 v2, s0, v16, v2
; GFX10-NEXT: v_cndmask_b32_e64 v16, 0, 1, s0
; GFX10-NEXT: v_add_co_u32 v10, s0, v10, v13
; GFX10-NEXT: v_cndmask_b32_e64 v10, 0, 1, s0
; GFX10-NEXT: v_add_co_u32 v0, s0, v0, v14
; GFX10-NEXT: v_cndmask_b32_e64 v13, 0, 1, s0
; GFX10-NEXT: v_add_co_u32 v7, s0, v7, v9
; GFX10-NEXT: v_add_nc_u32_e32 v9, v12, v10
; GFX10-NEXT: v_cndmask_b32_e64 v7, 0, 1, s0
; GFX10-NEXT: v_add_co_u32 v2, s0, v2, v17
; GFX10-NEXT: v_cndmask_b32_e64 v10, 0, 1, s0
; GFX10-NEXT: v_add_co_u32 v0, s0, v0, v9
; GFX10-NEXT: v_add_nc_u32_e32 v11, v11, v13
; GFX10-NEXT: v_cndmask_b32_e64 v9, 0, 1, s0
; GFX10-NEXT: v_add_nc_u32_e32 v7, v15, v7
; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, v5, v0
; GFX10-NEXT: v_add_nc_u32_e32 v10, v16, v10
; GFX10-NEXT: v_add3_u32 v1, v11, v9, v1
; GFX10-NEXT: v_add_co_u32 v2, s0, v2, v7
; GFX10-NEXT: v_cndmask_b32_e64 v7, 0, 1, s0
; GFX10-NEXT: v_mul_hi_u32 v5, s8, v0
; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, v4, v1, vcc_lo
; GFX10-NEXT: v_mul_lo_u32 v4, s9, v0
; GFX10-NEXT: v_add3_u32 v3, v10, v7, v3
; GFX10-NEXT: v_mul_hi_u32 v0, s9, v0
; GFX10-NEXT: v_mul_lo_u32 v7, s8, v1
; GFX10-NEXT: v_mul_lo_u32 v10, s9, v1
; GFX10-NEXT: v_add_co_u32 v2, vcc_lo, v8, v2
; GFX10-NEXT: v_add_co_ci_u32_e32 v3, vcc_lo, v6, v3, vcc_lo
; GFX10-NEXT: v_mul_hi_u32 v6, s8, v1
; GFX10-NEXT: v_mul_hi_u32 v1, s9, v1
; GFX10-NEXT: v_add_co_u32 v4, s0, v4, v7
; GFX10-NEXT: v_cndmask_b32_e64 v7, 0, 1, s0
; GFX10-NEXT: v_add_co_u32 v0, s0, v10, v0
; GFX10-NEXT: v_cndmask_b32_e64 v8, 0, 1, s0
; GFX10-NEXT: v_add_co_u32 v4, s0, v4, v5
; GFX10-NEXT: v_cndmask_b32_e64 v4, 0, 1, s0
; GFX10-NEXT: v_add_co_u32 v0, s0, v0, v6
; GFX10-NEXT: v_cndmask_b32_e64 v5, 0, 1, s0
; GFX10-NEXT: v_mul_lo_u32 v6, s11, v2
; GFX10-NEXT: v_add_nc_u32_e32 v4, v7, v4
; GFX10-NEXT: v_mul_lo_u32 v7, s10, v3
; GFX10-NEXT: v_mul_lo_u32 v10, s11, v3
; GFX10-NEXT: v_add_nc_u32_e32 v5, v8, v5
; GFX10-NEXT: v_mul_hi_u32 v8, s10, v2
; GFX10-NEXT: v_add_co_u32 v4, s0, v0, v4
; GFX10-NEXT: v_cndmask_b32_e64 v0, 0, 1, s0
; GFX10-NEXT: v_mul_hi_u32 v2, s11, v2
; GFX10-NEXT: v_mul_hi_u32 v11, s10, v3
; GFX10-NEXT: v_add_co_u32 v6, s0, v6, v7
; GFX10-NEXT: v_add3_u32 v5, v5, v0, v1
; GFX10-NEXT: v_cndmask_b32_e64 v7, 0, 1, s0
; GFX10-NEXT: v_mad_u64_u32 v[0:1], s0, s12, v4, 0
; GFX10-NEXT: v_mul_lo_u32 v12, s13, v4
; GFX10-NEXT: v_mul_lo_u32 v13, s12, v5
; GFX10-NEXT: v_add_co_u32 v2, s0, v10, v2
; GFX10-NEXT: v_cndmask_b32_e64 v10, 0, 1, s0
; GFX10-NEXT: v_add_co_u32 v6, s0, v6, v8
; GFX10-NEXT: v_cndmask_b32_e64 v6, 0, 1, s0
; GFX10-NEXT: v_add_co_u32 v2, s0, v2, v11
; GFX10-NEXT: v_cndmask_b32_e64 v8, 0, 1, s0
; GFX10-NEXT: v_add3_u32 v1, v1, v13, v12
; GFX10-NEXT: v_add_nc_u32_e32 v6, v7, v6
; GFX10-NEXT: v_mul_hi_u32 v3, s11, v3
; GFX10-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x0
; GFX10-NEXT: v_add_nc_u32_e32 v7, v10, v8
; GFX10-NEXT: v_sub_nc_u32_e32 v8, s9, v1
; GFX10-NEXT: v_sub_co_u32 v10, vcc_lo, s8, v0
; GFX10-NEXT: v_sub_co_ci_u32_e64 v11, s0, s9, v1, vcc_lo
; GFX10-NEXT: v_subrev_co_ci_u32_e32 v0, vcc_lo, s13, v8, vcc_lo
; GFX10-NEXT: v_cmp_le_u32_e32 vcc_lo, s12, v10
; GFX10-NEXT: v_mov_b32_e32 v9, 0
; GFX10-NEXT: v_cndmask_b32_e64 v1, 0, -1, vcc_lo
; GFX10-NEXT: v_sub_co_u32 v8, vcc_lo, v10, s12
; GFX10-NEXT: v_subrev_co_ci_u32_e64 v12, s0, 0, v0, vcc_lo
; GFX10-NEXT: v_cmp_le_u32_e64 s0, s13, v11
; GFX10-NEXT: v_subrev_co_ci_u32_e32 v0, vcc_lo, s13, v0, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e64 v13, 0, -1, s0
; GFX10-NEXT: v_cmp_le_u32_e64 s0, s12, v8
; GFX10-NEXT: v_cndmask_b32_e64 v14, 0, -1, s0
; GFX10-NEXT: v_cmp_le_u32_e64 s0, s13, v12
; GFX10-NEXT: v_cndmask_b32_e64 v15, 0, -1, s0
; GFX10-NEXT: v_add_co_u32 v16, s0, v4, 1
; GFX10-NEXT: v_add_co_ci_u32_e64 v17, s0, 0, v5, s0
; GFX10-NEXT: v_cmp_eq_u32_e64 s0, s13, v11
; GFX10-NEXT: v_cndmask_b32_e64 v13, v13, v1, s0
; GFX10-NEXT: v_cmp_eq_u32_e64 s0, s13, v12
; GFX10-NEXT: v_cndmask_b32_e64 v14, v15, v14, s0
; GFX10-NEXT: v_add_co_u32 v6, s0, v2, v6
; GFX10-NEXT: v_cndmask_b32_e64 v1, 0, 1, s0
; GFX10-NEXT: v_add_co_u32 v15, s0, v16, 1
; GFX10-NEXT: v_add_co_ci_u32_e64 v18, s0, 0, v17, s0
; GFX10-NEXT: v_add3_u32 v3, v7, v1, v3
; GFX10-NEXT: v_mad_u64_u32 v[1:2], s0, s14, v6, 0
; GFX10-NEXT: v_mul_lo_u32 v19, s15, v6
; GFX10-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v14
; GFX10-NEXT: v_mul_lo_u32 v7, s14, v3
; GFX10-NEXT: v_cndmask_b32_e32 v15, v16, v15, vcc_lo
; GFX10-NEXT: v_sub_co_u32 v16, s0, v8, s12
; GFX10-NEXT: v_subrev_co_ci_u32_e64 v20, s0, 0, v0, s0
; GFX10-NEXT: v_add3_u32 v2, v2, v7, v19
; GFX10-NEXT: v_sub_co_u32 v7, s0, s10, v1
; GFX10-NEXT: v_cndmask_b32_e32 v17, v17, v18, vcc_lo
; GFX10-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v13
; GFX10-NEXT: v_sub_co_ci_u32_e64 v13, s1, s11, v2, s0
; GFX10-NEXT: v_sub_nc_u32_e32 v2, s11, v2
; GFX10-NEXT: v_cmp_ne_u32_e64 s1, 0, v14
; GFX10-NEXT: v_cndmask_b32_e32 v0, v4, v15, vcc_lo
; GFX10-NEXT: v_cmp_le_u32_e64 s2, s15, v13
; GFX10-NEXT: v_cndmask_b32_e32 v1, v5, v17, vcc_lo
; GFX10-NEXT: v_subrev_co_ci_u32_e64 v2, s0, s15, v2, s0
; GFX10-NEXT: v_cmp_le_u32_e64 s0, s14, v7
; GFX10-NEXT: v_cndmask_b32_e64 v4, v8, v16, s1
; GFX10-NEXT: v_cndmask_b32_e64 v5, 0, -1, s2
; GFX10-NEXT: v_cndmask_b32_e64 v12, v12, v20, s1
; GFX10-NEXT: v_cmp_eq_u32_e64 s1, s15, v13
; GFX10-NEXT: v_cndmask_b32_e64 v8, 0, -1, s0
; GFX10-NEXT: v_sub_co_u32 v14, s0, v7, s14
; GFX10-NEXT: v_subrev_co_ci_u32_e64 v15, s2, 0, v2, s0
; GFX10-NEXT: v_cndmask_b32_e64 v5, v5, v8, s1
; GFX10-NEXT: v_cndmask_b32_e32 v4, v10, v4, vcc_lo
; GFX10-NEXT: v_subrev_co_ci_u32_e64 v2, s0, s15, v2, s0
; GFX10-NEXT: v_cmp_le_u32_e64 s1, s15, v15
; GFX10-NEXT: v_cndmask_b32_e64 v8, 0, -1, s1
; GFX10-NEXT: v_cmp_le_u32_e64 s1, s14, v14
; GFX10-NEXT: v_cndmask_b32_e64 v10, 0, -1, s1
; GFX10-NEXT: v_add_co_u32 v16, s1, v6, 1
; GFX10-NEXT: v_add_co_ci_u32_e64 v17, s1, 0, v3, s1
; GFX10-NEXT: v_cmp_eq_u32_e64 s1, s15, v15
; GFX10-NEXT: v_cndmask_b32_e64 v8, v8, v10, s1
; GFX10-NEXT: v_add_co_u32 v10, s1, v16, 1
; GFX10-NEXT: v_add_co_ci_u32_e64 v18, s1, 0, v17, s1
; GFX10-NEXT: v_cmp_ne_u32_e64 s0, 0, v8
; GFX10-NEXT: v_sub_co_u32 v8, s1, v14, s14
; GFX10-NEXT: v_subrev_co_ci_u32_e64 v2, s1, 0, v2, s1
; GFX10-NEXT: v_cndmask_b32_e64 v10, v16, v10, s0
; GFX10-NEXT: v_cndmask_b32_e64 v16, v17, v18, s0
; GFX10-NEXT: v_cmp_ne_u32_e64 s1, 0, v5
; GFX10-NEXT: v_cndmask_b32_e64 v8, v14, v8, s0
; GFX10-NEXT: v_cndmask_b32_e64 v14, v15, v2, s0
; GFX10-NEXT: v_cndmask_b32_e32 v5, v11, v12, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e64 v2, v6, v10, s1
; GFX10-NEXT: v_cndmask_b32_e64 v3, v3, v16, s1
; GFX10-NEXT: v_cndmask_b32_e64 v6, v7, v8, s1
; GFX10-NEXT: v_cndmask_b32_e64 v7, v13, v14, s1
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: global_store_dwordx4 v9, v[0:3], s[4:5]
; GFX10-NEXT: global_store_dwordx4 v9, v[4:7], s[6:7]
; GFX10-NEXT: s_endpgm
%div = udiv <2 x i64> %x, %y
store <2 x i64> %div, <2 x i64> addrspace(1)* %out0
%rem = urem <2 x i64> %x, %y
store <2 x i64> %rem, <2 x i64> addrspace(1)* %out1
ret void
}
define amdgpu_kernel void @udiv_i8(i8 addrspace(1)* %out0, i8 addrspace(1)* %out1, i8 %x, i8 %y) {
; GFX8-LABEL: udiv_i8:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_load_dword s6, s[4:5], 0x10
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
; GFX8-NEXT: s_bfe_u32 s7, s6, 0x80008
; GFX8-NEXT: v_cvt_f32_ubyte0_e32 v0, s7
; GFX8-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX8-NEXT: s_sub_i32 s0, 0, s7
; GFX8-NEXT: v_mul_f32_e32 v0, 0x4f7ffffe, v0
; GFX8-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX8-NEXT: v_mul_lo_u32 v1, s0, v0
; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x0
; GFX8-NEXT: s_and_b32 s4, s6, 0xff
; GFX8-NEXT: v_mul_hi_u32 v1, v0, v1
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v0, v1
; GFX8-NEXT: v_mul_hi_u32 v2, s4, v0
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
; GFX8-NEXT: v_mov_b32_e32 v0, s0
; GFX8-NEXT: v_mov_b32_e32 v1, s1
; GFX8-NEXT: v_mul_lo_u32 v3, v2, s7
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 1, v2
; GFX8-NEXT: v_sub_u32_e32 v3, vcc, s4, v3
; GFX8-NEXT: v_cmp_le_u32_e32 vcc, s7, v3
; GFX8-NEXT: v_cndmask_b32_e32 v2, v2, v4, vcc
; GFX8-NEXT: v_subrev_u32_e64 v4, s[0:1], s7, v3
; GFX8-NEXT: v_cndmask_b32_e32 v3, v3, v4, vcc
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 1, v2
; GFX8-NEXT: v_cmp_le_u32_e32 vcc, s7, v3
; GFX8-NEXT: v_cndmask_b32_e32 v2, v2, v4, vcc
; GFX8-NEXT: v_subrev_u32_e64 v4, s[0:1], s7, v3
; GFX8-NEXT: flat_store_byte v[0:1], v2
; GFX8-NEXT: v_mov_b32_e32 v0, s2
; GFX8-NEXT: v_cndmask_b32_e32 v3, v3, v4, vcc
; GFX8-NEXT: v_mov_b32_e32 v1, s3
; GFX8-NEXT: flat_store_byte v[0:1], v3
; GFX8-NEXT: s_endpgm
;
; GFX9-LABEL: udiv_i8:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_load_dword s0, s[4:5], 0x10
; GFX9-NEXT: v_mov_b32_e32 v2, 0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_bfe_u32 s6, s0, 0x80008
; GFX9-NEXT: v_cvt_f32_ubyte0_e32 v0, s6
; GFX9-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX9-NEXT: s_sub_i32 s1, 0, s6
; GFX9-NEXT: s_and_b32 s7, s0, 0xff
; GFX9-NEXT: v_mul_f32_e32 v0, 0x4f7ffffe, v0
; GFX9-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX9-NEXT: v_mul_lo_u32 v1, s1, v0
; GFX9-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x0
; GFX9-NEXT: v_mul_hi_u32 v1, v0, v1
; GFX9-NEXT: v_add_u32_e32 v0, v0, v1
; GFX9-NEXT: v_mul_hi_u32 v0, s7, v0
; GFX9-NEXT: v_mul_lo_u32 v1, v0, s6
; GFX9-NEXT: v_add_u32_e32 v3, 1, v0
; GFX9-NEXT: v_sub_u32_e32 v1, s7, v1
; GFX9-NEXT: v_cmp_le_u32_e32 vcc, s6, v1
; GFX9-NEXT: v_cndmask_b32_e32 v0, v0, v3, vcc
; GFX9-NEXT: v_subrev_u32_e32 v3, s6, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX9-NEXT: v_add_u32_e32 v3, 1, v0
; GFX9-NEXT: v_cmp_le_u32_e32 vcc, s6, v1
; GFX9-NEXT: v_cndmask_b32_e32 v0, v0, v3, vcc
; GFX9-NEXT: v_subrev_u32_e32 v3, s6, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: global_store_byte v2, v0, s[0:1]
; GFX9-NEXT: global_store_byte v2, v1, s[2:3]
; GFX9-NEXT: s_endpgm
;
; GFX10-LABEL: udiv_i8:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_load_dword s0, s[4:5], 0x10
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: s_bfe_u32 s6, s0, 0x80008
; GFX10-NEXT: s_and_b32 s0, s0, 0xff
; GFX10-NEXT: v_cvt_f32_ubyte0_e32 v0, s6
; GFX10-NEXT: s_sub_i32 s1, 0, s6
; GFX10-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX10-NEXT: v_mul_f32_e32 v0, 0x4f7ffffe, v0
; GFX10-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX10-NEXT: v_mul_lo_u32 v1, s1, v0
; GFX10-NEXT: v_mul_hi_u32 v1, v0, v1
; GFX10-NEXT: v_add_nc_u32_e32 v0, v0, v1
; GFX10-NEXT: v_mul_hi_u32 v0, s0, v0
; GFX10-NEXT: v_mul_lo_u32 v1, v0, s6
; GFX10-NEXT: v_add_nc_u32_e32 v2, 1, v0
; GFX10-NEXT: v_sub_nc_u32_e32 v1, s0, v1
; GFX10-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x0
; GFX10-NEXT: v_subrev_nc_u32_e32 v3, s6, v1
; GFX10-NEXT: v_cmp_le_u32_e32 vcc_lo, s6, v1
; GFX10-NEXT: v_cndmask_b32_e32 v0, v0, v2, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX10-NEXT: v_add_nc_u32_e32 v2, 1, v0
; GFX10-NEXT: v_cmp_le_u32_e32 vcc_lo, s6, v1
; GFX10-NEXT: v_subrev_nc_u32_e32 v3, s6, v1
; GFX10-NEXT: v_cndmask_b32_e32 v0, v0, v2, vcc_lo
; GFX10-NEXT: v_mov_b32_e32 v2, 0
; GFX10-NEXT: v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: global_store_byte v2, v0, s[0:1]
; GFX10-NEXT: global_store_byte v2, v1, s[2:3]
; GFX10-NEXT: s_endpgm
%div = udiv i8 %x, %y
store i8 %div, i8 addrspace(1)* %out0
%rem = urem i8 %x, %y
store i8 %rem, i8 addrspace(1)* %out1
ret void
}
define amdgpu_kernel void @udivrem_v2i8(<2 x i8> addrspace(1)* %out0, <2 x i8> addrspace(1)* %out1, <2 x i8> %x, <2 x i8> %y) {
; GFX8-LABEL: udivrem_v2i8:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_load_dword s0, s[4:5], 0x10
; GFX8-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x0
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
; GFX8-NEXT: s_bfe_u32 s2, s0, 0x80010
; GFX8-NEXT: v_cvt_f32_ubyte0_e32 v0, s2
; GFX8-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX8-NEXT: v_cvt_f32_ubyte3_e32 v1, s0
; GFX8-NEXT: v_rcp_iflag_f32_e32 v1, v1
; GFX8-NEXT: s_sub_i32 s1, 0, s2
; GFX8-NEXT: v_mul_f32_e32 v0, 0x4f7ffffe, v0
; GFX8-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX8-NEXT: v_mul_f32_e32 v1, 0x4f7ffffe, v1
; GFX8-NEXT: v_cvt_u32_f32_e32 v1, v1
; GFX8-NEXT: s_lshr_b32 s3, s0, 24
; GFX8-NEXT: v_mul_lo_u32 v2, s1, v0
; GFX8-NEXT: s_sub_i32 s1, 0, s3
; GFX8-NEXT: v_mul_lo_u32 v3, s1, v1
; GFX8-NEXT: s_and_b32 s1, s0, 0xff
; GFX8-NEXT: v_mul_hi_u32 v2, v0, v2
; GFX8-NEXT: s_bfe_u32 s8, s0, 0x80008
; GFX8-NEXT: v_mul_hi_u32 v3, v1, v3
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v0, v2
; GFX8-NEXT: v_mul_hi_u32 v0, s1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v3
; GFX8-NEXT: v_mul_hi_u32 v1, s8, v1
; GFX8-NEXT: v_mul_lo_u32 v2, v0, s2
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 1, v0
; GFX8-NEXT: v_sub_u32_e32 v2, vcc, s1, v2
; GFX8-NEXT: v_cmp_le_u32_e32 vcc, s2, v2
; GFX8-NEXT: v_cndmask_b32_e32 v0, v0, v3, vcc
; GFX8-NEXT: v_subrev_u32_e64 v3, s[0:1], s2, v2
; GFX8-NEXT: v_cndmask_b32_e32 v2, v2, v3, vcc
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 1, v0
; GFX8-NEXT: v_cmp_le_u32_e32 vcc, s2, v2
; GFX8-NEXT: v_cndmask_b32_e32 v0, v0, v3, vcc
; GFX8-NEXT: v_mul_lo_u32 v3, v1, s3
; GFX8-NEXT: v_subrev_u32_e64 v4, s[0:1], s2, v2
; GFX8-NEXT: v_cndmask_b32_e32 v2, v2, v4, vcc
; GFX8-NEXT: v_sub_u32_e32 v3, vcc, s8, v3
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 1, v1
; GFX8-NEXT: v_cmp_le_u32_e32 vcc, s3, v3
; GFX8-NEXT: v_cndmask_b32_e32 v1, v1, v4, vcc
; GFX8-NEXT: v_subrev_u32_e64 v4, s[0:1], s3, v3
; GFX8-NEXT: v_cndmask_b32_e32 v3, v3, v4, vcc
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 1, v1
; GFX8-NEXT: v_cmp_le_u32_e32 vcc, s3, v3
; GFX8-NEXT: v_cndmask_b32_e32 v1, v1, v4, vcc
; GFX8-NEXT: v_and_b32_e32 v1, 0xff, v1
; GFX8-NEXT: v_subrev_u32_e64 v4, s[0:1], s3, v3
; GFX8-NEXT: v_lshlrev_b16_e32 v1, 8, v1
; GFX8-NEXT: v_cndmask_b32_e32 v3, v3, v4, vcc
; GFX8-NEXT: v_or_b32_sdwa v4, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX8-NEXT: v_mov_b32_e32 v0, s4
; GFX8-NEXT: v_mov_b32_e32 v1, s5
; GFX8-NEXT: flat_store_short v[0:1], v4
; GFX8-NEXT: v_and_b32_e32 v0, 0xff, v3
; GFX8-NEXT: v_lshlrev_b16_e32 v0, 8, v0
; GFX8-NEXT: v_or_b32_sdwa v2, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX8-NEXT: v_mov_b32_e32 v0, s6
; GFX8-NEXT: v_mov_b32_e32 v1, s7
; GFX8-NEXT: flat_store_short v[0:1], v2
; GFX8-NEXT: s_endpgm
;
; GFX9-LABEL: udivrem_v2i8:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_load_dword s0, s[4:5], 0x10
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_bfe_u32 s6, s0, 0x80010
; GFX9-NEXT: v_cvt_f32_ubyte0_e32 v1, s6
; GFX9-NEXT: v_rcp_iflag_f32_e32 v1, v1
; GFX9-NEXT: v_cvt_f32_ubyte3_e32 v0, s0
; GFX9-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX9-NEXT: s_sub_i32 s1, 0, s6
; GFX9-NEXT: v_mul_f32_e32 v1, 0x4f7ffffe, v1
; GFX9-NEXT: v_cvt_u32_f32_e32 v1, v1
; GFX9-NEXT: v_mul_f32_e32 v0, 0x4f7ffffe, v0
; GFX9-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX9-NEXT: s_lshr_b32 s7, s0, 24
; GFX9-NEXT: v_mul_lo_u32 v3, s1, v1
; GFX9-NEXT: s_sub_i32 s2, 0, s7
; GFX9-NEXT: v_mul_lo_u32 v2, s2, v0
; GFX9-NEXT: s_and_b32 s8, s0, 0xff
; GFX9-NEXT: v_mul_hi_u32 v3, v1, v3
; GFX9-NEXT: s_bfe_u32 s9, s0, 0x80008
; GFX9-NEXT: v_mul_hi_u32 v2, v0, v2
; GFX9-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x0
; GFX9-NEXT: v_add_u32_e32 v1, v1, v3
; GFX9-NEXT: v_mul_hi_u32 v1, s8, v1
; GFX9-NEXT: v_add_u32_e32 v0, v0, v2
; GFX9-NEXT: v_mul_hi_u32 v0, s9, v0
; GFX9-NEXT: v_mul_lo_u32 v3, v1, s6
; GFX9-NEXT: v_add_u32_e32 v4, 1, v1
; GFX9-NEXT: v_mul_lo_u32 v2, v0, s7
; GFX9-NEXT: v_sub_u32_e32 v3, s8, v3
; GFX9-NEXT: v_cmp_le_u32_e32 vcc, s6, v3
; GFX9-NEXT: v_cndmask_b32_e32 v1, v1, v4, vcc
; GFX9-NEXT: v_subrev_u32_e32 v4, s6, v3
; GFX9-NEXT: v_cndmask_b32_e32 v3, v3, v4, vcc
; GFX9-NEXT: v_add_u32_e32 v4, 1, v1
; GFX9-NEXT: v_cmp_le_u32_e32 vcc, s6, v3
; GFX9-NEXT: v_sub_u32_e32 v2, s9, v2
; GFX9-NEXT: v_cndmask_b32_e32 v1, v1, v4, vcc
; GFX9-NEXT: v_subrev_u32_e32 v4, s6, v3
; GFX9-NEXT: v_cndmask_b32_e32 v3, v3, v4, vcc
; GFX9-NEXT: v_add_u32_e32 v4, 1, v0
; GFX9-NEXT: v_cmp_le_u32_e32 vcc, s7, v2
; GFX9-NEXT: v_cndmask_b32_e32 v0, v0, v4, vcc
; GFX9-NEXT: v_subrev_u32_e32 v4, s7, v2
; GFX9-NEXT: v_cndmask_b32_e32 v2, v2, v4, vcc
; GFX9-NEXT: v_add_u32_e32 v4, 1, v0
; GFX9-NEXT: v_cmp_le_u32_e32 vcc, s7, v2
; GFX9-NEXT: v_cndmask_b32_e32 v0, v0, v4, vcc
; GFX9-NEXT: v_and_b32_e32 v0, 0xff, v0
; GFX9-NEXT: v_subrev_u32_e32 v4, s7, v2
; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
; GFX9-NEXT: v_cndmask_b32_e32 v2, v2, v4, vcc
; GFX9-NEXT: v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT: v_mov_b32_e32 v1, 0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: global_store_short v1, v0, s[0:1]
; GFX9-NEXT: v_and_b32_e32 v0, 0xff, v2
; GFX9-NEXT: v_lshlrev_b16_e32 v0, 8, v0
; GFX9-NEXT: v_or_b32_sdwa v0, v3, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT: global_store_short v1, v0, s[2:3]
; GFX9-NEXT: s_endpgm
;
; GFX10-LABEL: udivrem_v2i8:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_load_dword s0, s[4:5], 0x10
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: v_cvt_f32_ubyte3_e32 v0, s0
; GFX10-NEXT: s_bfe_u32 s1, s0, 0x80010
; GFX10-NEXT: s_lshr_b32 s2, s0, 24
; GFX10-NEXT: v_cvt_f32_ubyte0_e32 v1, s1
; GFX10-NEXT: s_sub_i32 s3, 0, s2
; GFX10-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX10-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x0
; GFX10-NEXT: v_rcp_iflag_f32_e32 v1, v1
; GFX10-NEXT: v_mul_f32_e32 v0, 0x4f7ffffe, v0
; GFX10-NEXT: v_mul_f32_e32 v1, 0x4f7ffffe, v1
; GFX10-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX10-NEXT: v_cvt_u32_f32_e32 v1, v1
; GFX10-NEXT: v_mul_lo_u32 v2, s3, v0
; GFX10-NEXT: s_sub_i32 s3, 0, s1
; GFX10-NEXT: v_mul_lo_u32 v3, s3, v1
; GFX10-NEXT: s_bfe_u32 s3, s0, 0x80008
; GFX10-NEXT: s_and_b32 s0, s0, 0xff
; GFX10-NEXT: v_mul_hi_u32 v2, v0, v2
; GFX10-NEXT: v_mul_hi_u32 v3, v1, v3
; GFX10-NEXT: v_add_nc_u32_e32 v0, v0, v2
; GFX10-NEXT: v_add_nc_u32_e32 v1, v1, v3
; GFX10-NEXT: v_mul_hi_u32 v0, s3, v0
; GFX10-NEXT: v_mul_hi_u32 v1, s0, v1
; GFX10-NEXT: v_mul_lo_u32 v2, v0, s2
; GFX10-NEXT: v_add_nc_u32_e32 v4, 1, v0
; GFX10-NEXT: v_mul_lo_u32 v3, v1, s1
; GFX10-NEXT: v_add_nc_u32_e32 v6, 1, v1
; GFX10-NEXT: v_sub_nc_u32_e32 v2, s3, v2
; GFX10-NEXT: v_sub_nc_u32_e32 v3, s0, v3
; GFX10-NEXT: v_subrev_nc_u32_e32 v5, s2, v2
; GFX10-NEXT: v_cmp_le_u32_e32 vcc_lo, s2, v2
; GFX10-NEXT: v_cmp_le_u32_e64 s0, s1, v3
; GFX10-NEXT: v_subrev_nc_u32_e32 v7, s1, v3
; GFX10-NEXT: v_cndmask_b32_e32 v0, v0, v4, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v2, v2, v5, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e64 v1, v1, v6, s0
; GFX10-NEXT: v_cndmask_b32_e64 v3, v3, v7, s0
; GFX10-NEXT: v_add_nc_u32_e32 v4, 1, v0
; GFX10-NEXT: v_cmp_le_u32_e32 vcc_lo, s2, v2
; GFX10-NEXT: v_subrev_nc_u32_e32 v5, s2, v2
; GFX10-NEXT: v_add_nc_u32_e32 v6, 1, v1
; GFX10-NEXT: v_cmp_le_u32_e64 s0, s1, v3
; GFX10-NEXT: v_subrev_nc_u32_e32 v7, s1, v3
; GFX10-NEXT: v_cndmask_b32_e32 v0, v0, v4, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v2, v2, v5, vcc_lo
; GFX10-NEXT: s_movk_i32 s1, 0xff
; GFX10-NEXT: v_cndmask_b32_e64 v1, v1, v6, s0
; GFX10-NEXT: v_cndmask_b32_e64 v3, v3, v7, s0
; GFX10-NEXT: v_and_b32_sdwa v0, v0, s1 dst_sel:BYTE_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX10-NEXT: v_and_b32_sdwa v2, v2, s1 dst_sel:BYTE_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX10-NEXT: v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX10-NEXT: v_mov_b32_e32 v1, 0
; GFX10-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: global_store_short v1, v0, s[4:5]
; GFX10-NEXT: global_store_short v1, v2, s[6:7]
; GFX10-NEXT: s_endpgm
%div = udiv <2 x i8> %x, %y
store <2 x i8> %div, <2 x i8> addrspace(1)* %out0
%rem = urem <2 x i8> %x, %y
store <2 x i8> %rem, <2 x i8> addrspace(1)* %out1
ret void
}
define amdgpu_kernel void @udiv_i16(i16 addrspace(1)* %out0, i16 addrspace(1)* %out1, i16 %x, i16 %y) {
; GFX8-LABEL: udiv_i16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_load_dword s6, s[4:5], 0x10
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
; GFX8-NEXT: s_lshr_b32 s7, s6, 16
; GFX8-NEXT: v_cvt_f32_u32_e32 v0, s7
; GFX8-NEXT: s_sub_i32 s0, 0, s7
; GFX8-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX8-NEXT: v_mul_f32_e32 v0, 0x4f7ffffe, v0
; GFX8-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX8-NEXT: v_mul_lo_u32 v1, s0, v0
; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x0
; GFX8-NEXT: s_and_b32 s4, s6, 0xffff
; GFX8-NEXT: v_mul_hi_u32 v1, v0, v1
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v0, v1
; GFX8-NEXT: v_mul_hi_u32 v2, s4, v0
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
; GFX8-NEXT: v_mov_b32_e32 v0, s0
; GFX8-NEXT: v_mov_b32_e32 v1, s1
; GFX8-NEXT: v_mul_lo_u32 v3, v2, s7
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 1, v2
; GFX8-NEXT: v_sub_u32_e32 v3, vcc, s4, v3
; GFX8-NEXT: v_cmp_le_u32_e32 vcc, s7, v3
; GFX8-NEXT: v_cndmask_b32_e32 v2, v2, v4, vcc
; GFX8-NEXT: v_subrev_u32_e64 v4, s[0:1], s7, v3
; GFX8-NEXT: v_cndmask_b32_e32 v3, v3, v4, vcc
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 1, v2
; GFX8-NEXT: v_cmp_le_u32_e32 vcc, s7, v3
; GFX8-NEXT: v_cndmask_b32_e32 v2, v2, v4, vcc
; GFX8-NEXT: v_subrev_u32_e64 v4, s[0:1], s7, v3
; GFX8-NEXT: flat_store_short v[0:1], v2
; GFX8-NEXT: v_mov_b32_e32 v0, s2
; GFX8-NEXT: v_cndmask_b32_e32 v3, v3, v4, vcc
; GFX8-NEXT: v_mov_b32_e32 v1, s3
; GFX8-NEXT: flat_store_short v[0:1], v3
; GFX8-NEXT: s_endpgm
;
; GFX9-LABEL: udiv_i16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_load_dword s0, s[4:5], 0x10
; GFX9-NEXT: v_mov_b32_e32 v2, 0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_lshr_b32 s6, s0, 16
; GFX9-NEXT: v_cvt_f32_u32_e32 v0, s6
; GFX9-NEXT: s_sub_i32 s1, 0, s6
; GFX9-NEXT: s_and_b32 s7, s0, 0xffff
; GFX9-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX9-NEXT: v_mul_f32_e32 v0, 0x4f7ffffe, v0
; GFX9-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX9-NEXT: v_mul_lo_u32 v1, s1, v0
; GFX9-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x0
; GFX9-NEXT: v_mul_hi_u32 v1, v0, v1
; GFX9-NEXT: v_add_u32_e32 v0, v0, v1
; GFX9-NEXT: v_mul_hi_u32 v0, s7, v0
; GFX9-NEXT: v_mul_lo_u32 v1, v0, s6
; GFX9-NEXT: v_add_u32_e32 v3, 1, v0
; GFX9-NEXT: v_sub_u32_e32 v1, s7, v1
; GFX9-NEXT: v_cmp_le_u32_e32 vcc, s6, v1
; GFX9-NEXT: v_cndmask_b32_e32 v0, v0, v3, vcc
; GFX9-NEXT: v_subrev_u32_e32 v3, s6, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX9-NEXT: v_add_u32_e32 v3, 1, v0
; GFX9-NEXT: v_cmp_le_u32_e32 vcc, s6, v1
; GFX9-NEXT: v_cndmask_b32_e32 v0, v0, v3, vcc
; GFX9-NEXT: v_subrev_u32_e32 v3, s6, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: global_store_short v2, v0, s[0:1]
; GFX9-NEXT: global_store_short v2, v1, s[2:3]
; GFX9-NEXT: s_endpgm
;
; GFX10-LABEL: udiv_i16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_load_dword s0, s[4:5], 0x10
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: s_lshr_b32 s6, s0, 16
; GFX10-NEXT: s_and_b32 s0, s0, 0xffff
; GFX10-NEXT: v_cvt_f32_u32_e32 v0, s6
; GFX10-NEXT: s_sub_i32 s1, 0, s6
; GFX10-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX10-NEXT: v_mul_f32_e32 v0, 0x4f7ffffe, v0
; GFX10-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX10-NEXT: v_mul_lo_u32 v1, s1, v0
; GFX10-NEXT: v_mul_hi_u32 v1, v0, v1
; GFX10-NEXT: v_add_nc_u32_e32 v0, v0, v1
; GFX10-NEXT: v_mul_hi_u32 v0, s0, v0
; GFX10-NEXT: v_mul_lo_u32 v1, v0, s6
; GFX10-NEXT: v_add_nc_u32_e32 v2, 1, v0
; GFX10-NEXT: v_sub_nc_u32_e32 v1, s0, v1
; GFX10-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x0
; GFX10-NEXT: v_subrev_nc_u32_e32 v3, s6, v1
; GFX10-NEXT: v_cmp_le_u32_e32 vcc_lo, s6, v1
; GFX10-NEXT: v_cndmask_b32_e32 v0, v0, v2, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX10-NEXT: v_add_nc_u32_e32 v2, 1, v0
; GFX10-NEXT: v_cmp_le_u32_e32 vcc_lo, s6, v1
; GFX10-NEXT: v_subrev_nc_u32_e32 v3, s6, v1
; GFX10-NEXT: v_cndmask_b32_e32 v0, v0, v2, vcc_lo
; GFX10-NEXT: v_mov_b32_e32 v2, 0
; GFX10-NEXT: v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: global_store_short v2, v0, s[0:1]
; GFX10-NEXT: global_store_short v2, v1, s[2:3]
; GFX10-NEXT: s_endpgm
%div = udiv i16 %x, %y
store i16 %div, i16 addrspace(1)* %out0
%rem = urem i16 %x, %y
store i16 %rem, i16 addrspace(1)* %out1
ret void
}
define amdgpu_kernel void @udivrem_v2i16(<2 x i16> addrspace(1)* %out0, <2 x i16> addrspace(1)* %out1, <2 x i16> %x, <2 x i16> %y) {
; GFX8-LABEL: udivrem_v2i16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x10
; GFX8-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x0
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
; GFX8-NEXT: s_and_b32 s2, s1, 0xffff
; GFX8-NEXT: v_cvt_f32_u32_e32 v0, s2
; GFX8-NEXT: s_lshr_b32 s3, s1, 16
; GFX8-NEXT: v_cvt_f32_u32_e32 v1, s3
; GFX8-NEXT: s_sub_i32 s1, 0, s2
; GFX8-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX8-NEXT: s_lshr_b32 s8, s0, 16
; GFX8-NEXT: v_rcp_iflag_f32_e32 v1, v1
; GFX8-NEXT: s_and_b32 s0, s0, 0xffff
; GFX8-NEXT: v_mul_f32_e32 v0, 0x4f7ffffe, v0
; GFX8-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX8-NEXT: v_mul_f32_e32 v1, 0x4f7ffffe, v1
; GFX8-NEXT: v_cvt_u32_f32_e32 v1, v1
; GFX8-NEXT: v_mul_lo_u32 v2, s1, v0
; GFX8-NEXT: s_sub_i32 s1, 0, s3
; GFX8-NEXT: v_mul_lo_u32 v3, s1, v1
; GFX8-NEXT: v_mul_hi_u32 v2, v0, v2
; GFX8-NEXT: v_mul_hi_u32 v3, v1, v3
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v0, v2
; GFX8-NEXT: v_mul_hi_u32 v0, s0, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v3
; GFX8-NEXT: v_mul_hi_u32 v1, s8, v1
; GFX8-NEXT: v_mul_lo_u32 v2, v0, s2
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 1, v0
; GFX8-NEXT: v_sub_u32_e32 v2, vcc, s0, v2
; GFX8-NEXT: v_cmp_le_u32_e32 vcc, s2, v2
; GFX8-NEXT: v_cndmask_b32_e32 v0, v0, v3, vcc
; GFX8-NEXT: v_subrev_u32_e64 v3, s[0:1], s2, v2
; GFX8-NEXT: v_cndmask_b32_e32 v2, v2, v3, vcc
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 1, v0
; GFX8-NEXT: v_cmp_le_u32_e32 vcc, s2, v2
; GFX8-NEXT: v_cndmask_b32_e32 v0, v0, v3, vcc
; GFX8-NEXT: v_mul_lo_u32 v3, v1, s3
; GFX8-NEXT: v_subrev_u32_e64 v4, s[0:1], s2, v2
; GFX8-NEXT: v_cndmask_b32_e32 v2, v2, v4, vcc
; GFX8-NEXT: v_sub_u32_e32 v3, vcc, s8, v3
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 1, v1
; GFX8-NEXT: v_cmp_le_u32_e32 vcc, s3, v3
; GFX8-NEXT: v_cndmask_b32_e32 v1, v1, v4, vcc
; GFX8-NEXT: v_subrev_u32_e64 v4, s[0:1], s3, v3
; GFX8-NEXT: v_cndmask_b32_e32 v3, v3, v4, vcc
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 1, v1
; GFX8-NEXT: v_cmp_le_u32_e32 vcc, s3, v3
; GFX8-NEXT: v_cndmask_b32_e32 v1, v1, v4, vcc
; GFX8-NEXT: v_subrev_u32_e64 v4, s[0:1], s3, v3
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff, v1
; GFX8-NEXT: v_cndmask_b32_e32 v3, v3, v4, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_or_b32_sdwa v4, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff, v3
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_or_b32_sdwa v2, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: v_mov_b32_e32 v0, s4
; GFX8-NEXT: v_mov_b32_e32 v1, s5
; GFX8-NEXT: flat_store_dword v[0:1], v4
; GFX8-NEXT: v_mov_b32_e32 v0, s6
; GFX8-NEXT: v_mov_b32_e32 v1, s7
; GFX8-NEXT: flat_store_dword v[0:1], v2
; GFX8-NEXT: s_endpgm
;
; GFX9-LABEL: udivrem_v2i16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x10
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_and_b32 s3, s1, 0xffff
; GFX9-NEXT: v_cvt_f32_u32_e32 v0, s3
; GFX9-NEXT: s_lshr_b32 s2, s1, 16
; GFX9-NEXT: v_cvt_f32_u32_e32 v1, s2
; GFX9-NEXT: s_sub_i32 s1, 0, s3
; GFX9-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX9-NEXT: s_sub_i32 s6, 0, s2
; GFX9-NEXT: v_rcp_iflag_f32_e32 v1, v1
; GFX9-NEXT: v_mul_f32_e32 v0, 0x4f7ffffe, v0
; GFX9-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX9-NEXT: v_mul_f32_e32 v1, 0x4f7ffffe, v1
; GFX9-NEXT: v_cvt_u32_f32_e32 v1, v1
; GFX9-NEXT: v_mul_lo_u32 v2, s1, v0
; GFX9-NEXT: s_lshr_b32 s1, s0, 16
; GFX9-NEXT: v_mul_lo_u32 v3, s6, v1
; GFX9-NEXT: s_and_b32 s0, s0, 0xffff
; GFX9-NEXT: v_mul_hi_u32 v2, v0, v2
; GFX9-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x0
; GFX9-NEXT: v_mul_hi_u32 v3, v1, v3
; GFX9-NEXT: v_add_u32_e32 v0, v0, v2
; GFX9-NEXT: v_mul_hi_u32 v0, s0, v0
; GFX9-NEXT: v_add_u32_e32 v1, v1, v3
; GFX9-NEXT: v_mul_hi_u32 v1, s1, v1
; GFX9-NEXT: v_mul_lo_u32 v2, v0, s3
; GFX9-NEXT: v_add_u32_e32 v4, 1, v0
; GFX9-NEXT: v_mul_lo_u32 v3, v1, s2
; GFX9-NEXT: v_add_u32_e32 v5, 1, v1
; GFX9-NEXT: v_sub_u32_e32 v2, s0, v2
; GFX9-NEXT: v_cmp_le_u32_e32 vcc, s3, v2
; GFX9-NEXT: v_cndmask_b32_e32 v0, v0, v4, vcc
; GFX9-NEXT: v_subrev_u32_e32 v4, s3, v2
; GFX9-NEXT: v_cndmask_b32_e32 v2, v2, v4, vcc
; GFX9-NEXT: v_add_u32_e32 v4, 1, v0
; GFX9-NEXT: v_cmp_le_u32_e32 vcc, s3, v2
; GFX9-NEXT: v_sub_u32_e32 v3, s1, v3
; GFX9-NEXT: v_cndmask_b32_e32 v0, v0, v4, vcc
; GFX9-NEXT: v_subrev_u32_e32 v4, s3, v2
; GFX9-NEXT: v_cmp_le_u32_e64 s[0:1], s2, v3
; GFX9-NEXT: v_cndmask_b32_e32 v2, v2, v4, vcc
; GFX9-NEXT: v_subrev_u32_e32 v4, s2, v3
; GFX9-NEXT: v_cndmask_b32_e64 v1, v1, v5, s[0:1]
; GFX9-NEXT: v_cndmask_b32_e64 v3, v3, v4, s[0:1]
; GFX9-NEXT: v_add_u32_e32 v4, 1, v1
; GFX9-NEXT: v_cmp_le_u32_e32 vcc, s2, v3
; GFX9-NEXT: v_cndmask_b32_e32 v1, v1, v4, vcc
; GFX9-NEXT: v_subrev_u32_e32 v4, s2, v3
; GFX9-NEXT: v_cndmask_b32_e32 v3, v3, v4, vcc
; GFX9-NEXT: v_mov_b32_e32 v4, 0xffff
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT: v_and_or_b32 v0, v0, v4, v1
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v3
; GFX9-NEXT: v_and_or_b32 v1, v2, v4, v1
; GFX9-NEXT: v_mov_b32_e32 v2, 0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: global_store_dword v2, v0, s[4:5]
; GFX9-NEXT: global_store_dword v2, v1, s[6:7]
; GFX9-NEXT: s_endpgm
;
; GFX10-LABEL: udivrem_v2i16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x10
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: s_lshr_b32 s2, s1, 16
; GFX10-NEXT: s_and_b32 s1, s1, 0xffff
; GFX10-NEXT: v_cvt_f32_u32_e32 v0, s2
; GFX10-NEXT: v_cvt_f32_u32_e32 v1, s1
; GFX10-NEXT: s_sub_i32 s3, 0, s2
; GFX10-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x0
; GFX10-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX10-NEXT: v_rcp_iflag_f32_e32 v1, v1
; GFX10-NEXT: v_mul_f32_e32 v0, 0x4f7ffffe, v0
; GFX10-NEXT: v_mul_f32_e32 v1, 0x4f7ffffe, v1
; GFX10-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX10-NEXT: v_cvt_u32_f32_e32 v1, v1
; GFX10-NEXT: v_mul_lo_u32 v2, s3, v0
; GFX10-NEXT: s_sub_i32 s3, 0, s1
; GFX10-NEXT: v_mul_lo_u32 v3, s3, v1
; GFX10-NEXT: s_lshr_b32 s3, s0, 16
; GFX10-NEXT: s_and_b32 s0, s0, 0xffff
; GFX10-NEXT: v_mul_hi_u32 v2, v0, v2
; GFX10-NEXT: v_mul_hi_u32 v3, v1, v3
; GFX10-NEXT: v_add_nc_u32_e32 v0, v0, v2
; GFX10-NEXT: v_add_nc_u32_e32 v1, v1, v3
; GFX10-NEXT: v_mul_hi_u32 v0, s3, v0
; GFX10-NEXT: v_mul_hi_u32 v1, s0, v1
; GFX10-NEXT: v_mul_lo_u32 v2, v0, s2
; GFX10-NEXT: v_add_nc_u32_e32 v4, 1, v0
; GFX10-NEXT: v_mul_lo_u32 v3, v1, s1
; GFX10-NEXT: v_add_nc_u32_e32 v6, 1, v1
; GFX10-NEXT: v_sub_nc_u32_e32 v2, s3, v2
; GFX10-NEXT: v_sub_nc_u32_e32 v3, s0, v3
; GFX10-NEXT: v_subrev_nc_u32_e32 v5, s2, v2
; GFX10-NEXT: v_cmp_le_u32_e32 vcc_lo, s2, v2
; GFX10-NEXT: v_cmp_le_u32_e64 s0, s1, v3
; GFX10-NEXT: v_subrev_nc_u32_e32 v7, s1, v3
; GFX10-NEXT: v_cndmask_b32_e32 v0, v0, v4, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v2, v2, v5, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e64 v1, v1, v6, s0
; GFX10-NEXT: v_cndmask_b32_e64 v3, v3, v7, s0
; GFX10-NEXT: v_add_nc_u32_e32 v4, 1, v0
; GFX10-NEXT: v_cmp_le_u32_e32 vcc_lo, s2, v2
; GFX10-NEXT: v_subrev_nc_u32_e32 v5, s2, v2
; GFX10-NEXT: v_add_nc_u32_e32 v6, 1, v1
; GFX10-NEXT: v_cmp_le_u32_e64 s0, s1, v3
; GFX10-NEXT: v_subrev_nc_u32_e32 v7, s1, v3
; GFX10-NEXT: v_cndmask_b32_e32 v0, v0, v4, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v2, v2, v5, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e64 v1, v1, v6, s0
; GFX10-NEXT: v_cndmask_b32_e64 v3, v3, v7, s0
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
; GFX10-NEXT: v_and_or_b32 v0, 0xffff, v1, v0
; GFX10-NEXT: v_mov_b32_e32 v1, 0
; GFX10-NEXT: v_and_or_b32 v2, 0xffff, v3, v2
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: global_store_dword v1, v0, s[4:5]
; GFX10-NEXT: global_store_dword v1, v2, s[6:7]
; GFX10-NEXT: s_endpgm
%div = udiv <2 x i16> %x, %y
store <2 x i16> %div, <2 x i16> addrspace(1)* %out0
%rem = urem <2 x i16> %x, %y
store <2 x i16> %rem, <2 x i16> addrspace(1)* %out1
ret void
}
define amdgpu_kernel void @udivrem_i3(i3 addrspace(1)* %out0, i3 addrspace(1)* %out1, i3 %x, i3 %y) {
; GFX8-LABEL: udivrem_i3:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_load_dword s6, s[4:5], 0x10
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
; GFX8-NEXT: s_bfe_u32 s7, s6, 0x30008
; GFX8-NEXT: v_cvt_f32_ubyte0_e32 v0, s7
; GFX8-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX8-NEXT: s_sub_i32 s0, 0, s7
; GFX8-NEXT: v_mul_f32_e32 v0, 0x4f7ffffe, v0
; GFX8-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX8-NEXT: v_mul_lo_u32 v1, s0, v0
; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x0
; GFX8-NEXT: s_and_b32 s4, s6, 7
; GFX8-NEXT: v_mul_hi_u32 v1, v0, v1
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v0, v1
; GFX8-NEXT: v_mul_hi_u32 v2, s4, v0
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
; GFX8-NEXT: v_mov_b32_e32 v0, s0
; GFX8-NEXT: v_mov_b32_e32 v1, s1
; GFX8-NEXT: v_mul_lo_u32 v3, v2, s7
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 1, v2
; GFX8-NEXT: v_sub_u32_e32 v3, vcc, s4, v3
; GFX8-NEXT: v_cmp_le_u32_e32 vcc, s7, v3
; GFX8-NEXT: v_cndmask_b32_e32 v2, v2, v4, vcc
; GFX8-NEXT: v_subrev_u32_e64 v4, s[0:1], s7, v3
; GFX8-NEXT: v_cndmask_b32_e32 v3, v3, v4, vcc
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 1, v2
; GFX8-NEXT: v_cmp_le_u32_e32 vcc, s7, v3
; GFX8-NEXT: v_cndmask_b32_e32 v2, v2, v4, vcc
; GFX8-NEXT: v_subrev_u32_e64 v4, s[0:1], s7, v3
; GFX8-NEXT: v_and_b32_e32 v2, 7, v2
; GFX8-NEXT: v_cndmask_b32_e32 v3, v3, v4, vcc
; GFX8-NEXT: flat_store_byte v[0:1], v2
; GFX8-NEXT: v_mov_b32_e32 v0, s2
; GFX8-NEXT: v_and_b32_e32 v2, 7, v3
; GFX8-NEXT: v_mov_b32_e32 v1, s3
; GFX8-NEXT: flat_store_byte v[0:1], v2
; GFX8-NEXT: s_endpgm
;
; GFX9-LABEL: udivrem_i3:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_load_dword s0, s[4:5], 0x10
; GFX9-NEXT: v_mov_b32_e32 v2, 0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_bfe_u32 s6, s0, 0x30008
; GFX9-NEXT: v_cvt_f32_ubyte0_e32 v0, s6
; GFX9-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX9-NEXT: s_sub_i32 s1, 0, s6
; GFX9-NEXT: s_and_b32 s7, s0, 7
; GFX9-NEXT: v_mul_f32_e32 v0, 0x4f7ffffe, v0
; GFX9-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX9-NEXT: v_mul_lo_u32 v1, s1, v0
; GFX9-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x0
; GFX9-NEXT: v_mul_hi_u32 v1, v0, v1
; GFX9-NEXT: v_add_u32_e32 v0, v0, v1
; GFX9-NEXT: v_mul_hi_u32 v0, s7, v0
; GFX9-NEXT: v_mul_lo_u32 v1, v0, s6
; GFX9-NEXT: v_add_u32_e32 v3, 1, v0
; GFX9-NEXT: v_sub_u32_e32 v1, s7, v1
; GFX9-NEXT: v_cmp_le_u32_e32 vcc, s6, v1
; GFX9-NEXT: v_cndmask_b32_e32 v0, v0, v3, vcc
; GFX9-NEXT: v_subrev_u32_e32 v3, s6, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX9-NEXT: v_add_u32_e32 v3, 1, v0
; GFX9-NEXT: v_cmp_le_u32_e32 vcc, s6, v1
; GFX9-NEXT: v_cndmask_b32_e32 v0, v0, v3, vcc
; GFX9-NEXT: v_subrev_u32_e32 v3, s6, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX9-NEXT: v_and_b32_e32 v0, 7, v0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: global_store_byte v2, v0, s[0:1]
; GFX9-NEXT: v_and_b32_e32 v0, 7, v1
; GFX9-NEXT: global_store_byte v2, v0, s[2:3]
; GFX9-NEXT: s_endpgm
;
; GFX10-LABEL: udivrem_i3:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_load_dword s0, s[4:5], 0x10
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: s_bfe_u32 s6, s0, 0x30008
; GFX10-NEXT: s_and_b32 s0, s0, 7
; GFX10-NEXT: v_cvt_f32_ubyte0_e32 v0, s6
; GFX10-NEXT: s_sub_i32 s1, 0, s6
; GFX10-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX10-NEXT: v_mul_f32_e32 v0, 0x4f7ffffe, v0
; GFX10-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX10-NEXT: v_mul_lo_u32 v1, s1, v0
; GFX10-NEXT: v_mul_hi_u32 v1, v0, v1
; GFX10-NEXT: v_add_nc_u32_e32 v0, v0, v1
; GFX10-NEXT: v_mul_hi_u32 v0, s0, v0
; GFX10-NEXT: v_mul_lo_u32 v1, v0, s6
; GFX10-NEXT: v_add_nc_u32_e32 v2, 1, v0
; GFX10-NEXT: v_sub_nc_u32_e32 v1, s0, v1
; GFX10-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x0
; GFX10-NEXT: v_subrev_nc_u32_e32 v3, s6, v1
; GFX10-NEXT: v_cmp_le_u32_e32 vcc_lo, s6, v1
; GFX10-NEXT: v_cndmask_b32_e32 v0, v0, v2, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX10-NEXT: v_add_nc_u32_e32 v2, 1, v0
; GFX10-NEXT: v_cmp_le_u32_e32 vcc_lo, s6, v1
; GFX10-NEXT: v_subrev_nc_u32_e32 v3, s6, v1
; GFX10-NEXT: v_cndmask_b32_e32 v0, v0, v2, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX10-NEXT: v_mov_b32_e32 v2, 0
; GFX10-NEXT: v_and_b32_e32 v0, 7, v0
; GFX10-NEXT: v_and_b32_e32 v1, 7, v1
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: global_store_byte v2, v0, s[0:1]
; GFX10-NEXT: global_store_byte v2, v1, s[2:3]
; GFX10-NEXT: s_endpgm
%div = udiv i3 %x, %y
store i3 %div, i3 addrspace(1)* %out0
%rem = urem i3 %x, %y
store i3 %rem, i3 addrspace(1)* %out1
ret void
}
define amdgpu_kernel void @udivrem_i27(i27 addrspace(1)* %out0, i27 addrspace(1)* %out1, i27 %x, i27 %y) {
; GFX8-LABEL: udivrem_i27:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x10
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
; GFX8-NEXT: s_and_b32 s7, s7, 0x7ffffff
; GFX8-NEXT: v_cvt_f32_u32_e32 v0, s7
; GFX8-NEXT: s_sub_i32 s0, 0, s7
; GFX8-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX8-NEXT: v_mul_f32_e32 v0, 0x4f7ffffe, v0
; GFX8-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX8-NEXT: v_mul_lo_u32 v1, s0, v0
; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x0
; GFX8-NEXT: s_and_b32 s4, s6, 0x7ffffff
; GFX8-NEXT: v_mul_hi_u32 v1, v0, v1
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v0, v1
; GFX8-NEXT: v_mul_hi_u32 v2, s4, v0
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
; GFX8-NEXT: v_mov_b32_e32 v0, s0
; GFX8-NEXT: v_mov_b32_e32 v1, s1
; GFX8-NEXT: v_mul_lo_u32 v3, v2, s7
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 1, v2
; GFX8-NEXT: v_sub_u32_e32 v3, vcc, s4, v3
; GFX8-NEXT: v_cmp_le_u32_e32 vcc, s7, v3
; GFX8-NEXT: v_cndmask_b32_e32 v2, v2, v4, vcc
; GFX8-NEXT: v_subrev_u32_e64 v4, s[0:1], s7, v3
; GFX8-NEXT: v_cndmask_b32_e32 v3, v3, v4, vcc
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 1, v2
; GFX8-NEXT: v_cmp_le_u32_e32 vcc, s7, v3
; GFX8-NEXT: v_cndmask_b32_e32 v2, v2, v4, vcc
; GFX8-NEXT: v_subrev_u32_e64 v4, s[0:1], s7, v3
; GFX8-NEXT: v_and_b32_e32 v2, 0x7ffffff, v2
; GFX8-NEXT: v_cndmask_b32_e32 v3, v3, v4, vcc
; GFX8-NEXT: flat_store_dword v[0:1], v2
; GFX8-NEXT: v_mov_b32_e32 v0, s2
; GFX8-NEXT: v_and_b32_e32 v2, 0x7ffffff, v3
; GFX8-NEXT: v_mov_b32_e32 v1, s3
; GFX8-NEXT: flat_store_dword v[0:1], v2
; GFX8-NEXT: s_endpgm
;
; GFX9-LABEL: udivrem_i27:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x10
; GFX9-NEXT: v_mov_b32_e32 v2, 0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_and_b32 s6, s1, 0x7ffffff
; GFX9-NEXT: v_cvt_f32_u32_e32 v0, s6
; GFX9-NEXT: s_sub_i32 s1, 0, s6
; GFX9-NEXT: s_and_b32 s7, s0, 0x7ffffff
; GFX9-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX9-NEXT: v_mul_f32_e32 v0, 0x4f7ffffe, v0
; GFX9-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX9-NEXT: v_mul_lo_u32 v1, s1, v0
; GFX9-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x0
; GFX9-NEXT: v_mul_hi_u32 v1, v0, v1
; GFX9-NEXT: v_add_u32_e32 v0, v0, v1
; GFX9-NEXT: v_mul_hi_u32 v0, s7, v0
; GFX9-NEXT: v_mul_lo_u32 v1, v0, s6
; GFX9-NEXT: v_add_u32_e32 v3, 1, v0
; GFX9-NEXT: v_sub_u32_e32 v1, s7, v1
; GFX9-NEXT: v_cmp_le_u32_e32 vcc, s6, v1
; GFX9-NEXT: v_cndmask_b32_e32 v0, v0, v3, vcc
; GFX9-NEXT: v_subrev_u32_e32 v3, s6, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX9-NEXT: v_add_u32_e32 v3, 1, v0
; GFX9-NEXT: v_cmp_le_u32_e32 vcc, s6, v1
; GFX9-NEXT: v_cndmask_b32_e32 v0, v0, v3, vcc
; GFX9-NEXT: v_subrev_u32_e32 v3, s6, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX9-NEXT: v_and_b32_e32 v0, 0x7ffffff, v0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: global_store_dword v2, v0, s[0:1]
; GFX9-NEXT: v_and_b32_e32 v0, 0x7ffffff, v1
; GFX9-NEXT: global_store_dword v2, v0, s[2:3]
; GFX9-NEXT: s_endpgm
;
; GFX10-LABEL: udivrem_i27:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x10
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: s_and_b32 s6, s1, 0x7ffffff
; GFX10-NEXT: s_and_b32 s0, s0, 0x7ffffff
; GFX10-NEXT: v_cvt_f32_u32_e32 v0, s6
; GFX10-NEXT: s_sub_i32 s1, 0, s6
; GFX10-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX10-NEXT: v_mul_f32_e32 v0, 0x4f7ffffe, v0
; GFX10-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX10-NEXT: v_mul_lo_u32 v1, s1, v0
; GFX10-NEXT: v_mul_hi_u32 v1, v0, v1
; GFX10-NEXT: v_add_nc_u32_e32 v0, v0, v1
; GFX10-NEXT: v_mul_hi_u32 v0, s0, v0
; GFX10-NEXT: v_mul_lo_u32 v1, v0, s6
; GFX10-NEXT: v_add_nc_u32_e32 v2, 1, v0
; GFX10-NEXT: v_sub_nc_u32_e32 v1, s0, v1
; GFX10-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x0
; GFX10-NEXT: v_subrev_nc_u32_e32 v3, s6, v1
; GFX10-NEXT: v_cmp_le_u32_e32 vcc_lo, s6, v1
; GFX10-NEXT: v_cndmask_b32_e32 v0, v0, v2, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX10-NEXT: v_add_nc_u32_e32 v2, 1, v0
; GFX10-NEXT: v_cmp_le_u32_e32 vcc_lo, s6, v1
; GFX10-NEXT: v_subrev_nc_u32_e32 v3, s6, v1
; GFX10-NEXT: v_cndmask_b32_e32 v0, v0, v2, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX10-NEXT: v_mov_b32_e32 v2, 0
; GFX10-NEXT: v_and_b32_e32 v0, 0x7ffffff, v0
; GFX10-NEXT: v_and_b32_e32 v1, 0x7ffffff, v1
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: global_store_dword v2, v0, s[0:1]
; GFX10-NEXT: global_store_dword v2, v1, s[2:3]
; GFX10-NEXT: s_endpgm
%div = udiv i27 %x, %y
store i27 %div, i27 addrspace(1)* %out0
%rem = urem i27 %x, %y
store i27 %rem, i27 addrspace(1)* %out1
ret void
}