blob: 80407f5de04a162e55114bb30a180294ab972c77 [file] [edit]
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgpu6.01 < %s | FileCheck -check-prefixes=SI %s
; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgpu8.02 -mattr=-flat-for-global < %s | FileCheck -check-prefixes=VI %s
; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgpu9.00 -mattr=-flat-for-global < %s | FileCheck -check-prefixes=GFX9 %s
; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgpu10.10 -mattr=-flat-for-global < %s | FileCheck -check-prefixes=GFX10 %s
; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgpu11.00 -mattr=-flat-for-global < %s | FileCheck -check-prefixes=GFX11 %s
; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgpu12.00 -mattr=-flat-for-global < %s | FileCheck -check-prefixes=GFX12 %s
; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgpu12.50 -mattr=-flat-for-global < %s | FileCheck -check-prefixes=GFX1250 %s
; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=r600 -mcpu=redwood < %s | FileCheck -check-prefixes=EG %s
define amdgpu_kernel void @test_clmul_i32(ptr addrspace(1) %out, ptr addrspace(1) %in) {
; SI-LABEL: test_clmul_i32:
; SI: ; %bb.0:
; SI-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9
; SI-NEXT: s_mov_b32 s3, 0xf000
; SI-NEXT: s_mov_b32 s2, -1
; SI-NEXT: s_mov_b32 s10, s2
; SI-NEXT: s_mov_b32 s11, s3
; SI-NEXT: s_waitcnt lgkmcnt(0)
; SI-NEXT: s_mov_b32 s8, s6
; SI-NEXT: s_mov_b32 s9, s7
; SI-NEXT: buffer_load_dwordx2 v[0:1], off, s[8:11], 0
; SI-NEXT: s_mov_b32 s0, s4
; SI-NEXT: s_mov_b32 s1, s5
; SI-NEXT: s_waitcnt vmcnt(0)
; SI-NEXT: v_readfirstlane_b32 s4, v0
; SI-NEXT: v_readfirstlane_b32 s5, v1
; SI-NEXT: s_and_b32 s6, s5, 0x11111111
; SI-NEXT: s_and_b32 s7, s4, 0x22222222
; SI-NEXT: s_and_b32 s8, s5, 0x22222222
; SI-NEXT: s_and_b32 s9, s4, 0x11111111
; SI-NEXT: s_and_b32 s10, s5, 0x88888888
; SI-NEXT: s_and_b32 s11, s4, 0x44444444
; SI-NEXT: s_and_b32 s5, s5, 0x44444444
; SI-NEXT: s_and_b32 s4, s4, 0x88888888
; SI-NEXT: s_mul_i32 s12, s7, s6
; SI-NEXT: s_mul_i32 s13, s9, s8
; SI-NEXT: s_mul_i32 s16, s7, s10
; SI-NEXT: s_mul_i32 s17, s9, s6
; SI-NEXT: s_mul_i32 s14, s11, s10
; SI-NEXT: s_mul_i32 s15, s4, s5
; SI-NEXT: s_mul_i32 s18, s11, s5
; SI-NEXT: s_mul_i32 s19, s4, s8
; SI-NEXT: s_mul_i32 s20, s7, s8
; SI-NEXT: s_mul_i32 s21, s9, s5
; SI-NEXT: s_mul_i32 s22, s11, s6
; SI-NEXT: s_mul_i32 s23, s4, s10
; SI-NEXT: s_mul_i32 s7, s7, s5
; SI-NEXT: s_mul_i32 s4, s4, s6
; SI-NEXT: s_xor_b32 s5, s13, s12
; SI-NEXT: s_xor_b32 s6, s17, s16
; SI-NEXT: s_mul_i32 s9, s9, s10
; SI-NEXT: s_mul_i32 s11, s11, s8
; SI-NEXT: s_xor_b32 s8, s21, s20
; SI-NEXT: s_xor_b32 s5, s5, s14
; SI-NEXT: s_xor_b32 s6, s6, s18
; SI-NEXT: s_xor_b32 s7, s9, s7
; SI-NEXT: s_xor_b32 s8, s8, s22
; SI-NEXT: s_xor_b32 s5, s5, s15
; SI-NEXT: s_xor_b32 s6, s6, s19
; SI-NEXT: s_xor_b32 s7, s7, s11
; SI-NEXT: s_xor_b32 s8, s8, s23
; SI-NEXT: s_and_b32 s5, s5, 0x22222222
; SI-NEXT: s_and_b32 s6, s6, 0x11111111
; SI-NEXT: s_xor_b32 s4, s7, s4
; SI-NEXT: s_and_b32 s7, s8, 0x44444444
; SI-NEXT: s_or_b32 s5, s6, s5
; SI-NEXT: s_or_b32 s5, s5, s7
; SI-NEXT: s_and_b32 s4, s4, 0x88888888
; SI-NEXT: s_or_b32 s4, s5, s4
; SI-NEXT: v_mov_b32_e32 v0, s4
; SI-NEXT: buffer_store_dword v0, off, s[0:3], 0
; SI-NEXT: s_endpgm
;
; VI-LABEL: test_clmul_i32:
; VI: ; %bb.0:
; VI-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x24
; VI-NEXT: s_mov_b32 s3, 0xf000
; VI-NEXT: s_mov_b32 s2, -1
; VI-NEXT: s_mov_b32 s10, s2
; VI-NEXT: s_mov_b32 s11, s3
; VI-NEXT: s_waitcnt lgkmcnt(0)
; VI-NEXT: s_mov_b32 s8, s6
; VI-NEXT: s_mov_b32 s9, s7
; VI-NEXT: buffer_load_dwordx2 v[0:1], off, s[8:11], 0
; VI-NEXT: s_mov_b32 s0, s4
; VI-NEXT: s_mov_b32 s1, s5
; VI-NEXT: s_waitcnt vmcnt(0)
; VI-NEXT: v_readfirstlane_b32 s4, v0
; VI-NEXT: v_readfirstlane_b32 s5, v1
; VI-NEXT: s_and_b32 s6, s5, 0x11111111
; VI-NEXT: s_and_b32 s7, s4, 0x22222222
; VI-NEXT: s_and_b32 s8, s5, 0x22222222
; VI-NEXT: s_and_b32 s9, s4, 0x11111111
; VI-NEXT: s_and_b32 s10, s5, 0x88888888
; VI-NEXT: s_and_b32 s11, s4, 0x44444444
; VI-NEXT: s_and_b32 s5, s5, 0x44444444
; VI-NEXT: s_and_b32 s4, s4, 0x88888888
; VI-NEXT: s_mul_i32 s12, s7, s6
; VI-NEXT: s_mul_i32 s13, s9, s8
; VI-NEXT: s_mul_i32 s16, s7, s10
; VI-NEXT: s_mul_i32 s17, s9, s6
; VI-NEXT: s_mul_i32 s14, s11, s10
; VI-NEXT: s_mul_i32 s15, s4, s5
; VI-NEXT: s_mul_i32 s18, s11, s5
; VI-NEXT: s_mul_i32 s19, s4, s8
; VI-NEXT: s_mul_i32 s20, s7, s8
; VI-NEXT: s_mul_i32 s21, s9, s5
; VI-NEXT: s_mul_i32 s22, s11, s6
; VI-NEXT: s_mul_i32 s23, s4, s10
; VI-NEXT: s_mul_i32 s7, s7, s5
; VI-NEXT: s_mul_i32 s4, s4, s6
; VI-NEXT: s_xor_b32 s5, s13, s12
; VI-NEXT: s_xor_b32 s6, s17, s16
; VI-NEXT: s_mul_i32 s9, s9, s10
; VI-NEXT: s_mul_i32 s11, s11, s8
; VI-NEXT: s_xor_b32 s8, s21, s20
; VI-NEXT: s_xor_b32 s5, s5, s14
; VI-NEXT: s_xor_b32 s6, s6, s18
; VI-NEXT: s_xor_b32 s7, s9, s7
; VI-NEXT: s_xor_b32 s8, s8, s22
; VI-NEXT: s_xor_b32 s5, s5, s15
; VI-NEXT: s_xor_b32 s6, s6, s19
; VI-NEXT: s_xor_b32 s7, s7, s11
; VI-NEXT: s_xor_b32 s8, s8, s23
; VI-NEXT: s_and_b32 s5, s5, 0x22222222
; VI-NEXT: s_and_b32 s6, s6, 0x11111111
; VI-NEXT: s_xor_b32 s4, s7, s4
; VI-NEXT: s_and_b32 s7, s8, 0x44444444
; VI-NEXT: s_or_b32 s5, s6, s5
; VI-NEXT: s_or_b32 s5, s5, s7
; VI-NEXT: s_and_b32 s4, s4, 0x88888888
; VI-NEXT: s_or_b32 s4, s5, s4
; VI-NEXT: v_mov_b32_e32 v0, s4
; VI-NEXT: buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT: s_endpgm
;
; GFX9-LABEL: test_clmul_i32:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_load_dwordx4 s[8:11], s[4:5], 0x24
; GFX9-NEXT: s_mov_b32 s3, 0xf000
; GFX9-NEXT: s_mov_b32 s2, -1
; GFX9-NEXT: s_mov_b32 s6, s2
; GFX9-NEXT: s_mov_b32 s7, s3
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_mov_b32 s4, s10
; GFX9-NEXT: s_mov_b32 s5, s11
; GFX9-NEXT: buffer_load_dwordx2 v[0:1], off, s[4:7], 0
; GFX9-NEXT: s_mov_b32 s0, s8
; GFX9-NEXT: s_mov_b32 s1, s9
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_readfirstlane_b32 s4, v0
; GFX9-NEXT: v_readfirstlane_b32 s5, v1
; GFX9-NEXT: s_and_b32 s6, s5, 0x11111111
; GFX9-NEXT: s_and_b32 s7, s4, 0x22222222
; GFX9-NEXT: s_and_b32 s8, s5, 0x22222222
; GFX9-NEXT: s_and_b32 s9, s4, 0x11111111
; GFX9-NEXT: s_and_b32 s10, s5, 0x88888888
; GFX9-NEXT: s_and_b32 s11, s4, 0x44444444
; GFX9-NEXT: s_and_b32 s5, s5, 0x44444444
; GFX9-NEXT: s_and_b32 s4, s4, 0x88888888
; GFX9-NEXT: s_mul_i32 s12, s7, s6
; GFX9-NEXT: s_mul_i32 s13, s9, s8
; GFX9-NEXT: s_mul_i32 s16, s7, s10
; GFX9-NEXT: s_mul_i32 s17, s9, s6
; GFX9-NEXT: s_mul_i32 s14, s11, s10
; GFX9-NEXT: s_mul_i32 s15, s4, s5
; GFX9-NEXT: s_mul_i32 s18, s11, s5
; GFX9-NEXT: s_mul_i32 s19, s4, s8
; GFX9-NEXT: s_mul_i32 s20, s7, s8
; GFX9-NEXT: s_mul_i32 s21, s9, s5
; GFX9-NEXT: s_mul_i32 s22, s11, s6
; GFX9-NEXT: s_mul_i32 s23, s4, s10
; GFX9-NEXT: s_mul_i32 s7, s7, s5
; GFX9-NEXT: s_mul_i32 s4, s4, s6
; GFX9-NEXT: s_xor_b32 s5, s13, s12
; GFX9-NEXT: s_xor_b32 s6, s17, s16
; GFX9-NEXT: s_mul_i32 s9, s9, s10
; GFX9-NEXT: s_mul_i32 s11, s11, s8
; GFX9-NEXT: s_xor_b32 s8, s21, s20
; GFX9-NEXT: s_xor_b32 s5, s5, s14
; GFX9-NEXT: s_xor_b32 s6, s6, s18
; GFX9-NEXT: s_xor_b32 s7, s9, s7
; GFX9-NEXT: s_xor_b32 s8, s8, s22
; GFX9-NEXT: s_xor_b32 s5, s5, s15
; GFX9-NEXT: s_xor_b32 s6, s6, s19
; GFX9-NEXT: s_xor_b32 s7, s7, s11
; GFX9-NEXT: s_xor_b32 s8, s8, s23
; GFX9-NEXT: s_and_b32 s5, s5, 0x22222222
; GFX9-NEXT: s_and_b32 s6, s6, 0x11111111
; GFX9-NEXT: s_xor_b32 s4, s7, s4
; GFX9-NEXT: s_and_b32 s7, s8, 0x44444444
; GFX9-NEXT: s_or_b32 s5, s6, s5
; GFX9-NEXT: s_or_b32 s5, s5, s7
; GFX9-NEXT: s_and_b32 s4, s4, 0x88888888
; GFX9-NEXT: s_or_b32 s4, s5, s4
; GFX9-NEXT: v_mov_b32_e32 v0, s4
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], 0
; GFX9-NEXT: s_endpgm
;
; GFX10-LABEL: test_clmul_i32:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
; GFX10-NEXT: s_mov_b32 s6, -1
; GFX10-NEXT: s_mov_b32 s7, 0x31016000
; GFX10-NEXT: s_mov_b32 s10, s6
; GFX10-NEXT: s_mov_b32 s11, s7
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: s_mov_b32 s8, s2
; GFX10-NEXT: s_mov_b32 s9, s3
; GFX10-NEXT: buffer_load_dwordx2 v[0:1], off, s[8:11], 0
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: v_readfirstlane_b32 s2, v1
; GFX10-NEXT: v_readfirstlane_b32 s3, v0
; GFX10-NEXT: s_and_b32 s4, s2, 0x11111111
; GFX10-NEXT: s_and_b32 s5, s3, 0x22222222
; GFX10-NEXT: s_and_b32 s8, s2, 0x22222222
; GFX10-NEXT: s_and_b32 s9, s3, 0x11111111
; GFX10-NEXT: s_and_b32 s10, s2, 0x88888888
; GFX10-NEXT: s_and_b32 s11, s3, 0x44444444
; GFX10-NEXT: s_and_b32 s2, s2, 0x44444444
; GFX10-NEXT: s_and_b32 s3, s3, 0x88888888
; GFX10-NEXT: s_mul_i32 s12, s5, s4
; GFX10-NEXT: s_mul_i32 s13, s9, s8
; GFX10-NEXT: s_mul_i32 s16, s5, s10
; GFX10-NEXT: s_mul_i32 s17, s9, s4
; GFX10-NEXT: s_mul_i32 s14, s11, s10
; GFX10-NEXT: s_mul_i32 s15, s3, s2
; GFX10-NEXT: s_mul_i32 s18, s11, s2
; GFX10-NEXT: s_mul_i32 s19, s3, s8
; GFX10-NEXT: s_mul_i32 s20, s5, s8
; GFX10-NEXT: s_mul_i32 s21, s9, s2
; GFX10-NEXT: s_mul_i32 s22, s11, s4
; GFX10-NEXT: s_mul_i32 s23, s3, s10
; GFX10-NEXT: s_mul_i32 s5, s5, s2
; GFX10-NEXT: s_mul_i32 s3, s3, s4
; GFX10-NEXT: s_xor_b32 s2, s13, s12
; GFX10-NEXT: s_xor_b32 s4, s17, s16
; GFX10-NEXT: s_mul_i32 s9, s9, s10
; GFX10-NEXT: s_mul_i32 s11, s11, s8
; GFX10-NEXT: s_xor_b32 s8, s21, s20
; GFX10-NEXT: s_xor_b32 s2, s2, s14
; GFX10-NEXT: s_xor_b32 s4, s4, s18
; GFX10-NEXT: s_xor_b32 s5, s9, s5
; GFX10-NEXT: s_xor_b32 s8, s8, s22
; GFX10-NEXT: s_xor_b32 s2, s2, s15
; GFX10-NEXT: s_xor_b32 s4, s4, s19
; GFX10-NEXT: s_xor_b32 s5, s5, s11
; GFX10-NEXT: s_xor_b32 s8, s8, s23
; GFX10-NEXT: s_and_b32 s2, s2, 0x22222222
; GFX10-NEXT: s_and_b32 s4, s4, 0x11111111
; GFX10-NEXT: s_and_b32 s8, s8, 0x44444444
; GFX10-NEXT: s_or_b32 s2, s4, s2
; GFX10-NEXT: s_xor_b32 s3, s5, s3
; GFX10-NEXT: s_or_b32 s2, s2, s8
; GFX10-NEXT: s_and_b32 s3, s3, 0x88888888
; GFX10-NEXT: s_mov_b32 s4, s0
; GFX10-NEXT: s_or_b32 s2, s2, s3
; GFX10-NEXT: s_mov_b32 s5, s1
; GFX10-NEXT: v_mov_b32_e32 v0, s2
; GFX10-NEXT: buffer_store_dword v0, off, s[4:7], 0
; GFX10-NEXT: s_endpgm
;
; GFX11-LABEL: test_clmul_i32:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT: s_mov_b32 s6, -1
; GFX11-NEXT: s_mov_b32 s7, 0x31016000
; GFX11-NEXT: s_mov_b32 s10, s6
; GFX11-NEXT: s_mov_b32 s11, s7
; GFX11-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-NEXT: s_mov_b32 s8, s2
; GFX11-NEXT: s_mov_b32 s9, s3
; GFX11-NEXT: buffer_load_b64 v[0:1], off, s[8:11], 0
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: v_readfirstlane_b32 s2, v1
; GFX11-NEXT: v_readfirstlane_b32 s3, v0
; GFX11-NEXT: s_and_b32 s4, s2, 0x11111111
; GFX11-NEXT: s_and_b32 s5, s3, 0x22222222
; GFX11-NEXT: s_and_b32 s8, s2, 0x22222222
; GFX11-NEXT: s_and_b32 s9, s3, 0x11111111
; GFX11-NEXT: s_and_b32 s10, s2, 0x88888888
; GFX11-NEXT: s_and_b32 s11, s3, 0x44444444
; GFX11-NEXT: s_and_b32 s2, s2, 0x44444444
; GFX11-NEXT: s_and_b32 s3, s3, 0x88888888
; GFX11-NEXT: s_mul_i32 s12, s5, s4
; GFX11-NEXT: s_mul_i32 s13, s9, s8
; GFX11-NEXT: s_mul_i32 s16, s5, s10
; GFX11-NEXT: s_mul_i32 s17, s9, s4
; GFX11-NEXT: s_mul_i32 s14, s11, s10
; GFX11-NEXT: s_mul_i32 s15, s3, s2
; GFX11-NEXT: s_mul_i32 s18, s11, s2
; GFX11-NEXT: s_mul_i32 s19, s3, s8
; GFX11-NEXT: s_mul_i32 s20, s5, s8
; GFX11-NEXT: s_mul_i32 s21, s9, s2
; GFX11-NEXT: s_mul_i32 s22, s11, s4
; GFX11-NEXT: s_mul_i32 s23, s3, s10
; GFX11-NEXT: s_mul_i32 s5, s5, s2
; GFX11-NEXT: s_mul_i32 s3, s3, s4
; GFX11-NEXT: s_xor_b32 s2, s13, s12
; GFX11-NEXT: s_xor_b32 s4, s17, s16
; GFX11-NEXT: s_mul_i32 s9, s9, s10
; GFX11-NEXT: s_mul_i32 s11, s11, s8
; GFX11-NEXT: s_xor_b32 s8, s21, s20
; GFX11-NEXT: s_xor_b32 s2, s2, s14
; GFX11-NEXT: s_xor_b32 s4, s4, s18
; GFX11-NEXT: s_xor_b32 s5, s9, s5
; GFX11-NEXT: s_xor_b32 s8, s8, s22
; GFX11-NEXT: s_xor_b32 s2, s2, s15
; GFX11-NEXT: s_xor_b32 s4, s4, s19
; GFX11-NEXT: s_xor_b32 s5, s5, s11
; GFX11-NEXT: s_xor_b32 s8, s8, s23
; GFX11-NEXT: s_and_b32 s2, s2, 0x22222222
; GFX11-NEXT: s_and_b32 s4, s4, 0x11111111
; GFX11-NEXT: s_and_b32 s8, s8, 0x44444444
; GFX11-NEXT: s_or_b32 s2, s4, s2
; GFX11-NEXT: s_xor_b32 s3, s5, s3
; GFX11-NEXT: s_or_b32 s2, s2, s8
; GFX11-NEXT: s_and_b32 s3, s3, 0x88888888
; GFX11-NEXT: s_mov_b32 s4, s0
; GFX11-NEXT: s_or_b32 s2, s2, s3
; GFX11-NEXT: s_mov_b32 s5, s1
; GFX11-NEXT: v_mov_b32_e32 v0, s2
; GFX11-NEXT: buffer_store_b32 v0, off, s[4:7], 0
; GFX11-NEXT: s_endpgm
;
; GFX12-LABEL: test_clmul_i32:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24
; GFX12-NEXT: s_mov_b32 s6, -1
; GFX12-NEXT: s_mov_b32 s7, 0x31016000
; GFX12-NEXT: s_mov_b32 s10, s6
; GFX12-NEXT: s_mov_b32 s11, s7
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_mov_b32 s8, s2
; GFX12-NEXT: s_mov_b32 s9, s3
; GFX12-NEXT: buffer_load_b64 v[0:1], off, s[8:11], null
; GFX12-NEXT: s_wait_loadcnt 0x0
; GFX12-NEXT: v_readfirstlane_b32 s2, v1
; GFX12-NEXT: v_readfirstlane_b32 s3, v0
; GFX12-NEXT: s_and_b32 s4, s2, 0x11111111
; GFX12-NEXT: s_and_b32 s5, s3, 0x22222222
; GFX12-NEXT: s_and_b32 s8, s2, 0x22222222
; GFX12-NEXT: s_and_b32 s9, s3, 0x11111111
; GFX12-NEXT: s_and_b32 s10, s2, 0x88888888
; GFX12-NEXT: s_and_b32 s11, s3, 0x44444444
; GFX12-NEXT: s_and_b32 s2, s2, 0x44444444
; GFX12-NEXT: s_and_b32 s3, s3, 0x88888888
; GFX12-NEXT: s_mul_i32 s12, s5, s4
; GFX12-NEXT: s_mul_i32 s13, s9, s8
; GFX12-NEXT: s_mul_i32 s16, s5, s10
; GFX12-NEXT: s_mul_i32 s17, s9, s4
; GFX12-NEXT: s_mul_i32 s14, s11, s10
; GFX12-NEXT: s_mul_i32 s15, s3, s2
; GFX12-NEXT: s_mul_i32 s18, s11, s2
; GFX12-NEXT: s_mul_i32 s19, s3, s8
; GFX12-NEXT: s_mul_i32 s20, s5, s8
; GFX12-NEXT: s_mul_i32 s21, s9, s2
; GFX12-NEXT: s_mul_i32 s22, s11, s4
; GFX12-NEXT: s_mul_i32 s23, s3, s10
; GFX12-NEXT: s_mul_i32 s5, s5, s2
; GFX12-NEXT: s_mul_i32 s3, s3, s4
; GFX12-NEXT: s_xor_b32 s2, s13, s12
; GFX12-NEXT: s_xor_b32 s4, s17, s16
; GFX12-NEXT: s_mul_i32 s9, s9, s10
; GFX12-NEXT: s_mul_i32 s11, s11, s8
; GFX12-NEXT: s_xor_b32 s8, s21, s20
; GFX12-NEXT: s_xor_b32 s2, s2, s14
; GFX12-NEXT: s_xor_b32 s4, s4, s18
; GFX12-NEXT: s_xor_b32 s5, s9, s5
; GFX12-NEXT: s_xor_b32 s8, s8, s22
; GFX12-NEXT: s_xor_b32 s2, s2, s15
; GFX12-NEXT: s_xor_b32 s4, s4, s19
; GFX12-NEXT: s_xor_b32 s5, s5, s11
; GFX12-NEXT: s_xor_b32 s8, s8, s23
; GFX12-NEXT: s_and_b32 s2, s2, 0x22222222
; GFX12-NEXT: s_and_b32 s4, s4, 0x11111111
; GFX12-NEXT: s_and_b32 s8, s8, 0x44444444
; GFX12-NEXT: s_or_b32 s2, s4, s2
; GFX12-NEXT: s_xor_b32 s3, s5, s3
; GFX12-NEXT: s_or_b32 s2, s2, s8
; GFX12-NEXT: s_and_b32 s3, s3, 0x88888888
; GFX12-NEXT: s_mov_b32 s4, s0
; GFX12-NEXT: s_or_b32 s2, s2, s3
; GFX12-NEXT: s_mov_b32 s5, s1
; GFX12-NEXT: v_mov_b32_e32 v0, s2
; GFX12-NEXT: buffer_store_b32 v0, off, s[4:7], null
; GFX12-NEXT: s_endpgm
;
; GFX1250-LABEL: test_clmul_i32:
; GFX1250: ; %bb.0:
; GFX1250-NEXT: global_prefetch_b8 v0, s[0:1] scope:SCOPE_SE
; GFX1250-NEXT: v_nop
; GFX1250-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; msbs: dst=0 src0=0 src1=0 src2=0
; GFX1250-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 nv
; GFX1250-NEXT: s_mov_b32 s6, -1
; GFX1250-NEXT: s_mov_b32 s7, 0x31016000
; GFX1250-NEXT: s_mov_b32 s10, s6
; GFX1250-NEXT: s_mov_b32 s11, s7
; GFX1250-NEXT: s_wait_kmcnt 0x0
; GFX1250-NEXT: s_mov_b32 s8, s2
; GFX1250-NEXT: s_mov_b32 s9, s3
; GFX1250-NEXT: buffer_load_b64 v[0:1], off, s[8:11], null
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: v_readfirstlane_b32 s2, v1
; GFX1250-NEXT: v_readfirstlane_b32 s3, v0
; GFX1250-NEXT: s_and_b32 s4, s2, 0x11111111
; GFX1250-NEXT: s_and_b32 s5, s3, 0x22222222
; GFX1250-NEXT: s_and_b32 s8, s2, 0x22222222
; GFX1250-NEXT: s_and_b32 s9, s3, 0x11111111
; GFX1250-NEXT: s_and_b32 s10, s2, 0x88888888
; GFX1250-NEXT: s_and_b32 s11, s3, 0x44444444
; GFX1250-NEXT: s_and_b32 s2, s2, 0x44444444
; GFX1250-NEXT: s_and_b32 s3, s3, 0x88888888
; GFX1250-NEXT: s_mul_i32 s12, s5, s4
; GFX1250-NEXT: s_mul_i32 s13, s9, s8
; GFX1250-NEXT: s_mul_i32 s16, s5, s10
; GFX1250-NEXT: s_mul_i32 s17, s9, s4
; GFX1250-NEXT: s_mul_i32 s14, s11, s10
; GFX1250-NEXT: s_mul_i32 s15, s3, s2
; GFX1250-NEXT: s_mul_i32 s18, s11, s2
; GFX1250-NEXT: s_mul_i32 s19, s3, s8
; GFX1250-NEXT: s_mul_i32 s20, s5, s8
; GFX1250-NEXT: s_mul_i32 s21, s9, s2
; GFX1250-NEXT: s_mul_i32 s22, s11, s4
; GFX1250-NEXT: s_mul_i32 s23, s3, s10
; GFX1250-NEXT: s_mul_i32 s5, s5, s2
; GFX1250-NEXT: s_mul_i32 s3, s3, s4
; GFX1250-NEXT: s_xor_b32 s2, s13, s12
; GFX1250-NEXT: s_xor_b32 s4, s17, s16
; GFX1250-NEXT: s_mul_i32 s9, s9, s10
; GFX1250-NEXT: s_mul_i32 s11, s11, s8
; GFX1250-NEXT: s_xor_b32 s8, s21, s20
; GFX1250-NEXT: s_xor_b32 s2, s2, s14
; GFX1250-NEXT: s_xor_b32 s4, s4, s18
; GFX1250-NEXT: s_xor_b32 s5, s9, s5
; GFX1250-NEXT: s_xor_b32 s8, s8, s22
; GFX1250-NEXT: s_xor_b32 s2, s2, s15
; GFX1250-NEXT: s_xor_b32 s4, s4, s19
; GFX1250-NEXT: s_xor_b32 s5, s5, s11
; GFX1250-NEXT: s_xor_b32 s8, s8, s23
; GFX1250-NEXT: s_and_b32 s2, s2, 0x22222222
; GFX1250-NEXT: s_and_b32 s4, s4, 0x11111111
; GFX1250-NEXT: s_and_b32 s8, s8, 0x44444444
; GFX1250-NEXT: s_or_b32 s2, s4, s2
; GFX1250-NEXT: s_xor_b32 s3, s5, s3
; GFX1250-NEXT: s_or_b32 s2, s2, s8
; GFX1250-NEXT: s_and_b32 s3, s3, 0x88888888
; GFX1250-NEXT: s_mov_b32 s4, s0
; GFX1250-NEXT: s_or_b32 s2, s2, s3
; GFX1250-NEXT: s_mov_b32 s5, s1
; GFX1250-NEXT: v_mov_b32_e32 v0, s2
; GFX1250-NEXT: buffer_store_b32 v0, off, s[4:7], null
; GFX1250-NEXT: s_endpgm
;
; EG-LABEL: test_clmul_i32:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 0 @6
; EG-NEXT: ALU 53, @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: AND_INT T0.W, T0.X, literal.x,
; EG-NEXT: AND_INT * T1.W, T0.Y, literal.y,
; EG-NEXT: 1145324612(7.850667e+02), 572662306(2.197316e-18)
; EG-NEXT: AND_INT T2.W, T0.Y, literal.x,
; EG-NEXT: MULLO_INT * T0.Z, PV.W, PS,
; EG-NEXT: 286331153(1.144374e-28), 0(0.000000e+00)
; EG-NEXT: AND_INT T3.W, T0.X, literal.x,
; EG-NEXT: MULLO_INT * T1.X, T0.W, PV.W,
; EG-NEXT: -2004318072(-8.217301e-34), 0(0.000000e+00)
; EG-NEXT: AND_INT T4.W, T0.Y, literal.x,
; EG-NEXT: MULLO_INT * T1.Y, PV.W, T1.W,
; EG-NEXT: 1145324612(7.850667e+02), 0(0.000000e+00)
; EG-NEXT: AND_INT T1.Z, T0.Y, literal.x,
; EG-NEXT: AND_INT T5.W, T0.X, literal.y,
; EG-NEXT: MULLO_INT * T0.Y, T3.W, PV.W,
; EG-NEXT: -2004318072(-8.217301e-34), 286331153(1.144374e-28)
; EG-NEXT: MULLO_INT * T2.X, PV.W, PV.Z,
; EG-NEXT: MULLO_INT * T2.Y, T0.W, T1.Z,
; EG-NEXT: MULLO_INT * T0.W, T0.W, T4.W,
; EG-NEXT: MULLO_INT * T2.Z, T5.W, T4.W,
; EG-NEXT: AND_INT T6.W, T0.X, literal.x,
; EG-NEXT: MULLO_INT * T0.X, T5.W, T2.W,
; EG-NEXT: 572662306(2.197316e-18), 0(0.000000e+00)
; EG-NEXT: MULLO_INT * T3.X, PV.W, T1.Z,
; EG-NEXT: MULLO_INT * T3.Y, T5.W, T1.W,
; EG-NEXT: MULLO_INT * T3.Z, T6.W, T2.W,
; EG-NEXT: XOR_INT T3.Z, T3.Y, PS,
; EG-NEXT: XOR_INT T5.W, T0.X, T3.X,
; EG-NEXT: MULLO_INT * T0.X, T6.W, T1.W,
; EG-NEXT: XOR_INT T3.Y, T2.Z, PS,
; EG-NEXT: XOR_INT T2.Z, PV.W, T0.W, BS:VEC_102/SCL_221
; EG-NEXT: XOR_INT T0.W, PV.Z, T2.Y,
; EG-NEXT: MULLO_INT * T0.X, T6.W, T4.W,
; EG-NEXT: XOR_INT T0.X, T2.X, PS,
; EG-NEXT: XOR_INT T0.Y, PV.W, T0.Y,
; EG-NEXT: XOR_INT T2.Z, PV.Z, T1.Y, BS:VEC_021/SCL_122
; EG-NEXT: XOR_INT T0.W, PV.Y, T1.X,
; EG-NEXT: MULLO_INT * T1.X, T3.W, T1.Z,
; EG-NEXT: XOR_INT T1.X, PV.W, PS,
; EG-NEXT: AND_INT T1.Y, PV.Z, literal.x,
; EG-NEXT: AND_INT T1.Z, PV.Y, literal.y,
; EG-NEXT: XOR_INT T0.W, PV.X, T0.Z,
; EG-NEXT: MULLO_INT * T0.X, T3.W, T2.W,
; EG-NEXT: 286331153(1.144374e-28), 572662306(2.197316e-18)
; EG-NEXT: XOR_INT T0.Z, PV.W, PS,
; EG-NEXT: OR_INT T0.W, PV.Y, PV.Z,
; EG-NEXT: AND_INT * T1.W, PV.X, literal.x,
; EG-NEXT: 1145324612(7.850667e+02), 0(0.000000e+00)
; EG-NEXT: OR_INT T0.W, PV.W, PS,
; EG-NEXT: AND_INT * T1.W, PV.Z, literal.x,
; EG-NEXT: -2004318072(-8.217301e-34), 0(0.000000e+00)
; EG-NEXT: OR_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)
%b_ptr = getelementptr i32, ptr addrspace(1) %in, i32 1
%a = load i32, ptr addrspace(1) %in
%b = load i32, ptr addrspace(1) %b_ptr
%res = call i32 @llvm.clmul.i32(i32 %a, i32 %b)
store i32 %res, ptr addrspace(1) %out
ret void
}
define amdgpu_kernel void @test_clmulr_i32(ptr addrspace(1) %out, ptr addrspace(1) %in) {
; SI-LABEL: test_clmulr_i32:
; SI: ; %bb.0:
; SI-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9
; SI-NEXT: s_mov_b32 s3, 0xf000
; SI-NEXT: s_mov_b32 s2, -1
; SI-NEXT: s_mov_b32 s10, s2
; SI-NEXT: s_mov_b32 s11, s3
; SI-NEXT: s_waitcnt lgkmcnt(0)
; SI-NEXT: s_mov_b32 s8, s6
; SI-NEXT: s_mov_b32 s9, s7
; SI-NEXT: buffer_load_dwordx2 v[0:1], off, s[8:11], 0
; SI-NEXT: s_mov_b32 s7, 0
; SI-NEXT: s_mov_b32 s21, s7
; SI-NEXT: s_mov_b32 s0, s4
; SI-NEXT: s_mov_b32 s1, s5
; SI-NEXT: s_mov_b32 s9, s7
; SI-NEXT: s_mov_b32 s15, s7
; SI-NEXT: s_mov_b32 s11, s7
; SI-NEXT: s_mov_b32 s13, s7
; SI-NEXT: s_mov_b32 s17, s7
; SI-NEXT: s_mov_b32 s19, s7
; SI-NEXT: s_mov_b32 s23, s7
; SI-NEXT: s_mov_b32 s25, s7
; SI-NEXT: s_mov_b32 s27, s7
; SI-NEXT: s_mov_b32 s29, s7
; SI-NEXT: s_mov_b32 s31, s7
; SI-NEXT: s_mov_b32 s35, s7
; SI-NEXT: s_mov_b32 s37, s7
; SI-NEXT: s_mov_b32 s39, s7
; SI-NEXT: s_mov_b32 s41, s7
; SI-NEXT: s_mov_b32 s43, s7
; SI-NEXT: s_waitcnt vmcnt(0)
; SI-NEXT: v_readfirstlane_b32 s33, v1
; SI-NEXT: s_and_b32 s20, s33, 2
; SI-NEXT: v_readfirstlane_b32 s6, v0
; SI-NEXT: s_bfe_i32 s8, s33, 0x10000
; SI-NEXT: v_cmp_eq_u64_e64 s[20:21], s[20:21], 0
; SI-NEXT: s_lshl_b64 s[4:5], s[6:7], 1
; SI-NEXT: s_and_b32 s8, s8, s6
; SI-NEXT: s_and_b64 s[20:21], s[20:21], exec
; SI-NEXT: s_cselect_b32 s21, 0, s5
; SI-NEXT: s_cselect_b32 s20, 0, s4
; SI-NEXT: s_and_b32 s14, s33, 4
; SI-NEXT: s_xor_b64 s[20:21], s[8:9], s[20:21]
; SI-NEXT: v_cmp_eq_u64_e64 s[8:9], s[14:15], 0
; SI-NEXT: s_lshl_b64 s[14:15], s[6:7], 2
; SI-NEXT: s_and_b64 s[8:9], s[8:9], exec
; SI-NEXT: s_cselect_b32 s15, 0, s15
; SI-NEXT: s_cselect_b32 s14, 0, s14
; SI-NEXT: s_and_b32 s10, s33, 8
; SI-NEXT: v_cmp_eq_u64_e64 s[10:11], s[10:11], 0
; SI-NEXT: s_xor_b64 s[14:15], s[20:21], s[14:15]
; SI-NEXT: s_lshl_b64 s[20:21], s[6:7], 3
; SI-NEXT: s_and_b64 s[10:11], s[10:11], exec
; SI-NEXT: s_cselect_b32 s11, 0, s21
; SI-NEXT: s_cselect_b32 s10, 0, s20
; SI-NEXT: s_and_b32 s12, s33, 16
; SI-NEXT: v_cmp_eq_u64_e64 s[12:13], s[12:13], 0
; SI-NEXT: s_lshl_b64 s[20:21], s[6:7], 4
; SI-NEXT: s_xor_b64 s[10:11], s[14:15], s[10:11]
; SI-NEXT: s_and_b64 s[12:13], s[12:13], exec
; SI-NEXT: s_cselect_b32 s13, 0, s21
; SI-NEXT: s_cselect_b32 s12, 0, s20
; SI-NEXT: s_and_b32 s16, s33, 32
; SI-NEXT: v_cmp_eq_u64_e64 s[16:17], s[16:17], 0
; SI-NEXT: s_lshl_b64 s[14:15], s[6:7], 5
; SI-NEXT: s_xor_b64 s[10:11], s[10:11], s[12:13]
; SI-NEXT: s_and_b64 s[12:13], s[16:17], exec
; SI-NEXT: s_cselect_b32 s13, 0, s15
; SI-NEXT: s_cselect_b32 s12, 0, s14
; SI-NEXT: s_and_b32 s18, s33, 64
; SI-NEXT: v_cmp_eq_u64_e64 s[16:17], s[18:19], 0
; SI-NEXT: s_lshl_b64 s[14:15], s[6:7], 6
; SI-NEXT: s_xor_b64 s[10:11], s[10:11], s[12:13]
; SI-NEXT: s_and_b64 s[12:13], s[16:17], exec
; SI-NEXT: s_cselect_b32 s13, 0, s15
; SI-NEXT: s_cselect_b32 s12, 0, s14
; SI-NEXT: s_and_b32 s22, s33, 0x80
; SI-NEXT: v_cmp_eq_u64_e64 s[16:17], s[22:23], 0
; SI-NEXT: s_lshl_b64 s[14:15], s[6:7], 7
; SI-NEXT: s_xor_b64 s[10:11], s[10:11], s[12:13]
; SI-NEXT: s_and_b64 s[12:13], s[16:17], exec
; SI-NEXT: s_cselect_b32 s13, 0, s15
; SI-NEXT: s_cselect_b32 s12, 0, s14
; SI-NEXT: s_and_b32 s24, s33, 0x100
; SI-NEXT: v_cmp_eq_u64_e64 s[16:17], s[24:25], 0
; SI-NEXT: s_lshl_b64 s[14:15], s[6:7], 8
; SI-NEXT: s_xor_b64 s[10:11], s[10:11], s[12:13]
; SI-NEXT: s_and_b64 s[12:13], s[16:17], exec
; SI-NEXT: s_cselect_b32 s13, 0, s15
; SI-NEXT: s_cselect_b32 s12, 0, s14
; SI-NEXT: s_and_b32 s26, s33, 0x200
; SI-NEXT: v_cmp_eq_u64_e64 s[16:17], s[26:27], 0
; SI-NEXT: s_lshl_b64 s[14:15], s[6:7], 9
; SI-NEXT: s_xor_b64 s[10:11], s[10:11], s[12:13]
; SI-NEXT: s_and_b64 s[12:13], s[16:17], exec
; SI-NEXT: s_cselect_b32 s13, 0, s15
; SI-NEXT: s_cselect_b32 s12, 0, s14
; SI-NEXT: s_and_b32 s28, s33, 0x400
; SI-NEXT: v_cmp_eq_u64_e64 s[16:17], s[28:29], 0
; SI-NEXT: s_lshl_b64 s[14:15], s[6:7], 10
; SI-NEXT: s_xor_b64 s[10:11], s[10:11], s[12:13]
; SI-NEXT: s_and_b64 s[12:13], s[16:17], exec
; SI-NEXT: s_cselect_b32 s13, 0, s15
; SI-NEXT: s_cselect_b32 s12, 0, s14
; SI-NEXT: s_and_b32 s30, s33, 0x800
; SI-NEXT: v_cmp_eq_u64_e64 s[16:17], s[30:31], 0
; SI-NEXT: s_lshl_b64 s[14:15], s[6:7], 11
; SI-NEXT: s_xor_b64 s[10:11], s[10:11], s[12:13]
; SI-NEXT: s_and_b64 s[12:13], s[16:17], exec
; SI-NEXT: s_cselect_b32 s13, 0, s15
; SI-NEXT: s_cselect_b32 s12, 0, s14
; SI-NEXT: s_and_b32 s34, s33, 0x1000
; SI-NEXT: v_cmp_eq_u64_e64 s[16:17], s[34:35], 0
; SI-NEXT: s_lshl_b64 s[14:15], s[6:7], 12
; SI-NEXT: s_xor_b64 s[10:11], s[10:11], s[12:13]
; SI-NEXT: s_and_b64 s[12:13], s[16:17], exec
; SI-NEXT: s_cselect_b32 s13, 0, s15
; SI-NEXT: s_cselect_b32 s12, 0, s14
; SI-NEXT: s_and_b32 s36, s33, 0x2000
; SI-NEXT: v_cmp_eq_u64_e64 s[16:17], s[36:37], 0
; SI-NEXT: s_lshl_b64 s[14:15], s[6:7], 13
; SI-NEXT: s_xor_b64 s[10:11], s[10:11], s[12:13]
; SI-NEXT: s_and_b64 s[12:13], s[16:17], exec
; SI-NEXT: s_cselect_b32 s13, 0, s15
; SI-NEXT: s_cselect_b32 s12, 0, s14
; SI-NEXT: s_and_b32 s38, s33, 0x4000
; SI-NEXT: v_cmp_eq_u64_e64 s[16:17], s[38:39], 0
; SI-NEXT: s_lshl_b64 s[14:15], s[6:7], 14
; SI-NEXT: s_xor_b64 s[10:11], s[10:11], s[12:13]
; SI-NEXT: s_and_b64 s[12:13], s[16:17], exec
; SI-NEXT: s_cselect_b32 s13, 0, s15
; SI-NEXT: s_cselect_b32 s12, 0, s14
; SI-NEXT: s_and_b32 s40, s33, 0x8000
; SI-NEXT: v_cmp_eq_u64_e64 s[16:17], s[40:41], 0
; SI-NEXT: s_lshl_b64 s[14:15], s[6:7], 15
; SI-NEXT: s_xor_b64 s[10:11], s[10:11], s[12:13]
; SI-NEXT: s_and_b64 s[12:13], s[16:17], exec
; SI-NEXT: s_cselect_b32 s13, 0, s15
; SI-NEXT: s_cselect_b32 s12, 0, s14
; SI-NEXT: s_and_b32 s42, s33, 0x10000
; SI-NEXT: v_cmp_eq_u64_e64 s[16:17], s[42:43], 0
; SI-NEXT: s_lshl_b64 s[14:15], s[6:7], 16
; SI-NEXT: s_xor_b64 s[10:11], s[10:11], s[12:13]
; SI-NEXT: s_and_b64 s[12:13], s[16:17], exec
; SI-NEXT: s_mov_b32 s5, s7
; SI-NEXT: s_cselect_b32 s13, 0, s15
; SI-NEXT: s_cselect_b32 s12, 0, s14
; SI-NEXT: s_and_b32 s4, s33, 0x20000
; SI-NEXT: v_cmp_eq_u64_e64 s[4:5], s[4:5], 0
; SI-NEXT: s_lshl_b64 s[14:15], s[6:7], 17
; SI-NEXT: s_xor_b64 s[10:11], s[10:11], s[12:13]
; SI-NEXT: s_and_b64 s[4:5], s[4:5], exec
; SI-NEXT: s_mov_b32 s9, s7
; SI-NEXT: s_cselect_b32 s5, 0, s15
; SI-NEXT: s_cselect_b32 s4, 0, s14
; SI-NEXT: s_and_b32 s8, s33, 0x40000
; SI-NEXT: v_cmp_eq_u64_e64 s[8:9], s[8:9], 0
; SI-NEXT: s_lshl_b64 s[12:13], s[6:7], 18
; SI-NEXT: s_xor_b64 s[4:5], s[10:11], s[4:5]
; SI-NEXT: s_and_b64 s[8:9], s[8:9], exec
; SI-NEXT: s_cselect_b32 s9, 0, s13
; SI-NEXT: s_cselect_b32 s8, 0, s12
; SI-NEXT: s_and_b32 s10, s33, 0x80000
; SI-NEXT: s_mov_b32 s11, s7
; SI-NEXT: s_xor_b64 s[4:5], s[4:5], s[8:9]
; SI-NEXT: v_cmp_eq_u64_e64 s[8:9], s[10:11], 0
; SI-NEXT: s_lshl_b64 s[10:11], s[6:7], 19
; SI-NEXT: s_and_b64 s[8:9], s[8:9], exec
; SI-NEXT: s_cselect_b32 s9, 0, s11
; SI-NEXT: s_cselect_b32 s8, 0, s10
; SI-NEXT: s_xor_b64 s[4:5], s[4:5], s[8:9]
; SI-NEXT: s_and_b32 s8, s33, 0x100000
; SI-NEXT: s_mov_b32 s9, s7
; SI-NEXT: v_cmp_eq_u64_e64 s[8:9], s[8:9], 0
; SI-NEXT: s_lshl_b64 s[10:11], s[6:7], 20
; SI-NEXT: s_and_b64 s[8:9], s[8:9], exec
; SI-NEXT: s_cselect_b32 s9, 0, s11
; SI-NEXT: s_cselect_b32 s8, 0, s10
; SI-NEXT: s_xor_b64 s[4:5], s[4:5], s[8:9]
; SI-NEXT: s_and_b32 s8, s33, 0x200000
; SI-NEXT: s_mov_b32 s9, s7
; SI-NEXT: v_cmp_eq_u64_e64 s[8:9], s[8:9], 0
; SI-NEXT: s_lshl_b64 s[10:11], s[6:7], 21
; SI-NEXT: s_and_b64 s[8:9], s[8:9], exec
; SI-NEXT: s_cselect_b32 s9, 0, s11
; SI-NEXT: s_cselect_b32 s8, 0, s10
; SI-NEXT: s_xor_b64 s[4:5], s[4:5], s[8:9]
; SI-NEXT: s_and_b32 s8, s33, 0x400000
; SI-NEXT: s_mov_b32 s9, s7
; SI-NEXT: v_cmp_eq_u64_e64 s[8:9], s[8:9], 0
; SI-NEXT: s_lshl_b64 s[10:11], s[6:7], 22
; SI-NEXT: s_and_b64 s[8:9], s[8:9], exec
; SI-NEXT: s_cselect_b32 s9, 0, s11
; SI-NEXT: s_cselect_b32 s8, 0, s10
; SI-NEXT: s_xor_b64 s[4:5], s[4:5], s[8:9]
; SI-NEXT: s_and_b32 s8, s33, 0x800000
; SI-NEXT: s_mov_b32 s9, s7
; SI-NEXT: v_cmp_eq_u64_e64 s[8:9], s[8:9], 0
; SI-NEXT: s_lshl_b64 s[10:11], s[6:7], 23
; SI-NEXT: s_and_b64 s[8:9], s[8:9], exec
; SI-NEXT: s_cselect_b32 s9, 0, s11
; SI-NEXT: s_cselect_b32 s8, 0, s10
; SI-NEXT: s_xor_b64 s[4:5], s[4:5], s[8:9]
; SI-NEXT: s_and_b32 s8, s33, 0x1000000
; SI-NEXT: s_mov_b32 s9, s7
; SI-NEXT: v_cmp_eq_u64_e64 s[8:9], s[8:9], 0
; SI-NEXT: s_lshl_b64 s[10:11], s[6:7], 24
; SI-NEXT: s_and_b64 s[8:9], s[8:9], exec
; SI-NEXT: s_cselect_b32 s9, 0, s11
; SI-NEXT: s_cselect_b32 s8, 0, s10
; SI-NEXT: s_xor_b64 s[4:5], s[4:5], s[8:9]
; SI-NEXT: s_and_b32 s8, s33, 0x2000000
; SI-NEXT: s_mov_b32 s9, s7
; SI-NEXT: v_cmp_eq_u64_e64 s[8:9], s[8:9], 0
; SI-NEXT: s_lshl_b64 s[10:11], s[6:7], 25
; SI-NEXT: s_and_b64 s[8:9], s[8:9], exec
; SI-NEXT: s_cselect_b32 s9, 0, s11
; SI-NEXT: s_cselect_b32 s8, 0, s10
; SI-NEXT: s_xor_b64 s[4:5], s[4:5], s[8:9]
; SI-NEXT: s_and_b32 s8, s33, 0x4000000
; SI-NEXT: s_mov_b32 s9, s7
; SI-NEXT: v_cmp_eq_u64_e64 s[8:9], s[8:9], 0
; SI-NEXT: s_lshl_b64 s[10:11], s[6:7], 26
; SI-NEXT: s_and_b64 s[8:9], s[8:9], exec
; SI-NEXT: s_cselect_b32 s9, 0, s11
; SI-NEXT: s_cselect_b32 s8, 0, s10
; SI-NEXT: s_xor_b64 s[4:5], s[4:5], s[8:9]
; SI-NEXT: s_and_b32 s8, s33, 0x8000000
; SI-NEXT: s_mov_b32 s9, s7
; SI-NEXT: v_cmp_eq_u64_e64 s[8:9], s[8:9], 0
; SI-NEXT: s_lshl_b64 s[10:11], s[6:7], 27
; SI-NEXT: s_and_b64 s[8:9], s[8:9], exec
; SI-NEXT: s_cselect_b32 s9, 0, s11
; SI-NEXT: s_cselect_b32 s8, 0, s10
; SI-NEXT: s_xor_b64 s[4:5], s[4:5], s[8:9]
; SI-NEXT: s_and_b32 s8, s33, 0x10000000
; SI-NEXT: s_mov_b32 s9, s7
; SI-NEXT: v_cmp_eq_u64_e64 s[8:9], s[8:9], 0
; SI-NEXT: s_lshl_b64 s[10:11], s[6:7], 28
; SI-NEXT: s_and_b64 s[8:9], s[8:9], exec
; SI-NEXT: s_cselect_b32 s9, 0, s11
; SI-NEXT: s_cselect_b32 s8, 0, s10
; SI-NEXT: s_xor_b64 s[4:5], s[4:5], s[8:9]
; SI-NEXT: s_and_b32 s8, s33, 0x20000000
; SI-NEXT: s_mov_b32 s9, s7
; SI-NEXT: v_cmp_eq_u64_e64 s[8:9], s[8:9], 0
; SI-NEXT: s_lshl_b64 s[10:11], s[6:7], 29
; SI-NEXT: s_and_b64 s[8:9], s[8:9], exec
; SI-NEXT: s_cselect_b32 s9, 0, s11
; SI-NEXT: s_cselect_b32 s8, 0, s10
; SI-NEXT: s_xor_b64 s[4:5], s[4:5], s[8:9]
; SI-NEXT: s_and_b32 s8, s33, 2.0
; SI-NEXT: s_mov_b32 s9, s7
; SI-NEXT: v_cmp_eq_u64_e64 s[8:9], s[8:9], 0
; SI-NEXT: s_lshl_b64 s[10:11], s[6:7], 30
; SI-NEXT: s_and_b64 s[8:9], s[8:9], exec
; SI-NEXT: s_cselect_b32 s9, 0, s11
; SI-NEXT: s_cselect_b32 s8, 0, s10
; SI-NEXT: s_xor_b64 s[4:5], s[4:5], s[8:9]
; SI-NEXT: s_lshl_b64 s[6:7], s[6:7], 31
; SI-NEXT: s_cmp_gt_i32 s33, -1
; SI-NEXT: s_cselect_b32 s7, 0, s7
; SI-NEXT: s_cselect_b32 s6, 0, s6
; SI-NEXT: s_xor_b64 s[4:5], s[4:5], s[6:7]
; SI-NEXT: s_lshr_b64 s[4:5], s[4:5], 31
; SI-NEXT: v_mov_b32_e32 v0, s4
; SI-NEXT: buffer_store_dword v0, off, s[0:3], 0
; SI-NEXT: s_endpgm
;
; VI-LABEL: test_clmulr_i32:
; VI: ; %bb.0:
; VI-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x24
; VI-NEXT: s_mov_b32 s3, 0xf000
; VI-NEXT: s_mov_b32 s2, -1
; VI-NEXT: s_mov_b32 s10, s2
; VI-NEXT: s_mov_b32 s11, s3
; VI-NEXT: s_waitcnt lgkmcnt(0)
; VI-NEXT: s_mov_b32 s8, s6
; VI-NEXT: s_mov_b32 s9, s7
; VI-NEXT: buffer_load_dwordx2 v[0:1], off, s[8:11], 0
; VI-NEXT: s_mov_b32 s0, s4
; VI-NEXT: s_mov_b32 s7, 0
; VI-NEXT: s_mov_b32 s1, s5
; VI-NEXT: s_mov_b32 s11, s7
; VI-NEXT: s_mov_b32 s9, s7
; VI-NEXT: s_mov_b32 s13, s7
; VI-NEXT: s_mov_b32 s15, s7
; VI-NEXT: s_mov_b32 s17, s7
; VI-NEXT: s_mov_b32 s19, s7
; VI-NEXT: s_mov_b32 s21, s7
; VI-NEXT: s_mov_b32 s23, s7
; VI-NEXT: s_mov_b32 s25, s7
; VI-NEXT: s_mov_b32 s27, s7
; VI-NEXT: s_mov_b32 s29, s7
; VI-NEXT: s_mov_b32 s31, s7
; VI-NEXT: s_mov_b32 s35, s7
; VI-NEXT: s_mov_b32 s37, s7
; VI-NEXT: s_mov_b32 s39, s7
; VI-NEXT: s_mov_b32 s41, s7
; VI-NEXT: s_mov_b32 s43, s7
; VI-NEXT: s_mov_b32 s45, s7
; VI-NEXT: s_mov_b32 s47, s7
; VI-NEXT: s_waitcnt vmcnt(0)
; VI-NEXT: v_readfirstlane_b32 s4, v1
; VI-NEXT: v_readfirstlane_b32 s6, v0
; VI-NEXT: s_bfe_i32 s5, s4, 0x10000
; VI-NEXT: s_lshl_b64 s[48:49], s[6:7], 1
; VI-NEXT: s_and_b32 s10, s4, 2
; VI-NEXT: s_and_b32 s8, s5, s6
; VI-NEXT: s_cmp_eq_u64 s[10:11], 0
; VI-NEXT: s_cselect_b32 s11, 0, s49
; VI-NEXT: s_cselect_b32 s10, 0, s48
; VI-NEXT: s_lshl_b64 s[48:49], s[6:7], 2
; VI-NEXT: s_and_b32 s12, s4, 4
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_cmp_eq_u64 s[12:13], 0
; VI-NEXT: s_cselect_b32 s11, 0, s49
; VI-NEXT: s_cselect_b32 s10, 0, s48
; VI-NEXT: s_lshl_b64 s[12:13], s[6:7], 3
; VI-NEXT: s_and_b32 s14, s4, 8
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_cmp_eq_u64 s[14:15], 0
; VI-NEXT: s_cselect_b32 s11, 0, s13
; VI-NEXT: s_cselect_b32 s10, 0, s12
; VI-NEXT: s_lshl_b64 s[12:13], s[6:7], 4
; VI-NEXT: s_and_b32 s16, s4, 16
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_cmp_eq_u64 s[16:17], 0
; VI-NEXT: s_cselect_b32 s11, 0, s13
; VI-NEXT: s_cselect_b32 s10, 0, s12
; VI-NEXT: s_lshl_b64 s[12:13], s[6:7], 5
; VI-NEXT: s_and_b32 s18, s4, 32
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_cmp_eq_u64 s[18:19], 0
; VI-NEXT: s_cselect_b32 s11, 0, s13
; VI-NEXT: s_cselect_b32 s10, 0, s12
; VI-NEXT: s_lshl_b64 s[12:13], s[6:7], 6
; VI-NEXT: s_and_b32 s20, s4, 64
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_cmp_eq_u64 s[20:21], 0
; VI-NEXT: s_cselect_b32 s11, 0, s13
; VI-NEXT: s_cselect_b32 s10, 0, s12
; VI-NEXT: s_lshl_b64 s[12:13], s[6:7], 7
; VI-NEXT: s_and_b32 s22, s4, 0x80
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_cmp_eq_u64 s[22:23], 0
; VI-NEXT: s_cselect_b32 s11, 0, s13
; VI-NEXT: s_cselect_b32 s10, 0, s12
; VI-NEXT: s_lshl_b64 s[12:13], s[6:7], 8
; VI-NEXT: s_and_b32 s24, s4, 0x100
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_cmp_eq_u64 s[24:25], 0
; VI-NEXT: s_cselect_b32 s11, 0, s13
; VI-NEXT: s_cselect_b32 s10, 0, s12
; VI-NEXT: s_lshl_b64 s[12:13], s[6:7], 9
; VI-NEXT: s_and_b32 s26, s4, 0x200
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_cmp_eq_u64 s[26:27], 0
; VI-NEXT: s_cselect_b32 s11, 0, s13
; VI-NEXT: s_cselect_b32 s10, 0, s12
; VI-NEXT: s_lshl_b64 s[12:13], s[6:7], 10
; VI-NEXT: s_and_b32 s28, s4, 0x400
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_cmp_eq_u64 s[28:29], 0
; VI-NEXT: s_cselect_b32 s11, 0, s13
; VI-NEXT: s_cselect_b32 s10, 0, s12
; VI-NEXT: s_lshl_b64 s[12:13], s[6:7], 11
; VI-NEXT: s_and_b32 s30, s4, 0x800
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_cmp_eq_u64 s[30:31], 0
; VI-NEXT: s_cselect_b32 s11, 0, s13
; VI-NEXT: s_cselect_b32 s10, 0, s12
; VI-NEXT: s_lshl_b64 s[12:13], s[6:7], 12
; VI-NEXT: s_and_b32 s34, s4, 0x1000
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_cmp_eq_u64 s[34:35], 0
; VI-NEXT: s_cselect_b32 s11, 0, s13
; VI-NEXT: s_cselect_b32 s10, 0, s12
; VI-NEXT: s_lshl_b64 s[12:13], s[6:7], 13
; VI-NEXT: s_and_b32 s36, s4, 0x2000
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_cmp_eq_u64 s[36:37], 0
; VI-NEXT: s_cselect_b32 s11, 0, s13
; VI-NEXT: s_cselect_b32 s10, 0, s12
; VI-NEXT: s_lshl_b64 s[12:13], s[6:7], 14
; VI-NEXT: s_and_b32 s38, s4, 0x4000
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_cmp_eq_u64 s[38:39], 0
; VI-NEXT: s_cselect_b32 s11, 0, s13
; VI-NEXT: s_cselect_b32 s10, 0, s12
; VI-NEXT: s_lshl_b64 s[12:13], s[6:7], 15
; VI-NEXT: s_and_b32 s40, s4, 0x8000
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_cmp_eq_u64 s[40:41], 0
; VI-NEXT: s_cselect_b32 s11, 0, s13
; VI-NEXT: s_cselect_b32 s10, 0, s12
; VI-NEXT: s_lshl_b64 s[12:13], s[6:7], 16
; VI-NEXT: s_and_b32 s42, s4, 0x10000
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_cmp_eq_u64 s[42:43], 0
; VI-NEXT: s_cselect_b32 s11, 0, s13
; VI-NEXT: s_cselect_b32 s10, 0, s12
; VI-NEXT: s_lshl_b64 s[12:13], s[6:7], 17
; VI-NEXT: s_and_b32 s44, s4, 0x20000
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_cmp_eq_u64 s[44:45], 0
; VI-NEXT: s_cselect_b32 s11, 0, s13
; VI-NEXT: s_cselect_b32 s10, 0, s12
; VI-NEXT: s_lshl_b64 s[12:13], s[6:7], 18
; VI-NEXT: s_and_b32 s46, s4, 0x40000
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_cmp_eq_u64 s[46:47], 0
; VI-NEXT: s_cselect_b32 s11, 0, s13
; VI-NEXT: s_cselect_b32 s10, 0, s12
; VI-NEXT: s_lshl_b64 s[12:13], s[6:7], 19
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_and_b32 s10, s4, 0x80000
; VI-NEXT: s_mov_b32 s11, s7
; VI-NEXT: s_cmp_eq_u64 s[10:11], 0
; VI-NEXT: s_cselect_b32 s11, 0, s13
; VI-NEXT: s_cselect_b32 s10, 0, s12
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_lshl_b64 s[10:11], s[6:7], 20
; VI-NEXT: s_and_b32 s12, s4, 0x100000
; VI-NEXT: s_mov_b32 s13, s7
; VI-NEXT: s_cmp_eq_u64 s[12:13], 0
; VI-NEXT: s_cselect_b32 s11, 0, s11
; VI-NEXT: s_cselect_b32 s10, 0, s10
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_lshl_b64 s[10:11], s[6:7], 21
; VI-NEXT: s_and_b32 s12, s4, 0x200000
; VI-NEXT: s_cmp_eq_u64 s[12:13], 0
; VI-NEXT: s_cselect_b32 s11, 0, s11
; VI-NEXT: s_cselect_b32 s10, 0, s10
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_lshl_b64 s[10:11], s[6:7], 22
; VI-NEXT: s_and_b32 s12, s4, 0x400000
; VI-NEXT: s_cmp_eq_u64 s[12:13], 0
; VI-NEXT: s_cselect_b32 s11, 0, s11
; VI-NEXT: s_cselect_b32 s10, 0, s10
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_lshl_b64 s[10:11], s[6:7], 23
; VI-NEXT: s_and_b32 s12, s4, 0x800000
; VI-NEXT: s_cmp_eq_u64 s[12:13], 0
; VI-NEXT: s_cselect_b32 s11, 0, s11
; VI-NEXT: s_cselect_b32 s10, 0, s10
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_lshl_b64 s[10:11], s[6:7], 24
; VI-NEXT: s_and_b32 s12, s4, 0x1000000
; VI-NEXT: s_cmp_eq_u64 s[12:13], 0
; VI-NEXT: s_cselect_b32 s11, 0, s11
; VI-NEXT: s_cselect_b32 s10, 0, s10
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_lshl_b64 s[10:11], s[6:7], 25
; VI-NEXT: s_and_b32 s12, s4, 0x2000000
; VI-NEXT: s_cmp_eq_u64 s[12:13], 0
; VI-NEXT: s_cselect_b32 s11, 0, s11
; VI-NEXT: s_cselect_b32 s10, 0, s10
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_lshl_b64 s[10:11], s[6:7], 26
; VI-NEXT: s_and_b32 s12, s4, 0x4000000
; VI-NEXT: s_cmp_eq_u64 s[12:13], 0
; VI-NEXT: s_cselect_b32 s11, 0, s11
; VI-NEXT: s_cselect_b32 s10, 0, s10
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_lshl_b64 s[10:11], s[6:7], 27
; VI-NEXT: s_and_b32 s12, s4, 0x8000000
; VI-NEXT: s_cmp_eq_u64 s[12:13], 0
; VI-NEXT: s_cselect_b32 s11, 0, s11
; VI-NEXT: s_cselect_b32 s10, 0, s10
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_lshl_b64 s[10:11], s[6:7], 28
; VI-NEXT: s_and_b32 s12, s4, 0x10000000
; VI-NEXT: s_cmp_eq_u64 s[12:13], 0
; VI-NEXT: s_cselect_b32 s11, 0, s11
; VI-NEXT: s_cselect_b32 s10, 0, s10
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_lshl_b64 s[10:11], s[6:7], 29
; VI-NEXT: s_and_b32 s12, s4, 0x20000000
; VI-NEXT: s_cmp_eq_u64 s[12:13], 0
; VI-NEXT: s_cselect_b32 s11, 0, s11
; VI-NEXT: s_cselect_b32 s10, 0, s10
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_lshl_b64 s[10:11], s[6:7], 30
; VI-NEXT: s_and_b32 s12, s4, 2.0
; VI-NEXT: s_cmp_eq_u64 s[12:13], 0
; VI-NEXT: s_cselect_b32 s11, 0, s11
; VI-NEXT: s_cselect_b32 s10, 0, s10
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_lshl_b64 s[6:7], s[6:7], 31
; VI-NEXT: s_cmp_gt_i32 s4, -1
; VI-NEXT: s_cselect_b32 s5, 0, s7
; VI-NEXT: s_cselect_b32 s4, 0, s6
; VI-NEXT: s_xor_b64 s[4:5], s[8:9], s[4:5]
; VI-NEXT: s_lshr_b64 s[4:5], s[4:5], 31
; VI-NEXT: v_mov_b32_e32 v0, s4
; VI-NEXT: buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT: s_endpgm
;
; GFX9-LABEL: test_clmulr_i32:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_load_dwordx4 s[8:11], s[4:5], 0x24
; GFX9-NEXT: s_mov_b32 s3, 0xf000
; GFX9-NEXT: s_mov_b32 s2, -1
; GFX9-NEXT: s_mov_b32 s6, s2
; GFX9-NEXT: s_mov_b32 s7, s3
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_mov_b32 s4, s10
; GFX9-NEXT: s_mov_b32 s5, s11
; GFX9-NEXT: buffer_load_dwordx2 v[0:1], off, s[4:7], 0
; GFX9-NEXT: s_mov_b32 s0, s8
; GFX9-NEXT: s_mov_b32 s5, 0
; GFX9-NEXT: s_mov_b32 s11, s5
; GFX9-NEXT: s_mov_b32 s7, s5
; GFX9-NEXT: s_mov_b32 s13, s5
; GFX9-NEXT: s_mov_b32 s15, s5
; GFX9-NEXT: s_mov_b32 s17, s5
; GFX9-NEXT: s_mov_b32 s19, s5
; GFX9-NEXT: s_mov_b32 s21, s5
; GFX9-NEXT: s_mov_b32 s23, s5
; GFX9-NEXT: s_mov_b32 s25, s5
; GFX9-NEXT: s_mov_b32 s27, s5
; GFX9-NEXT: s_mov_b32 s29, s5
; GFX9-NEXT: s_mov_b32 s31, s5
; GFX9-NEXT: s_mov_b32 s35, s5
; GFX9-NEXT: s_mov_b32 s37, s5
; GFX9-NEXT: s_mov_b32 s39, s5
; GFX9-NEXT: s_mov_b32 s41, s5
; GFX9-NEXT: s_mov_b32 s43, s5
; GFX9-NEXT: s_mov_b32 s45, s5
; GFX9-NEXT: s_mov_b32 s47, s5
; GFX9-NEXT: s_mov_b32 s1, s9
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_readfirstlane_b32 s8, v1
; GFX9-NEXT: v_readfirstlane_b32 s4, v0
; GFX9-NEXT: s_bfe_i32 s6, s8, 0x10000
; GFX9-NEXT: s_lshl_b64 s[48:49], s[4:5], 1
; GFX9-NEXT: s_and_b32 s10, s8, 2
; GFX9-NEXT: s_and_b32 s6, s6, s4
; GFX9-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s49
; GFX9-NEXT: s_cselect_b32 s10, 0, s48
; GFX9-NEXT: s_lshl_b64 s[48:49], s[4:5], 2
; GFX9-NEXT: s_and_b32 s12, s8, 4
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_cmp_eq_u64 s[12:13], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s49
; GFX9-NEXT: s_cselect_b32 s10, 0, s48
; GFX9-NEXT: s_lshl_b64 s[12:13], s[4:5], 3
; GFX9-NEXT: s_and_b32 s14, s8, 8
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_cmp_eq_u64 s[14:15], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s13
; GFX9-NEXT: s_cselect_b32 s10, 0, s12
; GFX9-NEXT: s_lshl_b64 s[12:13], s[4:5], 4
; GFX9-NEXT: s_and_b32 s16, s8, 16
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_cmp_eq_u64 s[16:17], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s13
; GFX9-NEXT: s_cselect_b32 s10, 0, s12
; GFX9-NEXT: s_lshl_b64 s[12:13], s[4:5], 5
; GFX9-NEXT: s_and_b32 s18, s8, 32
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_cmp_eq_u64 s[18:19], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s13
; GFX9-NEXT: s_cselect_b32 s10, 0, s12
; GFX9-NEXT: s_lshl_b64 s[12:13], s[4:5], 6
; GFX9-NEXT: s_and_b32 s20, s8, 64
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_cmp_eq_u64 s[20:21], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s13
; GFX9-NEXT: s_cselect_b32 s10, 0, s12
; GFX9-NEXT: s_lshl_b64 s[12:13], s[4:5], 7
; GFX9-NEXT: s_and_b32 s22, s8, 0x80
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_cmp_eq_u64 s[22:23], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s13
; GFX9-NEXT: s_cselect_b32 s10, 0, s12
; GFX9-NEXT: s_lshl_b64 s[12:13], s[4:5], 8
; GFX9-NEXT: s_and_b32 s24, s8, 0x100
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_cmp_eq_u64 s[24:25], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s13
; GFX9-NEXT: s_cselect_b32 s10, 0, s12
; GFX9-NEXT: s_lshl_b64 s[12:13], s[4:5], 9
; GFX9-NEXT: s_and_b32 s26, s8, 0x200
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_cmp_eq_u64 s[26:27], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s13
; GFX9-NEXT: s_cselect_b32 s10, 0, s12
; GFX9-NEXT: s_lshl_b64 s[12:13], s[4:5], 10
; GFX9-NEXT: s_and_b32 s28, s8, 0x400
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_cmp_eq_u64 s[28:29], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s13
; GFX9-NEXT: s_cselect_b32 s10, 0, s12
; GFX9-NEXT: s_lshl_b64 s[12:13], s[4:5], 11
; GFX9-NEXT: s_and_b32 s30, s8, 0x800
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_cmp_eq_u64 s[30:31], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s13
; GFX9-NEXT: s_cselect_b32 s10, 0, s12
; GFX9-NEXT: s_lshl_b64 s[12:13], s[4:5], 12
; GFX9-NEXT: s_and_b32 s34, s8, 0x1000
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_cmp_eq_u64 s[34:35], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s13
; GFX9-NEXT: s_cselect_b32 s10, 0, s12
; GFX9-NEXT: s_lshl_b64 s[12:13], s[4:5], 13
; GFX9-NEXT: s_and_b32 s36, s8, 0x2000
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_cmp_eq_u64 s[36:37], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s13
; GFX9-NEXT: s_cselect_b32 s10, 0, s12
; GFX9-NEXT: s_lshl_b64 s[12:13], s[4:5], 14
; GFX9-NEXT: s_and_b32 s38, s8, 0x4000
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_cmp_eq_u64 s[38:39], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s13
; GFX9-NEXT: s_cselect_b32 s10, 0, s12
; GFX9-NEXT: s_lshl_b64 s[12:13], s[4:5], 15
; GFX9-NEXT: s_and_b32 s40, s8, 0x8000
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_cmp_eq_u64 s[40:41], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s13
; GFX9-NEXT: s_cselect_b32 s10, 0, s12
; GFX9-NEXT: s_lshl_b64 s[12:13], s[4:5], 16
; GFX9-NEXT: s_and_b32 s42, s8, 0x10000
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_cmp_eq_u64 s[42:43], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s13
; GFX9-NEXT: s_cselect_b32 s10, 0, s12
; GFX9-NEXT: s_lshl_b64 s[12:13], s[4:5], 17
; GFX9-NEXT: s_and_b32 s44, s8, 0x20000
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_cmp_eq_u64 s[44:45], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s13
; GFX9-NEXT: s_cselect_b32 s10, 0, s12
; GFX9-NEXT: s_lshl_b64 s[12:13], s[4:5], 18
; GFX9-NEXT: s_and_b32 s46, s8, 0x40000
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_cmp_eq_u64 s[46:47], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s13
; GFX9-NEXT: s_cselect_b32 s10, 0, s12
; GFX9-NEXT: s_lshl_b64 s[12:13], s[4:5], 19
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_and_b32 s10, s8, 0x80000
; GFX9-NEXT: s_mov_b32 s11, s5
; GFX9-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s13
; GFX9-NEXT: s_cselect_b32 s10, 0, s12
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_lshl_b64 s[10:11], s[4:5], 20
; GFX9-NEXT: s_and_b32 s12, s8, 0x100000
; GFX9-NEXT: s_mov_b32 s13, s5
; GFX9-NEXT: s_cmp_eq_u64 s[12:13], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s11
; GFX9-NEXT: s_cselect_b32 s10, 0, s10
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_lshl_b64 s[10:11], s[4:5], 21
; GFX9-NEXT: s_and_b32 s12, s8, 0x200000
; GFX9-NEXT: s_cmp_eq_u64 s[12:13], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s11
; GFX9-NEXT: s_cselect_b32 s10, 0, s10
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_lshl_b64 s[10:11], s[4:5], 22
; GFX9-NEXT: s_and_b32 s12, s8, 0x400000
; GFX9-NEXT: s_cmp_eq_u64 s[12:13], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s11
; GFX9-NEXT: s_cselect_b32 s10, 0, s10
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_lshl_b64 s[10:11], s[4:5], 23
; GFX9-NEXT: s_and_b32 s12, s8, 0x800000
; GFX9-NEXT: s_cmp_eq_u64 s[12:13], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s11
; GFX9-NEXT: s_cselect_b32 s10, 0, s10
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_lshl_b64 s[10:11], s[4:5], 24
; GFX9-NEXT: s_and_b32 s12, s8, 0x1000000
; GFX9-NEXT: s_cmp_eq_u64 s[12:13], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s11
; GFX9-NEXT: s_cselect_b32 s10, 0, s10
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_lshl_b64 s[10:11], s[4:5], 25
; GFX9-NEXT: s_and_b32 s12, s8, 0x2000000
; GFX9-NEXT: s_cmp_eq_u64 s[12:13], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s11
; GFX9-NEXT: s_cselect_b32 s10, 0, s10
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_lshl_b64 s[10:11], s[4:5], 26
; GFX9-NEXT: s_and_b32 s12, s8, 0x4000000
; GFX9-NEXT: s_cmp_eq_u64 s[12:13], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s11
; GFX9-NEXT: s_cselect_b32 s10, 0, s10
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_lshl_b64 s[10:11], s[4:5], 27
; GFX9-NEXT: s_and_b32 s12, s8, 0x8000000
; GFX9-NEXT: s_cmp_eq_u64 s[12:13], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s11
; GFX9-NEXT: s_cselect_b32 s10, 0, s10
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_lshl_b64 s[10:11], s[4:5], 28
; GFX9-NEXT: s_and_b32 s12, s8, 0x10000000
; GFX9-NEXT: s_cmp_eq_u64 s[12:13], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s11
; GFX9-NEXT: s_cselect_b32 s10, 0, s10
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_lshl_b64 s[10:11], s[4:5], 29
; GFX9-NEXT: s_and_b32 s12, s8, 0x20000000
; GFX9-NEXT: s_cmp_eq_u64 s[12:13], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s11
; GFX9-NEXT: s_cselect_b32 s10, 0, s10
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_lshl_b64 s[10:11], s[4:5], 30
; GFX9-NEXT: s_and_b32 s12, s8, 2.0
; GFX9-NEXT: s_cmp_eq_u64 s[12:13], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s11
; GFX9-NEXT: s_cselect_b32 s10, 0, s10
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_lshl_b64 s[4:5], s[4:5], 31
; GFX9-NEXT: s_cmp_gt_i32 s8, -1
; GFX9-NEXT: s_cselect_b32 s5, 0, s5
; GFX9-NEXT: s_cselect_b32 s4, 0, s4
; GFX9-NEXT: s_xor_b64 s[4:5], s[6:7], s[4:5]
; GFX9-NEXT: s_lshr_b64 s[4:5], s[4:5], 31
; GFX9-NEXT: v_mov_b32_e32 v0, s4
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], 0
; GFX9-NEXT: s_endpgm
;
; GFX10-LABEL: test_clmulr_i32:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
; GFX10-NEXT: s_mov_b32 s6, -1
; GFX10-NEXT: s_mov_b32 s7, 0x31016000
; GFX10-NEXT: s_mov_b32 s10, s6
; GFX10-NEXT: s_mov_b32 s11, s7
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: s_mov_b32 s8, s2
; GFX10-NEXT: s_mov_b32 s9, s3
; GFX10-NEXT: s_mov_b32 s3, 0
; GFX10-NEXT: buffer_load_dwordx2 v[0:1], off, s[8:11], 0
; GFX10-NEXT: s_waitcnt_depctr depctr_vm_vsrc(0)
; GFX10-NEXT: s_mov_b32 s11, s3
; GFX10-NEXT: s_mov_b32 s9, s3
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: v_readfirstlane_b32 s4, v1
; GFX10-NEXT: v_readfirstlane_b32 s2, v0
; GFX10-NEXT: s_bfe_i32 s5, s4, 0x10000
; GFX10-NEXT: s_and_b32 s10, s4, 2
; GFX10-NEXT: s_lshl_b64 s[12:13], s[2:3], 1
; GFX10-NEXT: s_and_b32 s8, s5, s2
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_mov_b32 s5, s1
; GFX10-NEXT: s_cselect_b32 s13, 0, s13
; GFX10-NEXT: s_cselect_b32 s12, 0, s12
; GFX10-NEXT: s_and_b32 s10, s4, 4
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 2
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 8
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 3
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 16
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 4
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 32
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 5
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 64
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 6
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x80
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 7
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x100
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 8
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x200
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 9
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x400
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 10
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x800
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 11
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x1000
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 12
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x2000
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 13
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x4000
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 14
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x8000
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 15
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x10000
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 16
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x20000
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 17
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x40000
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 18
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x80000
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 19
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x100000
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 20
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x200000
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 21
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x400000
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 22
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x800000
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 23
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x1000000
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 24
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x2000000
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 25
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x4000000
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 26
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x8000000
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 27
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x10000000
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 28
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x20000000
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 29
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 2.0
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 30
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s11, 0, s15
; GFX10-NEXT: s_cselect_b32 s10, 0, s14
; GFX10-NEXT: s_lshl_b64 s[2:3], s[2:3], 31
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; GFX10-NEXT: s_cmp_gt_i32 s4, -1
; GFX10-NEXT: s_mov_b32 s4, s0
; GFX10-NEXT: s_cselect_b32 s3, 0, s3
; GFX10-NEXT: s_cselect_b32 s2, 0, s2
; GFX10-NEXT: s_xor_b64 s[2:3], s[8:9], s[2:3]
; GFX10-NEXT: s_lshr_b64 s[2:3], s[2:3], 31
; GFX10-NEXT: v_mov_b32_e32 v0, s2
; GFX10-NEXT: buffer_store_dword v0, off, s[4:7], 0
; GFX10-NEXT: s_endpgm
;
; GFX11-LABEL: test_clmulr_i32:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT: s_mov_b32 s6, -1
; GFX11-NEXT: s_mov_b32 s7, 0x31016000
; GFX11-NEXT: s_mov_b32 s10, s6
; GFX11-NEXT: s_mov_b32 s11, s7
; GFX11-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-NEXT: s_mov_b32 s8, s2
; GFX11-NEXT: s_mov_b32 s9, s3
; GFX11-NEXT: s_mov_b32 s3, 0
; GFX11-NEXT: buffer_load_b64 v[0:1], off, s[8:11], 0
; GFX11-NEXT: s_mov_b32 s11, s3
; GFX11-NEXT: s_mov_b32 s9, s3
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: v_readfirstlane_b32 s4, v1
; GFX11-NEXT: v_readfirstlane_b32 s2, v0
; GFX11-NEXT: s_bfe_i32 s5, s4, 0x10000
; GFX11-NEXT: s_and_b32 s10, s4, 2
; GFX11-NEXT: s_lshl_b64 s[12:13], s[2:3], 1
; GFX11-NEXT: s_and_b32 s8, s5, s2
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_mov_b32 s5, s1
; GFX11-NEXT: s_cselect_b32 s13, 0, s13
; GFX11-NEXT: s_cselect_b32 s12, 0, s12
; GFX11-NEXT: s_and_b32 s10, s4, 4
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 2
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 8
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 3
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 16
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 4
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 32
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 5
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 64
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 6
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x80
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 7
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x100
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 8
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x200
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 9
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x400
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 10
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x800
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 11
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x1000
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 12
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x2000
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 13
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x4000
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 14
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x8000
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 15
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x10000
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 16
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x20000
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 17
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x40000
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 18
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x80000
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 19
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x100000
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 20
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x200000
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 21
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x400000
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 22
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x800000
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 23
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x1000000
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 24
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x2000000
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 25
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x4000000
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 26
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x8000000
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 27
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x10000000
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 28
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x20000000
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 29
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 2.0
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 30
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s11, 0, s15
; GFX11-NEXT: s_cselect_b32 s10, 0, s14
; GFX11-NEXT: s_lshl_b64 s[2:3], s[2:3], 31
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; GFX11-NEXT: s_cmp_gt_i32 s4, -1
; GFX11-NEXT: s_mov_b32 s4, s0
; GFX11-NEXT: s_cselect_b32 s3, 0, s3
; GFX11-NEXT: s_cselect_b32 s2, 0, s2
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
; GFX11-NEXT: s_xor_b64 s[2:3], s[8:9], s[2:3]
; GFX11-NEXT: s_lshr_b64 s[2:3], s[2:3], 31
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: v_mov_b32_e32 v0, s2
; GFX11-NEXT: buffer_store_b32 v0, off, s[4:7], 0
; GFX11-NEXT: s_endpgm
;
; GFX12-LABEL: test_clmulr_i32:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24
; GFX12-NEXT: s_mov_b32 s6, -1
; GFX12-NEXT: s_mov_b32 s7, 0x31016000
; GFX12-NEXT: s_mov_b32 s10, s6
; GFX12-NEXT: s_mov_b32 s11, s7
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_mov_b32 s8, s2
; GFX12-NEXT: s_mov_b32 s9, s3
; GFX12-NEXT: s_mov_b32 s3, 0
; GFX12-NEXT: buffer_load_b64 v[0:1], off, s[8:11], null
; GFX12-NEXT: s_mov_b32 s5, s3
; GFX12-NEXT: s_mov_b32 s9, s3
; GFX12-NEXT: s_mov_b32 s11, s3
; GFX12-NEXT: s_mov_b32 s13, s3
; GFX12-NEXT: s_mov_b32 s17, s3
; GFX12-NEXT: s_mov_b32 s15, s3
; GFX12-NEXT: s_mov_b32 s19, s3
; GFX12-NEXT: s_mov_b32 s20, s3
; GFX12-NEXT: s_mov_b32 s22, s3
; GFX12-NEXT: s_wait_loadcnt 0x0
; GFX12-NEXT: v_readfirstlane_b32 s16, v1
; GFX12-NEXT: v_readfirstlane_b32 s18, v0
; GFX12-NEXT: s_and_b32 s2, s16, 0x44444444
; GFX12-NEXT: s_and_b32 s4, s18, 0x22222222
; GFX12-NEXT: s_and_b32 s8, s16, 0x88888888
; GFX12-NEXT: s_and_b32 s10, s18, 0x11111111
; GFX12-NEXT: s_and_b32 s12, s16, 0x22222222
; GFX12-NEXT: s_and_b32 s16, s16, 0x11111111
; GFX12-NEXT: s_and_b32 s14, s18, 0x44444444
; GFX12-NEXT: s_and_b32 s18, s18, 0x88888888
; GFX12-NEXT: s_mul_u64 s[24:25], s[4:5], s[2:3]
; GFX12-NEXT: s_mul_u64 s[26:27], s[10:11], s[8:9]
; GFX12-NEXT: s_mul_u64 s[34:35], s[4:5], s[16:17]
; GFX12-NEXT: s_mul_u64 s[36:37], s[10:11], s[12:13]
; GFX12-NEXT: s_mul_u64 s[42:43], s[4:5], s[8:9]
; GFX12-NEXT: s_mul_u64 s[44:45], s[10:11], s[16:17]
; GFX12-NEXT: s_mul_u64 s[28:29], s[14:15], s[12:13]
; GFX12-NEXT: s_mul_u64 s[30:31], s[18:19], s[16:17]
; GFX12-NEXT: s_mul_u64 s[38:39], s[14:15], s[8:9]
; GFX12-NEXT: s_mul_u64 s[40:41], s[18:19], s[2:3]
; GFX12-NEXT: s_mul_u64 s[46:47], s[14:15], s[2:3]
; GFX12-NEXT: s_mul_u64 s[48:49], s[18:19], s[12:13]
; GFX12-NEXT: s_mul_u64 s[4:5], s[4:5], s[12:13]
; GFX12-NEXT: s_mul_u64 s[10:11], s[10:11], s[2:3]
; GFX12-NEXT: s_mul_u64 s[12:13], s[14:15], s[16:17]
; GFX12-NEXT: s_mul_u64 s[8:9], s[18:19], s[8:9]
; GFX12-NEXT: s_xor_b64 s[14:15], s[26:27], s[24:25]
; GFX12-NEXT: s_xor_b64 s[16:17], s[36:37], s[34:35]
; GFX12-NEXT: s_xor_b64 s[18:19], s[44:45], s[42:43]
; GFX12-NEXT: s_xor_b64 s[4:5], s[10:11], s[4:5]
; GFX12-NEXT: s_xor_b64 s[10:11], s[14:15], s[28:29]
; GFX12-NEXT: s_xor_b64 s[14:15], s[16:17], s[38:39]
; GFX12-NEXT: s_xor_b64 s[16:17], s[18:19], s[46:47]
; GFX12-NEXT: s_xor_b64 s[4:5], s[4:5], s[12:13]
; GFX12-NEXT: s_xor_b64 s[12:13], s[14:15], s[40:41]
; GFX12-NEXT: s_xor_b64 s[14:15], s[16:17], s[48:49]
; GFX12-NEXT: s_xor_b64 s[4:5], s[4:5], s[8:9]
; GFX12-NEXT: s_and_b32 s21, s13, 0x22222222
; GFX12-NEXT: s_and_b32 s23, s15, 0x11111111
; GFX12-NEXT: s_xor_b64 s[10:11], s[10:11], s[30:31]
; GFX12-NEXT: s_and_b32 s5, s5, 0x44444444
; GFX12-NEXT: s_or_b64 s[12:13], s[22:23], s[20:21]
; GFX12-NEXT: s_mov_b32 s4, s3
; GFX12-NEXT: s_and_b32 s9, s11, 0x8888888
; GFX12-NEXT: s_and_b32 s8, s10, 0x80000000
; GFX12-NEXT: s_or_b64 s[2:3], s[12:13], s[4:5]
; GFX12-NEXT: s_mov_b32 s4, s0
; GFX12-NEXT: s_or_b64 s[2:3], s[2:3], s[8:9]
; GFX12-NEXT: s_mov_b32 s5, s1
; GFX12-NEXT: s_lshr_b64 s[2:3], s[2:3], 31
; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX12-NEXT: v_mov_b32_e32 v0, s2
; GFX12-NEXT: buffer_store_b32 v0, off, s[4:7], null
; GFX12-NEXT: s_endpgm
;
; GFX1250-LABEL: test_clmulr_i32:
; GFX1250: ; %bb.0:
; GFX1250-NEXT: global_prefetch_b8 v0, s[0:1] scope:SCOPE_SE
; GFX1250-NEXT: v_nop
; GFX1250-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; msbs: dst=0 src0=0 src1=0 src2=0
; GFX1250-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 nv
; GFX1250-NEXT: s_mov_b32 s6, -1
; GFX1250-NEXT: s_mov_b32 s7, 0x31016000
; GFX1250-NEXT: s_mov_b32 s10, s6
; GFX1250-NEXT: s_mov_b32 s11, s7
; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: s_mov_b64 s[4:5], 0x88888888
; GFX1250-NEXT: s_wait_kmcnt 0x0
; GFX1250-NEXT: s_mov_b32 s8, s2
; GFX1250-NEXT: s_mov_b32 s9, s3
; GFX1250-NEXT: s_mov_b32 s3, 0
; GFX1250-NEXT: buffer_load_b64 v[0:1], off, s[8:11], null
; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: s_mov_b32 s9, s3
; GFX1250-NEXT: s_mov_b32 s13, s3
; GFX1250-NEXT: s_mov_b32 s15, s3
; GFX1250-NEXT: s_mov_b32 s11, s3
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: v_readfirstlane_b32 s2, v1
; GFX1250-NEXT: v_readfirstlane_b32 s10, v0
; GFX1250-NEXT: s_and_b64 s[16:17], s[2:3], 0x11111111
; GFX1250-NEXT: s_and_b32 s8, s10, 0x22222222
; GFX1250-NEXT: s_and_b64 s[18:19], s[2:3], 0x22222222
; GFX1250-NEXT: s_and_b32 s12, s10, 0x11111111
; GFX1250-NEXT: s_and_b32 s14, s2, 0x88888888
; GFX1250-NEXT: s_and_b64 s[20:21], s[10:11], 0x44444444
; GFX1250-NEXT: s_and_b32 s2, s2, 0x44444444
; GFX1250-NEXT: s_and_b64 s[4:5], s[10:11], s[4:5]
; GFX1250-NEXT: s_mul_u64 s[10:11], s[8:9], s[16:17]
; GFX1250-NEXT: s_mul_u64 s[22:23], s[12:13], s[18:19]
; GFX1250-NEXT: s_mul_u64 s[28:29], s[8:9], s[14:15]
; GFX1250-NEXT: s_mul_u64 s[30:31], s[12:13], s[16:17]
; GFX1250-NEXT: s_mul_u64 s[24:25], s[20:21], s[14:15]
; GFX1250-NEXT: s_mul_u64 s[26:27], s[4:5], s[2:3]
; GFX1250-NEXT: s_mul_u64 s[34:35], s[20:21], s[2:3]
; GFX1250-NEXT: s_mul_u64 s[38:39], s[8:9], s[18:19]
; GFX1250-NEXT: s_mul_u64 s[40:41], s[12:13], s[2:3]
; GFX1250-NEXT: s_mul_u64 s[44:45], s[4:5], s[14:15]
; GFX1250-NEXT: s_mul_u64 s[2:3], s[8:9], s[2:3]
; GFX1250-NEXT: s_mul_u64 s[8:9], s[12:13], s[14:15]
; GFX1250-NEXT: s_xor_b64 s[10:11], s[22:23], s[10:11]
; GFX1250-NEXT: s_xor_b64 s[14:15], s[30:31], s[28:29]
; GFX1250-NEXT: s_mul_u64 s[36:37], s[4:5], s[18:19]
; GFX1250-NEXT: s_mul_u64 s[42:43], s[20:21], s[16:17]
; GFX1250-NEXT: s_mul_u64 s[4:5], s[4:5], s[16:17]
; GFX1250-NEXT: s_xor_b64 s[16:17], s[40:41], s[38:39]
; GFX1250-NEXT: s_xor_b64 s[2:3], s[8:9], s[2:3]
; GFX1250-NEXT: s_xor_b64 s[8:9], s[10:11], s[24:25]
; GFX1250-NEXT: s_xor_b64 s[10:11], s[14:15], s[34:35]
; GFX1250-NEXT: s_mul_u64 s[12:13], s[20:21], s[18:19]
; GFX1250-NEXT: s_xor_b64 s[14:15], s[16:17], s[42:43]
; GFX1250-NEXT: s_xor_b64 s[8:9], s[8:9], s[26:27]
; GFX1250-NEXT: s_xor_b64 s[10:11], s[10:11], s[36:37]
; GFX1250-NEXT: s_xor_b64 s[2:3], s[2:3], s[12:13]
; GFX1250-NEXT: s_xor_b64 s[12:13], s[14:15], s[44:45]
; GFX1250-NEXT: s_and_b64 s[8:9], s[8:9], 0x2222222200000000
; GFX1250-NEXT: s_and_b64 s[10:11], s[10:11], 0x1111111100000000
; GFX1250-NEXT: s_and_b64 s[12:13], s[12:13], 0x4444444400000000
; GFX1250-NEXT: s_or_b64 s[8:9], s[10:11], s[8:9]
; GFX1250-NEXT: s_xor_b64 s[2:3], s[2:3], s[4:5]
; GFX1250-NEXT: s_or_b64 s[4:5], s[8:9], s[12:13]
; GFX1250-NEXT: s_and_b64 s[2:3], s[2:3], 0x888888880000000
; GFX1250-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX1250-NEXT: s_or_b64 s[2:3], s[4:5], s[2:3]
; GFX1250-NEXT: s_mov_b32 s4, s0
; GFX1250-NEXT: s_lshr_b64 s[2:3], s[2:3], 31
; GFX1250-NEXT: s_mov_b32 s5, s1
; GFX1250-NEXT: v_mov_b32_e32 v0, s2
; GFX1250-NEXT: buffer_store_b32 v0, off, s[4:7], null
; GFX1250-NEXT: s_endpgm
;
; EG-LABEL: test_clmulr_i32:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 0 @6
; EG-NEXT: ALU 92, @9, KC0[], KC1[]
; EG-NEXT: ALU 52, @102, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.X, T1.X, 1
; EG-NEXT: CF_END
; 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: AND_INT T0.W, T0.X, literal.x,
; EG-NEXT: LSHR * T1.W, T0.X, literal.y,
; EG-NEXT: 65280(9.147676e-41), 8(1.121039e-44)
; EG-NEXT: AND_INT T1.X, PS, literal.x,
; EG-NEXT: LSHR T1.Y, T0.X, literal.y,
; EG-NEXT: LSHL T0.Z, PV.W, literal.z,
; EG-NEXT: AND_INT T0.W, T0.Y, literal.x,
; EG-NEXT: LSHR * T1.W, T0.Y, literal.z,
; EG-NEXT: 65280(9.147676e-41), 24(3.363116e-44)
; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00)
; EG-NEXT: LSHL T0.X, T0.X, literal.x,
; EG-NEXT: AND_INT T2.Y, PS, literal.y,
; EG-NEXT: LSHR T1.Z, T0.Y, literal.x,
; EG-NEXT: LSHL T0.W, PV.W, literal.z,
; EG-NEXT: LSHL * T1.W, T0.Y, literal.x,
; EG-NEXT: 24(3.363116e-44), 65280(9.147676e-41)
; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00)
; EG-NEXT: OR_INT T0.Y, PS, PV.W,
; EG-NEXT: OR_INT T1.Z, PV.Y, PV.Z,
; EG-NEXT: OR_INT T0.W, PV.X, T0.Z,
; EG-NEXT: OR_INT * T1.W, T1.X, T1.Y,
; EG-NEXT: OR_INT T0.W, PV.W, PS,
; EG-NEXT: OR_INT * T1.W, PV.Y, PV.Z,
; EG-NEXT: AND_INT T0.Y, PS, literal.x,
; EG-NEXT: LSHR T0.Z, PS, literal.y,
; EG-NEXT: AND_INT T1.W, PV.W, literal.x,
; EG-NEXT: LSHR * T0.W, PV.W, literal.y,
; EG-NEXT: 252645135(7.053345e-30), 4(5.605194e-45)
; EG-NEXT: AND_INT T1.Y, PS, literal.x,
; EG-NEXT: LSHL T1.Z, PV.W, literal.y,
; EG-NEXT: AND_INT T0.W, PV.Z, literal.x,
; EG-NEXT: LSHL * T1.W, PV.Y, literal.y,
; EG-NEXT: 252645135(7.053345e-30), 4(5.605194e-45)
; EG-NEXT: OR_INT T0.W, PV.W, PS,
; EG-NEXT: OR_INT * T1.W, PV.Y, PV.Z,
; EG-NEXT: AND_INT T0.Y, PS, literal.x,
; EG-NEXT: LSHR T0.Z, PS, literal.y,
; EG-NEXT: AND_INT T1.W, PV.W, literal.x,
; EG-NEXT: LSHR * T0.W, PV.W, literal.y,
; EG-NEXT: 858993459(4.172325e-08), 2(2.802597e-45)
; EG-NEXT: AND_INT T1.Y, PS, literal.x,
; EG-NEXT: LSHL T1.Z, PV.W, literal.y,
; EG-NEXT: AND_INT T0.W, PV.Z, literal.x,
; EG-NEXT: LSHL * T1.W, PV.Y, literal.y,
; EG-NEXT: 858993459(4.172325e-08), 2(2.802597e-45)
; EG-NEXT: OR_INT T0.W, PV.W, PS,
; EG-NEXT: OR_INT * T1.W, PV.Y, PV.Z,
; EG-NEXT: AND_INT T0.Y, PS, literal.x,
; EG-NEXT: LSHR T0.Z, PS, 1,
; EG-NEXT: AND_INT T1.W, PV.W, literal.x,
; EG-NEXT: LSHR * T0.W, PV.W, 1,
; EG-NEXT: 1431655765(1.466015e+13), 0(0.000000e+00)
; EG-NEXT: AND_INT T1.Y, PS, literal.x,
; EG-NEXT: LSHL T1.Z, PV.W, 1,
; EG-NEXT: AND_INT T0.W, PV.Z, literal.x,
; EG-NEXT: LSHL * T1.W, PV.Y, 1,
; EG-NEXT: 1431655765(1.466015e+13), 0(0.000000e+00)
; EG-NEXT: OR_INT T0.W, PV.W, PS,
; EG-NEXT: OR_INT * T1.W, PV.Y, PV.Z,
; EG-NEXT: AND_INT T2.W, PS, literal.x,
; EG-NEXT: AND_INT * T3.W, PV.W, literal.y,
; EG-NEXT: 1145324612(7.850667e+02), 572662306(2.197316e-18)
; EG-NEXT: AND_INT T4.W, T0.W, literal.x,
; EG-NEXT: MULLO_INT * T0.X, PV.W, PS,
; EG-NEXT: 286331153(1.144374e-28), 0(0.000000e+00)
; EG-NEXT: AND_INT T5.W, T1.W, literal.x,
; EG-NEXT: MULLO_INT * T0.Y, T2.W, PV.W,
; EG-NEXT: -2004318072(-8.217301e-34), 0(0.000000e+00)
; EG-NEXT: AND_INT T6.W, T0.W, literal.x,
; EG-NEXT: MULLO_INT * T0.Z, PV.W, T3.W,
; EG-NEXT: 1145324612(7.850667e+02), 0(0.000000e+00)
; EG-NEXT: AND_INT T1.Z, T0.W, literal.x,
; EG-NEXT: AND_INT T0.W, T1.W, literal.y, BS:VEC_120/SCL_212
; EG-NEXT: MULLO_INT * T1.X, T5.W, PV.W,
; EG-NEXT: -2004318072(-8.217301e-34), 286331153(1.144374e-28)
; EG-NEXT: MULLO_INT * T1.Y, PV.W, PV.Z,
; EG-NEXT: MULLO_INT * T2.X, T2.W, T1.Z,
; EG-NEXT: MULLO_INT * T2.Y, T2.W, T6.W,
; EG-NEXT: MULLO_INT * T2.Z, T0.W, T6.W,
; EG-NEXT: AND_INT T1.W, T1.W, literal.x,
; EG-NEXT: MULLO_INT * T2.W, T0.W, T4.W,
; EG-NEXT: 572662306(2.197316e-18), 0(0.000000e+00)
; EG-NEXT: MULLO_INT * T3.X, PV.W, T1.Z,
; EG-NEXT: MULLO_INT * T0.W, T0.W, T3.W,
; EG-NEXT: MULLO_INT * T3.Y, T1.W, T4.W,
; EG-NEXT: XOR_INT T3.Z, T0.W, PS,
; EG-NEXT: XOR_INT * T0.W, T2.W, T3.X, BS:VEC_120/SCL_212
; EG-NEXT: MULLO_INT * T2.W, T1.W, T3.W,
; EG-NEXT: XOR_INT T3.Y, T2.Z, PS,
; EG-NEXT: XOR_INT T2.Z, T0.W, T2.Y,
; EG-NEXT: XOR_INT T0.W, T3.Z, T2.X, BS:VEC_120/SCL_212
; EG-NEXT: MULLO_INT * T1.W, T1.W, T6.W,
; EG-NEXT: XOR_INT * T2.X, T1.Y, PS,
; EG-NEXT: ALU clause starting at 102:
; EG-NEXT: XOR_INT T1.Y, T0.W, T1.X,
; EG-NEXT: XOR_INT T0.Z, T2.Z, T0.Z, BS:VEC_021/SCL_122
; EG-NEXT: XOR_INT T0.W, T3.Y, T0.Y,
; EG-NEXT: MULLO_INT * T0.Y, T5.W, T1.Z,
; EG-NEXT: XOR_INT T1.X, PV.W, PS,
; EG-NEXT: AND_INT T0.Y, PV.Z, literal.x,
; EG-NEXT: AND_INT T0.Z, PV.Y, literal.y,
; EG-NEXT: XOR_INT T0.W, T2.X, T0.X,
; EG-NEXT: MULLO_INT * T0.X, T5.W, T4.W,
; EG-NEXT: 286331153(1.144374e-28), 572662306(2.197316e-18)
; EG-NEXT: XOR_INT T1.Z, PV.W, PS,
; EG-NEXT: OR_INT T0.W, PV.Y, PV.Z,
; EG-NEXT: AND_INT * T1.W, PV.X, literal.x,
; EG-NEXT: 1145324612(7.850667e+02), 0(0.000000e+00)
; EG-NEXT: OR_INT T0.W, PV.W, PS,
; EG-NEXT: AND_INT * T1.W, PV.Z, literal.x,
; EG-NEXT: -2004318072(-8.217301e-34), 0(0.000000e+00)
; EG-NEXT: OR_INT * T0.W, PV.W, PS,
; EG-NEXT: AND_INT T1.W, PV.W, literal.x,
; EG-NEXT: LSHR * T2.W, PV.W, literal.y,
; EG-NEXT: 65280(9.147676e-41), 8(1.121039e-44)
; EG-NEXT: AND_INT T0.Y, PS, literal.x,
; EG-NEXT: LSHR T0.Z, T0.W, literal.y,
; EG-NEXT: LSHL T1.W, PV.W, literal.z,
; EG-NEXT: LSHL * T0.W, T0.W, literal.y,
; EG-NEXT: 65280(9.147676e-41), 24(3.363116e-44)
; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00)
; EG-NEXT: OR_INT T0.W, PS, PV.W,
; EG-NEXT: OR_INT * T1.W, PV.Y, PV.Z,
; EG-NEXT: OR_INT * T0.W, PV.W, PS,
; EG-NEXT: AND_INT T1.W, PV.W, literal.x,
; EG-NEXT: LSHR * T0.W, PV.W, literal.y,
; EG-NEXT: 252645135(7.053345e-30), 4(5.605194e-45)
; EG-NEXT: AND_INT T0.W, PS, literal.x,
; EG-NEXT: LSHL * T1.W, PV.W, literal.y,
; EG-NEXT: 252645135(7.053345e-30), 4(5.605194e-45)
; EG-NEXT: OR_INT * T0.W, PV.W, PS,
; EG-NEXT: AND_INT T1.W, PV.W, literal.x,
; EG-NEXT: LSHR * T0.W, PV.W, literal.y,
; EG-NEXT: 858993459(4.172325e-08), 2(2.802597e-45)
; EG-NEXT: AND_INT T0.W, PS, literal.x,
; EG-NEXT: LSHL * T1.W, PV.W, literal.y,
; EG-NEXT: 858993459(4.172325e-08), 2(2.802597e-45)
; EG-NEXT: OR_INT * T0.W, PV.W, PS,
; EG-NEXT: AND_INT T1.W, PV.W, literal.x,
; EG-NEXT: LSHR * T0.W, PV.W, 1,
; EG-NEXT: 1431655765(1.466015e+13), 0(0.000000e+00)
; EG-NEXT: AND_INT T0.W, PS, literal.x,
; EG-NEXT: LSHL * T1.W, PV.W, 1,
; EG-NEXT: 1431655765(1.466015e+13), 0(0.000000e+00)
; EG-NEXT: OR_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)
%b_ptr = getelementptr i32, ptr addrspace(1) %in, i32 1
%a = load i32, ptr addrspace(1) %in
%b = load i32, ptr addrspace(1) %b_ptr
%a.ext = zext i32 %a to i64
%b.ext = zext i32 %b to i64
%clmul = call i64 @llvm.clmul.i64(i64 %a.ext, i64 %b.ext)
%res.ext = lshr i64 %clmul, 31
%res = trunc i64 %res.ext to i32
store i32 %res, ptr addrspace(1) %out
ret void
}
define amdgpu_kernel void @test_clmulh_i32(ptr addrspace(1) %out, ptr addrspace(1) %in) {
; SI-LABEL: test_clmulh_i32:
; SI: ; %bb.0:
; SI-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9
; SI-NEXT: s_mov_b32 s3, 0xf000
; SI-NEXT: s_mov_b32 s2, -1
; SI-NEXT: s_mov_b32 s10, s2
; SI-NEXT: s_mov_b32 s11, s3
; SI-NEXT: s_waitcnt lgkmcnt(0)
; SI-NEXT: s_mov_b32 s8, s6
; SI-NEXT: s_mov_b32 s9, s7
; SI-NEXT: buffer_load_dwordx2 v[0:1], off, s[8:11], 0
; SI-NEXT: s_mov_b32 s7, 0
; SI-NEXT: s_mov_b32 s21, s7
; SI-NEXT: s_mov_b32 s0, s4
; SI-NEXT: s_mov_b32 s1, s5
; SI-NEXT: s_mov_b32 s9, s7
; SI-NEXT: s_mov_b32 s15, s7
; SI-NEXT: s_mov_b32 s11, s7
; SI-NEXT: s_mov_b32 s13, s7
; SI-NEXT: s_mov_b32 s17, s7
; SI-NEXT: s_mov_b32 s19, s7
; SI-NEXT: s_mov_b32 s23, s7
; SI-NEXT: s_mov_b32 s25, s7
; SI-NEXT: s_mov_b32 s27, s7
; SI-NEXT: s_mov_b32 s29, s7
; SI-NEXT: s_mov_b32 s31, s7
; SI-NEXT: s_mov_b32 s35, s7
; SI-NEXT: s_mov_b32 s37, s7
; SI-NEXT: s_mov_b32 s39, s7
; SI-NEXT: s_mov_b32 s41, s7
; SI-NEXT: s_mov_b32 s43, s7
; SI-NEXT: s_waitcnt vmcnt(0)
; SI-NEXT: v_readfirstlane_b32 s33, v1
; SI-NEXT: s_and_b32 s20, s33, 2
; SI-NEXT: v_readfirstlane_b32 s6, v0
; SI-NEXT: s_bfe_i32 s8, s33, 0x10000
; SI-NEXT: v_cmp_eq_u64_e64 s[20:21], s[20:21], 0
; SI-NEXT: s_lshl_b64 s[4:5], s[6:7], 1
; SI-NEXT: s_and_b32 s8, s8, s6
; SI-NEXT: s_and_b64 s[20:21], s[20:21], exec
; SI-NEXT: s_cselect_b32 s21, 0, s5
; SI-NEXT: s_cselect_b32 s20, 0, s4
; SI-NEXT: s_and_b32 s14, s33, 4
; SI-NEXT: s_xor_b64 s[20:21], s[8:9], s[20:21]
; SI-NEXT: v_cmp_eq_u64_e64 s[8:9], s[14:15], 0
; SI-NEXT: s_lshl_b64 s[14:15], s[6:7], 2
; SI-NEXT: s_and_b64 s[8:9], s[8:9], exec
; SI-NEXT: s_cselect_b32 s15, 0, s15
; SI-NEXT: s_cselect_b32 s14, 0, s14
; SI-NEXT: s_and_b32 s10, s33, 8
; SI-NEXT: v_cmp_eq_u64_e64 s[10:11], s[10:11], 0
; SI-NEXT: s_xor_b64 s[14:15], s[20:21], s[14:15]
; SI-NEXT: s_lshl_b64 s[20:21], s[6:7], 3
; SI-NEXT: s_and_b64 s[10:11], s[10:11], exec
; SI-NEXT: s_cselect_b32 s11, 0, s21
; SI-NEXT: s_cselect_b32 s10, 0, s20
; SI-NEXT: s_and_b32 s12, s33, 16
; SI-NEXT: v_cmp_eq_u64_e64 s[12:13], s[12:13], 0
; SI-NEXT: s_lshl_b64 s[20:21], s[6:7], 4
; SI-NEXT: s_xor_b64 s[10:11], s[14:15], s[10:11]
; SI-NEXT: s_and_b64 s[12:13], s[12:13], exec
; SI-NEXT: s_cselect_b32 s13, 0, s21
; SI-NEXT: s_cselect_b32 s12, 0, s20
; SI-NEXT: s_and_b32 s16, s33, 32
; SI-NEXT: v_cmp_eq_u64_e64 s[16:17], s[16:17], 0
; SI-NEXT: s_lshl_b64 s[14:15], s[6:7], 5
; SI-NEXT: s_xor_b64 s[10:11], s[10:11], s[12:13]
; SI-NEXT: s_and_b64 s[12:13], s[16:17], exec
; SI-NEXT: s_cselect_b32 s13, 0, s15
; SI-NEXT: s_cselect_b32 s12, 0, s14
; SI-NEXT: s_and_b32 s18, s33, 64
; SI-NEXT: v_cmp_eq_u64_e64 s[16:17], s[18:19], 0
; SI-NEXT: s_lshl_b64 s[14:15], s[6:7], 6
; SI-NEXT: s_xor_b64 s[10:11], s[10:11], s[12:13]
; SI-NEXT: s_and_b64 s[12:13], s[16:17], exec
; SI-NEXT: s_cselect_b32 s13, 0, s15
; SI-NEXT: s_cselect_b32 s12, 0, s14
; SI-NEXT: s_and_b32 s22, s33, 0x80
; SI-NEXT: v_cmp_eq_u64_e64 s[16:17], s[22:23], 0
; SI-NEXT: s_lshl_b64 s[14:15], s[6:7], 7
; SI-NEXT: s_xor_b64 s[10:11], s[10:11], s[12:13]
; SI-NEXT: s_and_b64 s[12:13], s[16:17], exec
; SI-NEXT: s_cselect_b32 s13, 0, s15
; SI-NEXT: s_cselect_b32 s12, 0, s14
; SI-NEXT: s_and_b32 s24, s33, 0x100
; SI-NEXT: v_cmp_eq_u64_e64 s[16:17], s[24:25], 0
; SI-NEXT: s_lshl_b64 s[14:15], s[6:7], 8
; SI-NEXT: s_xor_b64 s[10:11], s[10:11], s[12:13]
; SI-NEXT: s_and_b64 s[12:13], s[16:17], exec
; SI-NEXT: s_cselect_b32 s13, 0, s15
; SI-NEXT: s_cselect_b32 s12, 0, s14
; SI-NEXT: s_and_b32 s26, s33, 0x200
; SI-NEXT: v_cmp_eq_u64_e64 s[16:17], s[26:27], 0
; SI-NEXT: s_lshl_b64 s[14:15], s[6:7], 9
; SI-NEXT: s_xor_b64 s[10:11], s[10:11], s[12:13]
; SI-NEXT: s_and_b64 s[12:13], s[16:17], exec
; SI-NEXT: s_cselect_b32 s13, 0, s15
; SI-NEXT: s_cselect_b32 s12, 0, s14
; SI-NEXT: s_and_b32 s28, s33, 0x400
; SI-NEXT: v_cmp_eq_u64_e64 s[16:17], s[28:29], 0
; SI-NEXT: s_lshl_b64 s[14:15], s[6:7], 10
; SI-NEXT: s_xor_b64 s[10:11], s[10:11], s[12:13]
; SI-NEXT: s_and_b64 s[12:13], s[16:17], exec
; SI-NEXT: s_cselect_b32 s13, 0, s15
; SI-NEXT: s_cselect_b32 s12, 0, s14
; SI-NEXT: s_and_b32 s30, s33, 0x800
; SI-NEXT: v_cmp_eq_u64_e64 s[16:17], s[30:31], 0
; SI-NEXT: s_lshl_b64 s[14:15], s[6:7], 11
; SI-NEXT: s_xor_b64 s[10:11], s[10:11], s[12:13]
; SI-NEXT: s_and_b64 s[12:13], s[16:17], exec
; SI-NEXT: s_cselect_b32 s13, 0, s15
; SI-NEXT: s_cselect_b32 s12, 0, s14
; SI-NEXT: s_and_b32 s34, s33, 0x1000
; SI-NEXT: v_cmp_eq_u64_e64 s[16:17], s[34:35], 0
; SI-NEXT: s_lshl_b64 s[14:15], s[6:7], 12
; SI-NEXT: s_xor_b64 s[10:11], s[10:11], s[12:13]
; SI-NEXT: s_and_b64 s[12:13], s[16:17], exec
; SI-NEXT: s_cselect_b32 s13, 0, s15
; SI-NEXT: s_cselect_b32 s12, 0, s14
; SI-NEXT: s_and_b32 s36, s33, 0x2000
; SI-NEXT: v_cmp_eq_u64_e64 s[16:17], s[36:37], 0
; SI-NEXT: s_lshl_b64 s[14:15], s[6:7], 13
; SI-NEXT: s_xor_b64 s[10:11], s[10:11], s[12:13]
; SI-NEXT: s_and_b64 s[12:13], s[16:17], exec
; SI-NEXT: s_cselect_b32 s13, 0, s15
; SI-NEXT: s_cselect_b32 s12, 0, s14
; SI-NEXT: s_and_b32 s38, s33, 0x4000
; SI-NEXT: v_cmp_eq_u64_e64 s[16:17], s[38:39], 0
; SI-NEXT: s_lshl_b64 s[14:15], s[6:7], 14
; SI-NEXT: s_xor_b64 s[10:11], s[10:11], s[12:13]
; SI-NEXT: s_and_b64 s[12:13], s[16:17], exec
; SI-NEXT: s_cselect_b32 s13, 0, s15
; SI-NEXT: s_cselect_b32 s12, 0, s14
; SI-NEXT: s_and_b32 s40, s33, 0x8000
; SI-NEXT: v_cmp_eq_u64_e64 s[16:17], s[40:41], 0
; SI-NEXT: s_lshl_b64 s[14:15], s[6:7], 15
; SI-NEXT: s_xor_b64 s[10:11], s[10:11], s[12:13]
; SI-NEXT: s_and_b64 s[12:13], s[16:17], exec
; SI-NEXT: s_cselect_b32 s13, 0, s15
; SI-NEXT: s_cselect_b32 s12, 0, s14
; SI-NEXT: s_and_b32 s42, s33, 0x10000
; SI-NEXT: v_cmp_eq_u64_e64 s[16:17], s[42:43], 0
; SI-NEXT: s_lshl_b64 s[14:15], s[6:7], 16
; SI-NEXT: s_xor_b64 s[10:11], s[10:11], s[12:13]
; SI-NEXT: s_and_b64 s[12:13], s[16:17], exec
; SI-NEXT: s_mov_b32 s5, s7
; SI-NEXT: s_cselect_b32 s13, 0, s15
; SI-NEXT: s_cselect_b32 s12, 0, s14
; SI-NEXT: s_and_b32 s4, s33, 0x20000
; SI-NEXT: v_cmp_eq_u64_e64 s[4:5], s[4:5], 0
; SI-NEXT: s_lshl_b64 s[14:15], s[6:7], 17
; SI-NEXT: s_xor_b64 s[10:11], s[10:11], s[12:13]
; SI-NEXT: s_and_b64 s[4:5], s[4:5], exec
; SI-NEXT: s_mov_b32 s9, s7
; SI-NEXT: s_cselect_b32 s5, 0, s15
; SI-NEXT: s_cselect_b32 s4, 0, s14
; SI-NEXT: s_and_b32 s8, s33, 0x40000
; SI-NEXT: v_cmp_eq_u64_e64 s[8:9], s[8:9], 0
; SI-NEXT: s_lshl_b64 s[12:13], s[6:7], 18
; SI-NEXT: s_xor_b64 s[4:5], s[10:11], s[4:5]
; SI-NEXT: s_and_b64 s[8:9], s[8:9], exec
; SI-NEXT: s_cselect_b32 s9, 0, s13
; SI-NEXT: s_cselect_b32 s8, 0, s12
; SI-NEXT: s_and_b32 s10, s33, 0x80000
; SI-NEXT: s_mov_b32 s11, s7
; SI-NEXT: s_xor_b64 s[4:5], s[4:5], s[8:9]
; SI-NEXT: v_cmp_eq_u64_e64 s[8:9], s[10:11], 0
; SI-NEXT: s_lshl_b64 s[10:11], s[6:7], 19
; SI-NEXT: s_and_b64 s[8:9], s[8:9], exec
; SI-NEXT: s_cselect_b32 s9, 0, s11
; SI-NEXT: s_cselect_b32 s8, 0, s10
; SI-NEXT: s_xor_b64 s[4:5], s[4:5], s[8:9]
; SI-NEXT: s_and_b32 s8, s33, 0x100000
; SI-NEXT: s_mov_b32 s9, s7
; SI-NEXT: v_cmp_eq_u64_e64 s[8:9], s[8:9], 0
; SI-NEXT: s_lshl_b64 s[10:11], s[6:7], 20
; SI-NEXT: s_and_b64 s[8:9], s[8:9], exec
; SI-NEXT: s_cselect_b32 s9, 0, s11
; SI-NEXT: s_cselect_b32 s8, 0, s10
; SI-NEXT: s_xor_b64 s[4:5], s[4:5], s[8:9]
; SI-NEXT: s_and_b32 s8, s33, 0x200000
; SI-NEXT: s_mov_b32 s9, s7
; SI-NEXT: v_cmp_eq_u64_e64 s[8:9], s[8:9], 0
; SI-NEXT: s_lshl_b64 s[10:11], s[6:7], 21
; SI-NEXT: s_and_b64 s[8:9], s[8:9], exec
; SI-NEXT: s_cselect_b32 s9, 0, s11
; SI-NEXT: s_cselect_b32 s8, 0, s10
; SI-NEXT: s_xor_b64 s[4:5], s[4:5], s[8:9]
; SI-NEXT: s_and_b32 s8, s33, 0x400000
; SI-NEXT: s_mov_b32 s9, s7
; SI-NEXT: v_cmp_eq_u64_e64 s[8:9], s[8:9], 0
; SI-NEXT: s_lshl_b64 s[10:11], s[6:7], 22
; SI-NEXT: s_and_b64 s[8:9], s[8:9], exec
; SI-NEXT: s_cselect_b32 s9, 0, s11
; SI-NEXT: s_cselect_b32 s8, 0, s10
; SI-NEXT: s_xor_b64 s[4:5], s[4:5], s[8:9]
; SI-NEXT: s_and_b32 s8, s33, 0x800000
; SI-NEXT: s_mov_b32 s9, s7
; SI-NEXT: v_cmp_eq_u64_e64 s[8:9], s[8:9], 0
; SI-NEXT: s_lshl_b64 s[10:11], s[6:7], 23
; SI-NEXT: s_and_b64 s[8:9], s[8:9], exec
; SI-NEXT: s_cselect_b32 s9, 0, s11
; SI-NEXT: s_cselect_b32 s8, 0, s10
; SI-NEXT: s_xor_b64 s[4:5], s[4:5], s[8:9]
; SI-NEXT: s_and_b32 s8, s33, 0x1000000
; SI-NEXT: s_mov_b32 s9, s7
; SI-NEXT: v_cmp_eq_u64_e64 s[8:9], s[8:9], 0
; SI-NEXT: s_lshl_b64 s[10:11], s[6:7], 24
; SI-NEXT: s_and_b64 s[8:9], s[8:9], exec
; SI-NEXT: s_cselect_b32 s9, 0, s11
; SI-NEXT: s_cselect_b32 s8, 0, s10
; SI-NEXT: s_xor_b64 s[4:5], s[4:5], s[8:9]
; SI-NEXT: s_and_b32 s8, s33, 0x2000000
; SI-NEXT: s_mov_b32 s9, s7
; SI-NEXT: v_cmp_eq_u64_e64 s[8:9], s[8:9], 0
; SI-NEXT: s_lshl_b64 s[10:11], s[6:7], 25
; SI-NEXT: s_and_b64 s[8:9], s[8:9], exec
; SI-NEXT: s_cselect_b32 s9, 0, s11
; SI-NEXT: s_cselect_b32 s8, 0, s10
; SI-NEXT: s_xor_b64 s[4:5], s[4:5], s[8:9]
; SI-NEXT: s_and_b32 s8, s33, 0x4000000
; SI-NEXT: s_mov_b32 s9, s7
; SI-NEXT: v_cmp_eq_u64_e64 s[8:9], s[8:9], 0
; SI-NEXT: s_lshl_b64 s[10:11], s[6:7], 26
; SI-NEXT: s_and_b64 s[8:9], s[8:9], exec
; SI-NEXT: s_cselect_b32 s9, 0, s11
; SI-NEXT: s_cselect_b32 s8, 0, s10
; SI-NEXT: s_xor_b64 s[4:5], s[4:5], s[8:9]
; SI-NEXT: s_and_b32 s8, s33, 0x8000000
; SI-NEXT: s_mov_b32 s9, s7
; SI-NEXT: v_cmp_eq_u64_e64 s[8:9], s[8:9], 0
; SI-NEXT: s_lshl_b64 s[10:11], s[6:7], 27
; SI-NEXT: s_and_b64 s[8:9], s[8:9], exec
; SI-NEXT: s_cselect_b32 s9, 0, s11
; SI-NEXT: s_cselect_b32 s8, 0, s10
; SI-NEXT: s_xor_b64 s[4:5], s[4:5], s[8:9]
; SI-NEXT: s_and_b32 s8, s33, 0x10000000
; SI-NEXT: s_mov_b32 s9, s7
; SI-NEXT: v_cmp_eq_u64_e64 s[8:9], s[8:9], 0
; SI-NEXT: s_lshl_b64 s[10:11], s[6:7], 28
; SI-NEXT: s_and_b64 s[8:9], s[8:9], exec
; SI-NEXT: s_cselect_b32 s9, 0, s11
; SI-NEXT: s_cselect_b32 s8, 0, s10
; SI-NEXT: s_xor_b64 s[4:5], s[4:5], s[8:9]
; SI-NEXT: s_and_b32 s8, s33, 0x20000000
; SI-NEXT: s_mov_b32 s9, s7
; SI-NEXT: v_cmp_eq_u64_e64 s[8:9], s[8:9], 0
; SI-NEXT: s_lshl_b64 s[10:11], s[6:7], 29
; SI-NEXT: s_and_b64 s[8:9], s[8:9], exec
; SI-NEXT: s_cselect_b32 s9, 0, s11
; SI-NEXT: s_cselect_b32 s8, 0, s10
; SI-NEXT: s_xor_b64 s[4:5], s[4:5], s[8:9]
; SI-NEXT: s_and_b32 s8, s33, 2.0
; SI-NEXT: s_mov_b32 s9, s7
; SI-NEXT: v_cmp_eq_u64_e64 s[8:9], s[8:9], 0
; SI-NEXT: s_lshl_b64 s[10:11], s[6:7], 30
; SI-NEXT: s_and_b64 s[8:9], s[8:9], exec
; SI-NEXT: s_cselect_b32 s9, 0, s11
; SI-NEXT: s_cselect_b32 s8, 0, s10
; SI-NEXT: s_xor_b64 s[4:5], s[4:5], s[8:9]
; SI-NEXT: s_lshl_b64 s[6:7], s[6:7], 31
; SI-NEXT: s_cmp_gt_i32 s33, -1
; SI-NEXT: s_cselect_b32 s7, 0, s7
; SI-NEXT: s_cselect_b32 s6, 0, s6
; SI-NEXT: s_xor_b64 s[4:5], s[4:5], s[6:7]
; SI-NEXT: v_mov_b32_e32 v0, s5
; SI-NEXT: buffer_store_dword v0, off, s[0:3], 0
; SI-NEXT: s_endpgm
;
; VI-LABEL: test_clmulh_i32:
; VI: ; %bb.0:
; VI-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x24
; VI-NEXT: s_mov_b32 s3, 0xf000
; VI-NEXT: s_mov_b32 s2, -1
; VI-NEXT: s_mov_b32 s10, s2
; VI-NEXT: s_mov_b32 s11, s3
; VI-NEXT: s_waitcnt lgkmcnt(0)
; VI-NEXT: s_mov_b32 s8, s6
; VI-NEXT: s_mov_b32 s9, s7
; VI-NEXT: buffer_load_dwordx2 v[0:1], off, s[8:11], 0
; VI-NEXT: s_mov_b32 s0, s4
; VI-NEXT: s_mov_b32 s7, 0
; VI-NEXT: s_mov_b32 s1, s5
; VI-NEXT: s_mov_b32 s11, s7
; VI-NEXT: s_mov_b32 s9, s7
; VI-NEXT: s_mov_b32 s13, s7
; VI-NEXT: s_mov_b32 s15, s7
; VI-NEXT: s_mov_b32 s17, s7
; VI-NEXT: s_mov_b32 s19, s7
; VI-NEXT: s_mov_b32 s21, s7
; VI-NEXT: s_mov_b32 s23, s7
; VI-NEXT: s_mov_b32 s25, s7
; VI-NEXT: s_mov_b32 s27, s7
; VI-NEXT: s_mov_b32 s29, s7
; VI-NEXT: s_mov_b32 s31, s7
; VI-NEXT: s_mov_b32 s35, s7
; VI-NEXT: s_mov_b32 s37, s7
; VI-NEXT: s_mov_b32 s39, s7
; VI-NEXT: s_mov_b32 s41, s7
; VI-NEXT: s_mov_b32 s43, s7
; VI-NEXT: s_mov_b32 s45, s7
; VI-NEXT: s_mov_b32 s47, s7
; VI-NEXT: s_waitcnt vmcnt(0)
; VI-NEXT: v_readfirstlane_b32 s4, v1
; VI-NEXT: v_readfirstlane_b32 s6, v0
; VI-NEXT: s_bfe_i32 s5, s4, 0x10000
; VI-NEXT: s_lshl_b64 s[48:49], s[6:7], 1
; VI-NEXT: s_and_b32 s10, s4, 2
; VI-NEXT: s_and_b32 s8, s5, s6
; VI-NEXT: s_cmp_eq_u64 s[10:11], 0
; VI-NEXT: s_cselect_b32 s11, 0, s49
; VI-NEXT: s_cselect_b32 s10, 0, s48
; VI-NEXT: s_lshl_b64 s[48:49], s[6:7], 2
; VI-NEXT: s_and_b32 s12, s4, 4
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_cmp_eq_u64 s[12:13], 0
; VI-NEXT: s_cselect_b32 s11, 0, s49
; VI-NEXT: s_cselect_b32 s10, 0, s48
; VI-NEXT: s_lshl_b64 s[12:13], s[6:7], 3
; VI-NEXT: s_and_b32 s14, s4, 8
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_cmp_eq_u64 s[14:15], 0
; VI-NEXT: s_cselect_b32 s11, 0, s13
; VI-NEXT: s_cselect_b32 s10, 0, s12
; VI-NEXT: s_lshl_b64 s[12:13], s[6:7], 4
; VI-NEXT: s_and_b32 s16, s4, 16
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_cmp_eq_u64 s[16:17], 0
; VI-NEXT: s_cselect_b32 s11, 0, s13
; VI-NEXT: s_cselect_b32 s10, 0, s12
; VI-NEXT: s_lshl_b64 s[12:13], s[6:7], 5
; VI-NEXT: s_and_b32 s18, s4, 32
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_cmp_eq_u64 s[18:19], 0
; VI-NEXT: s_cselect_b32 s11, 0, s13
; VI-NEXT: s_cselect_b32 s10, 0, s12
; VI-NEXT: s_lshl_b64 s[12:13], s[6:7], 6
; VI-NEXT: s_and_b32 s20, s4, 64
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_cmp_eq_u64 s[20:21], 0
; VI-NEXT: s_cselect_b32 s11, 0, s13
; VI-NEXT: s_cselect_b32 s10, 0, s12
; VI-NEXT: s_lshl_b64 s[12:13], s[6:7], 7
; VI-NEXT: s_and_b32 s22, s4, 0x80
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_cmp_eq_u64 s[22:23], 0
; VI-NEXT: s_cselect_b32 s11, 0, s13
; VI-NEXT: s_cselect_b32 s10, 0, s12
; VI-NEXT: s_lshl_b64 s[12:13], s[6:7], 8
; VI-NEXT: s_and_b32 s24, s4, 0x100
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_cmp_eq_u64 s[24:25], 0
; VI-NEXT: s_cselect_b32 s11, 0, s13
; VI-NEXT: s_cselect_b32 s10, 0, s12
; VI-NEXT: s_lshl_b64 s[12:13], s[6:7], 9
; VI-NEXT: s_and_b32 s26, s4, 0x200
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_cmp_eq_u64 s[26:27], 0
; VI-NEXT: s_cselect_b32 s11, 0, s13
; VI-NEXT: s_cselect_b32 s10, 0, s12
; VI-NEXT: s_lshl_b64 s[12:13], s[6:7], 10
; VI-NEXT: s_and_b32 s28, s4, 0x400
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_cmp_eq_u64 s[28:29], 0
; VI-NEXT: s_cselect_b32 s11, 0, s13
; VI-NEXT: s_cselect_b32 s10, 0, s12
; VI-NEXT: s_lshl_b64 s[12:13], s[6:7], 11
; VI-NEXT: s_and_b32 s30, s4, 0x800
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_cmp_eq_u64 s[30:31], 0
; VI-NEXT: s_cselect_b32 s11, 0, s13
; VI-NEXT: s_cselect_b32 s10, 0, s12
; VI-NEXT: s_lshl_b64 s[12:13], s[6:7], 12
; VI-NEXT: s_and_b32 s34, s4, 0x1000
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_cmp_eq_u64 s[34:35], 0
; VI-NEXT: s_cselect_b32 s11, 0, s13
; VI-NEXT: s_cselect_b32 s10, 0, s12
; VI-NEXT: s_lshl_b64 s[12:13], s[6:7], 13
; VI-NEXT: s_and_b32 s36, s4, 0x2000
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_cmp_eq_u64 s[36:37], 0
; VI-NEXT: s_cselect_b32 s11, 0, s13
; VI-NEXT: s_cselect_b32 s10, 0, s12
; VI-NEXT: s_lshl_b64 s[12:13], s[6:7], 14
; VI-NEXT: s_and_b32 s38, s4, 0x4000
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_cmp_eq_u64 s[38:39], 0
; VI-NEXT: s_cselect_b32 s11, 0, s13
; VI-NEXT: s_cselect_b32 s10, 0, s12
; VI-NEXT: s_lshl_b64 s[12:13], s[6:7], 15
; VI-NEXT: s_and_b32 s40, s4, 0x8000
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_cmp_eq_u64 s[40:41], 0
; VI-NEXT: s_cselect_b32 s11, 0, s13
; VI-NEXT: s_cselect_b32 s10, 0, s12
; VI-NEXT: s_lshl_b64 s[12:13], s[6:7], 16
; VI-NEXT: s_and_b32 s42, s4, 0x10000
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_cmp_eq_u64 s[42:43], 0
; VI-NEXT: s_cselect_b32 s11, 0, s13
; VI-NEXT: s_cselect_b32 s10, 0, s12
; VI-NEXT: s_lshl_b64 s[12:13], s[6:7], 17
; VI-NEXT: s_and_b32 s44, s4, 0x20000
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_cmp_eq_u64 s[44:45], 0
; VI-NEXT: s_cselect_b32 s11, 0, s13
; VI-NEXT: s_cselect_b32 s10, 0, s12
; VI-NEXT: s_lshl_b64 s[12:13], s[6:7], 18
; VI-NEXT: s_and_b32 s46, s4, 0x40000
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_cmp_eq_u64 s[46:47], 0
; VI-NEXT: s_cselect_b32 s11, 0, s13
; VI-NEXT: s_cselect_b32 s10, 0, s12
; VI-NEXT: s_lshl_b64 s[12:13], s[6:7], 19
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_and_b32 s10, s4, 0x80000
; VI-NEXT: s_mov_b32 s11, s7
; VI-NEXT: s_cmp_eq_u64 s[10:11], 0
; VI-NEXT: s_cselect_b32 s11, 0, s13
; VI-NEXT: s_cselect_b32 s10, 0, s12
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_lshl_b64 s[10:11], s[6:7], 20
; VI-NEXT: s_and_b32 s12, s4, 0x100000
; VI-NEXT: s_mov_b32 s13, s7
; VI-NEXT: s_cmp_eq_u64 s[12:13], 0
; VI-NEXT: s_cselect_b32 s11, 0, s11
; VI-NEXT: s_cselect_b32 s10, 0, s10
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_lshl_b64 s[10:11], s[6:7], 21
; VI-NEXT: s_and_b32 s12, s4, 0x200000
; VI-NEXT: s_cmp_eq_u64 s[12:13], 0
; VI-NEXT: s_cselect_b32 s11, 0, s11
; VI-NEXT: s_cselect_b32 s10, 0, s10
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_lshl_b64 s[10:11], s[6:7], 22
; VI-NEXT: s_and_b32 s12, s4, 0x400000
; VI-NEXT: s_cmp_eq_u64 s[12:13], 0
; VI-NEXT: s_cselect_b32 s11, 0, s11
; VI-NEXT: s_cselect_b32 s10, 0, s10
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_lshl_b64 s[10:11], s[6:7], 23
; VI-NEXT: s_and_b32 s12, s4, 0x800000
; VI-NEXT: s_cmp_eq_u64 s[12:13], 0
; VI-NEXT: s_cselect_b32 s11, 0, s11
; VI-NEXT: s_cselect_b32 s10, 0, s10
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_lshl_b64 s[10:11], s[6:7], 24
; VI-NEXT: s_and_b32 s12, s4, 0x1000000
; VI-NEXT: s_cmp_eq_u64 s[12:13], 0
; VI-NEXT: s_cselect_b32 s11, 0, s11
; VI-NEXT: s_cselect_b32 s10, 0, s10
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_lshl_b64 s[10:11], s[6:7], 25
; VI-NEXT: s_and_b32 s12, s4, 0x2000000
; VI-NEXT: s_cmp_eq_u64 s[12:13], 0
; VI-NEXT: s_cselect_b32 s11, 0, s11
; VI-NEXT: s_cselect_b32 s10, 0, s10
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_lshl_b64 s[10:11], s[6:7], 26
; VI-NEXT: s_and_b32 s12, s4, 0x4000000
; VI-NEXT: s_cmp_eq_u64 s[12:13], 0
; VI-NEXT: s_cselect_b32 s11, 0, s11
; VI-NEXT: s_cselect_b32 s10, 0, s10
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_lshl_b64 s[10:11], s[6:7], 27
; VI-NEXT: s_and_b32 s12, s4, 0x8000000
; VI-NEXT: s_cmp_eq_u64 s[12:13], 0
; VI-NEXT: s_cselect_b32 s11, 0, s11
; VI-NEXT: s_cselect_b32 s10, 0, s10
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_lshl_b64 s[10:11], s[6:7], 28
; VI-NEXT: s_and_b32 s12, s4, 0x10000000
; VI-NEXT: s_cmp_eq_u64 s[12:13], 0
; VI-NEXT: s_cselect_b32 s11, 0, s11
; VI-NEXT: s_cselect_b32 s10, 0, s10
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_lshl_b64 s[10:11], s[6:7], 29
; VI-NEXT: s_and_b32 s12, s4, 0x20000000
; VI-NEXT: s_cmp_eq_u64 s[12:13], 0
; VI-NEXT: s_cselect_b32 s11, 0, s11
; VI-NEXT: s_cselect_b32 s10, 0, s10
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_lshl_b64 s[10:11], s[6:7], 30
; VI-NEXT: s_and_b32 s12, s4, 2.0
; VI-NEXT: s_cmp_eq_u64 s[12:13], 0
; VI-NEXT: s_cselect_b32 s11, 0, s11
; VI-NEXT: s_cselect_b32 s10, 0, s10
; VI-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; VI-NEXT: s_lshl_b64 s[6:7], s[6:7], 31
; VI-NEXT: s_cmp_gt_i32 s4, -1
; VI-NEXT: s_cselect_b32 s5, 0, s7
; VI-NEXT: s_cselect_b32 s4, 0, s6
; VI-NEXT: s_xor_b64 s[4:5], s[8:9], s[4:5]
; VI-NEXT: v_mov_b32_e32 v0, s5
; VI-NEXT: buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT: s_endpgm
;
; GFX9-LABEL: test_clmulh_i32:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_load_dwordx4 s[8:11], s[4:5], 0x24
; GFX9-NEXT: s_mov_b32 s3, 0xf000
; GFX9-NEXT: s_mov_b32 s2, -1
; GFX9-NEXT: s_mov_b32 s6, s2
; GFX9-NEXT: s_mov_b32 s7, s3
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_mov_b32 s4, s10
; GFX9-NEXT: s_mov_b32 s5, s11
; GFX9-NEXT: buffer_load_dwordx2 v[0:1], off, s[4:7], 0
; GFX9-NEXT: s_mov_b32 s0, s8
; GFX9-NEXT: s_mov_b32 s5, 0
; GFX9-NEXT: s_mov_b32 s11, s5
; GFX9-NEXT: s_mov_b32 s7, s5
; GFX9-NEXT: s_mov_b32 s13, s5
; GFX9-NEXT: s_mov_b32 s15, s5
; GFX9-NEXT: s_mov_b32 s17, s5
; GFX9-NEXT: s_mov_b32 s19, s5
; GFX9-NEXT: s_mov_b32 s21, s5
; GFX9-NEXT: s_mov_b32 s23, s5
; GFX9-NEXT: s_mov_b32 s25, s5
; GFX9-NEXT: s_mov_b32 s27, s5
; GFX9-NEXT: s_mov_b32 s29, s5
; GFX9-NEXT: s_mov_b32 s31, s5
; GFX9-NEXT: s_mov_b32 s35, s5
; GFX9-NEXT: s_mov_b32 s37, s5
; GFX9-NEXT: s_mov_b32 s39, s5
; GFX9-NEXT: s_mov_b32 s41, s5
; GFX9-NEXT: s_mov_b32 s43, s5
; GFX9-NEXT: s_mov_b32 s45, s5
; GFX9-NEXT: s_mov_b32 s47, s5
; GFX9-NEXT: s_mov_b32 s1, s9
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_readfirstlane_b32 s8, v1
; GFX9-NEXT: v_readfirstlane_b32 s4, v0
; GFX9-NEXT: s_bfe_i32 s6, s8, 0x10000
; GFX9-NEXT: s_lshl_b64 s[48:49], s[4:5], 1
; GFX9-NEXT: s_and_b32 s10, s8, 2
; GFX9-NEXT: s_and_b32 s6, s6, s4
; GFX9-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s49
; GFX9-NEXT: s_cselect_b32 s10, 0, s48
; GFX9-NEXT: s_lshl_b64 s[48:49], s[4:5], 2
; GFX9-NEXT: s_and_b32 s12, s8, 4
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_cmp_eq_u64 s[12:13], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s49
; GFX9-NEXT: s_cselect_b32 s10, 0, s48
; GFX9-NEXT: s_lshl_b64 s[12:13], s[4:5], 3
; GFX9-NEXT: s_and_b32 s14, s8, 8
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_cmp_eq_u64 s[14:15], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s13
; GFX9-NEXT: s_cselect_b32 s10, 0, s12
; GFX9-NEXT: s_lshl_b64 s[12:13], s[4:5], 4
; GFX9-NEXT: s_and_b32 s16, s8, 16
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_cmp_eq_u64 s[16:17], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s13
; GFX9-NEXT: s_cselect_b32 s10, 0, s12
; GFX9-NEXT: s_lshl_b64 s[12:13], s[4:5], 5
; GFX9-NEXT: s_and_b32 s18, s8, 32
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_cmp_eq_u64 s[18:19], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s13
; GFX9-NEXT: s_cselect_b32 s10, 0, s12
; GFX9-NEXT: s_lshl_b64 s[12:13], s[4:5], 6
; GFX9-NEXT: s_and_b32 s20, s8, 64
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_cmp_eq_u64 s[20:21], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s13
; GFX9-NEXT: s_cselect_b32 s10, 0, s12
; GFX9-NEXT: s_lshl_b64 s[12:13], s[4:5], 7
; GFX9-NEXT: s_and_b32 s22, s8, 0x80
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_cmp_eq_u64 s[22:23], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s13
; GFX9-NEXT: s_cselect_b32 s10, 0, s12
; GFX9-NEXT: s_lshl_b64 s[12:13], s[4:5], 8
; GFX9-NEXT: s_and_b32 s24, s8, 0x100
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_cmp_eq_u64 s[24:25], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s13
; GFX9-NEXT: s_cselect_b32 s10, 0, s12
; GFX9-NEXT: s_lshl_b64 s[12:13], s[4:5], 9
; GFX9-NEXT: s_and_b32 s26, s8, 0x200
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_cmp_eq_u64 s[26:27], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s13
; GFX9-NEXT: s_cselect_b32 s10, 0, s12
; GFX9-NEXT: s_lshl_b64 s[12:13], s[4:5], 10
; GFX9-NEXT: s_and_b32 s28, s8, 0x400
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_cmp_eq_u64 s[28:29], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s13
; GFX9-NEXT: s_cselect_b32 s10, 0, s12
; GFX9-NEXT: s_lshl_b64 s[12:13], s[4:5], 11
; GFX9-NEXT: s_and_b32 s30, s8, 0x800
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_cmp_eq_u64 s[30:31], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s13
; GFX9-NEXT: s_cselect_b32 s10, 0, s12
; GFX9-NEXT: s_lshl_b64 s[12:13], s[4:5], 12
; GFX9-NEXT: s_and_b32 s34, s8, 0x1000
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_cmp_eq_u64 s[34:35], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s13
; GFX9-NEXT: s_cselect_b32 s10, 0, s12
; GFX9-NEXT: s_lshl_b64 s[12:13], s[4:5], 13
; GFX9-NEXT: s_and_b32 s36, s8, 0x2000
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_cmp_eq_u64 s[36:37], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s13
; GFX9-NEXT: s_cselect_b32 s10, 0, s12
; GFX9-NEXT: s_lshl_b64 s[12:13], s[4:5], 14
; GFX9-NEXT: s_and_b32 s38, s8, 0x4000
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_cmp_eq_u64 s[38:39], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s13
; GFX9-NEXT: s_cselect_b32 s10, 0, s12
; GFX9-NEXT: s_lshl_b64 s[12:13], s[4:5], 15
; GFX9-NEXT: s_and_b32 s40, s8, 0x8000
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_cmp_eq_u64 s[40:41], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s13
; GFX9-NEXT: s_cselect_b32 s10, 0, s12
; GFX9-NEXT: s_lshl_b64 s[12:13], s[4:5], 16
; GFX9-NEXT: s_and_b32 s42, s8, 0x10000
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_cmp_eq_u64 s[42:43], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s13
; GFX9-NEXT: s_cselect_b32 s10, 0, s12
; GFX9-NEXT: s_lshl_b64 s[12:13], s[4:5], 17
; GFX9-NEXT: s_and_b32 s44, s8, 0x20000
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_cmp_eq_u64 s[44:45], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s13
; GFX9-NEXT: s_cselect_b32 s10, 0, s12
; GFX9-NEXT: s_lshl_b64 s[12:13], s[4:5], 18
; GFX9-NEXT: s_and_b32 s46, s8, 0x40000
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_cmp_eq_u64 s[46:47], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s13
; GFX9-NEXT: s_cselect_b32 s10, 0, s12
; GFX9-NEXT: s_lshl_b64 s[12:13], s[4:5], 19
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_and_b32 s10, s8, 0x80000
; GFX9-NEXT: s_mov_b32 s11, s5
; GFX9-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s13
; GFX9-NEXT: s_cselect_b32 s10, 0, s12
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_lshl_b64 s[10:11], s[4:5], 20
; GFX9-NEXT: s_and_b32 s12, s8, 0x100000
; GFX9-NEXT: s_mov_b32 s13, s5
; GFX9-NEXT: s_cmp_eq_u64 s[12:13], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s11
; GFX9-NEXT: s_cselect_b32 s10, 0, s10
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_lshl_b64 s[10:11], s[4:5], 21
; GFX9-NEXT: s_and_b32 s12, s8, 0x200000
; GFX9-NEXT: s_cmp_eq_u64 s[12:13], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s11
; GFX9-NEXT: s_cselect_b32 s10, 0, s10
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_lshl_b64 s[10:11], s[4:5], 22
; GFX9-NEXT: s_and_b32 s12, s8, 0x400000
; GFX9-NEXT: s_cmp_eq_u64 s[12:13], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s11
; GFX9-NEXT: s_cselect_b32 s10, 0, s10
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_lshl_b64 s[10:11], s[4:5], 23
; GFX9-NEXT: s_and_b32 s12, s8, 0x800000
; GFX9-NEXT: s_cmp_eq_u64 s[12:13], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s11
; GFX9-NEXT: s_cselect_b32 s10, 0, s10
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_lshl_b64 s[10:11], s[4:5], 24
; GFX9-NEXT: s_and_b32 s12, s8, 0x1000000
; GFX9-NEXT: s_cmp_eq_u64 s[12:13], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s11
; GFX9-NEXT: s_cselect_b32 s10, 0, s10
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_lshl_b64 s[10:11], s[4:5], 25
; GFX9-NEXT: s_and_b32 s12, s8, 0x2000000
; GFX9-NEXT: s_cmp_eq_u64 s[12:13], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s11
; GFX9-NEXT: s_cselect_b32 s10, 0, s10
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_lshl_b64 s[10:11], s[4:5], 26
; GFX9-NEXT: s_and_b32 s12, s8, 0x4000000
; GFX9-NEXT: s_cmp_eq_u64 s[12:13], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s11
; GFX9-NEXT: s_cselect_b32 s10, 0, s10
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_lshl_b64 s[10:11], s[4:5], 27
; GFX9-NEXT: s_and_b32 s12, s8, 0x8000000
; GFX9-NEXT: s_cmp_eq_u64 s[12:13], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s11
; GFX9-NEXT: s_cselect_b32 s10, 0, s10
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_lshl_b64 s[10:11], s[4:5], 28
; GFX9-NEXT: s_and_b32 s12, s8, 0x10000000
; GFX9-NEXT: s_cmp_eq_u64 s[12:13], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s11
; GFX9-NEXT: s_cselect_b32 s10, 0, s10
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_lshl_b64 s[10:11], s[4:5], 29
; GFX9-NEXT: s_and_b32 s12, s8, 0x20000000
; GFX9-NEXT: s_cmp_eq_u64 s[12:13], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s11
; GFX9-NEXT: s_cselect_b32 s10, 0, s10
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_lshl_b64 s[10:11], s[4:5], 30
; GFX9-NEXT: s_and_b32 s12, s8, 2.0
; GFX9-NEXT: s_cmp_eq_u64 s[12:13], 0
; GFX9-NEXT: s_cselect_b32 s11, 0, s11
; GFX9-NEXT: s_cselect_b32 s10, 0, s10
; GFX9-NEXT: s_xor_b64 s[6:7], s[6:7], s[10:11]
; GFX9-NEXT: s_lshl_b64 s[4:5], s[4:5], 31
; GFX9-NEXT: s_cmp_gt_i32 s8, -1
; GFX9-NEXT: s_cselect_b32 s5, 0, s5
; GFX9-NEXT: s_cselect_b32 s4, 0, s4
; GFX9-NEXT: s_xor_b64 s[4:5], s[6:7], s[4:5]
; GFX9-NEXT: v_mov_b32_e32 v0, s5
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], 0
; GFX9-NEXT: s_endpgm
;
; GFX10-LABEL: test_clmulh_i32:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
; GFX10-NEXT: s_mov_b32 s6, -1
; GFX10-NEXT: s_mov_b32 s7, 0x31016000
; GFX10-NEXT: s_mov_b32 s10, s6
; GFX10-NEXT: s_mov_b32 s11, s7
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: s_mov_b32 s8, s2
; GFX10-NEXT: s_mov_b32 s9, s3
; GFX10-NEXT: s_mov_b32 s3, 0
; GFX10-NEXT: buffer_load_dwordx2 v[0:1], off, s[8:11], 0
; GFX10-NEXT: s_waitcnt_depctr depctr_vm_vsrc(0)
; GFX10-NEXT: s_mov_b32 s11, s3
; GFX10-NEXT: s_mov_b32 s9, s3
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: v_readfirstlane_b32 s4, v1
; GFX10-NEXT: v_readfirstlane_b32 s2, v0
; GFX10-NEXT: s_bfe_i32 s5, s4, 0x10000
; GFX10-NEXT: s_and_b32 s10, s4, 2
; GFX10-NEXT: s_lshl_b64 s[12:13], s[2:3], 1
; GFX10-NEXT: s_and_b32 s8, s5, s2
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_mov_b32 s5, s1
; GFX10-NEXT: s_cselect_b32 s13, 0, s13
; GFX10-NEXT: s_cselect_b32 s12, 0, s12
; GFX10-NEXT: s_and_b32 s10, s4, 4
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 2
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 8
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 3
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 16
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 4
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 32
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 5
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 64
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 6
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x80
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 7
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x100
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 8
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x200
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 9
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x400
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 10
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x800
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 11
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x1000
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 12
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x2000
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 13
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x4000
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 14
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x8000
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 15
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x10000
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 16
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x20000
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 17
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x40000
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 18
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x80000
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 19
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x100000
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 20
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x200000
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 21
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x400000
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 22
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x800000
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 23
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x1000000
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 24
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x2000000
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 25
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x4000000
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 26
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x8000000
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 27
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x10000000
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 28
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 0x20000000
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 29
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s13, 0, s15
; GFX10-NEXT: s_cselect_b32 s12, 0, s14
; GFX10-NEXT: s_and_b32 s10, s4, 2.0
; GFX10-NEXT: s_lshl_b64 s[14:15], s[2:3], 30
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX10-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX10-NEXT: s_cselect_b32 s11, 0, s15
; GFX10-NEXT: s_cselect_b32 s10, 0, s14
; GFX10-NEXT: s_lshl_b64 s[2:3], s[2:3], 31
; GFX10-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; GFX10-NEXT: s_cmp_gt_i32 s4, -1
; GFX10-NEXT: s_mov_b32 s4, s0
; GFX10-NEXT: s_cselect_b32 s3, 0, s3
; GFX10-NEXT: s_cselect_b32 s2, 0, s2
; GFX10-NEXT: s_xor_b64 s[2:3], s[8:9], s[2:3]
; GFX10-NEXT: v_mov_b32_e32 v0, s3
; GFX10-NEXT: buffer_store_dword v0, off, s[4:7], 0
; GFX10-NEXT: s_endpgm
;
; GFX11-LABEL: test_clmulh_i32:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT: s_mov_b32 s6, -1
; GFX11-NEXT: s_mov_b32 s7, 0x31016000
; GFX11-NEXT: s_mov_b32 s10, s6
; GFX11-NEXT: s_mov_b32 s11, s7
; GFX11-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-NEXT: s_mov_b32 s8, s2
; GFX11-NEXT: s_mov_b32 s9, s3
; GFX11-NEXT: s_mov_b32 s3, 0
; GFX11-NEXT: buffer_load_b64 v[0:1], off, s[8:11], 0
; GFX11-NEXT: s_mov_b32 s11, s3
; GFX11-NEXT: s_mov_b32 s9, s3
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: v_readfirstlane_b32 s4, v1
; GFX11-NEXT: v_readfirstlane_b32 s2, v0
; GFX11-NEXT: s_bfe_i32 s5, s4, 0x10000
; GFX11-NEXT: s_and_b32 s10, s4, 2
; GFX11-NEXT: s_lshl_b64 s[12:13], s[2:3], 1
; GFX11-NEXT: s_and_b32 s8, s5, s2
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_mov_b32 s5, s1
; GFX11-NEXT: s_cselect_b32 s13, 0, s13
; GFX11-NEXT: s_cselect_b32 s12, 0, s12
; GFX11-NEXT: s_and_b32 s10, s4, 4
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 2
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 8
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 3
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 16
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 4
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 32
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 5
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 64
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 6
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x80
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 7
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x100
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 8
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x200
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 9
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x400
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 10
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x800
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 11
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x1000
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 12
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x2000
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 13
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x4000
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 14
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x8000
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 15
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x10000
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 16
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x20000
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 17
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x40000
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 18
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x80000
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 19
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x100000
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 20
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x200000
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 21
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x400000
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 22
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x800000
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 23
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x1000000
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 24
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x2000000
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 25
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x4000000
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 26
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x8000000
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 27
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x10000000
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 28
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 0x20000000
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 29
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s13, 0, s15
; GFX11-NEXT: s_cselect_b32 s12, 0, s14
; GFX11-NEXT: s_and_b32 s10, s4, 2.0
; GFX11-NEXT: s_lshl_b64 s[14:15], s[2:3], 30
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[12:13]
; GFX11-NEXT: s_cmp_eq_u64 s[10:11], 0
; GFX11-NEXT: s_cselect_b32 s11, 0, s15
; GFX11-NEXT: s_cselect_b32 s10, 0, s14
; GFX11-NEXT: s_lshl_b64 s[2:3], s[2:3], 31
; GFX11-NEXT: s_xor_b64 s[8:9], s[8:9], s[10:11]
; GFX11-NEXT: s_cmp_gt_i32 s4, -1
; GFX11-NEXT: s_mov_b32 s4, s0
; GFX11-NEXT: s_cselect_b32 s3, 0, s3
; GFX11-NEXT: s_cselect_b32 s2, 0, s2
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
; GFX11-NEXT: s_xor_b64 s[2:3], s[8:9], s[2:3]
; GFX11-NEXT: v_mov_b32_e32 v0, s3
; GFX11-NEXT: buffer_store_b32 v0, off, s[4:7], 0
; GFX11-NEXT: s_endpgm
;
; GFX12-LABEL: test_clmulh_i32:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x24
; GFX12-NEXT: s_mov_b32 s6, -1
; GFX12-NEXT: s_mov_b32 s7, 0x31016000
; GFX12-NEXT: s_mov_b32 s10, s6
; GFX12-NEXT: s_mov_b32 s11, s7
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_mov_b32 s8, s2
; GFX12-NEXT: s_mov_b32 s9, s3
; GFX12-NEXT: s_mov_b32 s3, 0
; GFX12-NEXT: buffer_load_b64 v[0:1], off, s[8:11], null
; GFX12-NEXT: s_mov_b32 s5, s3
; GFX12-NEXT: s_mov_b32 s9, s3
; GFX12-NEXT: s_mov_b32 s11, s3
; GFX12-NEXT: s_mov_b32 s13, s3
; GFX12-NEXT: s_mov_b32 s15, s3
; GFX12-NEXT: s_mov_b32 s17, s3
; GFX12-NEXT: s_mov_b32 s19, s3
; GFX12-NEXT: s_mov_b32 s20, s3
; GFX12-NEXT: s_mov_b32 s22, s3
; GFX12-NEXT: s_mov_b32 s24, s3
; GFX12-NEXT: s_wait_loadcnt 0x0
; GFX12-NEXT: v_readfirstlane_b32 s16, v1
; GFX12-NEXT: v_readfirstlane_b32 s18, v0
; GFX12-NEXT: s_and_b32 s2, s16, 0x11111111
; GFX12-NEXT: s_and_b32 s4, s18, 0x22222222
; GFX12-NEXT: s_and_b32 s8, s16, 0x22222222
; GFX12-NEXT: s_and_b32 s10, s18, 0x11111111
; GFX12-NEXT: s_and_b32 s12, s16, 0x88888888
; GFX12-NEXT: s_and_b32 s14, s18, 0x44444444
; GFX12-NEXT: s_and_b32 s16, s16, 0x44444444
; GFX12-NEXT: s_and_b32 s18, s18, 0x88888888
; GFX12-NEXT: s_mul_u64 s[26:27], s[4:5], s[2:3]
; GFX12-NEXT: s_mul_u64 s[28:29], s[10:11], s[8:9]
; GFX12-NEXT: s_mul_u64 s[36:37], s[4:5], s[12:13]
; GFX12-NEXT: s_mul_u64 s[38:39], s[10:11], s[2:3]
; GFX12-NEXT: s_mul_u64 s[30:31], s[14:15], s[12:13]
; GFX12-NEXT: s_mul_u64 s[34:35], s[18:19], s[16:17]
; GFX12-NEXT: s_mul_u64 s[40:41], s[14:15], s[16:17]
; GFX12-NEXT: s_mul_u64 s[42:43], s[18:19], s[8:9]
; GFX12-NEXT: s_mul_u64 s[44:45], s[4:5], s[8:9]
; GFX12-NEXT: s_mul_u64 s[46:47], s[10:11], s[16:17]
; GFX12-NEXT: s_mul_u64 s[48:49], s[14:15], s[2:3]
; GFX12-NEXT: s_mul_u64 s[4:5], s[4:5], s[16:17]
; GFX12-NEXT: s_mul_u64 s[10:11], s[10:11], s[12:13]
; GFX12-NEXT: s_mul_u64 s[8:9], s[14:15], s[8:9]
; GFX12-NEXT: s_xor_b64 s[14:15], s[28:29], s[26:27]
; GFX12-NEXT: s_xor_b64 s[16:17], s[38:39], s[36:37]
; GFX12-NEXT: s_mul_u64 s[50:51], s[18:19], s[12:13]
; GFX12-NEXT: s_mul_u64 s[12:13], s[18:19], s[2:3]
; GFX12-NEXT: s_xor_b64 s[18:19], s[46:47], s[44:45]
; GFX12-NEXT: s_xor_b64 s[4:5], s[10:11], s[4:5]
; GFX12-NEXT: s_xor_b64 s[10:11], s[14:15], s[30:31]
; GFX12-NEXT: s_xor_b64 s[14:15], s[16:17], s[40:41]
; GFX12-NEXT: s_xor_b64 s[16:17], s[18:19], s[48:49]
; GFX12-NEXT: s_xor_b64 s[4:5], s[4:5], s[8:9]
; GFX12-NEXT: s_xor_b64 s[8:9], s[10:11], s[34:35]
; GFX12-NEXT: s_xor_b64 s[10:11], s[14:15], s[42:43]
; GFX12-NEXT: s_xor_b64 s[14:15], s[16:17], s[50:51]
; GFX12-NEXT: s_and_b32 s21, s9, 0x22222222
; GFX12-NEXT: s_and_b32 s23, s11, 0x11111111
; GFX12-NEXT: s_xor_b64 s[4:5], s[4:5], s[12:13]
; GFX12-NEXT: s_and_b32 s25, s15, 0x44444444
; GFX12-NEXT: s_or_b64 s[8:9], s[22:23], s[20:21]
; GFX12-NEXT: s_and_b32 s5, s5, 0x88888888
; GFX12-NEXT: s_or_b64 s[8:9], s[8:9], s[24:25]
; GFX12-NEXT: s_mov_b32 s4, s3
; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX12-NEXT: s_or_b64 s[2:3], s[8:9], s[4:5]
; GFX12-NEXT: s_mov_b32 s4, s0
; GFX12-NEXT: v_mov_b32_e32 v0, s3
; GFX12-NEXT: s_mov_b32 s5, s1
; GFX12-NEXT: buffer_store_b32 v0, off, s[4:7], null
; GFX12-NEXT: s_endpgm
;
; GFX1250-LABEL: test_clmulh_i32:
; GFX1250: ; %bb.0:
; GFX1250-NEXT: global_prefetch_b8 v0, s[0:1] scope:SCOPE_SE
; GFX1250-NEXT: v_nop
; GFX1250-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; msbs: dst=0 src0=0 src1=0 src2=0
; GFX1250-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 nv
; GFX1250-NEXT: s_mov_b32 s6, -1
; GFX1250-NEXT: s_mov_b32 s7, 0x31016000
; GFX1250-NEXT: s_mov_b32 s10, s6
; GFX1250-NEXT: s_mov_b32 s11, s7
; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: s_mov_b64 s[4:5], 0x88888888
; GFX1250-NEXT: s_wait_kmcnt 0x0
; GFX1250-NEXT: s_mov_b32 s8, s2
; GFX1250-NEXT: s_mov_b32 s9, s3
; GFX1250-NEXT: s_mov_b32 s3, 0
; GFX1250-NEXT: buffer_load_b64 v[0:1], off, s[8:11], null
; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: s_mov_b32 s9, s3
; GFX1250-NEXT: s_mov_b32 s13, s3
; GFX1250-NEXT: s_mov_b32 s15, s3
; GFX1250-NEXT: s_mov_b32 s11, s3
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: v_readfirstlane_b32 s2, v1
; GFX1250-NEXT: v_readfirstlane_b32 s10, v0
; GFX1250-NEXT: s_and_b64 s[16:17], s[2:3], 0x11111111
; GFX1250-NEXT: s_and_b32 s8, s10, 0x22222222
; GFX1250-NEXT: s_and_b64 s[18:19], s[2:3], 0x22222222
; GFX1250-NEXT: s_and_b32 s12, s10, 0x11111111
; GFX1250-NEXT: s_and_b32 s14, s2, 0x88888888
; GFX1250-NEXT: s_and_b64 s[20:21], s[10:11], 0x44444444
; GFX1250-NEXT: s_and_b32 s2, s2, 0x44444444
; GFX1250-NEXT: s_and_b64 s[4:5], s[10:11], s[4:5]
; GFX1250-NEXT: s_mul_u64 s[10:11], s[8:9], s[16:17]
; GFX1250-NEXT: s_mul_u64 s[22:23], s[12:13], s[18:19]
; GFX1250-NEXT: s_mul_u64 s[28:29], s[8:9], s[14:15]
; GFX1250-NEXT: s_mul_u64 s[30:31], s[12:13], s[16:17]
; GFX1250-NEXT: s_mul_u64 s[24:25], s[20:21], s[14:15]
; GFX1250-NEXT: s_mul_u64 s[26:27], s[4:5], s[2:3]
; GFX1250-NEXT: s_mul_u64 s[34:35], s[20:21], s[2:3]
; GFX1250-NEXT: s_mul_u64 s[38:39], s[8:9], s[18:19]
; GFX1250-NEXT: s_mul_u64 s[40:41], s[12:13], s[2:3]
; GFX1250-NEXT: s_mul_u64 s[44:45], s[4:5], s[14:15]
; GFX1250-NEXT: s_mul_u64 s[2:3], s[8:9], s[2:3]
; GFX1250-NEXT: s_mul_u64 s[8:9], s[12:13], s[14:15]
; GFX1250-NEXT: s_xor_b64 s[10:11], s[22:23], s[10:11]
; GFX1250-NEXT: s_xor_b64 s[14:15], s[30:31], s[28:29]
; GFX1250-NEXT: s_mul_u64 s[36:37], s[4:5], s[18:19]
; GFX1250-NEXT: s_mul_u64 s[42:43], s[20:21], s[16:17]
; GFX1250-NEXT: s_mul_u64 s[4:5], s[4:5], s[16:17]
; GFX1250-NEXT: s_xor_b64 s[16:17], s[40:41], s[38:39]
; GFX1250-NEXT: s_xor_b64 s[2:3], s[8:9], s[2:3]
; GFX1250-NEXT: s_xor_b64 s[8:9], s[10:11], s[24:25]
; GFX1250-NEXT: s_xor_b64 s[10:11], s[14:15], s[34:35]
; GFX1250-NEXT: s_mul_u64 s[12:13], s[20:21], s[18:19]
; GFX1250-NEXT: s_xor_b64 s[14:15], s[16:17], s[42:43]
; GFX1250-NEXT: s_xor_b64 s[8:9], s[8:9], s[26:27]
; GFX1250-NEXT: s_xor_b64 s[10:11], s[10:11], s[36:37]
; GFX1250-NEXT: s_xor_b64 s[2:3], s[2:3], s[12:13]
; GFX1250-NEXT: s_xor_b64 s[12:13], s[14:15], s[44:45]
; GFX1250-NEXT: s_and_b64 s[8:9], s[8:9], 0x2222222200000000
; GFX1250-NEXT: s_and_b64 s[10:11], s[10:11], 0x1111111100000000
; GFX1250-NEXT: s_and_b64 s[12:13], s[12:13], 0x4444444400000000
; GFX1250-NEXT: s_or_b64 s[8:9], s[10:11], s[8:9]
; GFX1250-NEXT: s_xor_b64 s[2:3], s[2:3], s[4:5]
; GFX1250-NEXT: s_or_b64 s[4:5], s[8:9], s[12:13]
; GFX1250-NEXT: s_and_b64 s[2:3], s[2:3], 0x888888800000000
; GFX1250-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX1250-NEXT: s_or_b64 s[2:3], s[4:5], s[2:3]
; GFX1250-NEXT: s_mov_b32 s4, s0
; GFX1250-NEXT: v_mov_b32_e32 v0, s3
; GFX1250-NEXT: s_mov_b32 s5, s1
; GFX1250-NEXT: buffer_store_b32 v0, off, s[4:7], null
; GFX1250-NEXT: s_endpgm
;
; EG-LABEL: test_clmulh_i32:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 0 @6
; EG-NEXT: ALU 92, @9, KC0[], KC1[]
; EG-NEXT: ALU 53, @102, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.X, T1.X, 1
; EG-NEXT: CF_END
; 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: AND_INT T0.W, T0.X, literal.x,
; EG-NEXT: LSHR * T1.W, T0.X, literal.y,
; EG-NEXT: 65280(9.147676e-41), 8(1.121039e-44)
; EG-NEXT: AND_INT T1.X, PS, literal.x,
; EG-NEXT: LSHR T1.Y, T0.X, literal.y,
; EG-NEXT: LSHL T0.Z, PV.W, literal.z,
; EG-NEXT: AND_INT T0.W, T0.Y, literal.x,
; EG-NEXT: LSHR * T1.W, T0.Y, literal.z,
; EG-NEXT: 65280(9.147676e-41), 24(3.363116e-44)
; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00)
; EG-NEXT: LSHL T0.X, T0.X, literal.x,
; EG-NEXT: AND_INT T2.Y, PS, literal.y,
; EG-NEXT: LSHR T1.Z, T0.Y, literal.x,
; EG-NEXT: LSHL T0.W, PV.W, literal.z,
; EG-NEXT: LSHL * T1.W, T0.Y, literal.x,
; EG-NEXT: 24(3.363116e-44), 65280(9.147676e-41)
; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00)
; EG-NEXT: OR_INT T0.Y, PS, PV.W,
; EG-NEXT: OR_INT T1.Z, PV.Y, PV.Z,
; EG-NEXT: OR_INT T0.W, PV.X, T0.Z,
; EG-NEXT: OR_INT * T1.W, T1.X, T1.Y,
; EG-NEXT: OR_INT T0.W, PV.W, PS,
; EG-NEXT: OR_INT * T1.W, PV.Y, PV.Z,
; EG-NEXT: AND_INT T0.Y, PS, literal.x,
; EG-NEXT: LSHR T0.Z, PS, literal.y,
; EG-NEXT: AND_INT T1.W, PV.W, literal.x,
; EG-NEXT: LSHR * T0.W, PV.W, literal.y,
; EG-NEXT: 252645135(7.053345e-30), 4(5.605194e-45)
; EG-NEXT: AND_INT T1.Y, PS, literal.x,
; EG-NEXT: LSHL T1.Z, PV.W, literal.y,
; EG-NEXT: AND_INT T0.W, PV.Z, literal.x,
; EG-NEXT: LSHL * T1.W, PV.Y, literal.y,
; EG-NEXT: 252645135(7.053345e-30), 4(5.605194e-45)
; EG-NEXT: OR_INT T0.W, PV.W, PS,
; EG-NEXT: OR_INT * T1.W, PV.Y, PV.Z,
; EG-NEXT: AND_INT T0.Y, PS, literal.x,
; EG-NEXT: LSHR T0.Z, PS, literal.y,
; EG-NEXT: AND_INT T1.W, PV.W, literal.x,
; EG-NEXT: LSHR * T0.W, PV.W, literal.y,
; EG-NEXT: 858993459(4.172325e-08), 2(2.802597e-45)
; EG-NEXT: AND_INT T1.Y, PS, literal.x,
; EG-NEXT: LSHL T1.Z, PV.W, literal.y,
; EG-NEXT: AND_INT T0.W, PV.Z, literal.x,
; EG-NEXT: LSHL * T1.W, PV.Y, literal.y,
; EG-NEXT: 858993459(4.172325e-08), 2(2.802597e-45)
; EG-NEXT: OR_INT T0.W, PV.W, PS,
; EG-NEXT: OR_INT * T1.W, PV.Y, PV.Z,
; EG-NEXT: AND_INT T0.Y, PS, literal.x,
; EG-NEXT: LSHR T0.Z, PS, 1,
; EG-NEXT: AND_INT T1.W, PV.W, literal.x,
; EG-NEXT: LSHR * T0.W, PV.W, 1,
; EG-NEXT: 1431655765(1.466015e+13), 0(0.000000e+00)
; EG-NEXT: AND_INT T1.Y, PS, literal.x,
; EG-NEXT: LSHL T1.Z, PV.W, 1,
; EG-NEXT: AND_INT T0.W, PV.Z, literal.x,
; EG-NEXT: LSHL * T1.W, PV.Y, 1,
; EG-NEXT: 1431655765(1.466015e+13), 0(0.000000e+00)
; EG-NEXT: OR_INT T0.W, PV.W, PS,
; EG-NEXT: OR_INT * T1.W, PV.Y, PV.Z,
; EG-NEXT: AND_INT T2.W, PS, literal.x,
; EG-NEXT: AND_INT * T3.W, PV.W, literal.y,
; EG-NEXT: 1145324612(7.850667e+02), 572662306(2.197316e-18)
; EG-NEXT: AND_INT T4.W, T0.W, literal.x,
; EG-NEXT: MULLO_INT * T0.X, PV.W, PS,
; EG-NEXT: 286331153(1.144374e-28), 0(0.000000e+00)
; EG-NEXT: AND_INT T5.W, T1.W, literal.x,
; EG-NEXT: MULLO_INT * T0.Y, T2.W, PV.W,
; EG-NEXT: -2004318072(-8.217301e-34), 0(0.000000e+00)
; EG-NEXT: AND_INT T6.W, T0.W, literal.x,
; EG-NEXT: MULLO_INT * T0.Z, PV.W, T3.W,
; EG-NEXT: 1145324612(7.850667e+02), 0(0.000000e+00)
; EG-NEXT: AND_INT T1.Z, T0.W, literal.x,
; EG-NEXT: AND_INT T0.W, T1.W, literal.y, BS:VEC_120/SCL_212
; EG-NEXT: MULLO_INT * T1.X, T5.W, PV.W,
; EG-NEXT: -2004318072(-8.217301e-34), 286331153(1.144374e-28)
; EG-NEXT: MULLO_INT * T1.Y, PV.W, PV.Z,
; EG-NEXT: MULLO_INT * T2.X, T2.W, T1.Z,
; EG-NEXT: MULLO_INT * T2.Y, T2.W, T6.W,
; EG-NEXT: MULLO_INT * T2.Z, T0.W, T6.W,
; EG-NEXT: AND_INT T1.W, T1.W, literal.x,
; EG-NEXT: MULLO_INT * T2.W, T0.W, T4.W,
; EG-NEXT: 572662306(2.197316e-18), 0(0.000000e+00)
; EG-NEXT: MULLO_INT * T3.X, PV.W, T1.Z,
; EG-NEXT: MULLO_INT * T0.W, T0.W, T3.W,
; EG-NEXT: MULLO_INT * T3.Y, T1.W, T4.W,
; EG-NEXT: XOR_INT T3.Z, T0.W, PS,
; EG-NEXT: XOR_INT * T0.W, T2.W, T3.X, BS:VEC_120/SCL_212
; EG-NEXT: MULLO_INT * T2.W, T1.W, T3.W,
; EG-NEXT: XOR_INT T3.Y, T2.Z, PS,
; EG-NEXT: XOR_INT T2.Z, T0.W, T2.Y,
; EG-NEXT: XOR_INT T0.W, T3.Z, T2.X, BS:VEC_120/SCL_212
; EG-NEXT: MULLO_INT * T1.W, T1.W, T6.W,
; EG-NEXT: XOR_INT * T2.X, T1.Y, PS,
; EG-NEXT: ALU clause starting at 102:
; EG-NEXT: XOR_INT T1.Y, T0.W, T1.X,
; EG-NEXT: XOR_INT T0.Z, T2.Z, T0.Z, BS:VEC_021/SCL_122
; EG-NEXT: XOR_INT T0.W, T3.Y, T0.Y,
; EG-NEXT: MULLO_INT * T0.Y, T5.W, T1.Z,
; EG-NEXT: XOR_INT T1.X, PV.W, PS,
; EG-NEXT: AND_INT T0.Y, PV.Z, literal.x,
; EG-NEXT: AND_INT T0.Z, PV.Y, literal.y,
; EG-NEXT: XOR_INT T0.W, T2.X, T0.X,
; EG-NEXT: MULLO_INT * T0.X, T5.W, T4.W,
; EG-NEXT: 286331153(1.144374e-28), 572662306(2.197316e-18)
; EG-NEXT: XOR_INT T1.Z, PV.W, PS,
; EG-NEXT: OR_INT T0.W, PV.Y, PV.Z,
; EG-NEXT: AND_INT * T1.W, PV.X, literal.x,
; EG-NEXT: 1145324612(7.850667e+02), 0(0.000000e+00)
; EG-NEXT: OR_INT T0.W, PV.W, PS,
; EG-NEXT: AND_INT * T1.W, PV.Z, literal.x,
; EG-NEXT: -2004318072(-8.217301e-34), 0(0.000000e+00)
; EG-NEXT: OR_INT * T0.W, PV.W, PS,
; EG-NEXT: AND_INT T1.W, PV.W, literal.x,
; EG-NEXT: LSHR * T2.W, PV.W, literal.y,
; EG-NEXT: 65280(9.147676e-41), 8(1.121039e-44)
; EG-NEXT: AND_INT T0.Y, PS, literal.x,
; EG-NEXT: LSHR T0.Z, T0.W, literal.y,
; EG-NEXT: LSHL T1.W, PV.W, literal.z,
; EG-NEXT: LSHL * T0.W, T0.W, literal.y,
; EG-NEXT: 65280(9.147676e-41), 24(3.363116e-44)
; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00)
; EG-NEXT: OR_INT T0.W, PS, PV.W,
; EG-NEXT: OR_INT * T1.W, PV.Y, PV.Z,
; EG-NEXT: OR_INT * T0.W, PV.W, PS,
; EG-NEXT: AND_INT T1.W, PV.W, literal.x,
; EG-NEXT: LSHR * T0.W, PV.W, literal.y,
; EG-NEXT: 252645135(7.053345e-30), 4(5.605194e-45)
; EG-NEXT: AND_INT T0.W, PS, literal.x,
; EG-NEXT: LSHL * T1.W, PV.W, literal.y,
; EG-NEXT: 252645135(7.053345e-30), 4(5.605194e-45)
; EG-NEXT: OR_INT * T0.W, PV.W, PS,
; EG-NEXT: AND_INT T1.W, PV.W, literal.x,
; EG-NEXT: LSHR * T0.W, PV.W, literal.y,
; EG-NEXT: 858993459(4.172325e-08), 2(2.802597e-45)
; EG-NEXT: AND_INT T0.W, PS, literal.x,
; EG-NEXT: LSHL * T1.W, PV.W, literal.y,
; EG-NEXT: 858993459(4.172325e-08), 2(2.802597e-45)
; EG-NEXT: OR_INT * T0.W, PV.W, PS,
; EG-NEXT: AND_INT T1.W, PV.W, literal.x,
; EG-NEXT: LSHR * T0.W, PV.W, 1,
; EG-NEXT: 1431655765(1.466015e+13), 0(0.000000e+00)
; EG-NEXT: AND_INT T0.W, PS, literal.x,
; EG-NEXT: LSHL * T1.W, PV.W, 1,
; EG-NEXT: 1431655764(1.466015e+13), 0(0.000000e+00)
; EG-NEXT: OR_INT * T0.W, PV.W, PS,
; EG-NEXT: LSHR T0.X, PV.W, 1,
; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.x,
; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00)
%b_ptr = getelementptr i32, ptr addrspace(1) %in, i32 1
%a = load i32, ptr addrspace(1) %in
%b = load i32, ptr addrspace(1) %b_ptr
%a.ext = zext i32 %a to i64
%b.ext = zext i32 %b to i64
%clmul = call i64 @llvm.clmul.i64(i64 %a.ext, i64 %b.ext)
%res.ext = lshr i64 %clmul, 32
%res = trunc i64 %res.ext to i32
store i32 %res, ptr addrspace(1) %out
ret void
}
declare i32 @llvm.amdgcn.workitem.id.x() #1
attributes #0 = { nounwind }
attributes #1 = { nounwind readnone}