caio/clang-p2996
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
1e2c5481502b305d2e38ce94defffb43b82dbe3c
clang-p2996/llvm/test/CodeGen/AMDGPU/sdiv.ll
Alexander Timofeev fbdea5a2e9 [AMDGPU] Always select s_cselect_b32 for uniform 'select' SDNode 
This patch contains changes necessary to carry physical condition register (SCC) dependencies through the SDNode scheduler.  It adds the edge in the SDNodeScheduler dependency graph instead of inserting the SCC copy between each definition and use. This approach lets the scheduler place instructions in an optimal way placing the copy only when the dependency cannot be resolved.

Reviewed By: rampitec

Differential Revision: https://reviews.llvm.org/D133593
2022-09-15 22:03:56 +02:00

2370 lines
94 KiB
LLVM
Raw Blame History

; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc < %s -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=gfx600 | FileCheck %s --check-prefix=GCN
; RUN: llc < %s -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=tonga -mattr=-flat-for-global | FileCheck %s --check-prefix=TONGA
; RUN: llc < %s -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=gfx900 -mattr=-flat-for-global | FileCheck %s --check-prefix=GFX9
; RUN: llc < %s -amdgpu-scalarize-global-loads=false -march=r600 -mcpu=redwood | FileCheck %s --check-prefix=EG
; The code generated by sdiv is long and complex and may frequently change.
; The goal of this test is to make sure the ISel doesn't fail.
;
; This program was previously failing to compile when one of the selectcc
; opcodes generated by the sdiv lowering was being legalized and optimized to:
; selectcc Remainder -1, 0, -1, SETGT
; This was fixed by adding an additional pattern in R600Instructions.td to
; match this pattern with a CNDGE_INT.
define amdgpu_kernel void @sdiv_i32(i32 addrspace(1)* %out, i32 addrspace(1)* %in) {
; GCN-LABEL: sdiv_i32:
; GCN: ; %bb.0:
; GCN-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s6, -1
; GCN-NEXT: s_mov_b32 s10, s6
; GCN-NEXT: s_mov_b32 s11, s7
; GCN-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s8, s2
; GCN-NEXT: s_mov_b32 s9, s3
; GCN-NEXT: buffer_load_dwordx2 v[0:1], off, s[8:11], 0
; GCN-NEXT: s_mov_b32 s4, s0
; GCN-NEXT: s_mov_b32 s5, s1
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_ashrrev_i32_e32 v2, 31, v1
; GCN-NEXT: v_add_i32_e32 v1, vcc, v2, v1
; GCN-NEXT: v_xor_b32_e32 v1, v1, v2
; GCN-NEXT: v_cvt_f32_u32_e32 v3, v1
; GCN-NEXT: v_sub_i32_e32 v4, vcc, 0, v1
; GCN-NEXT: v_ashrrev_i32_e32 v5, 31, v0
; GCN-NEXT: v_rcp_iflag_f32_e32 v3, v3
; GCN-NEXT: v_add_i32_e32 v0, vcc, v5, v0
; GCN-NEXT: v_xor_b32_e32 v0, v0, v5
; GCN-NEXT: v_mul_f32_e32 v3, 0x4f7ffffe, v3
; GCN-NEXT: v_cvt_u32_f32_e32 v3, v3
; GCN-NEXT: v_xor_b32_e32 v2, v5, v2
; GCN-NEXT: v_mul_lo_u32 v4, v4, v3
; GCN-NEXT: v_mul_hi_u32 v4, v3, v4
; GCN-NEXT: v_add_i32_e32 v3, vcc, v3, v4
; GCN-NEXT: v_mul_hi_u32 v3, v0, v3
; GCN-NEXT: v_mul_lo_u32 v4, v3, v1
; GCN-NEXT: v_add_i32_e32 v5, vcc, 1, v3
; GCN-NEXT: v_sub_i32_e32 v0, vcc, v0, v4
; GCN-NEXT: v_subrev_i32_e32 v4, vcc, v1, v0
; GCN-NEXT: v_cmp_ge_u32_e32 vcc, v0, v1
; GCN-NEXT: v_cndmask_b32_e32 v3, v3, v5, vcc
; GCN-NEXT: v_cndmask_b32_e32 v0, v0, v4, vcc
; GCN-NEXT: v_add_i32_e32 v4, vcc, 1, v3
; GCN-NEXT: v_cmp_ge_u32_e32 vcc, v0, v1
; GCN-NEXT: v_cndmask_b32_e32 v0, v3, v4, vcc
; GCN-NEXT: v_xor_b32_e32 v0, v0, v2
; GCN-NEXT: v_subrev_i32_e32 v0, vcc, v2, v0
; GCN-NEXT: buffer_store_dword v0, off, s[4:7], 0
; GCN-NEXT: s_endpgm
;
; TONGA-LABEL: sdiv_i32:
; TONGA: ; %bb.0:
; TONGA-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24
; TONGA-NEXT: s_mov_b32 s7, 0xf000
; TONGA-NEXT: s_mov_b32 s6, -1
; TONGA-NEXT: s_mov_b32 s10, s6
; TONGA-NEXT: s_mov_b32 s11, s7
; TONGA-NEXT: s_waitcnt lgkmcnt(0)
; TONGA-NEXT: s_mov_b32 s8, s2
; TONGA-NEXT: s_mov_b32 s9, s3
; TONGA-NEXT: buffer_load_dwordx2 v[0:1], off, s[8:11], 0
; TONGA-NEXT: s_mov_b32 s4, s0
; TONGA-NEXT: s_mov_b32 s5, s1
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: v_ashrrev_i32_e32 v2, 31, v1
; TONGA-NEXT: v_add_u32_e32 v1, vcc, v2, v1
; TONGA-NEXT: v_xor_b32_e32 v1, v1, v2
; TONGA-NEXT: v_cvt_f32_u32_e32 v3, v1
; TONGA-NEXT: v_sub_u32_e32 v4, vcc, 0, v1
; TONGA-NEXT: v_ashrrev_i32_e32 v5, 31, v0
; TONGA-NEXT: v_rcp_iflag_f32_e32 v3, v3
; TONGA-NEXT: v_add_u32_e32 v0, vcc, v5, v0
; TONGA-NEXT: v_xor_b32_e32 v0, v0, v5
; TONGA-NEXT: v_mul_f32_e32 v3, 0x4f7ffffe, v3
; TONGA-NEXT: v_cvt_u32_f32_e32 v3, v3
; TONGA-NEXT: v_xor_b32_e32 v2, v5, v2
; TONGA-NEXT: v_mul_lo_u32 v4, v4, v3
; TONGA-NEXT: v_mul_hi_u32 v4, v3, v4
; TONGA-NEXT: v_add_u32_e32 v3, vcc, v3, v4
; TONGA-NEXT: v_mul_hi_u32 v3, v0, v3
; TONGA-NEXT: v_mul_lo_u32 v4, v3, v1
; TONGA-NEXT: v_add_u32_e32 v5, vcc, 1, v3
; TONGA-NEXT: v_sub_u32_e32 v0, vcc, v0, v4
; TONGA-NEXT: v_subrev_u32_e32 v4, vcc, v1, v0
; TONGA-NEXT: v_cmp_ge_u32_e32 vcc, v0, v1
; TONGA-NEXT: v_cndmask_b32_e32 v3, v3, v5, vcc
; TONGA-NEXT: v_cndmask_b32_e32 v0, v0, v4, vcc
; TONGA-NEXT: v_add_u32_e32 v4, vcc, 1, v3
; TONGA-NEXT: v_cmp_ge_u32_e32 vcc, v0, v1
; TONGA-NEXT: v_cndmask_b32_e32 v0, v3, v4, vcc
; TONGA-NEXT: v_xor_b32_e32 v0, v0, v2
; TONGA-NEXT: v_subrev_u32_e32 v0, vcc, v2, v0
; TONGA-NEXT: buffer_store_dword v0, off, s[4:7], 0
; TONGA-NEXT: s_endpgm
;
; GFX9-LABEL: sdiv_i32:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24
; GFX9-NEXT: s_mov_b32 s7, 0xf000
; GFX9-NEXT: s_mov_b32 s6, -1
; GFX9-NEXT: s_mov_b32 s10, s6
; GFX9-NEXT: s_mov_b32 s11, s7
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_mov_b32 s8, s2
; GFX9-NEXT: s_mov_b32 s9, s3
; GFX9-NEXT: buffer_load_dwordx2 v[0:1], off, s[8:11], 0
; GFX9-NEXT: s_mov_b32 s4, s0
; GFX9-NEXT: s_mov_b32 s5, s1
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_readfirstlane_b32 s2, v1
; GFX9-NEXT: s_ashr_i32 s3, s2, 31
; GFX9-NEXT: s_add_i32 s2, s2, s3
; GFX9-NEXT: s_xor_b32 s2, s2, s3
; GFX9-NEXT: v_cvt_f32_u32_e32 v1, s2
; GFX9-NEXT: v_readfirstlane_b32 s0, v0
; GFX9-NEXT: s_ashr_i32 s1, s0, 31
; GFX9-NEXT: s_add_i32 s0, s0, s1
; GFX9-NEXT: v_rcp_iflag_f32_e32 v1, v1
; GFX9-NEXT: s_xor_b32 s3, s1, s3
; GFX9-NEXT: s_xor_b32 s0, s0, s1
; GFX9-NEXT: s_sub_i32 s1, 0, s2
; GFX9-NEXT: v_mul_f32_e32 v0, 0x4f7ffffe, v1
; GFX9-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX9-NEXT: v_readfirstlane_b32 s8, v0
; GFX9-NEXT: s_mul_i32 s1, s1, s8
; GFX9-NEXT: s_mul_hi_u32 s1, s8, s1
; GFX9-NEXT: s_add_i32 s8, s8, s1
; GFX9-NEXT: s_mul_hi_u32 s1, s0, s8
; GFX9-NEXT: s_mul_i32 s8, s1, s2
; GFX9-NEXT: s_sub_i32 s0, s0, s8
; GFX9-NEXT: s_add_i32 s9, s1, 1
; GFX9-NEXT: s_sub_i32 s8, s0, s2
; GFX9-NEXT: s_cmp_ge_u32 s0, s2
; GFX9-NEXT: s_cselect_b32 s1, s9, s1
; GFX9-NEXT: s_cselect_b32 s0, s8, s0
; GFX9-NEXT: s_add_i32 s8, s1, 1
; GFX9-NEXT: s_cmp_ge_u32 s0, s2
; GFX9-NEXT: s_cselect_b32 s0, s8, s1
; GFX9-NEXT: s_xor_b32 s0, s0, s3
; GFX9-NEXT: s_sub_i32 s0, s0, s3
; GFX9-NEXT: v_mov_b32_e32 v0, s0
; GFX9-NEXT: buffer_store_dword v0, off, s[4:7], 0
; GFX9-NEXT: s_endpgm
;
; EG-LABEL: sdiv_i32:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 0 @6
; EG-NEXT: ALU 26, @9, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.X, T1.X, 1
; EG-NEXT: CF_END
; EG-NEXT: PAD
; EG-NEXT: Fetch clause starting at 6:
; EG-NEXT: VTX_READ_64 T0.XY, T0.X, 0, #1
; EG-NEXT: ALU clause starting at 8:
; EG-NEXT: MOV * T0.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 9:
; EG-NEXT: SETGT_INT * T0.W, 0.0, T0.Y,
; EG-NEXT: ADD_INT * T1.W, T0.Y, PV.W,
; EG-NEXT: XOR_INT * T1.W, PV.W, T0.W,
; EG-NEXT: SUB_INT T2.W, 0.0, PV.W,
; EG-NEXT: RECIP_UINT * T0.Y, PV.W,
; EG-NEXT: SETGT_INT T3.W, 0.0, T0.X,
; EG-NEXT: MULLO_INT * T0.Z, PV.W, PS,
; EG-NEXT: ADD_INT T2.W, T0.X, PV.W,
; EG-NEXT: MULHI * T0.X, T0.Y, PS,
; EG-NEXT: ADD_INT T4.W, T0.Y, PS,
; EG-NEXT: XOR_INT * T2.W, PV.W, T3.W,
; EG-NEXT: MULHI * T0.X, PS, PV.W,
; EG-NEXT: MULLO_INT * T0.Y, PS, T1.W,
; EG-NEXT: SUB_INT * T2.W, T2.W, PS,
; EG-NEXT: ADD_INT T0.Z, T0.X, 1,
; EG-NEXT: SETGE_UINT T4.W, PV.W, T1.W,
; EG-NEXT: SUB_INT * T5.W, PV.W, T1.W,
; EG-NEXT: CNDE_INT T2.W, PV.W, T2.W, PS,
; EG-NEXT: CNDE_INT * T4.W, PV.W, T0.X, PV.Z,
; EG-NEXT: ADD_INT T5.W, PS, 1,
; EG-NEXT: SETGE_UINT * T1.W, PV.W, T1.W,
; EG-NEXT: CNDE_INT T1.W, PS, T4.W, PV.W, BS:VEC_102/SCL_221
; EG-NEXT: XOR_INT * T0.W, T3.W, T0.W,
; EG-NEXT: XOR_INT * T1.W, PV.W, PS,
; EG-NEXT: SUB_INT T0.X, PV.W, T0.W,
; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.x,
; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00)
%den_ptr = getelementptr i32, i32 addrspace(1)* %in, i32 1
%num = load i32, i32 addrspace(1) * %in
%den = load i32, i32 addrspace(1) * %den_ptr
%result = sdiv i32 %num, %den
store i32 %result, i32 addrspace(1)* %out
ret void
}
define amdgpu_kernel void @sdiv_i32_4(i32 addrspace(1)* %out, i32 addrspace(1)* %in) {
; GCN-LABEL: sdiv_i32_4:
; GCN: ; %bb.0:
; GCN-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s6, -1
; GCN-NEXT: s_mov_b32 s10, s6
; GCN-NEXT: s_mov_b32 s11, s7
; GCN-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s8, s2
; GCN-NEXT: s_mov_b32 s9, s3
; GCN-NEXT: buffer_load_dword v0, off, s[8:11], 0
; GCN-NEXT: s_mov_b32 s4, s0
; GCN-NEXT: s_mov_b32 s5, s1
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_ashrrev_i32_e32 v1, 31, v0
; GCN-NEXT: v_lshrrev_b32_e32 v1, 30, v1
; GCN-NEXT: v_add_i32_e32 v0, vcc, v1, v0
; GCN-NEXT: v_ashrrev_i32_e32 v0, 2, v0
; GCN-NEXT: buffer_store_dword v0, off, s[4:7], 0
; GCN-NEXT: s_endpgm
;
; TONGA-LABEL: sdiv_i32_4:
; TONGA: ; %bb.0:
; TONGA-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24
; TONGA-NEXT: s_mov_b32 s7, 0xf000
; TONGA-NEXT: s_mov_b32 s6, -1
; TONGA-NEXT: s_mov_b32 s10, s6
; TONGA-NEXT: s_mov_b32 s11, s7
; TONGA-NEXT: s_waitcnt lgkmcnt(0)
; TONGA-NEXT: s_mov_b32 s8, s2
; TONGA-NEXT: s_mov_b32 s9, s3
; TONGA-NEXT: buffer_load_dword v0, off, s[8:11], 0
; TONGA-NEXT: s_mov_b32 s4, s0
; TONGA-NEXT: s_mov_b32 s5, s1
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: v_ashrrev_i32_e32 v1, 31, v0
; TONGA-NEXT: v_lshrrev_b32_e32 v1, 30, v1
; TONGA-NEXT: v_add_u32_e32 v0, vcc, v1, v0
; TONGA-NEXT: v_ashrrev_i32_e32 v0, 2, v0
; TONGA-NEXT: buffer_store_dword v0, off, s[4:7], 0
; TONGA-NEXT: s_endpgm
;
; GFX9-LABEL: sdiv_i32_4:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24
; GFX9-NEXT: s_mov_b32 s7, 0xf000
; GFX9-NEXT: s_mov_b32 s6, -1
; GFX9-NEXT: s_mov_b32 s10, s6
; GFX9-NEXT: s_mov_b32 s11, s7
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_mov_b32 s8, s2
; GFX9-NEXT: s_mov_b32 s9, s3
; GFX9-NEXT: buffer_load_dword v0, off, s[8:11], 0
; GFX9-NEXT: s_mov_b32 s4, s0
; GFX9-NEXT: s_mov_b32 s5, s1
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_ashrrev_i32_e32 v1, 31, v0
; GFX9-NEXT: v_lshrrev_b32_e32 v1, 30, v1
; GFX9-NEXT: v_add_u32_e32 v0, v0, v1
; GFX9-NEXT: v_ashrrev_i32_e32 v0, 2, v0
; GFX9-NEXT: buffer_store_dword v0, off, s[4:7], 0
; GFX9-NEXT: s_endpgm
;
; EG-LABEL: sdiv_i32_4:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 0 @6
; EG-NEXT: ALU 7, @9, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.X, T1.X, 1
; EG-NEXT: CF_END
; EG-NEXT: PAD
; EG-NEXT: Fetch clause starting at 6:
; EG-NEXT: VTX_READ_32 T0.X, T0.X, 0, #1
; EG-NEXT: ALU clause starting at 8:
; EG-NEXT: MOV * T0.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 9:
; EG-NEXT: ASHR * T0.W, T0.X, literal.x,
; EG-NEXT: 31(4.344025e-44), 0(0.000000e+00)
; EG-NEXT: LSHR * T0.W, PV.W, literal.x,
; EG-NEXT: 30(4.203895e-44), 0(0.000000e+00)
; EG-NEXT: ADD_INT * T0.W, T0.X, PV.W,
; EG-NEXT: ASHR T0.X, PV.W, literal.x,
; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.x,
; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00)
%num = load i32, i32 addrspace(1) * %in
%result = sdiv i32 %num, 4
store i32 %result, i32 addrspace(1)* %out
ret void
}
; Multiply by a weird constant to make sure setIntDivIsCheap is
; working.
define amdgpu_kernel void @slow_sdiv_i32_3435(i32 addrspace(1)* %out, i32 addrspace(1)* %in) {
; GCN-LABEL: slow_sdiv_i32_3435:
; GCN: ; %bb.0:
; GCN-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s6, -1
; GCN-NEXT: s_mov_b32 s10, s6
; GCN-NEXT: s_mov_b32 s11, s7
; GCN-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s8, s2
; GCN-NEXT: s_mov_b32 s9, s3
; GCN-NEXT: buffer_load_dword v0, off, s[8:11], 0
; GCN-NEXT: s_mov_b32 s2, 0x98a1930b
; GCN-NEXT: s_mov_b32 s4, s0
; GCN-NEXT: s_mov_b32 s5, s1
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_hi_i32 v1, v0, s2
; GCN-NEXT: v_add_i32_e32 v0, vcc, v1, v0
; GCN-NEXT: v_lshrrev_b32_e32 v1, 31, v0
; GCN-NEXT: v_ashrrev_i32_e32 v0, 11, v0
; GCN-NEXT: v_add_i32_e32 v0, vcc, v1, v0
; GCN-NEXT: buffer_store_dword v0, off, s[4:7], 0
; GCN-NEXT: s_endpgm
;
; TONGA-LABEL: slow_sdiv_i32_3435:
; TONGA: ; %bb.0:
; TONGA-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24
; TONGA-NEXT: s_mov_b32 s7, 0xf000
; TONGA-NEXT: s_mov_b32 s6, -1
; TONGA-NEXT: s_mov_b32 s10, s6
; TONGA-NEXT: s_mov_b32 s11, s7
; TONGA-NEXT: s_waitcnt lgkmcnt(0)
; TONGA-NEXT: s_mov_b32 s8, s2
; TONGA-NEXT: s_mov_b32 s9, s3
; TONGA-NEXT: buffer_load_dword v0, off, s[8:11], 0
; TONGA-NEXT: s_mov_b32 s2, 0x98a1930b
; TONGA-NEXT: s_mov_b32 s4, s0
; TONGA-NEXT: s_mov_b32 s5, s1
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: v_mul_hi_i32 v1, v0, s2
; TONGA-NEXT: v_add_u32_e32 v0, vcc, v1, v0
; TONGA-NEXT: v_lshrrev_b32_e32 v1, 31, v0
; TONGA-NEXT: v_ashrrev_i32_e32 v0, 11, v0
; TONGA-NEXT: v_add_u32_e32 v0, vcc, v1, v0
; TONGA-NEXT: buffer_store_dword v0, off, s[4:7], 0
; TONGA-NEXT: s_endpgm
;
; GFX9-LABEL: slow_sdiv_i32_3435:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24
; GFX9-NEXT: s_mov_b32 s7, 0xf000
; GFX9-NEXT: s_mov_b32 s6, -1
; GFX9-NEXT: s_mov_b32 s10, s6
; GFX9-NEXT: s_mov_b32 s11, s7
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_mov_b32 s8, s2
; GFX9-NEXT: s_mov_b32 s9, s3
; GFX9-NEXT: buffer_load_dword v0, off, s[8:11], 0
; GFX9-NEXT: s_mov_b32 s2, 0x98a1930b
; GFX9-NEXT: s_mov_b32 s4, s0
; GFX9-NEXT: s_mov_b32 s5, s1
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_mul_hi_i32 v1, v0, s2
; GFX9-NEXT: v_add_u32_e32 v0, v1, v0
; GFX9-NEXT: v_lshrrev_b32_e32 v1, 31, v0
; GFX9-NEXT: v_ashrrev_i32_e32 v0, 11, v0
; GFX9-NEXT: v_add_u32_e32 v0, v0, v1
; GFX9-NEXT: buffer_store_dword v0, off, s[4:7], 0
; GFX9-NEXT: s_endpgm
;
; EG-LABEL: slow_sdiv_i32_3435:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 0 @6
; EG-NEXT: ALU 8, @9, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.X, T1.X, 1
; EG-NEXT: CF_END
; EG-NEXT: PAD
; EG-NEXT: Fetch clause starting at 6:
; EG-NEXT: VTX_READ_32 T0.X, T0.X, 0, #1
; EG-NEXT: ALU clause starting at 8:
; EG-NEXT: MOV * T0.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 9:
; EG-NEXT: MULHI_INT * T0.Y, T0.X, literal.x,
; EG-NEXT: -1734241525(-4.176600e-24), 0(0.000000e+00)
; EG-NEXT: ADD_INT * T0.W, PS, T0.X,
; EG-NEXT: ASHR T1.W, PV.W, literal.x,
; EG-NEXT: LSHR * T0.W, PV.W, literal.y,
; EG-NEXT: 11(1.541428e-44), 31(4.344025e-44)
; EG-NEXT: ADD_INT T0.X, PV.W, PS,
; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.x,
; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00)
%num = load i32, i32 addrspace(1) * %in
%result = sdiv i32 %num, 3435
store i32 %result, i32 addrspace(1)* %out
ret void
}
define amdgpu_kernel void @sdiv_v2i32(<2 x i32> addrspace(1)* %out, <2 x i32> addrspace(1)* %in) {
; GCN-LABEL: sdiv_v2i32:
; GCN: ; %bb.0:
; GCN-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s6, -1
; GCN-NEXT: s_mov_b32 s10, s6
; GCN-NEXT: s_mov_b32 s11, s7
; GCN-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s8, s2
; GCN-NEXT: s_mov_b32 s9, s3
; GCN-NEXT: buffer_load_dwordx4 v[0:3], off, s[8:11], 0
; GCN-NEXT: s_mov_b32 s4, s0
; GCN-NEXT: s_mov_b32 s5, s1
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_ashrrev_i32_e32 v5, 31, v2
; GCN-NEXT: v_ashrrev_i32_e32 v7, 31, v3
; GCN-NEXT: v_add_i32_e32 v2, vcc, v5, v2
; GCN-NEXT: v_add_i32_e32 v3, vcc, v7, v3
; GCN-NEXT: v_ashrrev_i32_e32 v4, 31, v0
; GCN-NEXT: v_ashrrev_i32_e32 v6, 31, v1
; GCN-NEXT: v_xor_b32_e32 v2, v2, v5
; GCN-NEXT: v_xor_b32_e32 v3, v3, v7
; GCN-NEXT: v_xor_b32_e32 v8, v4, v5
; GCN-NEXT: v_xor_b32_e32 v9, v6, v7
; GCN-NEXT: v_cvt_f32_u32_e32 v5, v2
; GCN-NEXT: v_cvt_f32_u32_e32 v7, v3
; GCN-NEXT: v_sub_i32_e32 v10, vcc, 0, v2
; GCN-NEXT: v_rcp_iflag_f32_e32 v5, v5
; GCN-NEXT: v_rcp_iflag_f32_e32 v7, v7
; GCN-NEXT: v_sub_i32_e32 v11, vcc, 0, v3
; GCN-NEXT: v_mul_f32_e32 v5, 0x4f7ffffe, v5
; GCN-NEXT: v_mul_f32_e32 v7, 0x4f7ffffe, v7
; GCN-NEXT: v_cvt_u32_f32_e32 v5, v5
; GCN-NEXT: v_cvt_u32_f32_e32 v7, v7
; GCN-NEXT: v_add_i32_e32 v0, vcc, v4, v0
; GCN-NEXT: v_mul_lo_u32 v10, v10, v5
; GCN-NEXT: v_mul_lo_u32 v11, v11, v7
; GCN-NEXT: v_add_i32_e32 v1, vcc, v6, v1
; GCN-NEXT: v_xor_b32_e32 v0, v0, v4
; GCN-NEXT: v_mul_hi_u32 v4, v5, v10
; GCN-NEXT: v_xor_b32_e32 v1, v1, v6
; GCN-NEXT: v_mul_hi_u32 v6, v7, v11
; GCN-NEXT: v_add_i32_e32 v4, vcc, v5, v4
; GCN-NEXT: v_add_i32_e32 v5, vcc, v6, v7
; GCN-NEXT: v_mul_hi_u32 v4, v0, v4
; GCN-NEXT: v_mul_hi_u32 v5, v1, v5
; GCN-NEXT: v_mul_lo_u32 v6, v4, v2
; GCN-NEXT: v_mul_lo_u32 v10, v5, v3
; GCN-NEXT: v_add_i32_e32 v7, vcc, 1, v4
; GCN-NEXT: v_sub_i32_e32 v0, vcc, v0, v6
; GCN-NEXT: v_sub_i32_e32 v1, vcc, v1, v10
; GCN-NEXT: v_add_i32_e32 v11, vcc, 1, v5
; GCN-NEXT: v_cmp_ge_u32_e64 s[0:1], v0, v2
; GCN-NEXT: v_cmp_ge_u32_e64 s[2:3], v1, v3
; GCN-NEXT: v_subrev_i32_e32 v6, vcc, v2, v0
; GCN-NEXT: v_cndmask_b32_e64 v4, v4, v7, s[0:1]
; GCN-NEXT: v_subrev_i32_e32 v7, vcc, v3, v1
; GCN-NEXT: v_cndmask_b32_e64 v5, v5, v11, s[2:3]
; GCN-NEXT: v_cndmask_b32_e64 v0, v0, v6, s[0:1]
; GCN-NEXT: v_add_i32_e32 v6, vcc, 1, v4
; GCN-NEXT: v_cndmask_b32_e64 v1, v1, v7, s[2:3]
; GCN-NEXT: v_add_i32_e32 v7, vcc, 1, v5
; GCN-NEXT: v_cmp_ge_u32_e32 vcc, v0, v2
; GCN-NEXT: v_cndmask_b32_e32 v0, v4, v6, vcc
; GCN-NEXT: v_cmp_ge_u32_e32 vcc, v1, v3
; GCN-NEXT: v_cndmask_b32_e32 v1, v5, v7, vcc
; GCN-NEXT: v_xor_b32_e32 v0, v0, v8
; GCN-NEXT: v_xor_b32_e32 v1, v1, v9
; GCN-NEXT: v_subrev_i32_e32 v0, vcc, v8, v0
; GCN-NEXT: v_subrev_i32_e32 v1, vcc, v9, v1
; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0
; GCN-NEXT: s_endpgm
;
; TONGA-LABEL: sdiv_v2i32:
; TONGA: ; %bb.0:
; TONGA-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24
; TONGA-NEXT: s_mov_b32 s7, 0xf000
; TONGA-NEXT: s_mov_b32 s6, -1
; TONGA-NEXT: s_mov_b32 s10, s6
; TONGA-NEXT: s_mov_b32 s11, s7
; TONGA-NEXT: s_waitcnt lgkmcnt(0)
; TONGA-NEXT: s_mov_b32 s8, s2
; TONGA-NEXT: s_mov_b32 s9, s3
; TONGA-NEXT: buffer_load_dwordx4 v[0:3], off, s[8:11], 0
; TONGA-NEXT: s_mov_b32 s4, s0
; TONGA-NEXT: s_mov_b32 s5, s1
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: v_ashrrev_i32_e32 v5, 31, v2
; TONGA-NEXT: v_ashrrev_i32_e32 v7, 31, v3
; TONGA-NEXT: v_add_u32_e32 v2, vcc, v5, v2
; TONGA-NEXT: v_add_u32_e32 v3, vcc, v7, v3
; TONGA-NEXT: v_ashrrev_i32_e32 v4, 31, v0
; TONGA-NEXT: v_ashrrev_i32_e32 v6, 31, v1
; TONGA-NEXT: v_xor_b32_e32 v2, v2, v5
; TONGA-NEXT: v_xor_b32_e32 v3, v3, v7
; TONGA-NEXT: v_xor_b32_e32 v8, v4, v5
; TONGA-NEXT: v_xor_b32_e32 v9, v6, v7
; TONGA-NEXT: v_cvt_f32_u32_e32 v5, v2
; TONGA-NEXT: v_cvt_f32_u32_e32 v7, v3
; TONGA-NEXT: v_sub_u32_e32 v10, vcc, 0, v2
; TONGA-NEXT: v_rcp_iflag_f32_e32 v5, v5
; TONGA-NEXT: v_rcp_iflag_f32_e32 v7, v7
; TONGA-NEXT: v_sub_u32_e32 v11, vcc, 0, v3
; TONGA-NEXT: v_mul_f32_e32 v5, 0x4f7ffffe, v5
; TONGA-NEXT: v_mul_f32_e32 v7, 0x4f7ffffe, v7
; TONGA-NEXT: v_cvt_u32_f32_e32 v5, v5
; TONGA-NEXT: v_cvt_u32_f32_e32 v7, v7
; TONGA-NEXT: v_add_u32_e32 v0, vcc, v4, v0
; TONGA-NEXT: v_mul_lo_u32 v10, v10, v5
; TONGA-NEXT: v_mul_lo_u32 v11, v11, v7
; TONGA-NEXT: v_add_u32_e32 v1, vcc, v6, v1
; TONGA-NEXT: v_xor_b32_e32 v0, v0, v4
; TONGA-NEXT: v_mul_hi_u32 v4, v5, v10
; TONGA-NEXT: v_xor_b32_e32 v1, v1, v6
; TONGA-NEXT: v_mul_hi_u32 v6, v7, v11
; TONGA-NEXT: v_add_u32_e32 v4, vcc, v5, v4
; TONGA-NEXT: v_add_u32_e32 v5, vcc, v6, v7
; TONGA-NEXT: v_mul_hi_u32 v4, v0, v4
; TONGA-NEXT: v_mul_hi_u32 v5, v1, v5
; TONGA-NEXT: v_mul_lo_u32 v6, v4, v2
; TONGA-NEXT: v_mul_lo_u32 v10, v5, v3
; TONGA-NEXT: v_add_u32_e32 v7, vcc, 1, v4
; TONGA-NEXT: v_sub_u32_e32 v0, vcc, v0, v6
; TONGA-NEXT: v_sub_u32_e32 v1, vcc, v1, v10
; TONGA-NEXT: v_add_u32_e32 v11, vcc, 1, v5
; TONGA-NEXT: v_cmp_ge_u32_e64 s[0:1], v0, v2
; TONGA-NEXT: v_cmp_ge_u32_e64 s[2:3], v1, v3
; TONGA-NEXT: v_subrev_u32_e32 v6, vcc, v2, v0
; TONGA-NEXT: v_cndmask_b32_e64 v4, v4, v7, s[0:1]
; TONGA-NEXT: v_subrev_u32_e32 v7, vcc, v3, v1
; TONGA-NEXT: v_cndmask_b32_e64 v5, v5, v11, s[2:3]
; TONGA-NEXT: v_cndmask_b32_e64 v0, v0, v6, s[0:1]
; TONGA-NEXT: v_add_u32_e32 v6, vcc, 1, v4
; TONGA-NEXT: v_cndmask_b32_e64 v1, v1, v7, s[2:3]
; TONGA-NEXT: v_add_u32_e32 v7, vcc, 1, v5
; TONGA-NEXT: v_cmp_ge_u32_e32 vcc, v0, v2
; TONGA-NEXT: v_cndmask_b32_e32 v0, v4, v6, vcc
; TONGA-NEXT: v_cmp_ge_u32_e32 vcc, v1, v3
; TONGA-NEXT: v_cndmask_b32_e32 v1, v5, v7, vcc
; TONGA-NEXT: v_xor_b32_e32 v0, v0, v8
; TONGA-NEXT: v_xor_b32_e32 v1, v1, v9
; TONGA-NEXT: v_subrev_u32_e32 v0, vcc, v8, v0
; TONGA-NEXT: v_subrev_u32_e32 v1, vcc, v9, v1
; TONGA-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0
; TONGA-NEXT: s_endpgm
;
; GFX9-LABEL: sdiv_v2i32:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX9-NEXT: s_mov_b32 s3, 0xf000
; GFX9-NEXT: s_mov_b32 s2, -1
; GFX9-NEXT: s_mov_b32 s10, s2
; GFX9-NEXT: s_mov_b32 s11, s3
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_mov_b32 s8, s6
; GFX9-NEXT: s_mov_b32 s9, s7
; GFX9-NEXT: buffer_load_dwordx4 v[0:3], off, s[8:11], 0
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_readfirstlane_b32 s0, v2
; GFX9-NEXT: s_ashr_i32 s1, s0, 31
; GFX9-NEXT: s_add_i32 s0, s0, s1
; GFX9-NEXT: s_xor_b32 s6, s0, s1
; GFX9-NEXT: v_cvt_f32_u32_e32 v2, s6
; GFX9-NEXT: v_readfirstlane_b32 s7, v0
; GFX9-NEXT: s_ashr_i32 s8, s7, 31
; GFX9-NEXT: s_add_i32 s7, s7, s8
; GFX9-NEXT: v_rcp_iflag_f32_e32 v2, v2
; GFX9-NEXT: s_xor_b32 s9, s8, s1
; GFX9-NEXT: s_xor_b32 s1, s7, s8
; GFX9-NEXT: s_sub_i32 s7, 0, s6
; GFX9-NEXT: v_mul_f32_e32 v0, 0x4f7ffffe, v2
; GFX9-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX9-NEXT: s_mov_b32 s0, s4
; GFX9-NEXT: v_readfirstlane_b32 s4, v3
; GFX9-NEXT: v_readfirstlane_b32 s8, v0
; GFX9-NEXT: s_mul_i32 s7, s7, s8
; GFX9-NEXT: s_mul_hi_u32 s7, s8, s7
; GFX9-NEXT: s_add_i32 s8, s8, s7
; GFX9-NEXT: s_mul_hi_u32 s7, s1, s8
; GFX9-NEXT: s_mul_i32 s8, s7, s6
; GFX9-NEXT: s_sub_i32 s1, s1, s8
; GFX9-NEXT: s_add_i32 s10, s7, 1
; GFX9-NEXT: s_sub_i32 s8, s1, s6
; GFX9-NEXT: s_cmp_ge_u32 s1, s6
; GFX9-NEXT: s_cselect_b32 s7, s10, s7
; GFX9-NEXT: s_cselect_b32 s1, s8, s1
; GFX9-NEXT: s_add_i32 s8, s7, 1
; GFX9-NEXT: s_cmp_ge_u32 s1, s6
; GFX9-NEXT: s_cselect_b32 s6, s8, s7
; GFX9-NEXT: s_ashr_i32 s7, s4, 31
; GFX9-NEXT: s_add_i32 s4, s4, s7
; GFX9-NEXT: s_xor_b32 s4, s4, s7
; GFX9-NEXT: v_cvt_f32_u32_e32 v0, s4
; GFX9-NEXT: s_mov_b32 s1, s5
; GFX9-NEXT: v_readfirstlane_b32 s5, v1
; GFX9-NEXT: s_ashr_i32 s8, s5, 31
; GFX9-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX9-NEXT: s_xor_b32 s6, s6, s9
; GFX9-NEXT: s_add_i32 s5, s5, s8
; GFX9-NEXT: s_xor_b32 s7, s8, s7
; GFX9-NEXT: v_mul_f32_e32 v0, 0x4f7ffffe, v0
; GFX9-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX9-NEXT: s_sub_i32 s6, s6, s9
; GFX9-NEXT: s_xor_b32 s5, s5, s8
; GFX9-NEXT: s_sub_i32 s8, 0, s4
; GFX9-NEXT: v_readfirstlane_b32 s9, v0
; GFX9-NEXT: s_mul_i32 s8, s8, s9
; GFX9-NEXT: s_mul_hi_u32 s8, s9, s8
; GFX9-NEXT: s_add_i32 s9, s9, s8
; GFX9-NEXT: s_mul_hi_u32 s8, s5, s9
; GFX9-NEXT: s_mul_i32 s9, s8, s4
; GFX9-NEXT: s_sub_i32 s5, s5, s9
; GFX9-NEXT: s_add_i32 s10, s8, 1
; GFX9-NEXT: s_sub_i32 s9, s5, s4
; GFX9-NEXT: s_cmp_ge_u32 s5, s4
; GFX9-NEXT: s_cselect_b32 s8, s10, s8
; GFX9-NEXT: s_cselect_b32 s5, s9, s5
; GFX9-NEXT: s_add_i32 s9, s8, 1
; GFX9-NEXT: s_cmp_ge_u32 s5, s4
; GFX9-NEXT: s_cselect_b32 s4, s9, s8
; GFX9-NEXT: s_xor_b32 s4, s4, s7
; GFX9-NEXT: s_sub_i32 s4, s4, s7
; GFX9-NEXT: v_mov_b32_e32 v0, s6
; GFX9-NEXT: v_mov_b32_e32 v1, s4
; GFX9-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0
; GFX9-NEXT: s_endpgm
;
; EG-LABEL: sdiv_v2i32:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 0 @6
; EG-NEXT: ALU 51, @9, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.XY, T1.X, 1
; EG-NEXT: CF_END
; EG-NEXT: PAD
; EG-NEXT: Fetch clause starting at 6:
; EG-NEXT: VTX_READ_128 T0.XYZW, T0.X, 0, #1
; EG-NEXT: ALU clause starting at 8:
; EG-NEXT: MOV * T0.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 9:
; EG-NEXT: SETGT_INT * T1.W, 0.0, T0.W,
; EG-NEXT: ADD_INT T0.W, T0.W, PV.W,
; EG-NEXT: SETGT_INT * T2.W, 0.0, T0.Z,
; EG-NEXT: XOR_INT * T0.W, PV.W, T1.W,
; EG-NEXT: SUB_INT T1.Z, 0.0, PV.W,
; EG-NEXT: ADD_INT T3.W, T0.Z, T2.W,
; EG-NEXT: RECIP_UINT * T0.Z, PV.W,
; EG-NEXT: XOR_INT T3.W, PV.W, T2.W,
; EG-NEXT: MULLO_INT * T1.X, PV.Z, PS,
; EG-NEXT: SUB_INT T4.W, 0.0, PV.W,
; EG-NEXT: RECIP_UINT * T1.Y, PV.W,
; EG-NEXT: SETGT_INT T5.W, 0.0, T0.X,
; EG-NEXT: MULLO_INT * T1.Z, PV.W, PS,
; EG-NEXT: SETGT_INT T2.Z, 0.0, T0.Y,
; EG-NEXT: ADD_INT T4.W, T0.X, PV.W,
; EG-NEXT: MULHI * T0.X, T1.Y, PS,
; EG-NEXT: ADD_INT T1.Y, T1.Y, PS,
; EG-NEXT: XOR_INT T1.Z, PV.W, T5.W,
; EG-NEXT: ADD_INT T4.W, T0.Y, PV.Z, BS:VEC_120/SCL_212
; EG-NEXT: MULHI * T0.X, T0.Z, T1.X,
; EG-NEXT: ADD_INT T0.Z, T0.Z, PS,
; EG-NEXT: XOR_INT T4.W, PV.W, T2.Z,
; EG-NEXT: MULHI * T0.X, PV.Z, PV.Y,
; EG-NEXT: MULHI * T0.Y, PV.W, PV.Z,
; EG-NEXT: MULLO_INT * T0.Z, PS, T0.W,
; EG-NEXT: SUB_INT T4.W, T4.W, PS,
; EG-NEXT: MULLO_INT * T0.Z, T0.X, T3.W,
; EG-NEXT: SUB_INT T1.Y, T1.Z, PS,
; EG-NEXT: ADD_INT T0.Z, T0.Y, 1,
; EG-NEXT: SETGE_UINT T6.W, PV.W, T0.W,
; EG-NEXT: SUB_INT * T7.W, PV.W, T0.W,
; EG-NEXT: CNDE_INT T1.X, PV.W, T4.W, PS, BS:VEC_021/SCL_122
; EG-NEXT: CNDE_INT T0.Y, PV.W, T0.Y, PV.Z,
; EG-NEXT: ADD_INT T0.Z, T0.X, 1,
; EG-NEXT: SETGE_UINT T4.W, PV.Y, T3.W,
; EG-NEXT: SUB_INT * T6.W, PV.Y, T3.W,
; EG-NEXT: CNDE_INT T1.Y, PV.W, T1.Y, PS,
; EG-NEXT: CNDE_INT T0.Z, PV.W, T0.X, PV.Z,
; EG-NEXT: ADD_INT T4.W, PV.Y, 1,
; EG-NEXT: SETGE_UINT * T0.W, PV.X, T0.W,
; EG-NEXT: CNDE_INT T0.Y, PS, T0.Y, PV.W,
; EG-NEXT: XOR_INT T1.Z, T2.Z, T1.W, BS:VEC_021/SCL_122
; EG-NEXT: ADD_INT T0.W, PV.Z, 1,
; EG-NEXT: SETGE_UINT * T1.W, PV.Y, T3.W,
; EG-NEXT: CNDE_INT T0.Z, PS, T0.Z, PV.W,
; EG-NEXT: XOR_INT T0.W, T5.W, T2.W,
; EG-NEXT: XOR_INT * T1.W, PV.Y, PV.Z,
; EG-NEXT: SUB_INT T0.Y, PS, T1.Z,
; EG-NEXT: XOR_INT * T1.W, PV.Z, PV.W,
; EG-NEXT: SUB_INT T0.X, PV.W, T0.W,
; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.x,
; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00)
%den_ptr = getelementptr <2 x i32>, <2 x i32> addrspace(1)* %in, i32 1
%num = load <2 x i32>, <2 x i32> addrspace(1) * %in
%den = load <2 x i32>, <2 x i32> addrspace(1) * %den_ptr
%result = sdiv <2 x i32> %num, %den
store <2 x i32> %result, <2 x i32> addrspace(1)* %out
ret void
}
define amdgpu_kernel void @sdiv_v2i32_4(<2 x i32> addrspace(1)* %out, <2 x i32> addrspace(1)* %in) {
; GCN-LABEL: sdiv_v2i32_4:
; GCN: ; %bb.0:
; GCN-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s6, -1
; GCN-NEXT: s_mov_b32 s10, s6
; GCN-NEXT: s_mov_b32 s11, s7
; GCN-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s8, s2
; GCN-NEXT: s_mov_b32 s9, s3
; GCN-NEXT: buffer_load_dwordx2 v[0:1], off, s[8:11], 0
; GCN-NEXT: s_mov_b32 s4, s0
; GCN-NEXT: s_mov_b32 s5, s1
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_ashrrev_i32_e32 v2, 31, v0
; GCN-NEXT: v_ashrrev_i32_e32 v3, 31, v1
; GCN-NEXT: v_lshrrev_b32_e32 v2, 30, v2
; GCN-NEXT: v_lshrrev_b32_e32 v3, 30, v3
; GCN-NEXT: v_add_i32_e32 v0, vcc, v2, v0
; GCN-NEXT: v_add_i32_e32 v1, vcc, v3, v1
; GCN-NEXT: v_ashrrev_i32_e32 v0, 2, v0
; GCN-NEXT: v_ashrrev_i32_e32 v1, 2, v1
; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0
; GCN-NEXT: s_endpgm
;
; TONGA-LABEL: sdiv_v2i32_4:
; TONGA: ; %bb.0:
; TONGA-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24
; TONGA-NEXT: s_mov_b32 s7, 0xf000
; TONGA-NEXT: s_mov_b32 s6, -1
; TONGA-NEXT: s_mov_b32 s10, s6
; TONGA-NEXT: s_mov_b32 s11, s7
; TONGA-NEXT: s_waitcnt lgkmcnt(0)
; TONGA-NEXT: s_mov_b32 s8, s2
; TONGA-NEXT: s_mov_b32 s9, s3
; TONGA-NEXT: buffer_load_dwordx2 v[0:1], off, s[8:11], 0
; TONGA-NEXT: s_mov_b32 s4, s0
; TONGA-NEXT: s_mov_b32 s5, s1
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: v_ashrrev_i32_e32 v2, 31, v0
; TONGA-NEXT: v_ashrrev_i32_e32 v3, 31, v1
; TONGA-NEXT: v_lshrrev_b32_e32 v2, 30, v2
; TONGA-NEXT: v_lshrrev_b32_e32 v3, 30, v3
; TONGA-NEXT: v_add_u32_e32 v0, vcc, v2, v0
; TONGA-NEXT: v_add_u32_e32 v1, vcc, v3, v1
; TONGA-NEXT: v_ashrrev_i32_e32 v0, 2, v0
; TONGA-NEXT: v_ashrrev_i32_e32 v1, 2, v1
; TONGA-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0
; TONGA-NEXT: s_endpgm
;
; GFX9-LABEL: sdiv_v2i32_4:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24
; GFX9-NEXT: s_mov_b32 s7, 0xf000
; GFX9-NEXT: s_mov_b32 s6, -1
; GFX9-NEXT: s_mov_b32 s10, s6
; GFX9-NEXT: s_mov_b32 s11, s7
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_mov_b32 s8, s2
; GFX9-NEXT: s_mov_b32 s9, s3
; GFX9-NEXT: buffer_load_dwordx2 v[0:1], off, s[8:11], 0
; GFX9-NEXT: s_mov_b32 s4, s0
; GFX9-NEXT: s_mov_b32 s5, s1
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_ashrrev_i32_e32 v2, 31, v0
; GFX9-NEXT: v_ashrrev_i32_e32 v3, 31, v1
; GFX9-NEXT: v_lshrrev_b32_e32 v2, 30, v2
; GFX9-NEXT: v_lshrrev_b32_e32 v3, 30, v3
; GFX9-NEXT: v_add_u32_e32 v0, v0, v2
; GFX9-NEXT: v_add_u32_e32 v1, v1, v3
; GFX9-NEXT: v_ashrrev_i32_e32 v0, 2, v0
; GFX9-NEXT: v_ashrrev_i32_e32 v1, 2, v1
; GFX9-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0
; GFX9-NEXT: s_endpgm
;
; EG-LABEL: sdiv_v2i32_4:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 0 @6
; EG-NEXT: ALU 13, @9, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.XY, T1.X, 1
; EG-NEXT: CF_END
; EG-NEXT: PAD
; EG-NEXT: Fetch clause starting at 6:
; EG-NEXT: VTX_READ_64 T0.XY, T0.X, 0, #1
; EG-NEXT: ALU clause starting at 8:
; EG-NEXT: MOV * T0.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 9:
; EG-NEXT: ASHR * T0.W, T0.Y, literal.x,
; EG-NEXT: 31(4.344025e-44), 0(0.000000e+00)
; EG-NEXT: LSHR T0.W, PV.W, literal.x,
; EG-NEXT: ASHR * T1.W, T0.X, literal.y,
; EG-NEXT: 30(4.203895e-44), 31(4.344025e-44)
; EG-NEXT: LSHR T1.W, PS, literal.x,
; EG-NEXT: ADD_INT * T0.W, T0.Y, PV.W,
; EG-NEXT: 30(4.203895e-44), 0(0.000000e+00)
; EG-NEXT: ASHR T0.Y, PS, literal.x,
; EG-NEXT: ADD_INT * T0.W, T0.X, PV.W,
; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00)
; EG-NEXT: ASHR T0.X, PV.W, literal.x,
; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.x,
; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00)
%num = load <2 x i32>, <2 x i32> addrspace(1) * %in
%result = sdiv <2 x i32> %num, <i32 4, i32 4>
store <2 x i32> %result, <2 x i32> addrspace(1)* %out
ret void
}
define amdgpu_kernel void @sdiv_v4i32(<4 x i32> addrspace(1)* %out, <4 x i32> addrspace(1)* %in) {
; GCN-LABEL: sdiv_v4i32:
; GCN: ; %bb.0:
; GCN-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
; GCN-NEXT: s_mov_b32 s11, 0xf000
; GCN-NEXT: s_mov_b32 s10, -1
; GCN-NEXT: s_mov_b32 s6, s10
; GCN-NEXT: s_mov_b32 s7, s11
; GCN-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s4, s2
; GCN-NEXT: s_mov_b32 s5, s3
; GCN-NEXT: buffer_load_dwordx4 v[0:3], off, s[4:7], 0
; GCN-NEXT: buffer_load_dwordx4 v[4:7], off, s[4:7], 0 offset:16
; GCN-NEXT: s_mov_b32 s8, s0
; GCN-NEXT: s_mov_b32 s9, s1
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_ashrrev_i32_e32 v8, 31, v0
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_ashrrev_i32_e32 v9, 31, v4
; GCN-NEXT: v_add_i32_e32 v4, vcc, v9, v4
; GCN-NEXT: v_add_i32_e32 v0, vcc, v8, v0
; GCN-NEXT: v_xor_b32_e32 v4, v4, v9
; GCN-NEXT: v_xor_b32_e32 v15, v8, v9
; GCN-NEXT: v_xor_b32_e32 v0, v0, v8
; GCN-NEXT: v_cvt_f32_u32_e32 v8, v4
; GCN-NEXT: v_sub_i32_e32 v9, vcc, 0, v4
; GCN-NEXT: v_ashrrev_i32_e32 v11, 31, v5
; GCN-NEXT: v_rcp_iflag_f32_e32 v8, v8
; GCN-NEXT: v_ashrrev_i32_e32 v13, 31, v6
; GCN-NEXT: v_ashrrev_i32_e32 v10, 31, v1
; GCN-NEXT: v_add_i32_e32 v5, vcc, v11, v5
; GCN-NEXT: v_mul_f32_e32 v8, 0x4f7ffffe, v8
; GCN-NEXT: v_cvt_u32_f32_e32 v8, v8
; GCN-NEXT: v_add_i32_e32 v6, vcc, v13, v6
; GCN-NEXT: v_add_i32_e32 v1, vcc, v10, v1
; GCN-NEXT: v_mul_lo_u32 v9, v9, v8
; GCN-NEXT: v_xor_b32_e32 v5, v5, v11
; GCN-NEXT: v_xor_b32_e32 v6, v6, v13
; GCN-NEXT: v_xor_b32_e32 v16, v10, v11
; GCN-NEXT: v_mul_hi_u32 v9, v8, v9
; GCN-NEXT: v_xor_b32_e32 v1, v1, v10
; GCN-NEXT: v_cvt_f32_u32_e32 v10, v5
; GCN-NEXT: v_cvt_f32_u32_e32 v11, v6
; GCN-NEXT: v_add_i32_e32 v8, vcc, v8, v9
; GCN-NEXT: v_rcp_iflag_f32_e32 v10, v10
; GCN-NEXT: v_rcp_iflag_f32_e32 v11, v11
; GCN-NEXT: v_mul_hi_u32 v8, v0, v8
; GCN-NEXT: v_ashrrev_i32_e32 v12, 31, v2
; GCN-NEXT: v_mul_f32_e32 v9, 0x4f7ffffe, v10
; GCN-NEXT: v_mul_f32_e32 v10, 0x4f7ffffe, v11
; GCN-NEXT: v_mul_lo_u32 v11, v8, v4
; GCN-NEXT: v_cvt_u32_f32_e32 v9, v9
; GCN-NEXT: v_add_i32_e32 v2, vcc, v12, v2
; GCN-NEXT: v_xor_b32_e32 v17, v12, v13
; GCN-NEXT: v_xor_b32_e32 v2, v2, v12
; GCN-NEXT: v_sub_i32_e32 v12, vcc, 0, v5
; GCN-NEXT: v_sub_i32_e32 v0, vcc, v0, v11
; GCN-NEXT: v_cvt_u32_f32_e32 v10, v10
; GCN-NEXT: v_mul_lo_u32 v12, v12, v9
; GCN-NEXT: v_add_i32_e32 v11, vcc, 1, v8
; GCN-NEXT: v_cmp_ge_u32_e64 s[0:1], v0, v4
; GCN-NEXT: v_cndmask_b32_e64 v8, v8, v11, s[0:1]
; GCN-NEXT: v_subrev_i32_e32 v11, vcc, v4, v0
; GCN-NEXT: v_ashrrev_i32_e32 v14, 31, v7
; GCN-NEXT: v_cndmask_b32_e64 v0, v0, v11, s[0:1]
; GCN-NEXT: v_add_i32_e32 v7, vcc, v14, v7
; GCN-NEXT: v_cmp_ge_u32_e64 s[0:1], v0, v4
; GCN-NEXT: v_sub_i32_e32 v0, vcc, 0, v6
; GCN-NEXT: v_mul_lo_u32 v0, v0, v10
; GCN-NEXT: v_xor_b32_e32 v4, v7, v14
; GCN-NEXT: v_mul_hi_u32 v7, v9, v12
; GCN-NEXT: v_cvt_f32_u32_e32 v12, v4
; GCN-NEXT: v_mul_hi_u32 v0, v10, v0
; GCN-NEXT: v_add_i32_e32 v11, vcc, 1, v8
; GCN-NEXT: v_add_i32_e32 v7, vcc, v7, v9
; GCN-NEXT: v_mul_hi_u32 v7, v1, v7
; GCN-NEXT: v_add_i32_e32 v0, vcc, v10, v0
; GCN-NEXT: v_mul_hi_u32 v0, v2, v0
; GCN-NEXT: v_mul_lo_u32 v10, v7, v5
; GCN-NEXT: v_rcp_iflag_f32_e32 v12, v12
; GCN-NEXT: v_sub_i32_e32 v9, vcc, 0, v4
; GCN-NEXT: v_sub_i32_e32 v1, vcc, v1, v10
; GCN-NEXT: v_mul_lo_u32 v10, v0, v6
; GCN-NEXT: v_cmp_ge_u32_e64 s[2:3], v1, v5
; GCN-NEXT: v_mul_f32_e32 v12, 0x4f7ffffe, v12
; GCN-NEXT: v_cvt_u32_f32_e32 v12, v12
; GCN-NEXT: v_sub_i32_e32 v2, vcc, v2, v10
; GCN-NEXT: v_add_i32_e32 v10, vcc, 1, v7
; GCN-NEXT: v_cndmask_b32_e64 v7, v7, v10, s[2:3]
; GCN-NEXT: v_add_i32_e32 v10, vcc, 1, v0
; GCN-NEXT: v_cmp_ge_u32_e64 s[4:5], v2, v6
; GCN-NEXT: v_cndmask_b32_e64 v10, v0, v10, s[4:5]
; GCN-NEXT: v_subrev_i32_e32 v0, vcc, v5, v1
; GCN-NEXT: v_cndmask_b32_e64 v0, v1, v0, s[2:3]
; GCN-NEXT: v_subrev_i32_e32 v1, vcc, v6, v2
; GCN-NEXT: v_cndmask_b32_e64 v2, v2, v1, s[4:5]
; GCN-NEXT: v_add_i32_e32 v1, vcc, 1, v7
; GCN-NEXT: v_cmp_ge_u32_e32 vcc, v0, v5
; GCN-NEXT: v_cndmask_b32_e64 v8, v8, v11, s[0:1]
; GCN-NEXT: v_cndmask_b32_e32 v0, v7, v1, vcc
; GCN-NEXT: v_xor_b32_e32 v1, v8, v15
; GCN-NEXT: v_xor_b32_e32 v5, v0, v16
; GCN-NEXT: v_subrev_i32_e32 v0, vcc, v15, v1
; GCN-NEXT: v_subrev_i32_e32 v1, vcc, v16, v5
; GCN-NEXT: v_mul_lo_u32 v5, v9, v12
; GCN-NEXT: v_ashrrev_i32_e32 v8, 31, v3
; GCN-NEXT: v_add_i32_e32 v3, vcc, v8, v3
; GCN-NEXT: v_mul_hi_u32 v5, v12, v5
; GCN-NEXT: v_xor_b32_e32 v3, v3, v8
; GCN-NEXT: v_add_i32_e32 v7, vcc, 1, v10
; GCN-NEXT: v_add_i32_e32 v5, vcc, v5, v12
; GCN-NEXT: v_mul_hi_u32 v5, v3, v5
; GCN-NEXT: v_cmp_ge_u32_e32 vcc, v2, v6
; GCN-NEXT: v_cndmask_b32_e32 v2, v10, v7, vcc
; GCN-NEXT: v_xor_b32_e32 v2, v2, v17
; GCN-NEXT: v_mul_lo_u32 v6, v5, v4
; GCN-NEXT: v_subrev_i32_e32 v2, vcc, v17, v2
; GCN-NEXT: v_xor_b32_e32 v7, v8, v14
; GCN-NEXT: v_sub_i32_e32 v3, vcc, v3, v6
; GCN-NEXT: v_add_i32_e32 v6, vcc, 1, v5
; GCN-NEXT: v_subrev_i32_e32 v8, vcc, v4, v3
; GCN-NEXT: v_cmp_ge_u32_e32 vcc, v3, v4
; GCN-NEXT: v_cndmask_b32_e32 v5, v5, v6, vcc
; GCN-NEXT: v_cndmask_b32_e32 v3, v3, v8, vcc
; GCN-NEXT: v_add_i32_e32 v6, vcc, 1, v5
; GCN-NEXT: v_cmp_ge_u32_e32 vcc, v3, v4
; GCN-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
; GCN-NEXT: v_xor_b32_e32 v3, v3, v7
; GCN-NEXT: v_subrev_i32_e32 v3, vcc, v7, v3
; GCN-NEXT: buffer_store_dwordx4 v[0:3], off, s[8:11], 0
; GCN-NEXT: s_endpgm
;
; TONGA-LABEL: sdiv_v4i32:
; TONGA: ; %bb.0:
; TONGA-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24
; TONGA-NEXT: s_mov_b32 s11, 0xf000
; TONGA-NEXT: s_mov_b32 s10, -1
; TONGA-NEXT: s_mov_b32 s6, s10
; TONGA-NEXT: s_mov_b32 s7, s11
; TONGA-NEXT: s_waitcnt lgkmcnt(0)
; TONGA-NEXT: s_mov_b32 s4, s2
; TONGA-NEXT: s_mov_b32 s5, s3
; TONGA-NEXT: buffer_load_dwordx4 v[0:3], off, s[4:7], 0
; TONGA-NEXT: buffer_load_dwordx4 v[4:7], off, s[4:7], 0 offset:16
; TONGA-NEXT: s_mov_b32 s8, s0
; TONGA-NEXT: s_mov_b32 s9, s1
; TONGA-NEXT: s_waitcnt vmcnt(1)
; TONGA-NEXT: v_ashrrev_i32_e32 v8, 31, v0
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: v_ashrrev_i32_e32 v9, 31, v4
; TONGA-NEXT: v_add_u32_e32 v4, vcc, v9, v4
; TONGA-NEXT: v_add_u32_e32 v0, vcc, v8, v0
; TONGA-NEXT: v_xor_b32_e32 v4, v4, v9
; TONGA-NEXT: v_xor_b32_e32 v15, v8, v9
; TONGA-NEXT: v_xor_b32_e32 v0, v0, v8
; TONGA-NEXT: v_cvt_f32_u32_e32 v8, v4
; TONGA-NEXT: v_sub_u32_e32 v9, vcc, 0, v4
; TONGA-NEXT: v_ashrrev_i32_e32 v11, 31, v5
; TONGA-NEXT: v_rcp_iflag_f32_e32 v8, v8
; TONGA-NEXT: v_ashrrev_i32_e32 v13, 31, v6
; TONGA-NEXT: v_ashrrev_i32_e32 v10, 31, v1
; TONGA-NEXT: v_add_u32_e32 v5, vcc, v11, v5
; TONGA-NEXT: v_mul_f32_e32 v8, 0x4f7ffffe, v8
; TONGA-NEXT: v_cvt_u32_f32_e32 v8, v8
; TONGA-NEXT: v_add_u32_e32 v6, vcc, v13, v6
; TONGA-NEXT: v_add_u32_e32 v1, vcc, v10, v1
; TONGA-NEXT: v_mul_lo_u32 v9, v9, v8
; TONGA-NEXT: v_xor_b32_e32 v5, v5, v11
; TONGA-NEXT: v_xor_b32_e32 v6, v6, v13
; TONGA-NEXT: v_xor_b32_e32 v16, v10, v11
; TONGA-NEXT: v_mul_hi_u32 v9, v8, v9
; TONGA-NEXT: v_xor_b32_e32 v1, v1, v10
; TONGA-NEXT: v_cvt_f32_u32_e32 v10, v5
; TONGA-NEXT: v_cvt_f32_u32_e32 v11, v6
; TONGA-NEXT: v_add_u32_e32 v8, vcc, v8, v9
; TONGA-NEXT: v_rcp_iflag_f32_e32 v10, v10
; TONGA-NEXT: v_rcp_iflag_f32_e32 v11, v11
; TONGA-NEXT: v_mul_hi_u32 v8, v0, v8
; TONGA-NEXT: v_ashrrev_i32_e32 v12, 31, v2
; TONGA-NEXT: v_mul_f32_e32 v9, 0x4f7ffffe, v10
; TONGA-NEXT: v_mul_f32_e32 v10, 0x4f7ffffe, v11
; TONGA-NEXT: v_mul_lo_u32 v11, v8, v4
; TONGA-NEXT: v_cvt_u32_f32_e32 v9, v9
; TONGA-NEXT: v_add_u32_e32 v2, vcc, v12, v2
; TONGA-NEXT: v_xor_b32_e32 v17, v12, v13
; TONGA-NEXT: v_xor_b32_e32 v2, v2, v12
; TONGA-NEXT: v_sub_u32_e32 v12, vcc, 0, v5
; TONGA-NEXT: v_sub_u32_e32 v0, vcc, v0, v11
; TONGA-NEXT: v_cvt_u32_f32_e32 v10, v10
; TONGA-NEXT: v_mul_lo_u32 v12, v12, v9
; TONGA-NEXT: v_add_u32_e32 v11, vcc, 1, v8
; TONGA-NEXT: v_cmp_ge_u32_e64 s[0:1], v0, v4
; TONGA-NEXT: v_cndmask_b32_e64 v8, v8, v11, s[0:1]
; TONGA-NEXT: v_subrev_u32_e32 v11, vcc, v4, v0
; TONGA-NEXT: v_ashrrev_i32_e32 v14, 31, v7
; TONGA-NEXT: v_cndmask_b32_e64 v0, v0, v11, s[0:1]
; TONGA-NEXT: v_add_u32_e32 v7, vcc, v14, v7
; TONGA-NEXT: v_cmp_ge_u32_e64 s[0:1], v0, v4
; TONGA-NEXT: v_sub_u32_e32 v0, vcc, 0, v6
; TONGA-NEXT: v_mul_lo_u32 v0, v0, v10
; TONGA-NEXT: v_xor_b32_e32 v4, v7, v14
; TONGA-NEXT: v_mul_hi_u32 v7, v9, v12
; TONGA-NEXT: v_cvt_f32_u32_e32 v12, v4
; TONGA-NEXT: v_mul_hi_u32 v0, v10, v0
; TONGA-NEXT: v_add_u32_e32 v11, vcc, 1, v8
; TONGA-NEXT: v_add_u32_e32 v7, vcc, v7, v9
; TONGA-NEXT: v_mul_hi_u32 v7, v1, v7
; TONGA-NEXT: v_add_u32_e32 v0, vcc, v10, v0
; TONGA-NEXT: v_mul_hi_u32 v0, v2, v0
; TONGA-NEXT: v_mul_lo_u32 v10, v7, v5
; TONGA-NEXT: v_rcp_iflag_f32_e32 v12, v12
; TONGA-NEXT: v_sub_u32_e32 v9, vcc, 0, v4
; TONGA-NEXT: v_sub_u32_e32 v1, vcc, v1, v10
; TONGA-NEXT: v_mul_lo_u32 v10, v0, v6
; TONGA-NEXT: v_cmp_ge_u32_e64 s[2:3], v1, v5
; TONGA-NEXT: v_mul_f32_e32 v12, 0x4f7ffffe, v12
; TONGA-NEXT: v_cvt_u32_f32_e32 v12, v12
; TONGA-NEXT: v_sub_u32_e32 v2, vcc, v2, v10
; TONGA-NEXT: v_add_u32_e32 v10, vcc, 1, v7
; TONGA-NEXT: v_cndmask_b32_e64 v7, v7, v10, s[2:3]
; TONGA-NEXT: v_add_u32_e32 v10, vcc, 1, v0
; TONGA-NEXT: v_cmp_ge_u32_e64 s[4:5], v2, v6
; TONGA-NEXT: v_cndmask_b32_e64 v10, v0, v10, s[4:5]
; TONGA-NEXT: v_subrev_u32_e32 v0, vcc, v5, v1
; TONGA-NEXT: v_cndmask_b32_e64 v0, v1, v0, s[2:3]
; TONGA-NEXT: v_subrev_u32_e32 v1, vcc, v6, v2
; TONGA-NEXT: v_cndmask_b32_e64 v2, v2, v1, s[4:5]
; TONGA-NEXT: v_add_u32_e32 v1, vcc, 1, v7
; TONGA-NEXT: v_cmp_ge_u32_e32 vcc, v0, v5
; TONGA-NEXT: v_cndmask_b32_e64 v8, v8, v11, s[0:1]
; TONGA-NEXT: v_cndmask_b32_e32 v0, v7, v1, vcc
; TONGA-NEXT: v_xor_b32_e32 v1, v8, v15
; TONGA-NEXT: v_xor_b32_e32 v5, v0, v16
; TONGA-NEXT: v_subrev_u32_e32 v0, vcc, v15, v1
; TONGA-NEXT: v_subrev_u32_e32 v1, vcc, v16, v5
; TONGA-NEXT: v_mul_lo_u32 v5, v9, v12
; TONGA-NEXT: v_ashrrev_i32_e32 v8, 31, v3
; TONGA-NEXT: v_add_u32_e32 v3, vcc, v8, v3
; TONGA-NEXT: v_mul_hi_u32 v5, v12, v5
; TONGA-NEXT: v_xor_b32_e32 v3, v3, v8
; TONGA-NEXT: v_add_u32_e32 v7, vcc, 1, v10
; TONGA-NEXT: v_add_u32_e32 v5, vcc, v5, v12
; TONGA-NEXT: v_mul_hi_u32 v5, v3, v5
; TONGA-NEXT: v_cmp_ge_u32_e32 vcc, v2, v6
; TONGA-NEXT: v_cndmask_b32_e32 v2, v10, v7, vcc
; TONGA-NEXT: v_xor_b32_e32 v2, v2, v17
; TONGA-NEXT: v_mul_lo_u32 v6, v5, v4
; TONGA-NEXT: v_subrev_u32_e32 v2, vcc, v17, v2
; TONGA-NEXT: v_xor_b32_e32 v7, v8, v14
; TONGA-NEXT: v_sub_u32_e32 v3, vcc, v3, v6
; TONGA-NEXT: v_add_u32_e32 v6, vcc, 1, v5
; TONGA-NEXT: v_subrev_u32_e32 v8, vcc, v4, v3
; TONGA-NEXT: v_cmp_ge_u32_e32 vcc, v3, v4
; TONGA-NEXT: v_cndmask_b32_e32 v5, v5, v6, vcc
; TONGA-NEXT: v_cndmask_b32_e32 v3, v3, v8, vcc
; TONGA-NEXT: v_add_u32_e32 v6, vcc, 1, v5
; TONGA-NEXT: v_cmp_ge_u32_e32 vcc, v3, v4
; TONGA-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
; TONGA-NEXT: v_xor_b32_e32 v3, v3, v7
; TONGA-NEXT: v_subrev_u32_e32 v3, vcc, v7, v3
; TONGA-NEXT: buffer_store_dwordx4 v[0:3], off, s[8:11], 0
; TONGA-NEXT: s_endpgm
;
; GFX9-LABEL: sdiv_v4i32:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX9-NEXT: s_mov_b32 s3, 0xf000
; GFX9-NEXT: s_mov_b32 s2, -1
; GFX9-NEXT: s_mov_b32 s10, s2
; GFX9-NEXT: s_mov_b32 s11, s3
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_mov_b32 s8, s6
; GFX9-NEXT: s_mov_b32 s9, s7
; GFX9-NEXT: buffer_load_dwordx4 v[0:3], off, s[8:11], 0 offset:16
; GFX9-NEXT: buffer_load_dwordx4 v[4:7], off, s[8:11], 0
; GFX9-NEXT: s_mov_b32 s0, s4
; GFX9-NEXT: s_waitcnt vmcnt(1)
; GFX9-NEXT: v_readfirstlane_b32 s1, v0
; GFX9-NEXT: s_ashr_i32 s4, s1, 31
; GFX9-NEXT: s_add_i32 s1, s1, s4
; GFX9-NEXT: s_xor_b32 s6, s1, s4
; GFX9-NEXT: v_cvt_f32_u32_e32 v0, s6
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_readfirstlane_b32 s8, v4
; GFX9-NEXT: s_ashr_i32 s9, s8, 31
; GFX9-NEXT: s_add_i32 s8, s8, s9
; GFX9-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX9-NEXT: s_xor_b32 s4, s9, s4
; GFX9-NEXT: s_xor_b32 s8, s8, s9
; GFX9-NEXT: s_sub_i32 s9, 0, s6
; GFX9-NEXT: v_mul_f32_e32 v0, 0x4f7ffffe, v0
; GFX9-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX9-NEXT: v_readfirstlane_b32 s7, v1
; GFX9-NEXT: s_mov_b32 s1, s5
; GFX9-NEXT: v_readfirstlane_b32 s5, v3
; GFX9-NEXT: v_readfirstlane_b32 s10, v0
; GFX9-NEXT: s_mul_i32 s9, s9, s10
; GFX9-NEXT: s_mul_hi_u32 s9, s10, s9
; GFX9-NEXT: s_add_i32 s10, s10, s9
; GFX9-NEXT: s_mul_hi_u32 s9, s8, s10
; GFX9-NEXT: s_mul_i32 s10, s9, s6
; GFX9-NEXT: s_sub_i32 s8, s8, s10
; GFX9-NEXT: s_add_i32 s11, s9, 1
; GFX9-NEXT: s_sub_i32 s10, s8, s6
; GFX9-NEXT: s_cmp_ge_u32 s8, s6
; GFX9-NEXT: s_cselect_b32 s9, s11, s9
; GFX9-NEXT: s_cselect_b32 s8, s10, s8
; GFX9-NEXT: s_add_i32 s10, s9, 1
; GFX9-NEXT: s_cmp_ge_u32 s8, s6
; GFX9-NEXT: s_cselect_b32 s6, s10, s9
; GFX9-NEXT: s_ashr_i32 s8, s7, 31
; GFX9-NEXT: s_add_i32 s7, s7, s8
; GFX9-NEXT: s_xor_b32 s7, s7, s8
; GFX9-NEXT: v_cvt_f32_u32_e32 v0, s7
; GFX9-NEXT: v_readfirstlane_b32 s10, v5
; GFX9-NEXT: s_ashr_i32 s11, s10, 31
; GFX9-NEXT: s_xor_b32 s6, s6, s4
; GFX9-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX9-NEXT: s_add_i32 s10, s10, s11
; GFX9-NEXT: s_xor_b32 s8, s11, s8
; GFX9-NEXT: s_sub_i32 s4, s6, s4
; GFX9-NEXT: v_mul_f32_e32 v0, 0x4f7ffffe, v0
; GFX9-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX9-NEXT: s_xor_b32 s6, s10, s11
; GFX9-NEXT: s_sub_i32 s10, 0, s7
; GFX9-NEXT: v_readfirstlane_b32 s9, v2
; GFX9-NEXT: v_readfirstlane_b32 s11, v0
; GFX9-NEXT: s_mul_i32 s10, s10, s11
; GFX9-NEXT: s_mul_hi_u32 s10, s11, s10
; GFX9-NEXT: s_add_i32 s11, s11, s10
; GFX9-NEXT: s_mul_hi_u32 s10, s6, s11
; GFX9-NEXT: s_mul_i32 s11, s10, s7
; GFX9-NEXT: s_sub_i32 s6, s6, s11
; GFX9-NEXT: s_add_i32 s12, s10, 1
; GFX9-NEXT: s_sub_i32 s11, s6, s7
; GFX9-NEXT: s_cmp_ge_u32 s6, s7
; GFX9-NEXT: s_cselect_b32 s10, s12, s10
; GFX9-NEXT: s_cselect_b32 s6, s11, s6
; GFX9-NEXT: s_add_i32 s11, s10, 1
; GFX9-NEXT: s_cmp_ge_u32 s6, s7
; GFX9-NEXT: s_cselect_b32 s6, s11, s10
; GFX9-NEXT: s_ashr_i32 s7, s9, 31
; GFX9-NEXT: s_add_i32 s9, s9, s7
; GFX9-NEXT: s_xor_b32 s9, s9, s7
; GFX9-NEXT: v_cvt_f32_u32_e32 v0, s9
; GFX9-NEXT: v_readfirstlane_b32 s11, v6
; GFX9-NEXT: s_ashr_i32 s12, s11, 31
; GFX9-NEXT: s_xor_b32 s6, s6, s8
; GFX9-NEXT: v_rcp_iflag_f32_e32 v0, v0
; GFX9-NEXT: s_add_i32 s11, s11, s12
; GFX9-NEXT: s_xor_b32 s7, s12, s7
; GFX9-NEXT: s_sub_i32 s6, s6, s8
; GFX9-NEXT: v_mul_f32_e32 v0, 0x4f7ffffe, v0
; GFX9-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX9-NEXT: s_xor_b32 s8, s11, s12
; GFX9-NEXT: s_sub_i32 s11, 0, s9
; GFX9-NEXT: v_readfirstlane_b32 s10, v7
; GFX9-NEXT: v_readfirstlane_b32 s12, v0
; GFX9-NEXT: s_mul_i32 s11, s11, s12
; GFX9-NEXT: s_mul_hi_u32 s11, s12, s11
; GFX9-NEXT: s_add_i32 s12, s12, s11
; GFX9-NEXT: s_mul_hi_u32 s11, s8, s12
; GFX9-NEXT: s_mul_i32 s12, s11, s9
; GFX9-NEXT: s_sub_i32 s8, s8, s12
; GFX9-NEXT: s_add_i32 s13, s11, 1
; GFX9-NEXT: s_sub_i32 s12, s8, s9
; GFX9-NEXT: s_cmp_ge_u32 s8, s9
; GFX9-NEXT: s_cselect_b32 s11, s13, s11
; GFX9-NEXT: s_cselect_b32 s8, s12, s8
; GFX9-NEXT: s_add_i32 s12, s11, 1
; GFX9-NEXT: s_cmp_ge_u32 s8, s9
; GFX9-NEXT: s_cselect_b32 s8, s12, s11
; GFX9-NEXT: s_ashr_i32 s9, s5, 31
; GFX9-NEXT: s_add_i32 s5, s5, s9
; GFX9-NEXT: s_xor_b32 s5, s5, s9
; GFX9-NEXT: v_cvt_f32_u32_e32 v2, s5
; GFX9-NEXT: v_mov_b32_e32 v0, s4
; GFX9-NEXT: v_mov_b32_e32 v1, s6
; GFX9-NEXT: s_ashr_i32 s4, s10, 31
; GFX9-NEXT: v_rcp_iflag_f32_e32 v2, v2
; GFX9-NEXT: s_xor_b32 s6, s8, s7
; GFX9-NEXT: s_xor_b32 s8, s4, s9
; GFX9-NEXT: s_sub_i32 s6, s6, s7
; GFX9-NEXT: v_mul_f32_e32 v2, 0x4f7ffffe, v2
; GFX9-NEXT: v_cvt_u32_f32_e32 v2, v2
; GFX9-NEXT: s_sub_i32 s7, 0, s5
; GFX9-NEXT: s_add_i32 s10, s10, s4
; GFX9-NEXT: s_xor_b32 s4, s10, s4
; GFX9-NEXT: v_readfirstlane_b32 s9, v2
; GFX9-NEXT: s_mul_i32 s7, s7, s9
; GFX9-NEXT: s_mul_hi_u32 s7, s9, s7
; GFX9-NEXT: s_add_i32 s9, s9, s7
; GFX9-NEXT: s_mul_hi_u32 s7, s4, s9
; GFX9-NEXT: s_mul_i32 s9, s7, s5
; GFX9-NEXT: s_sub_i32 s4, s4, s9
; GFX9-NEXT: s_add_i32 s10, s7, 1
; GFX9-NEXT: s_sub_i32 s9, s4, s5
; GFX9-NEXT: s_cmp_ge_u32 s4, s5
; GFX9-NEXT: s_cselect_b32 s7, s10, s7
; GFX9-NEXT: s_cselect_b32 s4, s9, s4
; GFX9-NEXT: s_add_i32 s9, s7, 1
; GFX9-NEXT: s_cmp_ge_u32 s4, s5
; GFX9-NEXT: s_cselect_b32 s4, s9, s7
; GFX9-NEXT: s_xor_b32 s4, s4, s8
; GFX9-NEXT: s_sub_i32 s4, s4, s8
; GFX9-NEXT: v_mov_b32_e32 v2, s6
; GFX9-NEXT: v_mov_b32_e32 v3, s4
; GFX9-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0
; GFX9-NEXT: s_endpgm
;
; EG-LABEL: sdiv_v4i32:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @10, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 1 @6
; EG-NEXT: ALU 101, @11, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T2.XYZW, T0.X, 1
; EG-NEXT: CF_END
; EG-NEXT: PAD
; EG-NEXT: Fetch clause starting at 6:
; EG-NEXT: VTX_READ_128 T1.XYZW, T0.X, 16, #1
; EG-NEXT: VTX_READ_128 T0.XYZW, T0.X, 0, #1
; EG-NEXT: ALU clause starting at 10:
; EG-NEXT: MOV * T0.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 11:
; EG-NEXT: SETGT_INT * T2.W, 0.0, T1.W,
; EG-NEXT: ADD_INT * T1.W, T1.W, PV.W,
; EG-NEXT: XOR_INT * T1.W, PV.W, T2.W,
; EG-NEXT: SUB_INT T3.W, 0.0, PV.W,
; EG-NEXT: RECIP_UINT * T2.X, PV.W,
; EG-NEXT: SETGT_INT T4.W, 0.0, T0.W,
; EG-NEXT: MULLO_INT * T2.Y, PV.W, PS,
; EG-NEXT: SETGT_INT T2.Z, 0.0, T1.Y,
; EG-NEXT: ADD_INT T0.W, T0.W, PV.W,
; EG-NEXT: MULHI * T2.Y, T2.X, PS,
; EG-NEXT: ADD_INT T3.Z, T2.X, PS,
; EG-NEXT: XOR_INT T0.W, PV.W, T4.W,
; EG-NEXT: ADD_INT * T3.W, T1.Y, PV.Z,
; EG-NEXT: XOR_INT T3.W, PS, T2.Z,
; EG-NEXT: MULHI * T1.Y, PV.W, PV.Z,
; EG-NEXT: SUB_INT T5.W, 0.0, PV.W,
; EG-NEXT: RECIP_UINT * T2.X, PV.W,
; EG-NEXT: SETGT_INT T6.W, 0.0, T0.Y,
; EG-NEXT: MULLO_INT * T2.Y, PV.W, PS,
; EG-NEXT: ADD_INT T5.W, T0.Y, PV.W,
; EG-NEXT: MULHI * T0.Y, T2.X, PS,
; EG-NEXT: ADD_INT T0.Y, T2.X, PS,
; EG-NEXT: XOR_INT T3.Z, PV.W, T6.W, BS:VEC_021/SCL_122
; EG-NEXT: SETGT_INT T5.W, 0.0, T1.Z,
; EG-NEXT: MULLO_INT * T2.X, T1.Y, T1.W,
; EG-NEXT: ADD_INT T7.W, T1.Z, PV.W,
; EG-NEXT: MULHI * T0.Y, PV.Z, PV.Y,
; EG-NEXT: XOR_INT T7.W, PV.W, T5.W, BS:VEC_021/SCL_122
; EG-NEXT: MULLO_INT * T1.Z, PS, T3.W,
; EG-NEXT: SUB_INT T4.Z, 0.0, PV.W,
; EG-NEXT: SETGT_INT T8.W, 0.0, T1.X,
; EG-NEXT: RECIP_UINT * T2.Y, PV.W,
; EG-NEXT: ADD_INT T9.W, T1.X, PV.W,
; EG-NEXT: MULLO_INT * T1.X, PV.Z, PS,
; EG-NEXT: SETGT_INT T4.Z, 0.0, T0.Z,
; EG-NEXT: XOR_INT T9.W, PV.W, T8.W,
; EG-NEXT: MULHI * T1.X, T2.Y, PS,
; EG-NEXT: ADD_INT T1.X, T2.Y, PS,
; EG-NEXT: SUB_INT T2.Y, 0.0, PV.W,
; EG-NEXT: SUB_INT T1.Z, T3.Z, T1.Z,
; EG-NEXT: ADD_INT T10.W, T0.Z, PV.Z, BS:VEC_201
; EG-NEXT: RECIP_UINT * T0.Z, PV.W,
; EG-NEXT: XOR_INT T3.X, PV.W, T4.Z,
; EG-NEXT: ADD_INT T3.Y, T0.Y, 1,
; EG-NEXT: SETGE_UINT T3.Z, PV.Z, T3.W,
; EG-NEXT: SUB_INT T10.W, PV.Z, T3.W,
; EG-NEXT: MULLO_INT * T2.Y, PV.Y, PS,
; EG-NEXT: CNDE_INT T1.Z, PV.Z, T1.Z, PV.W,
; EG-NEXT: CNDE_INT T10.W, PV.Z, T0.Y, PV.Y,
; EG-NEXT: MULHI * T0.Y, PV.X, T1.X,
; EG-NEXT: SETGT_INT T3.Y, 0.0, T0.X,
; EG-NEXT: ADD_INT T3.Z, PV.W, 1,
; EG-NEXT: SETGE_UINT T3.W, PV.Z, T3.W, BS:VEC_021/SCL_122
; EG-NEXT: MULLO_INT * T1.X, PS, T7.W,
; EG-NEXT: CNDE_INT T4.Y, PV.W, T10.W, PV.Z,
; EG-NEXT: ADD_INT T1.Z, T0.X, PV.Y,
; EG-NEXT: SUB_INT T3.W, T3.X, PS, BS:VEC_120/SCL_212
; EG-NEXT: MULHI * T0.X, T0.Z, T2.Y,
; EG-NEXT: ADD_INT T1.X, T0.Y, 1,
; EG-NEXT: SETGE_UINT T2.Y, PV.W, T7.W,
; EG-NEXT: ADD_INT T0.Z, T0.Z, PS,
; EG-NEXT: XOR_INT T10.W, PV.Z, T3.Y,
; EG-NEXT: SUB_INT * T0.W, T0.W, T2.X,
; EG-NEXT: SUB_INT T0.X, T3.W, T7.W,
; EG-NEXT: ADD_INT T5.Y, T1.Y, 1,
; EG-NEXT: SETGE_UINT T1.Z, PS, T1.W, BS:VEC_021/SCL_122
; EG-NEXT: SUB_INT T11.W, PS, T1.W, BS:VEC_021/SCL_122
; EG-NEXT: MULHI * T0.Z, PV.W, PV.Z,
; EG-NEXT: CNDE_INT T2.X, PV.Z, T0.W, PV.W, BS:VEC_021/SCL_122
; EG-NEXT: CNDE_INT T1.Y, PV.Z, T1.Y, PV.Y,
; EG-NEXT: CNDE_INT T1.Z, T2.Y, T3.W, PV.X, BS:VEC_201
; EG-NEXT: CNDE_INT T0.W, T2.Y, T0.Y, T1.X, BS:VEC_201
; EG-NEXT: MULLO_INT * T0.X, PS, T9.W,
; EG-NEXT: ADD_INT T1.X, PV.W, 1,
; EG-NEXT: SETGE_UINT T0.Y, PV.Z, T7.W,
; EG-NEXT: ADD_INT T1.Z, PV.Y, 1,
; EG-NEXT: SETGE_UINT T1.W, PV.X, T1.W, BS:VEC_102/SCL_221
; EG-NEXT: SUB_INT * T3.W, T10.W, PS,
; EG-NEXT: ADD_INT T0.X, T0.Z, 1,
; EG-NEXT: SETGE_UINT T2.Y, PS, T9.W, BS:VEC_102/SCL_221
; EG-NEXT: SUB_INT T3.Z, PS, T9.W, BS:VEC_102/SCL_221
; EG-NEXT: CNDE_INT T1.W, PV.W, T1.Y, PV.Z,
; EG-NEXT: XOR_INT * T2.W, T4.W, T2.W,
; EG-NEXT: XOR_INT T2.X, PV.W, PS,
; EG-NEXT: CNDE_INT T1.Y, PV.Y, T3.W, PV.Z, BS:VEC_021/SCL_122
; EG-NEXT: CNDE_INT T0.Z, PV.Y, T0.Z, PV.X,
; EG-NEXT: CNDE_INT T0.W, T0.Y, T0.W, T1.X, BS:VEC_102/SCL_221
; EG-NEXT: XOR_INT * T1.W, T4.Z, T5.W,
; EG-NEXT: XOR_INT T0.X, T6.W, T2.Z,
; EG-NEXT: XOR_INT T0.Y, PV.W, PS,
; EG-NEXT: ADD_INT T1.Z, PV.Z, 1,
; EG-NEXT: SETGE_UINT T0.W, PV.Y, T9.W, BS:VEC_021/SCL_122
; EG-NEXT: SUB_INT * T2.W, PV.X, T2.W,
; EG-NEXT: CNDE_INT T1.Y, PV.W, T0.Z, PV.Z,
; EG-NEXT: SUB_INT T2.Z, PV.Y, T1.W,
; EG-NEXT: XOR_INT T0.W, T3.Y, T8.W, BS:VEC_021/SCL_122
; EG-NEXT: XOR_INT * T1.W, T4.Y, PV.X,
; EG-NEXT: SUB_INT T2.Y, PS, T0.X,
; EG-NEXT: XOR_INT * T1.W, PV.Y, PV.W,
; EG-NEXT: SUB_INT T2.X, PV.W, T0.W,
; EG-NEXT: LSHR * T0.X, KC0[2].Y, literal.x,
; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00)
%den_ptr = getelementptr <4 x i32>, <4 x i32> addrspace(1)* %in, i32 1
%num = load <4 x i32>, <4 x i32> addrspace(1) * %in
%den = load <4 x i32>, <4 x i32> addrspace(1) * %den_ptr
%result = sdiv <4 x i32> %num, %den
store <4 x i32> %result, <4 x i32> addrspace(1)* %out
ret void
}
define amdgpu_kernel void @sdiv_v4i32_4(<4 x i32> addrspace(1)* %out, <4 x i32> addrspace(1)* %in) {
; GCN-LABEL: sdiv_v4i32_4:
; GCN: ; %bb.0:
; GCN-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s6, -1
; GCN-NEXT: s_mov_b32 s10, s6
; GCN-NEXT: s_mov_b32 s11, s7
; GCN-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s8, s2
; GCN-NEXT: s_mov_b32 s9, s3
; GCN-NEXT: buffer_load_dwordx4 v[0:3], off, s[8:11], 0
; GCN-NEXT: s_mov_b32 s4, s0
; GCN-NEXT: s_mov_b32 s5, s1
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_ashrrev_i32_e32 v4, 31, v0
; GCN-NEXT: v_ashrrev_i32_e32 v5, 31, v1
; GCN-NEXT: v_ashrrev_i32_e32 v6, 31, v2
; GCN-NEXT: v_ashrrev_i32_e32 v7, 31, v3
; GCN-NEXT: v_lshrrev_b32_e32 v4, 30, v4
; GCN-NEXT: v_lshrrev_b32_e32 v5, 30, v5
; GCN-NEXT: v_lshrrev_b32_e32 v6, 30, v6
; GCN-NEXT: v_lshrrev_b32_e32 v7, 30, v7
; GCN-NEXT: v_add_i32_e32 v0, vcc, v4, v0
; GCN-NEXT: v_add_i32_e32 v1, vcc, v5, v1
; GCN-NEXT: v_add_i32_e32 v2, vcc, v6, v2
; GCN-NEXT: v_add_i32_e32 v3, vcc, v7, v3
; GCN-NEXT: v_ashrrev_i32_e32 v0, 2, v0
; GCN-NEXT: v_ashrrev_i32_e32 v1, 2, v1
; GCN-NEXT: v_ashrrev_i32_e32 v2, 2, v2
; GCN-NEXT: v_ashrrev_i32_e32 v3, 2, v3
; GCN-NEXT: buffer_store_dwordx4 v[0:3], off, s[4:7], 0
; GCN-NEXT: s_endpgm
;
; TONGA-LABEL: sdiv_v4i32_4:
; TONGA: ; %bb.0:
; TONGA-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24
; TONGA-NEXT: s_mov_b32 s7, 0xf000
; TONGA-NEXT: s_mov_b32 s6, -1
; TONGA-NEXT: s_mov_b32 s10, s6
; TONGA-NEXT: s_mov_b32 s11, s7
; TONGA-NEXT: s_waitcnt lgkmcnt(0)
; TONGA-NEXT: s_mov_b32 s8, s2
; TONGA-NEXT: s_mov_b32 s9, s3
; TONGA-NEXT: buffer_load_dwordx4 v[0:3], off, s[8:11], 0
; TONGA-NEXT: s_mov_b32 s4, s0
; TONGA-NEXT: s_mov_b32 s5, s1
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: v_ashrrev_i32_e32 v4, 31, v0
; TONGA-NEXT: v_ashrrev_i32_e32 v5, 31, v1
; TONGA-NEXT: v_ashrrev_i32_e32 v6, 31, v2
; TONGA-NEXT: v_ashrrev_i32_e32 v7, 31, v3
; TONGA-NEXT: v_lshrrev_b32_e32 v4, 30, v4
; TONGA-NEXT: v_lshrrev_b32_e32 v5, 30, v5
; TONGA-NEXT: v_lshrrev_b32_e32 v6, 30, v6
; TONGA-NEXT: v_lshrrev_b32_e32 v7, 30, v7
; TONGA-NEXT: v_add_u32_e32 v0, vcc, v4, v0
; TONGA-NEXT: v_add_u32_e32 v1, vcc, v5, v1
; TONGA-NEXT: v_add_u32_e32 v2, vcc, v6, v2
; TONGA-NEXT: v_add_u32_e32 v3, vcc, v7, v3
; TONGA-NEXT: v_ashrrev_i32_e32 v0, 2, v0
; TONGA-NEXT: v_ashrrev_i32_e32 v1, 2, v1
; TONGA-NEXT: v_ashrrev_i32_e32 v2, 2, v2
; TONGA-NEXT: v_ashrrev_i32_e32 v3, 2, v3
; TONGA-NEXT: buffer_store_dwordx4 v[0:3], off, s[4:7], 0
; TONGA-NEXT: s_endpgm
;
; GFX9-LABEL: sdiv_v4i32_4:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24
; GFX9-NEXT: s_mov_b32 s7, 0xf000
; GFX9-NEXT: s_mov_b32 s6, -1
; GFX9-NEXT: s_mov_b32 s10, s6
; GFX9-NEXT: s_mov_b32 s11, s7
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_mov_b32 s8, s2
; GFX9-NEXT: s_mov_b32 s9, s3
; GFX9-NEXT: buffer_load_dwordx4 v[0:3], off, s[8:11], 0
; GFX9-NEXT: s_mov_b32 s4, s0
; GFX9-NEXT: s_mov_b32 s5, s1
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_ashrrev_i32_e32 v4, 31, v0
; GFX9-NEXT: v_ashrrev_i32_e32 v5, 31, v1
; GFX9-NEXT: v_ashrrev_i32_e32 v6, 31, v2
; GFX9-NEXT: v_ashrrev_i32_e32 v7, 31, v3
; GFX9-NEXT: v_lshrrev_b32_e32 v4, 30, v4
; GFX9-NEXT: v_lshrrev_b32_e32 v5, 30, v5
; GFX9-NEXT: v_lshrrev_b32_e32 v6, 30, v6
; GFX9-NEXT: v_lshrrev_b32_e32 v7, 30, v7
; GFX9-NEXT: v_add_u32_e32 v0, v0, v4
; GFX9-NEXT: v_add_u32_e32 v1, v1, v5
; GFX9-NEXT: v_add_u32_e32 v2, v2, v6
; GFX9-NEXT: v_add_u32_e32 v3, v3, v7
; GFX9-NEXT: v_ashrrev_i32_e32 v0, 2, v0
; GFX9-NEXT: v_ashrrev_i32_e32 v1, 2, v1
; GFX9-NEXT: v_ashrrev_i32_e32 v2, 2, v2
; GFX9-NEXT: v_ashrrev_i32_e32 v3, 2, v3
; GFX9-NEXT: buffer_store_dwordx4 v[0:3], off, s[4:7], 0
; GFX9-NEXT: s_endpgm
;
; EG-LABEL: sdiv_v4i32_4:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 0 @6
; EG-NEXT: ALU 24, @9, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T1.XYZW, T0.X, 1
; EG-NEXT: CF_END
; EG-NEXT: PAD
; EG-NEXT: Fetch clause starting at 6:
; EG-NEXT: VTX_READ_128 T0.XYZW, T0.X, 0, #1
; EG-NEXT: ALU clause starting at 8:
; EG-NEXT: MOV * T0.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 9:
; EG-NEXT: ASHR T1.W, T0.W, literal.x,
; EG-NEXT: ASHR * T2.W, T0.Z, literal.x,
; EG-NEXT: 31(4.344025e-44), 0(0.000000e+00)
; EG-NEXT: LSHR * T1.W, PV.W, literal.x,
; EG-NEXT: 30(4.203895e-44), 0(0.000000e+00)
; EG-NEXT: ADD_INT T1.Z, T0.W, PV.W,
; EG-NEXT: LSHR T0.W, T2.W, literal.x, BS:VEC_120/SCL_212
; EG-NEXT: ASHR * T1.W, T0.Y, literal.y,
; EG-NEXT: 30(4.203895e-44), 31(4.344025e-44)
; EG-NEXT: LSHR T1.Y, PS, literal.x,
; EG-NEXT: ASHR T2.Z, T0.X, literal.y,
; EG-NEXT: ADD_INT T0.W, T0.Z, PV.W,
; EG-NEXT: ASHR * T1.W, PV.Z, literal.z,
; EG-NEXT: 30(4.203895e-44), 31(4.344025e-44)
; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00)
; EG-NEXT: ASHR T1.Z, PV.W, literal.x,
; EG-NEXT: LSHR T0.W, PV.Z, literal.y,
; EG-NEXT: ADD_INT * T2.W, T0.Y, PV.Y,
; EG-NEXT: 2(2.802597e-45), 30(4.203895e-44)
; EG-NEXT: ASHR T1.Y, PS, literal.x,
; EG-NEXT: ADD_INT * T0.W, T0.X, PV.W,
; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00)
; EG-NEXT: ASHR T1.X, PV.W, literal.x,
; EG-NEXT: LSHR * T0.X, KC0[2].Y, literal.x,
; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00)
%num = load <4 x i32>, <4 x i32> addrspace(1) * %in
%result = sdiv <4 x i32> %num, <i32 4, i32 4, i32 4, i32 4>
store <4 x i32> %result, <4 x i32> addrspace(1)* %out
ret void
}
define amdgpu_kernel void @v_sdiv_i8(i32 addrspace(1)* %out, i8 addrspace(1)* %in) {
; GCN-LABEL: v_sdiv_i8:
; GCN: ; %bb.0:
; GCN-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s6, -1
; GCN-NEXT: s_mov_b32 s10, s6
; GCN-NEXT: s_mov_b32 s11, s7
; GCN-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s8, s2
; GCN-NEXT: s_mov_b32 s9, s3
; GCN-NEXT: buffer_load_sbyte v0, off, s[8:11], 0 offset:1
; GCN-NEXT: buffer_load_sbyte v1, off, s[8:11], 0
; GCN-NEXT: s_mov_b32 s4, s0
; GCN-NEXT: s_mov_b32 s5, s1
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_cvt_f32_i32_e32 v2, v0
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_cvt_f32_i32_e32 v3, v1
; GCN-NEXT: v_xor_b32_e32 v0, v1, v0
; GCN-NEXT: v_ashrrev_i32_e32 v0, 30, v0
; GCN-NEXT: v_rcp_iflag_f32_e32 v4, v2
; GCN-NEXT: v_or_b32_e32 v0, 1, v0
; GCN-NEXT: v_mul_f32_e32 v1, v3, v4
; GCN-NEXT: v_trunc_f32_e32 v1, v1
; GCN-NEXT: v_mad_f32 v3, -v1, v2, v3
; GCN-NEXT: v_cvt_i32_f32_e32 v1, v1
; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v3|, |v2|
; GCN-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v1
; GCN-NEXT: v_bfe_i32 v0, v0, 0, 8
; GCN-NEXT: buffer_store_dword v0, off, s[4:7], 0
; GCN-NEXT: s_endpgm
;
; TONGA-LABEL: v_sdiv_i8:
; TONGA: ; %bb.0:
; TONGA-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24
; TONGA-NEXT: s_mov_b32 s7, 0xf000
; TONGA-NEXT: s_mov_b32 s6, -1
; TONGA-NEXT: s_mov_b32 s10, s6
; TONGA-NEXT: s_mov_b32 s11, s7
; TONGA-NEXT: s_waitcnt lgkmcnt(0)
; TONGA-NEXT: s_mov_b32 s8, s2
; TONGA-NEXT: s_mov_b32 s9, s3
; TONGA-NEXT: buffer_load_sbyte v0, off, s[8:11], 0 offset:1
; TONGA-NEXT: buffer_load_sbyte v1, off, s[8:11], 0
; TONGA-NEXT: s_mov_b32 s4, s0
; TONGA-NEXT: s_mov_b32 s5, s1
; TONGA-NEXT: s_waitcnt vmcnt(1)
; TONGA-NEXT: v_cvt_f32_i32_e32 v2, v0
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: v_cvt_f32_i32_e32 v3, v1
; TONGA-NEXT: v_xor_b32_e32 v0, v1, v0
; TONGA-NEXT: v_ashrrev_i32_e32 v0, 30, v0
; TONGA-NEXT: v_rcp_iflag_f32_e32 v4, v2
; TONGA-NEXT: v_or_b32_e32 v0, 1, v0
; TONGA-NEXT: v_mul_f32_e32 v1, v3, v4
; TONGA-NEXT: v_trunc_f32_e32 v1, v1
; TONGA-NEXT: v_mad_f32 v3, -v1, v2, v3
; TONGA-NEXT: v_cvt_i32_f32_e32 v1, v1
; TONGA-NEXT: v_cmp_ge_f32_e64 vcc, |v3|, |v2|
; TONGA-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
; TONGA-NEXT: v_add_u32_e32 v0, vcc, v0, v1
; TONGA-NEXT: v_bfe_i32 v0, v0, 0, 8
; TONGA-NEXT: buffer_store_dword v0, off, s[4:7], 0
; TONGA-NEXT: s_endpgm
;
; GFX9-LABEL: v_sdiv_i8:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24
; GFX9-NEXT: s_mov_b32 s7, 0xf000
; GFX9-NEXT: s_mov_b32 s6, -1
; GFX9-NEXT: s_mov_b32 s10, s6
; GFX9-NEXT: s_mov_b32 s11, s7
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_mov_b32 s8, s2
; GFX9-NEXT: s_mov_b32 s9, s3
; GFX9-NEXT: buffer_load_sbyte v0, off, s[8:11], 0 offset:1
; GFX9-NEXT: buffer_load_sbyte v1, off, s[8:11], 0
; GFX9-NEXT: s_mov_b32 s4, s0
; GFX9-NEXT: s_mov_b32 s5, s1
; GFX9-NEXT: s_waitcnt vmcnt(1)
; GFX9-NEXT: v_cvt_f32_i32_e32 v2, v0
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_cvt_f32_i32_e32 v3, v1
; GFX9-NEXT: v_xor_b32_e32 v0, v1, v0
; GFX9-NEXT: v_ashrrev_i32_e32 v0, 30, v0
; GFX9-NEXT: v_rcp_iflag_f32_e32 v4, v2
; GFX9-NEXT: v_or_b32_e32 v0, 1, v0
; GFX9-NEXT: v_mul_f32_e32 v1, v3, v4
; GFX9-NEXT: v_trunc_f32_e32 v1, v1
; GFX9-NEXT: v_cvt_i32_f32_e32 v4, v1
; GFX9-NEXT: v_mad_f32 v1, -v1, v2, v3
; GFX9-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, |v2|
; GFX9-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
; GFX9-NEXT: v_add_u32_e32 v0, v4, v0
; GFX9-NEXT: v_bfe_i32 v0, v0, 0, 8
; GFX9-NEXT: buffer_store_dword v0, off, s[4:7], 0
; GFX9-NEXT: s_endpgm
;
; EG-LABEL: v_sdiv_i8:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @10, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 1 @6
; EG-NEXT: ALU 21, @11, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.X, T1.X, 1
; EG-NEXT: CF_END
; EG-NEXT: PAD
; EG-NEXT: Fetch clause starting at 6:
; EG-NEXT: VTX_READ_8 T1.X, T0.X, 1, #1
; EG-NEXT: VTX_READ_8 T0.X, T0.X, 0, #1
; EG-NEXT: ALU clause starting at 10:
; EG-NEXT: MOV * T0.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 11:
; EG-NEXT: BFE_INT * T0.W, T1.X, 0.0, literal.x,
; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00)
; EG-NEXT: INT_TO_FLT * T0.Y, PV.W,
; EG-NEXT: BFE_INT T1.W, T0.X, 0.0, literal.x,
; EG-NEXT: RECIP_IEEE * T0.X, PS,
; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00)
; EG-NEXT: INT_TO_FLT * T0.Z, PV.W,
; EG-NEXT: MUL_IEEE * T2.W, PS, T0.X,
; EG-NEXT: TRUNC T2.W, PV.W,
; EG-NEXT: XOR_INT * T0.W, T1.W, T0.W,
; EG-NEXT: ASHR T0.W, PS, literal.x,
; EG-NEXT: MULADD_IEEE * T1.W, -PV.W, T0.Y, T0.Z,
; EG-NEXT: 30(4.203895e-44), 0(0.000000e+00)
; EG-NEXT: TRUNC T0.Z, T2.W,
; EG-NEXT: SETGE T1.W, |PS|, |T0.Y|,
; EG-NEXT: OR_INT * T0.W, PV.W, 1,
; EG-NEXT: CNDE T0.W, PV.W, 0.0, PS,
; EG-NEXT: FLT_TO_INT * T1.W, PV.Z,
; EG-NEXT: ADD_INT * T0.W, PS, PV.W,
; EG-NEXT: BFE_INT T0.X, PV.W, 0.0, literal.x,
; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.y,
; EG-NEXT: 8(1.121039e-44), 2(2.802597e-45)
%den_ptr = getelementptr i8, i8 addrspace(1)* %in, i8 1
%num = load i8, i8 addrspace(1) * %in
%den = load i8, i8 addrspace(1) * %den_ptr
%result = sdiv i8 %num, %den
%result.ext = sext i8 %result to i32
store i32 %result.ext, i32 addrspace(1)* %out
ret void
}
define amdgpu_kernel void @v_sdiv_i23(i32 addrspace(1)* %out, i23 addrspace(1)* %in) {
; GCN-LABEL: v_sdiv_i23:
; GCN: ; %bb.0:
; GCN-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s6, -1
; GCN-NEXT: s_mov_b32 s10, s6
; GCN-NEXT: s_mov_b32 s11, s7
; GCN-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s8, s2
; GCN-NEXT: s_mov_b32 s9, s3
; GCN-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 offset:2
; GCN-NEXT: buffer_load_ubyte v1, off, s[8:11], 0 offset:6
; GCN-NEXT: buffer_load_ushort v2, off, s[8:11], 0 offset:4
; GCN-NEXT: buffer_load_ushort v3, off, s[8:11], 0
; GCN-NEXT: s_mov_b32 s4, s0
; GCN-NEXT: s_mov_b32 s5, s1
; GCN-NEXT: s_waitcnt vmcnt(3)
; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GCN-NEXT: s_waitcnt vmcnt(2)
; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_or_b32_e32 v1, v2, v1
; GCN-NEXT: v_bfe_i32 v1, v1, 0, 23
; GCN-NEXT: v_cvt_f32_i32_e32 v2, v1
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_or_b32_e32 v0, v3, v0
; GCN-NEXT: v_bfe_i32 v0, v0, 0, 23
; GCN-NEXT: v_cvt_f32_i32_e32 v3, v0
; GCN-NEXT: v_rcp_iflag_f32_e32 v4, v2
; GCN-NEXT: v_xor_b32_e32 v0, v0, v1
; GCN-NEXT: v_ashrrev_i32_e32 v0, 30, v0
; GCN-NEXT: v_or_b32_e32 v0, 1, v0
; GCN-NEXT: v_mul_f32_e32 v1, v3, v4
; GCN-NEXT: v_trunc_f32_e32 v1, v1
; GCN-NEXT: v_mad_f32 v3, -v1, v2, v3
; GCN-NEXT: v_cvt_i32_f32_e32 v1, v1
; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v3|, |v2|
; GCN-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v1
; GCN-NEXT: v_bfe_i32 v0, v0, 0, 23
; GCN-NEXT: buffer_store_dword v0, off, s[4:7], 0
; GCN-NEXT: s_endpgm
;
; TONGA-LABEL: v_sdiv_i23:
; TONGA: ; %bb.0:
; TONGA-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24
; TONGA-NEXT: s_mov_b32 s7, 0xf000
; TONGA-NEXT: s_mov_b32 s6, -1
; TONGA-NEXT: s_mov_b32 s10, s6
; TONGA-NEXT: s_mov_b32 s11, s7
; TONGA-NEXT: s_waitcnt lgkmcnt(0)
; TONGA-NEXT: s_mov_b32 s8, s2
; TONGA-NEXT: s_mov_b32 s9, s3
; TONGA-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 offset:2
; TONGA-NEXT: buffer_load_ubyte v1, off, s[8:11], 0 offset:6
; TONGA-NEXT: buffer_load_ushort v2, off, s[8:11], 0 offset:4
; TONGA-NEXT: buffer_load_ushort v3, off, s[8:11], 0
; TONGA-NEXT: s_mov_b32 s4, s0
; TONGA-NEXT: s_mov_b32 s5, s1
; TONGA-NEXT: s_waitcnt vmcnt(3)
; TONGA-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; TONGA-NEXT: s_waitcnt vmcnt(2)
; TONGA-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; TONGA-NEXT: s_waitcnt vmcnt(1)
; TONGA-NEXT: v_or_b32_e32 v1, v2, v1
; TONGA-NEXT: v_bfe_i32 v1, v1, 0, 23
; TONGA-NEXT: v_cvt_f32_i32_e32 v2, v1
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: v_or_b32_e32 v0, v3, v0
; TONGA-NEXT: v_bfe_i32 v0, v0, 0, 23
; TONGA-NEXT: v_cvt_f32_i32_e32 v3, v0
; TONGA-NEXT: v_rcp_iflag_f32_e32 v4, v2
; TONGA-NEXT: v_xor_b32_e32 v0, v0, v1
; TONGA-NEXT: v_ashrrev_i32_e32 v0, 30, v0
; TONGA-NEXT: v_or_b32_e32 v0, 1, v0
; TONGA-NEXT: v_mul_f32_e32 v1, v3, v4
; TONGA-NEXT: v_trunc_f32_e32 v1, v1
; TONGA-NEXT: v_mad_f32 v3, -v1, v2, v3
; TONGA-NEXT: v_cvt_i32_f32_e32 v1, v1
; TONGA-NEXT: v_cmp_ge_f32_e64 vcc, |v3|, |v2|
; TONGA-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
; TONGA-NEXT: v_add_u32_e32 v0, vcc, v0, v1
; TONGA-NEXT: v_bfe_i32 v0, v0, 0, 23
; TONGA-NEXT: buffer_store_dword v0, off, s[4:7], 0
; TONGA-NEXT: s_endpgm
;
; GFX9-LABEL: v_sdiv_i23:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24
; GFX9-NEXT: s_mov_b32 s7, 0xf000
; GFX9-NEXT: s_mov_b32 s6, -1
; GFX9-NEXT: s_mov_b32 s10, s6
; GFX9-NEXT: s_mov_b32 s11, s7
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_mov_b32 s8, s2
; GFX9-NEXT: s_mov_b32 s9, s3
; GFX9-NEXT: buffer_load_ubyte v0, off, s[8:11], 0 offset:2
; GFX9-NEXT: buffer_load_ubyte v1, off, s[8:11], 0 offset:6
; GFX9-NEXT: buffer_load_ushort v2, off, s[8:11], 0 offset:4
; GFX9-NEXT: buffer_load_ushort v3, off, s[8:11], 0
; GFX9-NEXT: s_mov_b32 s4, s0
; GFX9-NEXT: s_mov_b32 s5, s1
; GFX9-NEXT: s_waitcnt vmcnt(3)
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: s_waitcnt vmcnt(2)
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT: s_waitcnt vmcnt(1)
; GFX9-NEXT: v_or_b32_e32 v1, v2, v1
; GFX9-NEXT: v_bfe_i32 v1, v1, 0, 23
; GFX9-NEXT: v_cvt_f32_i32_e32 v2, v1
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_or_b32_e32 v0, v3, v0
; GFX9-NEXT: v_bfe_i32 v0, v0, 0, 23
; GFX9-NEXT: v_cvt_f32_i32_e32 v3, v0
; GFX9-NEXT: v_rcp_iflag_f32_e32 v4, v2
; GFX9-NEXT: v_xor_b32_e32 v0, v0, v1
; GFX9-NEXT: v_ashrrev_i32_e32 v0, 30, v0
; GFX9-NEXT: v_or_b32_e32 v0, 1, v0
; GFX9-NEXT: v_mul_f32_e32 v1, v3, v4
; GFX9-NEXT: v_trunc_f32_e32 v1, v1
; GFX9-NEXT: v_cvt_i32_f32_e32 v4, v1
; GFX9-NEXT: v_mad_f32 v1, -v1, v2, v3
; GFX9-NEXT: v_cmp_ge_f32_e64 vcc, |v1|, |v2|
; GFX9-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
; GFX9-NEXT: v_add_u32_e32 v0, v4, v0
; GFX9-NEXT: v_bfe_i32 v0, v0, 0, 23
; GFX9-NEXT: buffer_store_dword v0, off, s[4:7], 0
; GFX9-NEXT: s_endpgm
;
; EG-LABEL: v_sdiv_i23:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @14, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 3 @6
; EG-NEXT: ALU 33, @15, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.X, T1.X, 1
; EG-NEXT: CF_END
; EG-NEXT: PAD
; EG-NEXT: Fetch clause starting at 6:
; EG-NEXT: VTX_READ_8 T1.X, T0.X, 6, #1
; EG-NEXT: VTX_READ_16 T2.X, T0.X, 0, #1
; EG-NEXT: VTX_READ_8 T3.X, T0.X, 2, #1
; EG-NEXT: VTX_READ_16 T0.X, T0.X, 4, #1
; EG-NEXT: ALU clause starting at 14:
; EG-NEXT: MOV * T0.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 15:
; EG-NEXT: LSHL * T0.W, T1.X, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: OR_INT T0.W, T0.X, PV.W,
; EG-NEXT: LSHL * T1.W, T3.X, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: LSHL * T0.W, PV.W, literal.x,
; EG-NEXT: 9(1.261169e-44), 0(0.000000e+00)
; EG-NEXT: ASHR T0.W, PV.W, literal.x,
; EG-NEXT: OR_INT * T1.W, T2.X, T1.W,
; EG-NEXT: 9(1.261169e-44), 0(0.000000e+00)
; EG-NEXT: LSHL T1.W, PS, literal.x,
; EG-NEXT: INT_TO_FLT * T0.X, PV.W,
; EG-NEXT: 9(1.261169e-44), 0(0.000000e+00)
; EG-NEXT: ASHR T1.W, PV.W, literal.x,
; EG-NEXT: RECIP_IEEE * T0.Y, PS,
; EG-NEXT: 9(1.261169e-44), 0(0.000000e+00)
; EG-NEXT: INT_TO_FLT * T0.Z, PV.W,
; EG-NEXT: MUL_IEEE * T2.W, PS, T0.Y,
; EG-NEXT: TRUNC T2.W, PV.W,
; EG-NEXT: XOR_INT * T0.W, T1.W, T0.W,
; EG-NEXT: ASHR T0.W, PS, literal.x,
; EG-NEXT: MULADD_IEEE * T1.W, -PV.W, T0.X, T0.Z,
; EG-NEXT: 30(4.203895e-44), 0(0.000000e+00)
; EG-NEXT: TRUNC T0.Z, T2.W,
; EG-NEXT: SETGE T1.W, |PS|, |T0.X|,
; EG-NEXT: OR_INT * T0.W, PV.W, 1,
; EG-NEXT: CNDE T0.W, PV.W, 0.0, PS,
; EG-NEXT: FLT_TO_INT * T1.W, PV.Z,
; EG-NEXT: ADD_INT * T0.W, PS, PV.W,
; EG-NEXT: LSHL * T0.W, PV.W, literal.x,
; EG-NEXT: 9(1.261169e-44), 0(0.000000e+00)
; EG-NEXT: ASHR T0.X, PV.W, literal.x,
; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.y,
; EG-NEXT: 9(1.261169e-44), 2(2.802597e-45)
%den_ptr = getelementptr i23, i23 addrspace(1)* %in, i23 1
%num = load i23, i23 addrspace(1) * %in
%den = load i23, i23 addrspace(1) * %den_ptr
%result = sdiv i23 %num, %den
%result.ext = sext i23 %result to i32
store i32 %result.ext, i32 addrspace(1)* %out
ret void
}
define amdgpu_kernel void @v_sdiv_i24(i32 addrspace(1)* %out, i24 addrspace(1)* %in) {
; GCN-LABEL: v_sdiv_i24:
; GCN: ; %bb.0:
; GCN-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s6, -1
; GCN-NEXT: s_mov_b32 s10, s6
; GCN-NEXT: s_mov_b32 s11, s7
; GCN-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s8, s2
; GCN-NEXT: s_mov_b32 s9, s3
; GCN-NEXT: buffer_load_sbyte v0, off, s[8:11], 0 offset:6
; GCN-NEXT: buffer_load_ushort v1, off, s[8:11], 0 offset:4
; GCN-NEXT: buffer_load_sbyte v2, off, s[8:11], 0 offset:2
; GCN-NEXT: buffer_load_ushort v3, off, s[8:11], 0
; GCN-NEXT: s_mov_b32 s4, s0
; GCN-NEXT: s_mov_b32 s5, s1
; GCN-NEXT: s_waitcnt vmcnt(3)
; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v0
; GCN-NEXT: s_waitcnt vmcnt(2)
; GCN-NEXT: v_or_b32_e32 v1, v1, v4
; GCN-NEXT: v_cvt_f32_i32_e32 v1, v1
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v2
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_or_b32_e32 v3, v3, v4
; GCN-NEXT: v_cvt_f32_i32_e32 v3, v3
; GCN-NEXT: v_rcp_iflag_f32_e32 v4, v1
; GCN-NEXT: v_xor_b32_e32 v0, v2, v0
; GCN-NEXT: v_ashrrev_i32_e32 v0, 30, v0
; GCN-NEXT: v_or_b32_e32 v0, 1, v0
; GCN-NEXT: v_mul_f32_e32 v2, v3, v4
; GCN-NEXT: v_trunc_f32_e32 v2, v2
; GCN-NEXT: v_mad_f32 v3, -v2, v1, v3
; GCN-NEXT: v_cvt_i32_f32_e32 v2, v2
; GCN-NEXT: v_cmp_ge_f32_e64 vcc, |v3|, |v1|
; GCN-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v2
; GCN-NEXT: v_bfe_i32 v0, v0, 0, 24
; GCN-NEXT: buffer_store_dword v0, off, s[4:7], 0
; GCN-NEXT: s_endpgm
;
; TONGA-LABEL: v_sdiv_i24:
; TONGA: ; %bb.0:
; TONGA-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24
; TONGA-NEXT: s_mov_b32 s7, 0xf000
; TONGA-NEXT: s_mov_b32 s6, -1
; TONGA-NEXT: s_mov_b32 s10, s6
; TONGA-NEXT: s_mov_b32 s11, s7
; TONGA-NEXT: s_waitcnt lgkmcnt(0)
; TONGA-NEXT: s_mov_b32 s8, s2
; TONGA-NEXT: s_mov_b32 s9, s3
; TONGA-NEXT: buffer_load_sbyte v0, off, s[8:11], 0 offset:6
; TONGA-NEXT: buffer_load_ushort v1, off, s[8:11], 0 offset:4
; TONGA-NEXT: buffer_load_sbyte v2, off, s[8:11], 0 offset:2
; TONGA-NEXT: buffer_load_ushort v3, off, s[8:11], 0
; TONGA-NEXT: s_mov_b32 s4, s0
; TONGA-NEXT: s_mov_b32 s5, s1
; TONGA-NEXT: s_waitcnt vmcnt(3)
; TONGA-NEXT: v_lshlrev_b32_e32 v4, 16, v0
; TONGA-NEXT: s_waitcnt vmcnt(2)
; TONGA-NEXT: v_or_b32_e32 v1, v1, v4
; TONGA-NEXT: v_cvt_f32_i32_e32 v1, v1
; TONGA-NEXT: s_waitcnt vmcnt(1)
; TONGA-NEXT: v_lshlrev_b32_e32 v4, 16, v2
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: v_or_b32_e32 v3, v3, v4
; TONGA-NEXT: v_cvt_f32_i32_e32 v3, v3
; TONGA-NEXT: v_rcp_iflag_f32_e32 v4, v1
; TONGA-NEXT: v_xor_b32_e32 v0, v2, v0
; TONGA-NEXT: v_ashrrev_i32_e32 v0, 30, v0
; TONGA-NEXT: v_or_b32_e32 v0, 1, v0
; TONGA-NEXT: v_mul_f32_e32 v2, v3, v4
; TONGA-NEXT: v_trunc_f32_e32 v2, v2
; TONGA-NEXT: v_mad_f32 v3, -v2, v1, v3
; TONGA-NEXT: v_cvt_i32_f32_e32 v2, v2
; TONGA-NEXT: v_cmp_ge_f32_e64 vcc, |v3|, |v1|
; TONGA-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
; TONGA-NEXT: v_add_u32_e32 v0, vcc, v0, v2
; TONGA-NEXT: v_bfe_i32 v0, v0, 0, 24
; TONGA-NEXT: buffer_store_dword v0, off, s[4:7], 0
; TONGA-NEXT: s_endpgm
;
; GFX9-LABEL: v_sdiv_i24:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24
; GFX9-NEXT: s_mov_b32 s7, 0xf000
; GFX9-NEXT: s_mov_b32 s6, -1
; GFX9-NEXT: s_mov_b32 s10, s6
; GFX9-NEXT: s_mov_b32 s11, s7
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_mov_b32 s8, s2
; GFX9-NEXT: s_mov_b32 s9, s3
; GFX9-NEXT: buffer_load_sbyte v0, off, s[8:11], 0 offset:6
; GFX9-NEXT: buffer_load_ushort v1, off, s[8:11], 0 offset:4
; GFX9-NEXT: buffer_load_sbyte v2, off, s[8:11], 0 offset:2
; GFX9-NEXT: buffer_load_ushort v3, off, s[8:11], 0
; GFX9-NEXT: s_mov_b32 s4, s0
; GFX9-NEXT: s_mov_b32 s5, s1
; GFX9-NEXT: s_waitcnt vmcnt(3)
; GFX9-NEXT: v_lshlrev_b32_e32 v4, 16, v0
; GFX9-NEXT: s_waitcnt vmcnt(2)
; GFX9-NEXT: v_or_b32_e32 v1, v1, v4
; GFX9-NEXT: v_cvt_f32_i32_e32 v1, v1
; GFX9-NEXT: s_waitcnt vmcnt(1)
; GFX9-NEXT: v_lshlrev_b32_e32 v4, 16, v2
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_or_b32_e32 v3, v3, v4
; GFX9-NEXT: v_cvt_f32_i32_e32 v3, v3
; GFX9-NEXT: v_rcp_iflag_f32_e32 v4, v1
; GFX9-NEXT: v_xor_b32_e32 v0, v2, v0
; GFX9-NEXT: v_ashrrev_i32_e32 v0, 30, v0
; GFX9-NEXT: v_or_b32_e32 v0, 1, v0
; GFX9-NEXT: v_mul_f32_e32 v2, v3, v4
; GFX9-NEXT: v_trunc_f32_e32 v2, v2
; GFX9-NEXT: v_cvt_i32_f32_e32 v4, v2
; GFX9-NEXT: v_mad_f32 v2, -v2, v1, v3
; GFX9-NEXT: v_cmp_ge_f32_e64 vcc, |v2|, |v1|
; GFX9-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
; GFX9-NEXT: v_add_u32_e32 v0, v4, v0
; GFX9-NEXT: v_bfe_i32 v0, v0, 0, 24
; GFX9-NEXT: buffer_store_dword v0, off, s[4:7], 0
; GFX9-NEXT: s_endpgm
;
; EG-LABEL: v_sdiv_i24:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @14, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 3 @6
; EG-NEXT: ALU 29, @15, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.X, T1.X, 1
; EG-NEXT: CF_END
; EG-NEXT: PAD
; EG-NEXT: Fetch clause starting at 6:
; EG-NEXT: VTX_READ_8 T1.X, T0.X, 6, #1
; EG-NEXT: VTX_READ_16 T2.X, T0.X, 0, #1
; EG-NEXT: VTX_READ_16 T3.X, T0.X, 4, #1
; EG-NEXT: VTX_READ_8 T0.X, T0.X, 2, #1
; EG-NEXT: ALU clause starting at 14:
; EG-NEXT: MOV * T0.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 15:
; EG-NEXT: BFE_INT * T0.W, T1.X, 0.0, literal.x,
; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00)
; EG-NEXT: LSHL * T1.W, PV.W, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: BFE_INT T2.W, T0.X, 0.0, literal.x,
; EG-NEXT: OR_INT * T1.W, T3.X, PV.W,
; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00)
; EG-NEXT: LSHL T3.W, PV.W, literal.x,
; EG-NEXT: INT_TO_FLT * T0.X, PS,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: OR_INT T1.W, T2.X, PV.W,
; EG-NEXT: RECIP_IEEE * T0.Y, PS,
; EG-NEXT: INT_TO_FLT * T0.Z, PV.W,
; EG-NEXT: MUL_IEEE * T1.W, PS, T0.Y,
; EG-NEXT: TRUNC T1.W, PV.W,
; EG-NEXT: XOR_INT * T0.W, T2.W, T0.W,
; EG-NEXT: ASHR T0.W, PS, literal.x,
; EG-NEXT: MULADD_IEEE * T2.W, -PV.W, T0.X, T0.Z,
; EG-NEXT: 30(4.203895e-44), 0(0.000000e+00)
; EG-NEXT: TRUNC T0.Z, T1.W,
; EG-NEXT: SETGE T1.W, |PS|, |T0.X|,
; EG-NEXT: OR_INT * T0.W, PV.W, 1,
; EG-NEXT: CNDE T0.W, PV.W, 0.0, PS,
; EG-NEXT: FLT_TO_INT * T1.W, PV.Z,
; EG-NEXT: ADD_INT * T0.W, PS, PV.W,
; EG-NEXT: LSHL * T0.W, PV.W, literal.x,
; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00)
; EG-NEXT: ASHR T0.X, PV.W, literal.x,
; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.y,
; EG-NEXT: 8(1.121039e-44), 2(2.802597e-45)
%den_ptr = getelementptr i24, i24 addrspace(1)* %in, i24 1
%num = load i24, i24 addrspace(1) * %in
%den = load i24, i24 addrspace(1) * %den_ptr
%result = sdiv i24 %num, %den
%result.ext = sext i24 %result to i32
store i32 %result.ext, i32 addrspace(1)* %out
ret void
}
define amdgpu_kernel void @v_sdiv_i25(i32 addrspace(1)* %out, i25 addrspace(1)* %in) {
; GCN-LABEL: v_sdiv_i25:
; GCN: ; %bb.0:
; GCN-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s6, -1
; GCN-NEXT: s_mov_b32 s10, s6
; GCN-NEXT: s_mov_b32 s11, s7
; GCN-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s8, s2
; GCN-NEXT: s_mov_b32 s9, s3
; GCN-NEXT: buffer_load_dwordx2 v[0:1], off, s[8:11], 0
; GCN-NEXT: s_mov_b32 s4, s0
; GCN-NEXT: s_mov_b32 s5, s1
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_bfe_i32 v2, v1, 0, 25
; GCN-NEXT: v_bfe_i32 v1, v1, 24, 1
; GCN-NEXT: v_add_i32_e32 v2, vcc, v2, v1
; GCN-NEXT: v_xor_b32_e32 v2, v2, v1
; GCN-NEXT: v_cvt_f32_u32_e32 v3, v2
; GCN-NEXT: v_sub_i32_e32 v4, vcc, 0, v2
; GCN-NEXT: v_bfe_i32 v5, v0, 0, 25
; GCN-NEXT: v_rcp_iflag_f32_e32 v3, v3
; GCN-NEXT: v_bfe_i32 v0, v0, 24, 1
; GCN-NEXT: v_add_i32_e32 v5, vcc, v5, v0
; GCN-NEXT: v_mul_f32_e32 v3, 0x4f7ffffe, v3
; GCN-NEXT: v_cvt_u32_f32_e32 v3, v3
; GCN-NEXT: v_xor_b32_e32 v5, v5, v0
; GCN-NEXT: v_xor_b32_e32 v0, v0, v1
; GCN-NEXT: v_mul_lo_u32 v4, v4, v3
; GCN-NEXT: v_mul_hi_u32 v4, v3, v4
; GCN-NEXT: v_add_i32_e32 v3, vcc, v3, v4
; GCN-NEXT: v_mul_hi_u32 v3, v5, v3
; GCN-NEXT: v_mul_lo_u32 v1, v3, v2
; GCN-NEXT: v_add_i32_e32 v4, vcc, 1, v3
; GCN-NEXT: v_sub_i32_e32 v1, vcc, v5, v1
; GCN-NEXT: v_subrev_i32_e32 v5, vcc, v2, v1
; GCN-NEXT: v_cmp_ge_u32_e32 vcc, v1, v2
; GCN-NEXT: v_cndmask_b32_e32 v3, v3, v4, vcc
; GCN-NEXT: v_cndmask_b32_e32 v1, v1, v5, vcc
; GCN-NEXT: v_add_i32_e32 v4, vcc, 1, v3
; GCN-NEXT: v_cmp_ge_u32_e32 vcc, v1, v2
; GCN-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
; GCN-NEXT: v_xor_b32_e32 v1, v1, v0
; GCN-NEXT: v_subrev_i32_e32 v0, vcc, v0, v1
; GCN-NEXT: v_bfe_i32 v0, v0, 0, 25
; GCN-NEXT: buffer_store_dword v0, off, s[4:7], 0
; GCN-NEXT: s_endpgm
;
; TONGA-LABEL: v_sdiv_i25:
; TONGA: ; %bb.0:
; TONGA-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24
; TONGA-NEXT: s_mov_b32 s7, 0xf000
; TONGA-NEXT: s_mov_b32 s6, -1
; TONGA-NEXT: s_mov_b32 s10, s6
; TONGA-NEXT: s_mov_b32 s11, s7
; TONGA-NEXT: s_waitcnt lgkmcnt(0)
; TONGA-NEXT: s_mov_b32 s8, s2
; TONGA-NEXT: s_mov_b32 s9, s3
; TONGA-NEXT: buffer_load_dwordx2 v[0:1], off, s[8:11], 0
; TONGA-NEXT: s_mov_b32 s4, s0
; TONGA-NEXT: s_mov_b32 s5, s1
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: v_bfe_i32 v2, v1, 0, 25
; TONGA-NEXT: v_bfe_i32 v1, v1, 24, 1
; TONGA-NEXT: v_add_u32_e32 v2, vcc, v2, v1
; TONGA-NEXT: v_xor_b32_e32 v2, v2, v1
; TONGA-NEXT: v_cvt_f32_u32_e32 v3, v2
; TONGA-NEXT: v_sub_u32_e32 v4, vcc, 0, v2
; TONGA-NEXT: v_bfe_i32 v5, v0, 0, 25
; TONGA-NEXT: v_rcp_iflag_f32_e32 v3, v3
; TONGA-NEXT: v_bfe_i32 v0, v0, 24, 1
; TONGA-NEXT: v_add_u32_e32 v5, vcc, v5, v0
; TONGA-NEXT: v_mul_f32_e32 v3, 0x4f7ffffe, v3
; TONGA-NEXT: v_cvt_u32_f32_e32 v3, v3
; TONGA-NEXT: v_xor_b32_e32 v5, v5, v0
; TONGA-NEXT: v_xor_b32_e32 v0, v0, v1
; TONGA-NEXT: v_mul_lo_u32 v4, v4, v3
; TONGA-NEXT: v_mul_hi_u32 v4, v3, v4
; TONGA-NEXT: v_add_u32_e32 v3, vcc, v3, v4
; TONGA-NEXT: v_mul_hi_u32 v3, v5, v3
; TONGA-NEXT: v_mul_lo_u32 v1, v3, v2
; TONGA-NEXT: v_add_u32_e32 v4, vcc, 1, v3
; TONGA-NEXT: v_sub_u32_e32 v1, vcc, v5, v1
; TONGA-NEXT: v_subrev_u32_e32 v5, vcc, v2, v1
; TONGA-NEXT: v_cmp_ge_u32_e32 vcc, v1, v2
; TONGA-NEXT: v_cndmask_b32_e32 v3, v3, v4, vcc
; TONGA-NEXT: v_cndmask_b32_e32 v1, v1, v5, vcc
; TONGA-NEXT: v_add_u32_e32 v4, vcc, 1, v3
; TONGA-NEXT: v_cmp_ge_u32_e32 vcc, v1, v2
; TONGA-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
; TONGA-NEXT: v_xor_b32_e32 v1, v1, v0
; TONGA-NEXT: v_subrev_u32_e32 v0, vcc, v0, v1
; TONGA-NEXT: v_bfe_i32 v0, v0, 0, 25
; TONGA-NEXT: buffer_store_dword v0, off, s[4:7], 0
; TONGA-NEXT: s_endpgm
;
; GFX9-LABEL: v_sdiv_i25:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24
; GFX9-NEXT: s_mov_b32 s7, 0xf000
; GFX9-NEXT: s_mov_b32 s6, -1
; GFX9-NEXT: s_mov_b32 s10, s6
; GFX9-NEXT: s_mov_b32 s11, s7
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_mov_b32 s8, s2
; GFX9-NEXT: s_mov_b32 s9, s3
; GFX9-NEXT: buffer_load_dwordx2 v[0:1], off, s[8:11], 0
; GFX9-NEXT: s_mov_b32 s4, s0
; GFX9-NEXT: s_mov_b32 s5, s1
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_readfirstlane_b32 s2, v1
; GFX9-NEXT: s_bfe_i32 s3, s2, 0x190000
; GFX9-NEXT: s_bfe_i32 s2, s2, 0x10018
; GFX9-NEXT: s_add_i32 s3, s3, s2
; GFX9-NEXT: s_xor_b32 s3, s3, s2
; GFX9-NEXT: v_cvt_f32_u32_e32 v1, s3
; GFX9-NEXT: v_readfirstlane_b32 s0, v0
; GFX9-NEXT: s_bfe_i32 s1, s0, 0x190000
; GFX9-NEXT: s_bfe_i32 s0, s0, 0x10018
; GFX9-NEXT: v_rcp_iflag_f32_e32 v0, v1
; GFX9-NEXT: s_add_i32 s1, s1, s0
; GFX9-NEXT: s_xor_b32 s2, s0, s2
; GFX9-NEXT: s_xor_b32 s0, s1, s0
; GFX9-NEXT: v_mul_f32_e32 v0, 0x4f7ffffe, v0
; GFX9-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX9-NEXT: s_sub_i32 s1, 0, s3
; GFX9-NEXT: v_readfirstlane_b32 s8, v0
; GFX9-NEXT: s_mul_i32 s1, s1, s8
; GFX9-NEXT: s_mul_hi_u32 s1, s8, s1
; GFX9-NEXT: s_add_i32 s8, s8, s1
; GFX9-NEXT: s_mul_hi_u32 s1, s0, s8
; GFX9-NEXT: s_mul_i32 s8, s1, s3
; GFX9-NEXT: s_sub_i32 s0, s0, s8
; GFX9-NEXT: s_add_i32 s9, s1, 1
; GFX9-NEXT: s_sub_i32 s8, s0, s3
; GFX9-NEXT: s_cmp_ge_u32 s0, s3
; GFX9-NEXT: s_cselect_b32 s1, s9, s1
; GFX9-NEXT: s_cselect_b32 s0, s8, s0
; GFX9-NEXT: s_add_i32 s8, s1, 1
; GFX9-NEXT: s_cmp_ge_u32 s0, s3
; GFX9-NEXT: s_cselect_b32 s0, s8, s1
; GFX9-NEXT: s_xor_b32 s0, s0, s2
; GFX9-NEXT: s_sub_i32 s0, s0, s2
; GFX9-NEXT: s_bfe_i32 s0, s0, 0x190000
; GFX9-NEXT: v_mov_b32_e32 v0, s0
; GFX9-NEXT: buffer_store_dword v0, off, s[4:7], 0
; GFX9-NEXT: s_endpgm
;
; EG-LABEL: v_sdiv_i25:
; EG: ; %bb.0:
; EG-NEXT: ALU 1, @10, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 1 @6
; EG-NEXT: ALU 37, @12, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.X, T1.X, 1
; EG-NEXT: CF_END
; EG-NEXT: PAD
; EG-NEXT: Fetch clause starting at 6:
; EG-NEXT: VTX_READ_32 T0.X, T0.X, 4, #1
; EG-NEXT: VTX_READ_32 T1.X, T1.X, 0, #1
; EG-NEXT: ALU clause starting at 10:
; EG-NEXT: MOV * T0.X, KC0[2].Z,
; EG-NEXT: MOV * T1.X, PV.X,
; EG-NEXT: ALU clause starting at 12:
; EG-NEXT: LSHL * T0.W, T0.X, literal.x,
; EG-NEXT: 7(9.809089e-45), 0(0.000000e+00)
; EG-NEXT: ASHR * T0.W, PV.W, literal.x,
; EG-NEXT: 7(9.809089e-45), 0(0.000000e+00)
; EG-NEXT: SETGT_INT * T1.W, 0.0, PV.W,
; EG-NEXT: ADD_INT T0.W, T0.W, PV.W,
; EG-NEXT: LSHL * T2.W, T1.X, literal.x,
; EG-NEXT: 7(9.809089e-45), 0(0.000000e+00)
; EG-NEXT: XOR_INT * T0.W, PV.W, T1.W,
; EG-NEXT: SUB_INT T0.Z, 0.0, PV.W,
; EG-NEXT: ASHR T2.W, T2.W, literal.x,
; EG-NEXT: RECIP_UINT * T0.X, PV.W,
; EG-NEXT: 7(9.809089e-45), 0(0.000000e+00)
; EG-NEXT: SETGT_INT T3.W, 0.0, PV.W,
; EG-NEXT: MULLO_INT * T0.Y, PV.Z, PS,
; EG-NEXT: ADD_INT T2.W, T2.W, PV.W,
; EG-NEXT: MULHI * T0.Y, T0.X, PS,
; EG-NEXT: ADD_INT T4.W, T0.X, PS,
; EG-NEXT: XOR_INT * T2.W, PV.W, T3.W,
; EG-NEXT: MULHI * T0.X, PS, PV.W,
; EG-NEXT: MULLO_INT * T0.Y, PS, T0.W,
; EG-NEXT: SUB_INT * T2.W, T2.W, PS,
; EG-NEXT: ADD_INT T0.Z, T0.X, 1,
; EG-NEXT: SETGE_UINT T4.W, PV.W, T0.W,
; EG-NEXT: SUB_INT * T5.W, PV.W, T0.W,
; EG-NEXT: CNDE_INT T2.W, PV.W, T2.W, PS,
; EG-NEXT: CNDE_INT * T4.W, PV.W, T0.X, PV.Z,
; EG-NEXT: ADD_INT T5.W, PS, 1,
; EG-NEXT: SETGE_UINT * T0.W, PV.W, T0.W,
; EG-NEXT: CNDE_INT T0.W, PS, T4.W, PV.W, BS:VEC_102/SCL_221
; EG-NEXT: XOR_INT * T1.W, T3.W, T1.W,
; EG-NEXT: XOR_INT * T0.W, PV.W, PS,
; EG-NEXT: SUB_INT * T0.W, PV.W, T1.W,
; EG-NEXT: LSHL * T0.W, PV.W, literal.x,
; EG-NEXT: 7(9.809089e-45), 0(0.000000e+00)
; EG-NEXT: ASHR T0.X, PV.W, literal.x,
; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.y,
; EG-NEXT: 7(9.809089e-45), 2(2.802597e-45)
%den_ptr = getelementptr i25, i25 addrspace(1)* %in, i25 1
%num = load i25, i25 addrspace(1) * %in
%den = load i25, i25 addrspace(1) * %den_ptr
%result = sdiv i25 %num, %den
%result.ext = sext i25 %result to i32
store i32 %result.ext, i32 addrspace(1)* %out
ret void
}
; Tests for 64-bit divide bypass.
; define amdgpu_kernel void @test_get_quotient(i64 addrspace(1)* %out, i64 %a, i64 %b) nounwind {
; %result = sdiv i64 %a, %b
; store i64 %result, i64 addrspace(1)* %out, align 8
; ret void
; }
; define amdgpu_kernel void @test_get_remainder(i64 addrspace(1)* %out, i64 %a, i64 %b) nounwind {
; %result = srem i64 %a, %b
; store i64 %result, i64 addrspace(1)* %out, align 8
; ret void
; }
; define amdgpu_kernel void @test_get_quotient_and_remainder(i64 addrspace(1)* %out, i64 %a, i64 %b) nounwind {
; %resultdiv = sdiv i64 %a, %b
; %resultrem = srem i64 %a, %b
; %result = add i64 %resultdiv, %resultrem
; store i64 %result, i64 addrspace(1)* %out, align 8
; ret void
; }
define amdgpu_kernel void @scalarize_mulhs_4xi32(<4 x i32> addrspace(1)* nocapture readonly %in, <4 x i32> addrspace(1)* nocapture %out) {
; GCN-LABEL: scalarize_mulhs_4xi32:
; GCN: ; %bb.0:
; GCN-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s6, -1
; GCN-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s4, s0
; GCN-NEXT: s_mov_b32 s5, s1
; GCN-NEXT: buffer_load_dwordx4 v[0:3], off, s[4:7], 0
; GCN-NEXT: s_mov_b32 s0, 0x1389c755
; GCN-NEXT: s_mov_b32 s4, s2
; GCN-NEXT: s_mov_b32 s5, s3
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_hi_i32 v0, v0, s0
; GCN-NEXT: v_mul_hi_i32 v1, v1, s0
; GCN-NEXT: v_mul_hi_i32 v2, v2, s0
; GCN-NEXT: v_mul_hi_i32 v3, v3, s0
; GCN-NEXT: v_lshrrev_b32_e32 v4, 31, v0
; GCN-NEXT: v_ashrrev_i32_e32 v0, 12, v0
; GCN-NEXT: v_lshrrev_b32_e32 v5, 31, v1
; GCN-NEXT: v_ashrrev_i32_e32 v1, 12, v1
; GCN-NEXT: v_lshrrev_b32_e32 v6, 31, v2
; GCN-NEXT: v_ashrrev_i32_e32 v2, 12, v2
; GCN-NEXT: v_lshrrev_b32_e32 v7, 31, v3
; GCN-NEXT: v_ashrrev_i32_e32 v3, 12, v3
; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v4
; GCN-NEXT: v_add_i32_e32 v1, vcc, v1, v5
; GCN-NEXT: v_add_i32_e32 v2, vcc, v6, v2
; GCN-NEXT: v_add_i32_e32 v3, vcc, v3, v7
; GCN-NEXT: buffer_store_dwordx4 v[0:3], off, s[4:7], 0
; GCN-NEXT: s_endpgm
;
; TONGA-LABEL: scalarize_mulhs_4xi32:
; TONGA: ; %bb.0:
; TONGA-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24
; TONGA-NEXT: s_mov_b32 s7, 0xf000
; TONGA-NEXT: s_mov_b32 s6, -1
; TONGA-NEXT: s_waitcnt lgkmcnt(0)
; TONGA-NEXT: s_mov_b32 s4, s0
; TONGA-NEXT: s_mov_b32 s5, s1
; TONGA-NEXT: buffer_load_dwordx4 v[0:3], off, s[4:7], 0
; TONGA-NEXT: s_mov_b32 s0, 0x1389c755
; TONGA-NEXT: s_mov_b32 s4, s2
; TONGA-NEXT: s_mov_b32 s5, s3
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: v_mul_hi_i32 v0, v0, s0
; TONGA-NEXT: v_mul_hi_i32 v1, v1, s0
; TONGA-NEXT: v_mul_hi_i32 v2, v2, s0
; TONGA-NEXT: v_mul_hi_i32 v3, v3, s0
; TONGA-NEXT: v_lshrrev_b32_e32 v4, 31, v0
; TONGA-NEXT: v_ashrrev_i32_e32 v0, 12, v0
; TONGA-NEXT: v_lshrrev_b32_e32 v5, 31, v1
; TONGA-NEXT: v_ashrrev_i32_e32 v1, 12, v1
; TONGA-NEXT: v_lshrrev_b32_e32 v6, 31, v2
; TONGA-NEXT: v_ashrrev_i32_e32 v2, 12, v2
; TONGA-NEXT: v_lshrrev_b32_e32 v7, 31, v3
; TONGA-NEXT: v_ashrrev_i32_e32 v3, 12, v3
; TONGA-NEXT: v_add_u32_e32 v0, vcc, v0, v4
; TONGA-NEXT: v_add_u32_e32 v1, vcc, v1, v5
; TONGA-NEXT: v_add_u32_e32 v2, vcc, v6, v2
; TONGA-NEXT: v_add_u32_e32 v3, vcc, v3, v7
; TONGA-NEXT: buffer_store_dwordx4 v[0:3], off, s[4:7], 0
; TONGA-NEXT: s_endpgm
;
; GFX9-LABEL: scalarize_mulhs_4xi32:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24
; GFX9-NEXT: s_mov_b32 s7, 0xf000
; GFX9-NEXT: s_mov_b32 s6, -1
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_mov_b32 s4, s0
; GFX9-NEXT: s_mov_b32 s5, s1
; GFX9-NEXT: buffer_load_dwordx4 v[0:3], off, s[4:7], 0
; GFX9-NEXT: s_mov_b32 s0, 0x1389c755
; GFX9-NEXT: s_mov_b32 s4, s2
; GFX9-NEXT: s_mov_b32 s5, s3
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_mul_hi_i32 v0, v0, s0
; GFX9-NEXT: v_mul_hi_i32 v1, v1, s0
; GFX9-NEXT: v_mul_hi_i32 v2, v2, s0
; GFX9-NEXT: v_mul_hi_i32 v3, v3, s0
; GFX9-NEXT: v_lshrrev_b32_e32 v4, 31, v0
; GFX9-NEXT: v_ashrrev_i32_e32 v0, 12, v0
; GFX9-NEXT: v_lshrrev_b32_e32 v5, 31, v1
; GFX9-NEXT: v_ashrrev_i32_e32 v1, 12, v1
; GFX9-NEXT: v_lshrrev_b32_e32 v6, 31, v2
; GFX9-NEXT: v_ashrrev_i32_e32 v2, 12, v2
; GFX9-NEXT: v_lshrrev_b32_e32 v7, 31, v3
; GFX9-NEXT: v_ashrrev_i32_e32 v3, 12, v3
; GFX9-NEXT: v_add_u32_e32 v0, v0, v4
; GFX9-NEXT: v_add_u32_e32 v1, v1, v5
; GFX9-NEXT: v_add_u32_e32 v2, v2, v6
; GFX9-NEXT: v_add_u32_e32 v3, v3, v7
; GFX9-NEXT: buffer_store_dwordx4 v[0:3], off, s[4:7], 0
; GFX9-NEXT: s_endpgm
;
; EG-LABEL: scalarize_mulhs_4xi32:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 0 @6
; EG-NEXT: ALU 25, @9, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.XYZW, T1.X, 1
; EG-NEXT: CF_END
; EG-NEXT: PAD
; EG-NEXT: Fetch clause starting at 6:
; EG-NEXT: VTX_READ_128 T0.XYZW, T0.X, 0, #1
; EG-NEXT: ALU clause starting at 8:
; EG-NEXT: MOV * T0.X, KC0[2].Y,
; EG-NEXT: ALU clause starting at 9:
; EG-NEXT: MULHI_INT * T0.W, T0.W, literal.x,
; EG-NEXT: 327796565(3.478022e-27), 0(0.000000e+00)
; EG-NEXT: ASHR T1.Z, PS, literal.x,
; EG-NEXT: LSHR T0.W, PS, literal.y,
; EG-NEXT: MULHI_INT * T0.Z, T0.Z, literal.z,
; EG-NEXT: 12(1.681558e-44), 31(4.344025e-44)
; EG-NEXT: 327796565(3.478022e-27), 0(0.000000e+00)
; EG-NEXT: ASHR T1.Y, PS, literal.x,
; EG-NEXT: LSHR T0.Z, PS, literal.y,
; EG-NEXT: ADD_INT T0.W, PV.Z, PV.W,
; EG-NEXT: MULHI_INT * T0.Y, T0.Y, literal.z,
; EG-NEXT: 12(1.681558e-44), 31(4.344025e-44)
; EG-NEXT: 327796565(3.478022e-27), 0(0.000000e+00)
; EG-NEXT: ASHR T2.Y, PS, literal.x,
; EG-NEXT: ADD_INT T0.Z, PV.Y, PV.Z,
; EG-NEXT: LSHR T1.W, PS, literal.y,
; EG-NEXT: MULHI_INT * T0.X, T0.X, literal.z,
; EG-NEXT: 12(1.681558e-44), 31(4.344025e-44)
; EG-NEXT: 327796565(3.478022e-27), 0(0.000000e+00)
; EG-NEXT: ADD_INT T0.Y, PV.Y, PV.W,
; EG-NEXT: ASHR T1.W, PS, literal.x,
; EG-NEXT: LSHR * T2.W, PS, literal.y,
; EG-NEXT: 12(1.681558e-44), 31(4.344025e-44)
; EG-NEXT: ADD_INT T0.X, PV.W, PS,
; EG-NEXT: LSHR * T1.X, KC0[2].Z, literal.x,
; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00)
%1 = load <4 x i32>, <4 x i32> addrspace(1)* %in, align 16
%2 = sdiv <4 x i32> %1, <i32 53668, i32 53668, i32 53668, i32 53668>
store <4 x i32> %2, <4 x i32> addrspace(1)* %out, align 16
ret void
}
Reference in New Issue View Git Blame Copy Permalink