[AMDGPU] Fix some tests that did not specify -mcpu

[tpr]

Summary

This fixes some tests that did not specify -mcpu. Doing that disables
all subtarget features, which gives behavior that (a) does not
necessarily correspond to any actual target, and (b) can change as we
add new subtarget features.

Change-Id: I8c0fe4fa03e9a93ef8bb722cd42d22e064526309

[tpr]

V2: Also added gfx1010 to memtime test.

[rampitec]

LGTM

[tpr]

This got landed in a slightly different form because of conflicts with other sdiv64 etc tests.

llvm/test/CodeGen/AMDGPU/clamp-modifier.ll

-; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,SI %s
+; RUN: llc -march=amdgcn -mcpu=tahiti -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,SI %s
 ; RUN: llc -march=amdgcn -mcpu=fiji -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GFX89 %s
 ; RUN: llc -march=amdgcn -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GFX89,GFX9 %s
 
 ; GCN-LABEL: {{^}}v_clamp_add_src_f32:
 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]]
 ; GCN-NOT: [[A]]
 ; GCN: v_add_f32_e64 v{{[0-9]+}}, [[A]], 1.0 clamp{{$}}
 define amdgpu_kernel void @v_clamp_add_src_f32(float addrspace(1)* %out, float addrspace(1)* %aptr) #0 {

llvm/test/CodeGen/AMDGPU/divrem24-assume.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
-; RUN: opt -S -mtriple=amdgcn-- -amdgpu-codegenprepare %s | FileCheck %s
+; RUN: opt -S -mtriple=amdgcn-- -mcpu=tonga -amdgpu-codegenprepare %s | FileCheck %s
 
 define amdgpu_kernel void @divrem24_assume(i32 addrspace(1)* %arg, i32 %arg1) {
 ; CHECK-LABEL: @divrem24_assume(
 ; CHECK-NEXT:  bb:
 ; CHECK-NEXT:    [[TMP:%.*]] = tail call i32 @llvm.amdgcn.workitem.id.x(), !range !0
 ; CHECK-NEXT:    [[TMP2:%.*]] = icmp ult i32 [[ARG1:%.*]], 42
 ; CHECK-NEXT:    tail call void @llvm.assume(i1 [[TMP2]])
 ; CHECK-NEXT:    [[TMP0:%.*]] = uitofp i32 [[TMP]] to float

llvm/test/CodeGen/AMDGPU/fadd-fma-fmul-combine.ll

-; RUN: llc -march=amdgcn -mattr=+fast-fmaf,-fp32-denormals -enable-unsafe-fp-math -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=GCN-FLUSH %s
-; RUN: llc -march=amdgcn -mattr=-fast-fmaf,-fp32-denormals -enable-unsafe-fp-math -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=GCN-FLUSH %s
+; RUN: llc -march=amdgcn -mcpu=tahiti -mattr=+fast-fmaf,-fp32-denormals -enable-unsafe-fp-math -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=GCN-FLUSH %s
+; RUN: llc -march=amdgcn -mcpu=tahiti -mattr=-fast-fmaf,-fp32-denormals -enable-unsafe-fp-math -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=GCN-FLUSH %s
 
-; RUN: llc -march=amdgcn -mattr=+fast-fmaf,+fp32-denormals -enable-unsafe-fp-math -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=GCN-FASTFMA %s
-; RUN: llc -march=amdgcn -mattr=-fast-fmaf,+fp32-denormals -enable-unsafe-fp-math -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=GCN-SLOWFMA %s
+; RUN: llc -march=amdgcn -mcpu=tahiti -mattr=+fast-fmaf,+fp32-denormals -enable-unsafe-fp-math -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=GCN-FASTFMA %s
+; RUN: llc -march=amdgcn -mcpu=tahiti -mattr=-fast-fmaf,+fp32-denormals -enable-unsafe-fp-math -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=GCN-SLOWFMA %s
 
 ; FIXME: This should also fold when fma is actually fast if an FMA
 ; exists in the original program.
 
 ; (fadd (fma x, y, (fmul u, v), z) -> (fma x, y (fma u, v, z))
 
 ; GCN-LABEL: {{^}}fast_add_fmuladd_fmul:
 ; GCN: buffer_load_dword [[X:v[0-9]+]]

llvm/test/CodeGen/AMDGPU/frem.ll

-; RUN:  llc -amdgpu-scalarize-global-loads=false  -march=amdgcn -verify-machineinstrs  < %s | FileCheck -check-prefix=SI -check-prefix=GCN -check-prefix=FUNC %s
+; RUN:  llc -amdgpu-scalarize-global-loads=false  -march=amdgcn -mcpu=tahiti -verify-machineinstrs  < %s | FileCheck -check-prefix=SI -check-prefix=GCN -check-prefix=FUNC %s
 ; RUN:  llc -amdgpu-scalarize-global-loads=false  -march=amdgcn -mcpu=bonaire -verify-machineinstrs < %s | FileCheck -check-prefix=CI -check-prefix=GCN -check-prefix=FUNC %s
 ; RUN:  llc -amdgpu-scalarize-global-loads=false  -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefix=CI -check-prefix=GCN -check-prefix=FUNC %s
 
 ; FUNC-LABEL: {{^}}frem_f32:
 ; GCN-DAG: buffer_load_dword [[X:v[0-9]+]], {{.*$}}
 ; GCN-DAG: buffer_load_dword [[Y:v[0-9]+]], {{.*}} offset:16
 ; GCN: v_div_scale_f32
 

llvm/test/CodeGen/AMDGPU/llvm.amdgcn.s.memtime.ll

-; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=GCN %s
-; RUN: llc -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefix=VI -check-prefix=GCN %s
+; RUN: llc -march=amdgcn -mcpu=tahiti -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=SIVI -check-prefix=GCN %s
+; RUN: llc -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefix=VI -check-prefix=SIVI -check-prefix=GCN %s
+; RUN: llc -march=amdgcn -mcpu=gfx1010 -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
 
 declare i64 @llvm.amdgcn.s.memtime() #0
 
 ; GCN-LABEL: {{^}}test_s_memtime:
 ; GCN-DAG: s_memtime s{{\[[0-9]+:[0-9]+\]}}
 ; GCN-DAG: s_load_dwordx2
 ; GCN: lgkmcnt
-; GCN: buffer_store_dwordx2
-; GCN-NOT: lgkmcnt
+; GCN: {{buffer|global}}_store_dwordx2
+; SIVI-NOT: lgkmcnt
 ; GCN: s_memtime s{{\[[0-9]+:[0-9]+\]}}
-; GCN: buffer_store_dwordx2
+; GCN: {{buffer|global}}_store_dwordx2
 define amdgpu_kernel void @test_s_memtime(i64 addrspace(1)* %out) #0 {
   %cycle0 = call i64 @llvm.amdgcn.s.memtime()
   store volatile i64 %cycle0, i64 addrspace(1)* %out
 
   %cycle1 = call i64 @llvm.amdgcn.s.memtime()
   store volatile i64 %cycle1, i64 addrspace(1)* %out
   ret void
 }
 
 attributes #0 = { nounwind }

llvm/test/CodeGen/AMDGPU/madmk.ll

-; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=SI %s
+; RUN: llc -march=amdgcn -mcpu=tahiti -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=SI %s
 ; XUN: llc -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VI %s
 
  ; FIXME: None of these trigger madmk emission anymore. It is still
  ; possible, but requires the correct registers to be used which is
  ; hard to trigger.
 
 declare i32 @llvm.amdgcn.workitem.id.x() nounwind readnone
 declare float @llvm.fabs.f32(float) nounwind readnone
define amdgpu_kernel void @s_s_madmk_f32(float addrspace(1)* noalias %out, [8 x i32], float %a, [8 x i32], float %b) nounwind {
 
   %mul = fmul float %a, 10.0
   %madmk = fadd float %mul, %b
   store float %madmk, float addrspace(1)* %out.gep, align 4
   ret void
 }
 
 ; GCN-LABEL: {{^}}v_s_madmk_f32:
-; GCN: s_load_dword [[SREG:s[0-9]+]]
-; GCN: buffer_load_dword [[VREG1:v[0-9]+]]
+; GCN-DAG: s_load_dword [[SREG:s[0-9]+]]
+; GCN-DAG: buffer_load_dword [[VREG1:v[0-9]+]]
 ; GCN: v_mov_b32_e32 [[VREG2:v[0-9]+]], [[SREG]]
 ; GCN: v_mac_f32_e32 [[VREG2]], 0x41200000, [[VREG1]]
 ; GCN: s_endpgm
 define amdgpu_kernel void @v_s_madmk_f32(float addrspace(1)* noalias %out, float addrspace(1)* noalias %in, float %b) nounwind {
   %tid = tail call i32 @llvm.amdgcn.workitem.id.x() nounwind readnone
   %gep.0 = getelementptr float, float addrspace(1)* %in, i32 %tid
   %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
   %a = load float, float addrspace(1)* %gep.0, align 4

llvm/test/CodeGen/AMDGPU/omod.ll

-; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=SI %s
+; RUN: llc -march=amdgcn -mcpu=tahiti -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=SI %s
 ; RUN: llc -march=amdgcn -mcpu=fiji -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VI %s
 
 ; IEEE bit enabled for compute kernel, no shouldn't use.
 ; GCN-LABEL: {{^}}v_omod_div2_f32_enable_ieee_signed_zeros:
 ; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
 ; GCN: v_add_f32_e32 [[ADD:v[0-9]+]], 1.0, [[A]]{{$}}
 ; GCN: v_mul_f32_e32 v{{[0-9]+}}, 0.5, [[ADD]]{{$}}
 define amdgpu_kernel void @v_omod_div2_f32_enable_ieee_signed_zeros(float addrspace(1)* %out, float addrspace(1)* %aptr) #4 {

llvm/test/CodeGen/AMDGPU/operand-folding.ll

-; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck %s
+; RUN: llc -march=amdgcn -mcpu=tahiti -verify-machineinstrs < %s | FileCheck %s
 
 ; CHECK-LABEL: {{^}}fold_sgpr:
 ; CHECK: v_add_i32_e32 v{{[0-9]+}}, vcc, s
 define amdgpu_kernel void @fold_sgpr(i32 addrspace(1)* %out, i32 %fold) {
 entry:
   %tmp0 = icmp ne i32 %fold, 0
   br i1 %tmp0, label %if, label %endif
 

llvm/test/CodeGen/AMDGPU/sdiv.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc < %s -amdgpu-scalarize-global-loads=false -march=amdgcn | FileCheck %s -check-prefixes=FUNC,SI,GCN
+; RUN: llc < %s -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=gfx600 | FileCheck %s -check-prefixes=FUNC,SI,GCN
 ; RUN: llc < %s -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=tonga -mattr=-flat-for-global | FileCheck %s -check-prefixes=FUNC,SI,TONGA
 ; RUN: llc < %s -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=gfx900 -mattr=-flat-for-global | FileCheck %s -check-prefixes=FUNC,SI,GFX9
 ; RUN: llc < %s -amdgpu-scalarize-global-loads=false -march=r600 -mcpu=redwood | FileCheck %s -check-prefixes=FUNC,EG
 
 ; The code generated by sdiv is long and complex and may frequently change.
 ; The goal of this test is to make sure the ISel doesn't fail.
 ;
 ; This program was previously failing to compile when one of the selectcc
 ; opcodes generated by the sdiv lowering was being legalized and optimized to:
 ; selectcc Remainder -1, 0, -1, SETGT
 ; This was fixed by adding an additional pattern in R600Instructions.td to
 ; match this pattern with a CNDGE_INT.
 
 define amdgpu_kernel void @sdiv_i32(i32 addrspace(1)* %out, i32 addrspace(1)* %in) {
 ; GCN-LABEL: sdiv_i32:
 ; GCN:       ; %bb.0:
-; GCN-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; GCN-NEXT:    s_load_dwordx4 s[8:11], s[0:1], 0x9
 ; GCN-NEXT:    s_mov_b32 s7, 0xf000
 ; GCN-NEXT:    s_mov_b32 s6, -1
-; GCN-NEXT:    s_mov_b32 s10, s6
-; GCN-NEXT:    s_mov_b32 s11, s7
+; GCN-NEXT:    s_mov_b32 s2, s6
+; GCN-NEXT:    s_mov_b32 s3, s7
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_mov_b32 s8, s2
-; GCN-NEXT:    s_mov_b32 s9, s3
-; GCN-NEXT:    buffer_load_dwordx2 v[0:1], off, s[8:11], 0
-; GCN-NEXT:    s_mov_b32 s4, s0
-; GCN-NEXT:    s_mov_b32 s5, s1
+; GCN-NEXT:    s_mov_b32 s0, s10
+; GCN-NEXT:    s_mov_b32 s1, s11
+; GCN-NEXT:    buffer_load_dwordx2 v[0:1], off, s[0:3], 0
+; GCN-NEXT:    s_mov_b32 s4, s8
+; GCN-NEXT:    s_mov_b32 s5, s9
 ; GCN-NEXT:    s_waitcnt vmcnt(0)
-; GCN-NEXT:    v_ashrrev_i32_e32 v2, 31, v0
-; GCN-NEXT:    v_ashrrev_i32_e32 v3, 31, v1
-; GCN-NEXT:    v_xor_b32_e32 v4, v2, v3
-; GCN-NEXT:    v_add_i32_e32 v0, vcc, v2, v0
-; GCN-NEXT:    v_add_i32_e32 v1, vcc, v3, v1
-; GCN-NEXT:    v_xor_b32_e32 v0, v0, v2
-; GCN-NEXT:    v_xor_b32_e32 v1, v1, v3
-; GCN-NEXT:    v_cvt_f32_u32_e32 v2, v1
-; GCN-NEXT:    v_rcp_iflag_f32_e32 v2, v2
-; GCN-NEXT:    v_mul_f32_e32 v2, 0x4f800000, v2
-; GCN-NEXT:    v_cvt_u32_f32_e32 v2, v2
-; GCN-NEXT:    v_mul_hi_u32 v3, v2, v1
-; GCN-NEXT:    v_mul_lo_u32 v5, v2, v1
-; GCN-NEXT:    v_sub_i32_e32 v6, vcc, 0, v5
-; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v3
-; GCN-NEXT:    v_cndmask_b32_e64 v3, v5, v6, s[0:1]
-; GCN-NEXT:    v_mul_hi_u32 v3, v3, v2
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, v3, v2
-; GCN-NEXT:    v_subrev_i32_e32 v2, vcc, v3, v2
-; GCN-NEXT:    v_cndmask_b32_e64 v2, v2, v5, s[0:1]
-; GCN-NEXT:    v_mul_hi_u32 v2, v2, v0
-; GCN-NEXT:    v_mul_lo_u32 v3, v2, v1
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, 1, v2
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, -1, v2
-; GCN-NEXT:    v_subrev_i32_e32 v7, vcc, v3, v0
-; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v0, v3
+; GCN-NEXT:    v_ashrrev_i32_e32 v2, 31, v1
+; GCN-NEXT:    v_add_i32_e32 v1, vcc, v2, v1
+; GCN-NEXT:    v_xor_b32_e32 v1, v1, v2
+; GCN-NEXT:    v_cvt_f32_u32_e32 v3, v1
+; GCN-NEXT:    v_ashrrev_i32_e32 v6, 31, v0
+; GCN-NEXT:    v_add_i32_e32 v0, vcc, v6, v0
+; GCN-NEXT:    v_xor_b32_e32 v0, v0, v6
+; GCN-NEXT:    v_rcp_iflag_f32_e32 v3, v3
+; GCN-NEXT:    v_xor_b32_e32 v2, v6, v2
+; GCN-NEXT:    v_mul_f32_e32 v3, 0x4f800000, v3
+; GCN-NEXT:    v_cvt_u32_f32_e32 v3, v3
+; GCN-NEXT:    v_mul_lo_u32 v4, v3, v1
+; GCN-NEXT:    v_mul_hi_u32 v5, v3, v1
+; GCN-NEXT:    v_sub_i32_e32 v7, vcc, 0, v4
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v5
+; GCN-NEXT:    v_cndmask_b32_e64 v4, v4, v7, s[0:1]
+; GCN-NEXT:    v_mul_hi_u32 v4, v4, v3
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v4, v3
+; GCN-NEXT:    v_subrev_i32_e32 v3, vcc, v4, v3
+; GCN-NEXT:    v_cndmask_b32_e64 v3, v3, v5, s[0:1]
+; GCN-NEXT:    v_mul_hi_u32 v3, v3, v0
+; GCN-NEXT:    v_mul_lo_u32 v4, v3, v1
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, 1, v3
+; GCN-NEXT:    v_add_i32_e32 v6, vcc, -1, v3
+; GCN-NEXT:    v_subrev_i32_e32 v7, vcc, v4, v0
+; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v0, v4
 ; GCN-NEXT:    v_cmp_ge_u32_e64 s[0:1], v7, v1
 ; GCN-NEXT:    s_and_b64 s[0:1], s[0:1], vcc
-; GCN-NEXT:    v_cndmask_b32_e64 v0, v2, v5, s[0:1]
+; GCN-NEXT:    v_cndmask_b32_e64 v0, v3, v5, s[0:1]
 ; GCN-NEXT:    v_cndmask_b32_e32 v0, v6, v0, vcc
-; GCN-NEXT:    v_xor_b32_e32 v0, v0, v4
-; GCN-NEXT:    v_sub_i32_e32 v0, vcc, v0, v4
+; GCN-NEXT:    v_xor_b32_e32 v0, v0, v2
+; GCN-NEXT:    v_sub_i32_e32 v0, vcc, v0, v2
 ; GCN-NEXT:    buffer_store_dword v0, off, s[4:7], 0
 ; GCN-NEXT:    s_endpgm
 ;
 ; TONGA-LABEL: sdiv_i32:
 ; TONGA:       ; %bb.0:
 ; TONGA-NEXT:    s_load_dwordx4 s[8:11], s[0:1], 0x24
 ; TONGA-NEXT:    s_mov_b32 s7, 0xf000
 ; TONGA-NEXT:    s_mov_b32 s6, -1
; EG-NEXT:    2(2.802597e-45), 0(0.000000e+00)
   %result = sdiv i32 %num, %den
   store i32 %result, i32 addrspace(1)* %out
   ret void
 }
 
 define amdgpu_kernel void @sdiv_i32_4(i32 addrspace(1)* %out, i32 addrspace(1)* %in) {
 ; GCN-LABEL: sdiv_i32_4:
 ; GCN:       ; %bb.0:
-; GCN-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
-; GCN-NEXT:    s_mov_b32 s7, 0xf000
-; GCN-NEXT:    s_mov_b32 s6, -1
-; GCN-NEXT:    s_mov_b32 s10, s6
-; GCN-NEXT:    s_mov_b32 s11, s7
+; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN-NEXT:    s_mov_b32 s3, 0xf000
+; GCN-NEXT:    s_mov_b32 s2, -1
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_mov_b32 s8, s2
-; GCN-NEXT:    s_mov_b32 s9, s3
-; GCN-NEXT:    buffer_load_dword v0, off, s[8:11], 0
-; GCN-NEXT:    s_mov_b32 s4, s0
-; GCN-NEXT:    s_mov_b32 s5, s1
+; GCN-NEXT:    s_mov_b32 s0, s4
+; GCN-NEXT:    s_mov_b32 s1, s5
+; GCN-NEXT:    s_mov_b32 s4, s6
+; GCN-NEXT:    s_mov_b32 s5, s7
+; GCN-NEXT:    s_mov_b32 s6, s2
+; GCN-NEXT:    s_mov_b32 s7, s3
+; GCN-NEXT:    buffer_load_dword v0, off, s[4:7], 0
 ; GCN-NEXT:    s_waitcnt vmcnt(0)
 ; GCN-NEXT:    v_ashrrev_i32_e32 v1, 31, v0
 ; GCN-NEXT:    v_lshrrev_b32_e32 v1, 30, v1
 ; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v1
 ; GCN-NEXT:    v_ashrrev_i32_e32 v0, 2, v0
-; GCN-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; GCN-NEXT:    s_endpgm
 ;
 ; TONGA-LABEL: sdiv_i32_4:
 ; TONGA:       ; %bb.0:
 ; TONGA-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; TONGA-NEXT:    s_mov_b32 s3, 0xf000
 ; TONGA-NEXT:    s_mov_b32 s2, -1
 ; TONGA-NEXT:    s_waitcnt lgkmcnt(0)

 }
 
 ; Multiply by a weird constant to make sure setIntDivIsCheap is
 ; working.
 
 define amdgpu_kernel void @slow_sdiv_i32_3435(i32 addrspace(1)* %out, i32 addrspace(1)* %in) {
 ; GCN-LABEL: slow_sdiv_i32_3435:
 ; GCN:       ; %bb.0:
-; GCN-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
-; GCN-NEXT:    s_mov_b32 s7, 0xf000
-; GCN-NEXT:    s_mov_b32 s6, -1
-; GCN-NEXT:    s_mov_b32 s10, s6
-; GCN-NEXT:    s_mov_b32 s11, s7
+; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN-NEXT:    s_mov_b32 s3, 0xf000
+; GCN-NEXT:    s_mov_b32 s2, -1
+; GCN-NEXT:    s_mov_b32 s10, s2
+; GCN-NEXT:    s_mov_b32 s11, s3
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_mov_b32 s8, s2
-; GCN-NEXT:    s_mov_b32 s9, s3
+; GCN-NEXT:    s_mov_b32 s8, s6
+; GCN-NEXT:    s_mov_b32 s9, s7
 ; GCN-NEXT:    buffer_load_dword v0, off, s[8:11], 0
-; GCN-NEXT:    s_mov_b32 s2, 0x98a1930b
-; GCN-NEXT:    s_mov_b32 s4, s0
-; GCN-NEXT:    s_mov_b32 s5, s1
+; GCN-NEXT:    s_mov_b32 s0, 0x98a1930b
+; GCN-NEXT:    s_mov_b32 s1, s5
 ; GCN-NEXT:    s_waitcnt vmcnt(0)
-; GCN-NEXT:    v_mul_hi_i32 v1, v0, s2
+; GCN-NEXT:    v_mul_hi_i32 v1, v0, s0
+; GCN-NEXT:    s_mov_b32 s0, s4
 ; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v1
 ; GCN-NEXT:    v_lshrrev_b32_e32 v1, 31, v0
 ; GCN-NEXT:    v_ashrrev_i32_e32 v0, 11, v0
 ; GCN-NEXT:    v_add_i32_e32 v0, vcc, v1, v0
-; GCN-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; GCN-NEXT:    s_endpgm
 ;
 ; TONGA-LABEL: slow_sdiv_i32_3435:
 ; TONGA:       ; %bb.0:
 ; TONGA-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; TONGA-NEXT:    s_mov_b32 s3, 0xf000
 ; TONGA-NEXT:    s_mov_b32 s2, -1
 ; TONGA-NEXT:    s_mov_b32 s10, s2

 }
 
 define amdgpu_kernel void @sdiv_v2i32(<2 x i32> addrspace(1)* %out, <2 x i32> addrspace(1)* %in) {
 ; GCN-LABEL: sdiv_v2i32:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
 ; GCN-NEXT:    s_mov_b32 s11, 0xf000
 ; GCN-NEXT:    s_mov_b32 s10, -1
-; GCN-NEXT:    s_mov_b32 s6, s10
-; GCN-NEXT:    s_mov_b32 s7, s11
+; GCN-NEXT:    s_mov_b32 s4, 0x4f800000
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_mov_b32 s4, s2
-; GCN-NEXT:    s_mov_b32 s5, s3
-; GCN-NEXT:    buffer_load_dwordx4 v[0:3], off, s[4:7], 0
-; GCN-NEXT:    s_mov_b32 s2, 0x4f800000
 ; GCN-NEXT:    s_mov_b32 s8, s0
 ; GCN-NEXT:    s_mov_b32 s9, s1
+; GCN-NEXT:    s_mov_b32 s0, s2
+; GCN-NEXT:    s_mov_b32 s1, s3
+; GCN-NEXT:    s_mov_b32 s2, s10
+; GCN-NEXT:    s_mov_b32 s3, s11
+; GCN-NEXT:    buffer_load_dwordx4 v[0:3], off, s[0:3], 0
 ; GCN-NEXT:    s_waitcnt vmcnt(0)
-; GCN-NEXT:    v_ashrrev_i32_e32 v4, 31, v0
 ; GCN-NEXT:    v_ashrrev_i32_e32 v5, 31, v2
-; GCN-NEXT:    v_ashrrev_i32_e32 v6, 31, v1
 ; GCN-NEXT:    v_ashrrev_i32_e32 v7, 31, v3
-; GCN-NEXT:    v_xor_b32_e32 v8, v4, v5
-; GCN-NEXT:    v_add_i32_e32 v0, vcc, v4, v0
 ; GCN-NEXT:    v_add_i32_e32 v2, vcc, v5, v2
-; GCN-NEXT:    v_xor_b32_e32 v9, v6, v7
-; GCN-NEXT:    v_add_i32_e32 v1, vcc, v6, v1
+; GCN-NEXT:    v_ashrrev_i32_e32 v4, 31, v0
 ; GCN-NEXT:    v_add_i32_e32 v3, vcc, v7, v3
-; GCN-NEXT:    v_xor_b32_e32 v0, v0, v4
 ; GCN-NEXT:    v_xor_b32_e32 v2, v2, v5
-; GCN-NEXT:    v_xor_b32_e32 v1, v1, v6
+; GCN-NEXT:    v_ashrrev_i32_e32 v6, 31, v1
+; GCN-NEXT:    v_xor_b32_e32 v8, v4, v5
+; GCN-NEXT:    v_cvt_f32_u32_e32 v5, v2
 ; GCN-NEXT:    v_xor_b32_e32 v3, v3, v7
-; GCN-NEXT:    v_cvt_f32_u32_e32 v4, v2
-; GCN-NEXT:    v_cvt_f32_u32_e32 v5, v3
-; GCN-NEXT:    v_rcp_iflag_f32_e32 v4, v4
+; GCN-NEXT:    v_xor_b32_e32 v9, v6, v7
+; GCN-NEXT:    v_cvt_f32_u32_e32 v7, v3
 ; GCN-NEXT:    v_rcp_iflag_f32_e32 v5, v5
-; GCN-NEXT:    v_mul_f32_e32 v4, s2, v4
-; GCN-NEXT:    v_mul_f32_e32 v5, s2, v5
+; GCN-NEXT:    v_add_i32_e32 v0, vcc, v4, v0
+; GCN-NEXT:    v_xor_b32_e32 v0, v0, v4
+; GCN-NEXT:    v_rcp_iflag_f32_e32 v7, v7
+; GCN-NEXT:    v_mul_f32_e32 v4, s4, v5
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v4, v4
+; GCN-NEXT:    v_add_i32_e32 v1, vcc, v6, v1
+; GCN-NEXT:    v_mul_f32_e32 v5, s4, v7
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v5, v5
+; GCN-NEXT:    v_xor_b32_e32 v1, v1, v6
 ; GCN-NEXT:    v_mul_hi_u32 v6, v4, v2
 ; GCN-NEXT:    v_mul_lo_u32 v7, v4, v2
 ; GCN-NEXT:    v_mul_hi_u32 v10, v5, v3
 ; GCN-NEXT:    v_mul_lo_u32 v11, v5, v3
-; GCN-NEXT:    v_sub_i32_e32 v12, vcc, 0, v7
-; GCN-NEXT:    v_sub_i32_e32 v13, vcc, 0, v11
 ; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v6
+; GCN-NEXT:    v_sub_i32_e32 v12, vcc, 0, v7
 ; GCN-NEXT:    v_cndmask_b32_e64 v6, v7, v12, s[0:1]
+; GCN-NEXT:    v_sub_i32_e32 v13, vcc, 0, v11
 ; GCN-NEXT:    v_cmp_eq_u32_e64 s[2:3], 0, v10
 ; GCN-NEXT:    v_cndmask_b32_e64 v7, v11, v13, s[2:3]
 ; GCN-NEXT:    v_mul_hi_u32 v6, v6, v4
 ; GCN-NEXT:    v_mul_hi_u32 v7, v7, v5
 ; GCN-NEXT:    v_add_i32_e32 v10, vcc, v6, v4
 ; GCN-NEXT:    v_subrev_i32_e32 v4, vcc, v6, v4
+; GCN-NEXT:    v_cndmask_b32_e64 v4, v4, v10, s[0:1]
 ; GCN-NEXT:    v_add_i32_e32 v6, vcc, v7, v5
 ; GCN-NEXT:    v_subrev_i32_e32 v5, vcc, v7, v5
-; GCN-NEXT:    v_cndmask_b32_e64 v4, v4, v10, s[0:1]
-; GCN-NEXT:    v_cndmask_b32_e64 v5, v5, v6, s[2:3]
 ; GCN-NEXT:    v_mul_hi_u32 v4, v4, v0
+; GCN-NEXT:    v_cndmask_b32_e64 v5, v5, v6, s[2:3]
 ; GCN-NEXT:    v_mul_hi_u32 v5, v5, v1
 ; GCN-NEXT:    v_mul_lo_u32 v6, v4, v2
 ; GCN-NEXT:    v_add_i32_e32 v7, vcc, 1, v4
-; GCN-NEXT:    v_add_i32_e32 v10, vcc, -1, v4
 ; GCN-NEXT:    v_mul_lo_u32 v11, v5, v3
-; GCN-NEXT:    v_add_i32_e32 v12, vcc, 1, v5
-; GCN-NEXT:    v_add_i32_e32 v13, vcc, -1, v5
+; GCN-NEXT:    v_add_i32_e32 v10, vcc, -1, v4
 ; GCN-NEXT:    v_subrev_i32_e32 v14, vcc, v6, v0
 ; GCN-NEXT:    v_cmp_ge_u32_e64 s[0:1], v0, v6
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[2:3], v14, v2
 ; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, v11, v1
+; GCN-NEXT:    v_add_i32_e32 v12, vcc, 1, v5
+; GCN-NEXT:    v_add_i32_e32 v13, vcc, -1, v5
 ; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v1, v11
-; GCN-NEXT:    v_cmp_ge_u32_e64 s[2:3], v14, v2
 ; GCN-NEXT:    v_cmp_ge_u32_e64 s[4:5], v0, v3
 ; GCN-NEXT:    s_and_b64 s[2:3], s[2:3], s[0:1]
 ; GCN-NEXT:    v_cndmask_b32_e64 v0, v4, v7, s[2:3]
 ; GCN-NEXT:    s_and_b64 s[2:3], s[4:5], vcc
 ; GCN-NEXT:    v_cndmask_b32_e64 v1, v5, v12, s[2:3]
 ; GCN-NEXT:    v_cndmask_b32_e64 v0, v10, v0, s[0:1]
 ; GCN-NEXT:    v_cndmask_b32_e32 v1, v13, v1, vcc
 ; GCN-NEXT:    v_xor_b32_e32 v0, v0, v8
; EG-NEXT:    2(2.802597e-45), 0(0.000000e+00)
   %result = sdiv <2 x i32> %num, %den
   store <2 x i32> %result, <2 x i32> addrspace(1)* %out
   ret void
 }
 
 define amdgpu_kernel void @sdiv_v2i32_4(<2 x i32> addrspace(1)* %out, <2 x i32> addrspace(1)* %in) {
 ; GCN-LABEL: sdiv_v2i32_4:
 ; GCN:       ; %bb.0:
-; GCN-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
-; GCN-NEXT:    s_mov_b32 s7, 0xf000
-; GCN-NEXT:    s_mov_b32 s6, -1
-; GCN-NEXT:    s_mov_b32 s10, s6
-; GCN-NEXT:    s_mov_b32 s11, s7
+; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN-NEXT:    s_mov_b32 s3, 0xf000
+; GCN-NEXT:    s_mov_b32 s2, -1
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_mov_b32 s8, s2
-; GCN-NEXT:    s_mov_b32 s9, s3
-; GCN-NEXT:    buffer_load_dwordx2 v[0:1], off, s[8:11], 0
-; GCN-NEXT:    s_mov_b32 s4, s0
-; GCN-NEXT:    s_mov_b32 s5, s1
+; GCN-NEXT:    s_mov_b32 s0, s4
+; GCN-NEXT:    s_mov_b32 s1, s5
+; GCN-NEXT:    s_mov_b32 s4, s6
+; GCN-NEXT:    s_mov_b32 s5, s7
+; GCN-NEXT:    s_mov_b32 s6, s2
+; GCN-NEXT:    s_mov_b32 s7, s3
+; GCN-NEXT:    buffer_load_dwordx2 v[0:1], off, s[4:7], 0
 ; GCN-NEXT:    s_waitcnt vmcnt(0)
 ; GCN-NEXT:    v_ashrrev_i32_e32 v2, 31, v0
 ; GCN-NEXT:    v_ashrrev_i32_e32 v3, 31, v1
 ; GCN-NEXT:    v_lshrrev_b32_e32 v2, 30, v2
 ; GCN-NEXT:    v_lshrrev_b32_e32 v3, 30, v3
 ; GCN-NEXT:    v_add_i32_e32 v0, vcc, v2, v0
 ; GCN-NEXT:    v_add_i32_e32 v1, vcc, v3, v1
 ; GCN-NEXT:    v_ashrrev_i32_e32 v0, 2, v0
 ; GCN-NEXT:    v_ashrrev_i32_e32 v1, 2, v1
-; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
+; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 ; GCN-NEXT:    s_endpgm
 ;
 ; TONGA-LABEL: sdiv_v2i32_4:
 ; TONGA:       ; %bb.0:
 ; TONGA-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; TONGA-NEXT:    s_mov_b32 s3, 0xf000
 ; TONGA-NEXT:    s_mov_b32 s2, -1
 ; TONGA-NEXT:    s_waitcnt lgkmcnt(0)
; EG-NEXT:    2(2.802597e-45), 0(0.000000e+00)
   %result = sdiv <2 x i32> %num, <i32 4, i32 4>
   store <2 x i32> %result, <2 x i32> addrspace(1)* %out
   ret void
 }
 
 define amdgpu_kernel void @sdiv_v4i32(<4 x i32> addrspace(1)* %out, <4 x i32> addrspace(1)* %in) {
 ; GCN-LABEL: sdiv_v4i32:
 ; GCN:       ; %bb.0:
-; GCN-NEXT:    s_load_dwordx4 s[8:11], s[0:1], 0x9
-; GCN-NEXT:    s_mov_b32 s19, 0xf000
-; GCN-NEXT:    s_mov_b32 s18, -1
-; GCN-NEXT:    s_mov_b32 s2, s18
-; GCN-NEXT:    s_mov_b32 s3, s19
+; GCN-NEXT:    s_load_dwordx4 s[12:15], s[0:1], 0x9
+; GCN-NEXT:    s_mov_b32 s11, 0xf000
+; GCN-NEXT:    s_mov_b32 s10, -1
+; GCN-NEXT:    s_mov_b32 s2, s10
+; GCN-NEXT:    s_mov_b32 s3, s11
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_mov_b32 s0, s10
-; GCN-NEXT:    s_mov_b32 s1, s11
+; GCN-NEXT:    s_mov_b32 s0, s14
+; GCN-NEXT:    s_mov_b32 s1, s15
 ; GCN-NEXT:    buffer_load_dwordx4 v[0:3], off, s[0:3], 0
 ; GCN-NEXT:    buffer_load_dwordx4 v[4:7], off, s[0:3], 0 offset:16
-; GCN-NEXT:    s_mov_b32 s6, 0x4f800000
+; GCN-NEXT:    s_mov_b32 s14, 0x4f800000
+; GCN-NEXT:    s_mov_b32 s8, s12
+; GCN-NEXT:    s_mov_b32 s9, s13
 ; GCN-NEXT:    s_waitcnt vmcnt(1)
 ; GCN-NEXT:    v_ashrrev_i32_e32 v8, 31, v0
 ; GCN-NEXT:    s_waitcnt vmcnt(0)
 ; GCN-NEXT:    v_ashrrev_i32_e32 v9, 31, v4
-; GCN-NEXT:    v_ashrrev_i32_e32 v10, 31, v1
-; GCN-NEXT:    v_ashrrev_i32_e32 v11, 31, v5
-; GCN-NEXT:    v_ashrrev_i32_e32 v12, 31, v2
-; GCN-NEXT:    v_ashrrev_i32_e32 v13, 31, v6
-; GCN-NEXT:    v_ashrrev_i32_e32 v14, 31, v3
-; GCN-NEXT:    v_ashrrev_i32_e32 v15, 31, v7
-; GCN-NEXT:    v_xor_b32_e32 v16, v8, v9
-; GCN-NEXT:    v_add_i32_e32 v0, vcc, v8, v0
-; GCN-NEXT:    v_add_i32_e32 v1, vcc, v10, v1
-; GCN-NEXT:    v_add_i32_e32 v2, vcc, v12, v2
-; GCN-NEXT:    v_add_i32_e32 v3, vcc, v14, v3
 ; GCN-NEXT:    v_add_i32_e32 v4, vcc, v9, v4
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, v11, v5
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, v13, v6
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v15, v7
-; GCN-NEXT:    v_xor_b32_e32 v17, v10, v11
-; GCN-NEXT:    v_xor_b32_e32 v18, v12, v13
-; GCN-NEXT:    v_xor_b32_e32 v19, v14, v15
-; GCN-NEXT:    v_xor_b32_e32 v0, v0, v8
 ; GCN-NEXT:    v_xor_b32_e32 v4, v4, v9
-; GCN-NEXT:    v_xor_b32_e32 v1, v1, v10
+; GCN-NEXT:    v_xor_b32_e32 v15, v8, v9
+; GCN-NEXT:    v_cvt_f32_u32_e32 v9, v4
+; GCN-NEXT:    v_ashrrev_i32_e32 v11, 31, v5
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v11, v5
+; GCN-NEXT:    v_add_i32_e32 v0, vcc, v8, v0
+; GCN-NEXT:    v_rcp_iflag_f32_e32 v9, v9
 ; GCN-NEXT:    v_xor_b32_e32 v5, v5, v11
-; GCN-NEXT:    v_xor_b32_e32 v2, v2, v12
-; GCN-NEXT:    v_xor_b32_e32 v6, v6, v13
-; GCN-NEXT:    v_xor_b32_e32 v3, v3, v14
-; GCN-NEXT:    v_xor_b32_e32 v7, v7, v15
-; GCN-NEXT:    v_cvt_f32_u32_e32 v8, v4
-; GCN-NEXT:    v_cvt_f32_u32_e32 v9, v5
-; GCN-NEXT:    v_cvt_f32_u32_e32 v10, v6
+; GCN-NEXT:    v_xor_b32_e32 v0, v0, v8
+; GCN-NEXT:    v_cvt_f32_u32_e32 v8, v5
+; GCN-NEXT:    v_mul_f32_e32 v9, s14, v9
+; GCN-NEXT:    v_cvt_u32_f32_e32 v9, v9
+; GCN-NEXT:    v_ashrrev_i32_e32 v10, 31, v1
 ; GCN-NEXT:    v_rcp_iflag_f32_e32 v8, v8
-; GCN-NEXT:    v_rcp_iflag_f32_e32 v9, v9
-; GCN-NEXT:    v_rcp_iflag_f32_e32 v10, v10
-; GCN-NEXT:    v_mul_f32_e32 v8, s6, v8
-; GCN-NEXT:    v_mul_f32_e32 v9, s6, v9
-; GCN-NEXT:    v_mul_f32_e32 v10, s6, v10
+; GCN-NEXT:    v_add_i32_e32 v1, vcc, v10, v1
+; GCN-NEXT:    v_xor_b32_e32 v16, v10, v11
+; GCN-NEXT:    v_xor_b32_e32 v1, v1, v10
+; GCN-NEXT:    v_mul_f32_e32 v8, s14, v8
+; GCN-NEXT:    v_mul_hi_u32 v11, v9, v4
+; GCN-NEXT:    v_mul_lo_u32 v10, v9, v4
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v8, v8
-; GCN-NEXT:    v_cvt_u32_f32_e32 v9, v9
-; GCN-NEXT:    v_cvt_u32_f32_e32 v10, v10
-; GCN-NEXT:    v_mul_hi_u32 v11, v8, v4
-; GCN-NEXT:    v_mul_lo_u32 v12, v8, v4
-; GCN-NEXT:    v_mul_hi_u32 v13, v9, v5
-; GCN-NEXT:    v_mul_lo_u32 v14, v9, v5
-; GCN-NEXT:    v_sub_i32_e32 v15, vcc, 0, v12
+; GCN-NEXT:    v_ashrrev_i32_e32 v12, 31, v2
+; GCN-NEXT:    v_ashrrev_i32_e32 v13, 31, v6
+; GCN-NEXT:    v_add_i32_e32 v2, vcc, v12, v2
 ; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v11
-; GCN-NEXT:    v_mul_hi_u32 v11, v10, v6
-; GCN-NEXT:    v_cndmask_b32_e64 v12, v12, v15, s[0:1]
-; GCN-NEXT:    v_sub_i32_e32 v15, vcc, 0, v14
-; GCN-NEXT:    v_cmp_eq_u32_e64 s[2:3], 0, v13
-; GCN-NEXT:    v_mul_lo_u32 v13, v10, v6
-; GCN-NEXT:    v_cndmask_b32_e64 v14, v14, v15, s[2:3]
-; GCN-NEXT:    v_sub_i32_e32 v15, vcc, 0, v13
-; GCN-NEXT:    v_cmp_eq_u32_e64 s[4:5], 0, v11
-; GCN-NEXT:    v_cvt_f32_u32_e32 v11, v7
-; GCN-NEXT:    v_rcp_iflag_f32_e32 v11, v11
-; GCN-NEXT:    v_mul_f32_e32 v11, s6, v11
-; GCN-NEXT:    v_cvt_u32_f32_e32 v11, v11
-; GCN-NEXT:    v_cndmask_b32_e64 v13, v13, v15, s[4:5]
-; GCN-NEXT:    v_mul_hi_u32 v15, v11, v7
-; GCN-NEXT:    v_mul_lo_u32 v20, v11, v7
-; GCN-NEXT:    v_sub_i32_e32 v21, vcc, 0, v20
-; GCN-NEXT:    v_cmp_eq_u32_e64 s[6:7], 0, v15
-; GCN-NEXT:    v_cndmask_b32_e64 v15, v20, v21, s[6:7]
-; GCN-NEXT:    v_mul_hi_u32 v12, v12, v8
-; GCN-NEXT:    v_add_i32_e32 v20, vcc, v12, v8
-; GCN-NEXT:    v_subrev_i32_e32 v8, vcc, v12, v8
-; GCN-NEXT:    v_mul_hi_u32 v12, v14, v9
-; GCN-NEXT:    v_add_i32_e32 v14, vcc, v12, v9
-; GCN-NEXT:    v_subrev_i32_e32 v9, vcc, v12, v9
-; GCN-NEXT:    v_mul_hi_u32 v12, v13, v10
-; GCN-NEXT:    v_add_i32_e32 v13, vcc, v12, v10
-; GCN-NEXT:    v_subrev_i32_e32 v10, vcc, v12, v10
-; GCN-NEXT:    v_mul_hi_u32 v12, v15, v11
-; GCN-NEXT:    v_add_i32_e32 v15, vcc, v12, v11
-; GCN-NEXT:    v_subrev_i32_e32 v11, vcc, v12, v11
-; GCN-NEXT:    s_mov_b32 s16, s8
-; GCN-NEXT:    s_mov_b32 s17, s9
-; GCN-NEXT:    v_cndmask_b32_e64 v8, v8, v20, s[0:1]
-; GCN-NEXT:    v_cndmask_b32_e64 v9, v9, v14, s[2:3]
-; GCN-NEXT:    v_cndmask_b32_e64 v10, v10, v13, s[4:5]
-; GCN-NEXT:    v_cndmask_b32_e64 v11, v11, v15, s[6:7]
-; GCN-NEXT:    v_mul_hi_u32 v8, v8, v0
-; GCN-NEXT:    v_mul_hi_u32 v9, v9, v1
-; GCN-NEXT:    v_mul_hi_u32 v10, v10, v2
-; GCN-NEXT:    v_mul_hi_u32 v11, v11, v3
-; GCN-NEXT:    v_mul_lo_u32 v12, v8, v4
-; GCN-NEXT:    v_add_i32_e32 v13, vcc, 1, v8
-; GCN-NEXT:    v_add_i32_e32 v14, vcc, -1, v8
-; GCN-NEXT:    v_mul_lo_u32 v15, v9, v5
-; GCN-NEXT:    v_cmp_ge_u32_e64 s[0:1], v0, v12
-; GCN-NEXT:    v_sub_i32_e32 v0, vcc, v0, v12
-; GCN-NEXT:    v_add_i32_e32 v12, vcc, 1, v9
+; GCN-NEXT:    v_xor_b32_e32 v17, v12, v13
+; GCN-NEXT:    v_xor_b32_e32 v2, v2, v12
+; GCN-NEXT:    v_sub_i32_e32 v12, vcc, 0, v10
+; GCN-NEXT:    v_cndmask_b32_e64 v10, v10, v12, s[0:1]
+; GCN-NEXT:    v_mul_hi_u32 v12, v8, v5
+; GCN-NEXT:    v_add_i32_e32 v6, vcc, v13, v6
+; GCN-NEXT:    v_xor_b32_e32 v6, v6, v13
+; GCN-NEXT:    v_mul_lo_u32 v11, v8, v5
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[2:3], 0, v12
+; GCN-NEXT:    v_cvt_f32_u32_e32 v12, v6
+; GCN-NEXT:    v_mul_hi_u32 v10, v10, v9
+; GCN-NEXT:    v_sub_i32_e32 v13, vcc, 0, v11
+; GCN-NEXT:    v_cndmask_b32_e64 v11, v11, v13, s[2:3]
+; GCN-NEXT:    v_rcp_iflag_f32_e32 v12, v12
+; GCN-NEXT:    v_ashrrev_i32_e32 v14, 31, v7
+; GCN-NEXT:    v_add_i32_e32 v7, vcc, v14, v7
+; GCN-NEXT:    v_xor_b32_e32 v7, v7, v14
+; GCN-NEXT:    v_mul_f32_e32 v12, s14, v12
+; GCN-NEXT:    v_cvt_u32_f32_e32 v12, v12
+; GCN-NEXT:    v_mul_hi_u32 v18, v12, v6
+; GCN-NEXT:    v_mul_lo_u32 v13, v12, v6
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[4:5], 0, v18
+; GCN-NEXT:    v_add_i32_e32 v18, vcc, v10, v9
+; GCN-NEXT:    v_subrev_i32_e32 v9, vcc, v10, v9
+; GCN-NEXT:    v_mul_hi_u32 v10, v11, v8
+; GCN-NEXT:    v_cndmask_b32_e64 v9, v9, v18, s[0:1]
+; GCN-NEXT:    v_mul_hi_u32 v9, v9, v0
+; GCN-NEXT:    v_sub_i32_e32 v19, vcc, 0, v13
+; GCN-NEXT:    v_add_i32_e32 v11, vcc, v10, v8
+; GCN-NEXT:    v_subrev_i32_e32 v8, vcc, v10, v8
+; GCN-NEXT:    v_cndmask_b32_e64 v13, v13, v19, s[4:5]
+; GCN-NEXT:    v_cndmask_b32_e64 v8, v8, v11, s[2:3]
+; GCN-NEXT:    v_mul_hi_u32 v10, v13, v12
+; GCN-NEXT:    v_mul_lo_u32 v11, v9, v4
+; GCN-NEXT:    v_mul_hi_u32 v8, v8, v1
+; GCN-NEXT:    v_add_i32_e32 v13, vcc, v10, v12
+; GCN-NEXT:    v_subrev_i32_e32 v10, vcc, v10, v12
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[0:1], v0, v11
+; GCN-NEXT:    v_sub_i32_e32 v0, vcc, v0, v11
 ; GCN-NEXT:    v_cmp_ge_u32_e64 s[2:3], v0, v4
-; GCN-NEXT:    v_add_i32_e32 v0, vcc, -1, v9
-; GCN-NEXT:    v_mul_lo_u32 v4, v10, v6
-; GCN-NEXT:    v_cmp_ge_u32_e64 s[4:5], v1, v15
-; GCN-NEXT:    v_sub_i32_e32 v1, vcc, v1, v15
-; GCN-NEXT:    v_add_i32_e32 v15, vcc, 1, v10
-; GCN-NEXT:    v_cmp_ge_u32_e64 s[6:7], v1, v5
-; GCN-NEXT:    v_add_i32_e32 v1, vcc, -1, v10
-; GCN-NEXT:    v_mul_lo_u32 v5, v11, v7
-; GCN-NEXT:    v_cmp_ge_u32_e64 s[8:9], v2, v4
-; GCN-NEXT:    v_sub_i32_e32 v2, vcc, v2, v4
-; GCN-NEXT:    v_add_i32_e32 v4, vcc, -1, v11
-; GCN-NEXT:    v_cmp_ge_u32_e64 s[10:11], v3, v5
-; GCN-NEXT:    v_sub_i32_e32 v3, vcc, v3, v5
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, 1, v11
-; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v2, v6
-; GCN-NEXT:    v_cmp_ge_u32_e64 s[12:13], v3, v7
-; GCN-NEXT:    s_and_b64 s[2:3], s[2:3], s[0:1]
-; GCN-NEXT:    v_cndmask_b32_e64 v2, v8, v13, s[2:3]
-; GCN-NEXT:    s_and_b64 s[2:3], s[6:7], s[4:5]
-; GCN-NEXT:    v_cndmask_b32_e64 v3, v9, v12, s[2:3]
-; GCN-NEXT:    s_and_b64 vcc, vcc, s[8:9]
-; GCN-NEXT:    v_cndmask_b32_e32 v6, v10, v15, vcc
-; GCN-NEXT:    s_and_b64 vcc, s[12:13], s[10:11]
-; GCN-NEXT:    v_cndmask_b32_e32 v5, v11, v5, vcc
-; GCN-NEXT:    v_cndmask_b32_e64 v2, v14, v2, s[0:1]
-; GCN-NEXT:    v_cndmask_b32_e64 v0, v0, v3, s[4:5]
-; GCN-NEXT:    v_cndmask_b32_e64 v1, v1, v6, s[8:9]
-; GCN-NEXT:    v_cndmask_b32_e64 v3, v4, v5, s[10:11]
-; GCN-NEXT:    v_xor_b32_e32 v2, v2, v16
-; GCN-NEXT:    v_xor_b32_e32 v4, v0, v17
-; GCN-NEXT:    v_xor_b32_e32 v5, v1, v18
-; GCN-NEXT:    v_xor_b32_e32 v3, v3, v19
-; GCN-NEXT:    v_sub_i32_e32 v0, vcc, v2, v16
-; GCN-NEXT:    v_sub_i32_e32 v1, vcc, v4, v17
-; GCN-NEXT:    v_sub_i32_e32 v2, vcc, v5, v18
-; GCN-NEXT:    v_sub_i32_e32 v3, vcc, v3, v19
-; GCN-NEXT:    buffer_store_dwordx4 v[0:3], off, s[16:19], 0
+; GCN-NEXT:    v_cndmask_b32_e64 v10, v10, v13, s[4:5]
+; GCN-NEXT:    v_mul_lo_u32 v0, v8, v5
+; GCN-NEXT:    v_mul_hi_u32 v4, v10, v2
+; GCN-NEXT:    v_add_i32_e32 v12, vcc, -1, v9
+; GCN-NEXT:    v_add_i32_e32 v10, vcc, -1, v8
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[4:5], v1, v0
+; GCN-NEXT:    v_sub_i32_e32 v0, vcc, v1, v0
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[6:7], v0, v5
+; GCN-NEXT:    v_mul_lo_u32 v5, v4, v6
+; GCN-NEXT:    v_add_i32_e32 v1, vcc, 1, v9
+; GCN-NEXT:    v_add_i32_e32 v0, vcc, 1, v8
+; GCN-NEXT:    s_and_b64 vcc, s[2:3], s[0:1]
+; GCN-NEXT:    v_cndmask_b32_e32 v1, v9, v1, vcc
+; GCN-NEXT:    v_sub_i32_e32 v9, vcc, v2, v5
+; GCN-NEXT:    s_and_b64 vcc, s[6:7], s[4:5]
+; GCN-NEXT:    v_cvt_f32_u32_e32 v11, v7
+; GCN-NEXT:    v_cndmask_b32_e32 v0, v8, v0, vcc
+; GCN-NEXT:    v_cndmask_b32_e64 v1, v12, v1, s[0:1]
+; GCN-NEXT:    v_cndmask_b32_e64 v0, v10, v0, s[4:5]
+; GCN-NEXT:    v_xor_b32_e32 v1, v1, v15
+; GCN-NEXT:    v_xor_b32_e32 v8, v0, v16
+; GCN-NEXT:    v_sub_i32_e32 v0, vcc, v1, v15
+; GCN-NEXT:    v_sub_i32_e32 v1, vcc, v8, v16
+; GCN-NEXT:    v_rcp_iflag_f32_e32 v8, v11
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[0:1], v9, v6
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[2:3], v2, v5
+; GCN-NEXT:    v_ashrrev_i32_e32 v10, 31, v3
+; GCN-NEXT:    v_mul_f32_e32 v8, s14, v8
+; GCN-NEXT:    v_cvt_u32_f32_e32 v8, v8
+; GCN-NEXT:    v_add_i32_e32 v3, vcc, v10, v3
+; GCN-NEXT:    v_xor_b32_e32 v3, v3, v10
+; GCN-NEXT:    v_add_i32_e32 v6, vcc, -1, v4
+; GCN-NEXT:    v_mul_lo_u32 v5, v8, v7
+; GCN-NEXT:    v_mul_hi_u32 v9, v8, v7
+; GCN-NEXT:    v_add_i32_e32 v2, vcc, 1, v4
+; GCN-NEXT:    v_sub_i32_e32 v11, vcc, 0, v5
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[4:5], 0, v9
+; GCN-NEXT:    v_cndmask_b32_e64 v5, v5, v11, s[4:5]
+; GCN-NEXT:    v_mul_hi_u32 v5, v5, v8
+; GCN-NEXT:    v_add_i32_e32 v9, vcc, v5, v8
+; GCN-NEXT:    v_subrev_i32_e32 v5, vcc, v5, v8
+; GCN-NEXT:    v_cndmask_b32_e64 v5, v5, v9, s[4:5]
+; GCN-NEXT:    v_mul_hi_u32 v5, v5, v3
+; GCN-NEXT:    s_and_b64 vcc, s[0:1], s[2:3]
+; GCN-NEXT:    v_cndmask_b32_e32 v2, v4, v2, vcc
+; GCN-NEXT:    v_cndmask_b32_e64 v2, v6, v2, s[2:3]
+; GCN-NEXT:    v_mul_lo_u32 v4, v5, v7
+; GCN-NEXT:    v_xor_b32_e32 v2, v2, v17
+; GCN-NEXT:    v_sub_i32_e32 v2, vcc, v2, v17
+; GCN-NEXT:    v_xor_b32_e32 v6, v10, v14
+; GCN-NEXT:    v_sub_i32_e32 v8, vcc, v3, v4
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[0:1], v8, v7
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[2:3], v3, v4
+; GCN-NEXT:    v_add_i32_e32 v7, vcc, -1, v5
+; GCN-NEXT:    v_add_i32_e32 v3, vcc, 1, v5
+; GCN-NEXT:    s_and_b64 vcc, s[0:1], s[2:3]
+; GCN-NEXT:    v_cndmask_b32_e32 v3, v5, v3, vcc
+; GCN-NEXT:    v_cndmask_b32_e64 v3, v7, v3, s[2:3]
+; GCN-NEXT:    v_xor_b32_e32 v3, v3, v6
+; GCN-NEXT:    v_sub_i32_e32 v3, vcc, v3, v6
+; GCN-NEXT:    buffer_store_dwordx4 v[0:3], off, s[8:11], 0
 ; GCN-NEXT:    s_endpgm
 ;
 ; TONGA-LABEL: sdiv_v4i32:
 ; TONGA:       ; %bb.0:
 ; TONGA-NEXT:    s_load_dwordx4 s[12:15], s[0:1], 0x24
 ; TONGA-NEXT:    s_mov_b32 s11, 0xf000
 ; TONGA-NEXT:    s_mov_b32 s10, -1
 ; TONGA-NEXT:    s_mov_b32 s2, s10
; EG-NEXT:    2(2.802597e-45), 0(0.000000e+00)
   %result = sdiv <4 x i32> %num, %den
   store <4 x i32> %result, <4 x i32> addrspace(1)* %out
   ret void
 }
 
 define amdgpu_kernel void @sdiv_v4i32_4(<4 x i32> addrspace(1)* %out, <4 x i32> addrspace(1)* %in) {
 ; GCN-LABEL: sdiv_v4i32_4:
 ; GCN:       ; %bb.0:
-; GCN-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
-; GCN-NEXT:    s_mov_b32 s7, 0xf000
-; GCN-NEXT:    s_mov_b32 s6, -1
-; GCN-NEXT:    s_mov_b32 s10, s6
-; GCN-NEXT:    s_mov_b32 s11, s7
+; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN-NEXT:    s_mov_b32 s3, 0xf000
+; GCN-NEXT:    s_mov_b32 s2, -1
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_mov_b32 s8, s2
-; GCN-NEXT:    s_mov_b32 s9, s3
-; GCN-NEXT:    buffer_load_dwordx4 v[0:3], off, s[8:11], 0
-; GCN-NEXT:    s_mov_b32 s4, s0
-; GCN-NEXT:    s_mov_b32 s5, s1
+; GCN-NEXT:    s_mov_b32 s0, s4
+; GCN-NEXT:    s_mov_b32 s1, s5
+; GCN-NEXT:    s_mov_b32 s4, s6
+; GCN-NEXT:    s_mov_b32 s5, s7
+; GCN-NEXT:    s_mov_b32 s6, s2
+; GCN-NEXT:    s_mov_b32 s7, s3
+; GCN-NEXT:    buffer_load_dwordx4 v[0:3], off, s[4:7], 0
 ; GCN-NEXT:    s_waitcnt vmcnt(0)
 ; GCN-NEXT:    v_ashrrev_i32_e32 v4, 31, v0
 ; GCN-NEXT:    v_ashrrev_i32_e32 v5, 31, v1
 ; GCN-NEXT:    v_ashrrev_i32_e32 v6, 31, v2
 ; GCN-NEXT:    v_ashrrev_i32_e32 v7, 31, v3
 ; GCN-NEXT:    v_lshrrev_b32_e32 v4, 30, v4
 ; GCN-NEXT:    v_lshrrev_b32_e32 v5, 30, v5
 ; GCN-NEXT:    v_lshrrev_b32_e32 v6, 30, v6
 ; GCN-NEXT:    v_lshrrev_b32_e32 v7, 30, v7
 ; GCN-NEXT:    v_add_i32_e32 v0, vcc, v4, v0
 ; GCN-NEXT:    v_add_i32_e32 v1, vcc, v5, v1
 ; GCN-NEXT:    v_add_i32_e32 v2, vcc, v6, v2
 ; GCN-NEXT:    v_add_i32_e32 v3, vcc, v7, v3
 ; GCN-NEXT:    v_ashrrev_i32_e32 v0, 2, v0
 ; GCN-NEXT:    v_ashrrev_i32_e32 v1, 2, v1
 ; GCN-NEXT:    v_ashrrev_i32_e32 v2, 2, v2
 ; GCN-NEXT:    v_ashrrev_i32_e32 v3, 2, v3
-; GCN-NEXT:    buffer_store_dwordx4 v[0:3], off, s[4:7], 0
+; GCN-NEXT:    buffer_store_dwordx4 v[0:3], off, s[0:3], 0
 ; GCN-NEXT:    s_endpgm
 ;
 ; TONGA-LABEL: sdiv_v4i32_4:
 ; TONGA:       ; %bb.0:
 ; TONGA-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; TONGA-NEXT:    s_mov_b32 s3, 0xf000
 ; TONGA-NEXT:    s_mov_b32 s2, -1
 ; TONGA-NEXT:    s_waitcnt lgkmcnt(0)
; EG-NEXT:    2(2.802597e-45), 0(0.000000e+00)
   %result = sdiv <4 x i32> %num, <i32 4, i32 4, i32 4, i32 4>
   store <4 x i32> %result, <4 x i32> addrspace(1)* %out
   ret void
 }
 
 define amdgpu_kernel void @v_sdiv_i8(i32 addrspace(1)* %out, i8 addrspace(1)* %in) {
 ; GCN-LABEL: v_sdiv_i8:
 ; GCN:       ; %bb.0:
-; GCN-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
-; GCN-NEXT:    s_mov_b32 s7, 0xf000
-; GCN-NEXT:    s_mov_b32 s6, -1
-; GCN-NEXT:    s_mov_b32 s10, s6
-; GCN-NEXT:    s_mov_b32 s11, s7
+; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN-NEXT:    s_mov_b32 s3, 0xf000
+; GCN-NEXT:    s_mov_b32 s2, -1
+; GCN-NEXT:    s_mov_b32 s10, s2
+; GCN-NEXT:    s_mov_b32 s11, s3
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_mov_b32 s8, s2
-; GCN-NEXT:    s_mov_b32 s9, s3
+; GCN-NEXT:    s_mov_b32 s8, s6
+; GCN-NEXT:    s_mov_b32 s9, s7
 ; GCN-NEXT:    buffer_load_sbyte v0, off, s[8:11], 0
 ; GCN-NEXT:    buffer_load_sbyte v1, off, s[8:11], 0 offset:1
-; GCN-NEXT:    s_mov_b32 s4, s0
-; GCN-NEXT:    s_mov_b32 s5, s1
+; GCN-NEXT:    s_mov_b32 s0, s4
+; GCN-NEXT:    s_mov_b32 s1, s5
+; GCN-NEXT:    s_waitcnt vmcnt(1)
+; GCN-NEXT:    v_cvt_f32_i32_e32 v3, v0
 ; GCN-NEXT:    s_waitcnt vmcnt(0)
-; GCN-NEXT:    v_xor_b32_e32 v2, v0, v1
-; GCN-NEXT:    v_cvt_f32_i32_e32 v0, v0
-; GCN-NEXT:    v_cvt_f32_i32_e32 v1, v1
-; GCN-NEXT:    v_ashrrev_i32_e32 v2, 30, v2
-; GCN-NEXT:    v_rcp_iflag_f32_e32 v3, v1
-; GCN-NEXT:    v_or_b32_e32 v2, 1, v2
-; GCN-NEXT:    v_mul_f32_e32 v3, v0, v3
-; GCN-NEXT:    v_trunc_f32_e32 v3, v3
-; GCN-NEXT:    v_mad_f32 v0, -v3, v1, v0
-; GCN-NEXT:    v_cvt_i32_f32_e32 v3, v3
-; GCN-NEXT:    v_cmp_ge_f32_e64 vcc, |v0|, |v1|
-; GCN-NEXT:    v_cndmask_b32_e32 v0, 0, v2, vcc
-; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v3
+; GCN-NEXT:    v_cvt_f32_i32_e32 v2, v1
+; GCN-NEXT:    v_xor_b32_e32 v0, v0, v1
+; GCN-NEXT:    v_ashrrev_i32_e32 v0, 30, v0
+; GCN-NEXT:    v_or_b32_e32 v0, 1, v0
+; GCN-NEXT:    v_rcp_iflag_f32_e32 v4, v2
+; GCN-NEXT:    v_mul_f32_e32 v1, v3, v4
+; GCN-NEXT:    v_trunc_f32_e32 v1, v1
+; GCN-NEXT:    v_mad_f32 v3, -v1, v2, v3
+; GCN-NEXT:    v_cvt_i32_f32_e32 v1, v1
+; GCN-NEXT:    v_cmp_ge_f32_e64 vcc, |v3|, |v2|
+; GCN-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
+; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v1
 ; GCN-NEXT:    v_bfe_i32 v0, v0, 0, 8
-; GCN-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; GCN-NEXT:    s_endpgm
 ;
 ; TONGA-LABEL: v_sdiv_i8:
 ; TONGA:       ; %bb.0:
 ; TONGA-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; TONGA-NEXT:    s_mov_b32 s3, 0xf000
 ; TONGA-NEXT:    s_mov_b32 s2, -1
 ; TONGA-NEXT:    s_mov_b32 s10, s2
; EG-NEXT:    8(1.121039e-44), 2(2.802597e-45)
   %result.ext = sext i8 %result to i32
   store i32 %result.ext, i32 addrspace(1)* %out
   ret void
 }
 
 define amdgpu_kernel void @v_sdiv_i23(i32 addrspace(1)* %out, i23 addrspace(1)* %in) {
 ; GCN-LABEL: v_sdiv_i23:
 ; GCN:       ; %bb.0:
-; GCN-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
-; GCN-NEXT:    s_mov_b32 s7, 0xf000
-; GCN-NEXT:    s_mov_b32 s6, -1
-; GCN-NEXT:    s_mov_b32 s10, s6
-; GCN-NEXT:    s_mov_b32 s11, s7
+; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN-NEXT:    s_mov_b32 s3, 0xf000
+; GCN-NEXT:    s_mov_b32 s2, -1
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_mov_b32 s8, s2
-; GCN-NEXT:    s_mov_b32 s9, s3
-; GCN-NEXT:    buffer_load_ubyte v1, off, s[8:11], 0 offset:2
-; GCN-NEXT:    buffer_load_ubyte v3, off, s[8:11], 0 offset:6
-; GCN-NEXT:    buffer_load_ushort v0, off, s[8:11], 0
-; GCN-NEXT:    buffer_load_ushort v2, off, s[8:11], 0 offset:4
-; GCN-NEXT:    s_mov_b32 s4, s0
-; GCN-NEXT:    s_mov_b32 s5, s1
-; GCN-NEXT:    s_waitcnt vmcnt(3)
-; GCN-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; GCN-NEXT:    s_mov_b32 s0, s4
+; GCN-NEXT:    s_mov_b32 s1, s5
+; GCN-NEXT:    s_mov_b32 s4, s6
+; GCN-NEXT:    s_mov_b32 s5, s7
+; GCN-NEXT:    s_mov_b32 s6, s2
+; GCN-NEXT:    s_mov_b32 s7, s3
+; GCN-NEXT:    buffer_load_ushort v0, off, s[4:7], 0
+; GCN-NEXT:    buffer_load_ubyte v1, off, s[4:7], 0 offset:2
+; GCN-NEXT:    buffer_load_ushort v2, off, s[4:7], 0 offset:4
+; GCN-NEXT:    buffer_load_ubyte v3, off, s[4:7], 0 offset:6
 ; GCN-NEXT:    s_waitcnt vmcnt(2)
-; GCN-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT:    s_waitcnt vmcnt(1)
+; GCN-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GCN-NEXT:    v_or_b32_e32 v0, v0, v1
 ; GCN-NEXT:    s_waitcnt vmcnt(0)
-; GCN-NEXT:    v_or_b32_e32 v1, v2, v3
+; GCN-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GCN-NEXT:    v_or_b32_e32 v2, v2, v3
+; GCN-NEXT:    v_bfe_i32 v2, v2, 0, 23
+; GCN-NEXT:    v_cvt_f32_i32_e32 v3, v2
 ; GCN-NEXT:    v_bfe_i32 v0, v0, 0, 23
-; GCN-NEXT:    v_bfe_i32 v1, v1, 0, 23
-; GCN-NEXT:    v_xor_b32_e32 v2, v0, v1
-; GCN-NEXT:    v_cvt_f32_i32_e32 v0, v0
-; GCN-NEXT:    v_cvt_f32_i32_e32 v1, v1
-; GCN-NEXT:    v_ashrrev_i32_e32 v2, 30, v2
-; GCN-NEXT:    v_rcp_iflag_f32_e32 v3, v1
-; GCN-NEXT:    v_or_b32_e32 v2, 1, v2
-; GCN-NEXT:    v_mul_f32_e32 v3, v0, v3
-; GCN-NEXT:    v_trunc_f32_e32 v3, v3
-; GCN-NEXT:    v_mad_f32 v0, -v3, v1, v0
-; GCN-NEXT:    v_cvt_i32_f32_e32 v3, v3
-; GCN-NEXT:    v_cmp_ge_f32_e64 vcc, |v0|, |v1|
-; GCN-NEXT:    v_cndmask_b32_e32 v0, 0, v2, vcc
-; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v3
+; GCN-NEXT:    v_cvt_f32_i32_e32 v1, v0
+; GCN-NEXT:    v_xor_b32_e32 v0, v0, v2
+; GCN-NEXT:    v_rcp_iflag_f32_e32 v4, v3
+; GCN-NEXT:    v_ashrrev_i32_e32 v0, 30, v0
+; GCN-NEXT:    v_or_b32_e32 v0, 1, v0
+; GCN-NEXT:    v_mul_f32_e32 v2, v1, v4
+; GCN-NEXT:    v_trunc_f32_e32 v2, v2
+; GCN-NEXT:    v_mad_f32 v1, -v2, v3, v1
+; GCN-NEXT:    v_cvt_i32_f32_e32 v2, v2
+; GCN-NEXT:    v_cmp_ge_f32_e64 vcc, |v1|, |v3|
+; GCN-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
+; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v2
 ; GCN-NEXT:    v_bfe_i32 v0, v0, 0, 23
-; GCN-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; GCN-NEXT:    s_endpgm
 ;
 ; TONGA-LABEL: v_sdiv_i23:
 ; TONGA:       ; %bb.0:
 ; TONGA-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; TONGA-NEXT:    s_mov_b32 s3, 0xf000
 ; TONGA-NEXT:    s_mov_b32 s2, -1
 ; TONGA-NEXT:    s_waitcnt lgkmcnt(0)
; EG-NEXT:    9(1.261169e-44), 2(2.802597e-45)
   %result.ext = sext i23 %result to i32
   store i32 %result.ext, i32 addrspace(1)* %out
   ret void
 }
 
 define amdgpu_kernel void @v_sdiv_i24(i32 addrspace(1)* %out, i24 addrspace(1)* %in) {
 ; GCN-LABEL: v_sdiv_i24:
 ; GCN:       ; %bb.0:
-; GCN-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
-; GCN-NEXT:    s_mov_b32 s7, 0xf000
-; GCN-NEXT:    s_mov_b32 s6, -1
-; GCN-NEXT:    s_mov_b32 s10, s6
-; GCN-NEXT:    s_mov_b32 s11, s7
+; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN-NEXT:    s_mov_b32 s3, 0xf000
+; GCN-NEXT:    s_mov_b32 s2, -1
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_mov_b32 s8, s2
-; GCN-NEXT:    s_mov_b32 s9, s3
-; GCN-NEXT:    buffer_load_sbyte v1, off, s[8:11], 0 offset:2
-; GCN-NEXT:    buffer_load_sbyte v3, off, s[8:11], 0 offset:6
-; GCN-NEXT:    buffer_load_ushort v0, off, s[8:11], 0
-; GCN-NEXT:    buffer_load_ushort v2, off, s[8:11], 0 offset:4
-; GCN-NEXT:    s_mov_b32 s4, s0
-; GCN-NEXT:    s_mov_b32 s5, s1
-; GCN-NEXT:    s_waitcnt vmcnt(3)
-; GCN-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; GCN-NEXT:    s_mov_b32 s0, s4
+; GCN-NEXT:    s_mov_b32 s1, s5
+; GCN-NEXT:    s_mov_b32 s4, s6
+; GCN-NEXT:    s_mov_b32 s5, s7
+; GCN-NEXT:    s_mov_b32 s6, s2
+; GCN-NEXT:    s_mov_b32 s7, s3
+; GCN-NEXT:    buffer_load_ushort v0, off, s[4:7], 0
+; GCN-NEXT:    buffer_load_sbyte v1, off, s[4:7], 0 offset:2
+; GCN-NEXT:    buffer_load_ushort v2, off, s[4:7], 0 offset:4
+; GCN-NEXT:    buffer_load_sbyte v3, off, s[4:7], 0 offset:6
 ; GCN-NEXT:    s_waitcnt vmcnt(2)
-; GCN-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; GCN-NEXT:    s_waitcnt vmcnt(1)
+; GCN-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GCN-NEXT:    v_or_b32_e32 v0, v0, v1
 ; GCN-NEXT:    s_waitcnt vmcnt(0)
+; GCN-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; GCN-NEXT:    v_or_b32_e32 v2, v2, v3
+; GCN-NEXT:    v_cvt_f32_i32_e32 v2, v2
+; GCN-NEXT:    v_cvt_f32_i32_e32 v0, v0
 ; GCN-NEXT:    v_xor_b32_e32 v1, v1, v3
 ; GCN-NEXT:    v_ashrrev_i32_e32 v1, 30, v1
-; GCN-NEXT:    v_cvt_f32_i32_e32 v0, v0
-; GCN-NEXT:    v_cvt_f32_i32_e32 v2, v2
+; GCN-NEXT:    v_rcp_iflag_f32_e32 v4, v2
 ; GCN-NEXT:    v_or_b32_e32 v1, 1, v1
-; GCN-NEXT:    v_rcp_iflag_f32_e32 v3, v2
-; GCN-NEXT:    v_mul_f32_e32 v3, v0, v3
+; GCN-NEXT:    v_mul_f32_e32 v3, v0, v4
 ; GCN-NEXT:    v_trunc_f32_e32 v3, v3
 ; GCN-NEXT:    v_mad_f32 v0, -v3, v2, v0
 ; GCN-NEXT:    v_cvt_i32_f32_e32 v3, v3
 ; GCN-NEXT:    v_cmp_ge_f32_e64 vcc, |v0|, |v2|
 ; GCN-NEXT:    v_cndmask_b32_e32 v0, 0, v1, vcc
 ; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v3
 ; GCN-NEXT:    v_bfe_i32 v0, v0, 0, 24
-; GCN-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; GCN-NEXT:    s_endpgm
 ;
 ; TONGA-LABEL: v_sdiv_i24:
 ; TONGA:       ; %bb.0:
 ; TONGA-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; TONGA-NEXT:    s_mov_b32 s3, 0xf000
 ; TONGA-NEXT:    s_mov_b32 s2, -1
 ; TONGA-NEXT:    s_waitcnt lgkmcnt(0)
; EG-NEXT:    8(1.121039e-44), 2(2.802597e-45)
   %result.ext = sext i24 %result to i32
   store i32 %result.ext, i32 addrspace(1)* %out
   ret void
 }
 
 define amdgpu_kernel void @v_sdiv_i25(i32 addrspace(1)* %out, i25 addrspace(1)* %in) {
 ; GCN-LABEL: v_sdiv_i25:
 ; GCN:       ; %bb.0:
-; GCN-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; GCN-NEXT:    s_load_dwordx4 s[8:11], s[0:1], 0x9
 ; GCN-NEXT:    s_mov_b32 s7, 0xf000
 ; GCN-NEXT:    s_mov_b32 s6, -1
-; GCN-NEXT:    s_mov_b32 s10, s6
-; GCN-NEXT:    s_mov_b32 s11, s7
+; GCN-NEXT:    s_mov_b32 s2, s6
+; GCN-NEXT:    s_mov_b32 s3, s7
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_mov_b32 s8, s2
-; GCN-NEXT:    s_mov_b32 s9, s3
-; GCN-NEXT:    buffer_load_dwordx2 v[0:1], off, s[8:11], 0
-; GCN-NEXT:    s_mov_b32 s4, s0
-; GCN-NEXT:    s_mov_b32 s5, s1
+; GCN-NEXT:    s_mov_b32 s0, s10
+; GCN-NEXT:    s_mov_b32 s1, s11
+; GCN-NEXT:    buffer_load_dwordx2 v[0:1], off, s[0:3], 0
+; GCN-NEXT:    s_mov_b32 s4, s8
+; GCN-NEXT:    s_mov_b32 s5, s9
 ; GCN-NEXT:    s_waitcnt vmcnt(0)
-; GCN-NEXT:    v_bfe_i32 v2, v0, 0, 25
-; GCN-NEXT:    v_bfe_i32 v3, v1, 0, 25
-; GCN-NEXT:    v_bfe_i32 v0, v0, 24, 1
+; GCN-NEXT:    v_bfe_i32 v2, v1, 0, 25
 ; GCN-NEXT:    v_bfe_i32 v1, v1, 24, 1
-; GCN-NEXT:    v_xor_b32_e32 v4, v0, v1
-; GCN-NEXT:    v_add_i32_e32 v2, vcc, v0, v2
-; GCN-NEXT:    v_add_i32_e32 v3, vcc, v1, v3
-; GCN-NEXT:    v_xor_b32_e32 v0, v2, v0
-; GCN-NEXT:    v_xor_b32_e32 v1, v3, v1
-; GCN-NEXT:    v_cvt_f32_u32_e32 v2, v1
-; GCN-NEXT:    v_rcp_iflag_f32_e32 v2, v2
-; GCN-NEXT:    v_mul_f32_e32 v2, 0x4f800000, v2
-; GCN-NEXT:    v_cvt_u32_f32_e32 v2, v2
-; GCN-NEXT:    v_mul_hi_u32 v3, v2, v1
-; GCN-NEXT:    v_mul_lo_u32 v5, v2, v1
-; GCN-NEXT:    v_sub_i32_e32 v6, vcc, 0, v5
-; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v3
-; GCN-NEXT:    v_cndmask_b32_e64 v3, v5, v6, s[0:1]
-; GCN-NEXT:    v_mul_hi_u32 v3, v3, v2
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, v3, v2
-; GCN-NEXT:    v_subrev_i32_e32 v2, vcc, v3, v2
-; GCN-NEXT:    v_cndmask_b32_e64 v2, v2, v5, s[0:1]
-; GCN-NEXT:    v_mul_hi_u32 v2, v2, v0
-; GCN-NEXT:    v_mul_lo_u32 v3, v2, v1
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, 1, v2
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, -1, v2
-; GCN-NEXT:    v_subrev_i32_e32 v7, vcc, v3, v0
-; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v0, v3
-; GCN-NEXT:    v_cmp_ge_u32_e64 s[0:1], v7, v1
+; GCN-NEXT:    v_add_i32_e32 v2, vcc, v1, v2
+; GCN-NEXT:    v_xor_b32_e32 v2, v2, v1
+; GCN-NEXT:    v_cvt_f32_u32_e32 v3, v2
+; GCN-NEXT:    v_bfe_i32 v4, v0, 0, 25
+; GCN-NEXT:    v_bfe_i32 v0, v0, 24, 1
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v0, v4
+; GCN-NEXT:    v_rcp_iflag_f32_e32 v3, v3
+; GCN-NEXT:    v_xor_b32_e32 v4, v4, v0
+; GCN-NEXT:    v_xor_b32_e32 v0, v0, v1
+; GCN-NEXT:    v_mul_f32_e32 v3, 0x4f800000, v3
+; GCN-NEXT:    v_cvt_u32_f32_e32 v3, v3
+; GCN-NEXT:    v_mul_lo_u32 v5, v3, v2
+; GCN-NEXT:    v_mul_hi_u32 v6, v3, v2
+; GCN-NEXT:    v_sub_i32_e32 v7, vcc, 0, v5
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v6
+; GCN-NEXT:    v_cndmask_b32_e64 v5, v5, v7, s[0:1]
+; GCN-NEXT:    v_mul_hi_u32 v5, v5, v3
+; GCN-NEXT:    v_add_i32_e32 v6, vcc, v5, v3
+; GCN-NEXT:    v_subrev_i32_e32 v3, vcc, v5, v3
+; GCN-NEXT:    v_cndmask_b32_e64 v3, v3, v6, s[0:1]
+; GCN-NEXT:    v_mul_hi_u32 v3, v3, v4
+; GCN-NEXT:    v_mul_lo_u32 v1, v3, v2
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, 1, v3
+; GCN-NEXT:    v_add_i32_e32 v6, vcc, -1, v3
+; GCN-NEXT:    v_subrev_i32_e32 v7, vcc, v1, v4
+; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v4, v1
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[0:1], v7, v2
 ; GCN-NEXT:    s_and_b64 s[0:1], s[0:1], vcc
-; GCN-NEXT:    v_cndmask_b32_e64 v0, v2, v5, s[0:1]
-; GCN-NEXT:    v_cndmask_b32_e32 v0, v6, v0, vcc
-; GCN-NEXT:    v_xor_b32_e32 v0, v0, v4
-; GCN-NEXT:    v_sub_i32_e32 v0, vcc, v0, v4
+; GCN-NEXT:    v_cndmask_b32_e64 v1, v3, v5, s[0:1]
+; GCN-NEXT:    v_cndmask_b32_e32 v1, v6, v1, vcc
+; GCN-NEXT:    v_xor_b32_e32 v1, v1, v0
+; GCN-NEXT:    v_sub_i32_e32 v0, vcc, v1, v0
 ; GCN-NEXT:    v_bfe_i32 v0, v0, 0, 25
 ; GCN-NEXT:    buffer_store_dword v0, off, s[4:7], 0
 ; GCN-NEXT:    s_endpgm
 ;
 ; TONGA-LABEL: v_sdiv_i25:
 ; TONGA:       ; %bb.0:
 ; TONGA-NEXT:    s_load_dwordx4 s[8:11], s[0:1], 0x24
 ; TONGA-NEXT:    s_mov_b32 s7, 0xf000

 ;   %result = add i64 %resultdiv, %resultrem
 ;   store i64 %result, i64 addrspace(1)* %out, align 8
 ;   ret void
 ; }
 
 define amdgpu_kernel void @scalarize_mulhs_4xi32(<4 x i32> addrspace(1)* nocapture readonly %in, <4 x i32> addrspace(1)* nocapture %out) {
 ; GCN-LABEL: scalarize_mulhs_4xi32:
 ; GCN:       ; %bb.0:
-; GCN-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
-; GCN-NEXT:    s_mov_b32 s7, 0xf000
-; GCN-NEXT:    s_mov_b32 s6, -1
+; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN-NEXT:    s_mov_b32 s3, 0xf000
+; GCN-NEXT:    s_mov_b32 s2, -1
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_mov_b32 s4, s0
-; GCN-NEXT:    s_mov_b32 s5, s1
-; GCN-NEXT:    buffer_load_dwordx4 v[0:3], off, s[4:7], 0
+; GCN-NEXT:    s_mov_b32 s0, s4
+; GCN-NEXT:    s_mov_b32 s1, s5
+; GCN-NEXT:    buffer_load_dwordx4 v[0:3], off, s[0:3], 0
 ; GCN-NEXT:    s_mov_b32 s0, 0x1389c755
-; GCN-NEXT:    s_mov_b32 s4, s2
-; GCN-NEXT:    s_mov_b32 s5, s3
+; GCN-NEXT:    s_mov_b32 s4, s6
+; GCN-NEXT:    s_mov_b32 s5, s7
+; GCN-NEXT:    s_mov_b32 s6, s2
+; GCN-NEXT:    s_mov_b32 s7, s3
 ; GCN-NEXT:    s_waitcnt vmcnt(0)
 ; GCN-NEXT:    v_mul_hi_i32 v0, v0, s0
 ; GCN-NEXT:    v_mul_hi_i32 v1, v1, s0
 ; GCN-NEXT:    v_mul_hi_i32 v2, v2, s0
 ; GCN-NEXT:    v_mul_hi_i32 v3, v3, s0
 ; GCN-NEXT:    v_lshrrev_b32_e32 v4, 31, v0
 ; GCN-NEXT:    v_ashrrev_i32_e32 v0, 12, v0
 ; GCN-NEXT:    v_lshrrev_b32_e32 v5, 31, v1

llvm/test/CodeGen/AMDGPU/sdiv64.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
+; RUN: llc -march=amdgcn -mcpu=gfx600 -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
 
 define amdgpu_kernel void @s_test_sdiv(i64 addrspace(1)* %out, i64 %x, i64 %y) {
 ; GCN-LABEL: s_test_sdiv:
 ; GCN:       ; %bb.0:
-; GCN-NEXT:    s_load_dwordx4 s[8:11], s[0:1], 0x9
 ; GCN-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xd
+; GCN-NEXT:    v_mov_b32_e32 v7, 0
+; GCN-NEXT:    s_load_dwordx4 s[8:11], s[0:1], 0x9
 ; GCN-NEXT:    s_mov_b32 s7, 0xf000
 ; GCN-NEXT:    s_mov_b32 s6, -1
-; GCN-NEXT:    v_mov_b32_e32 v0, 0
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_ashr_i32 s0, s3, 31
+; GCN-NEXT:    s_ashr_i32 s12, s3, 31
+; GCN-NEXT:    s_add_u32 s2, s2, s12
+; GCN-NEXT:    s_mov_b32 s13, s12
+; GCN-NEXT:    s_addc_u32 s3, s3, s12
+; GCN-NEXT:    s_xor_b64 s[2:3], s[2:3], s[12:13]
+; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s2
+; GCN-NEXT:    v_cvt_f32_u32_e32 v1, s3
+; GCN-NEXT:    s_sub_u32 s4, 0, s2
+; GCN-NEXT:    s_subb_u32 s5, 0, s3
+; GCN-NEXT:    s_ashr_i32 s14, s11, 31
+; GCN-NEXT:    v_mac_f32_e32 v0, 0x4f800000, v1
+; GCN-NEXT:    v_rcp_f32_e32 v0, v0
 ; GCN-NEXT:    v_mov_b32_e32 v1, 0
-; GCN-NEXT:    s_mov_b32 s4, s8
-; GCN-NEXT:    s_mov_b32 s5, s9
-; GCN-NEXT:    s_mov_b32 s1, s0
-; GCN-NEXT:    s_add_u32 s2, s2, s0
-; GCN-NEXT:    s_addc_u32 s3, s3, s0
-; GCN-NEXT:    s_xor_b64 s[2:3], s[2:3], s[0:1]
-; GCN-NEXT:    v_cvt_f32_u32_e32 v2, s2
-; GCN-NEXT:    v_cvt_f32_u32_e32 v3, s3
-; GCN-NEXT:    s_sub_u32 s14, 0, s2
-; GCN-NEXT:    v_mov_b32_e32 v4, s3
-; GCN-NEXT:    v_mac_f32_e32 v2, 0x4f800000, v3
-; GCN-NEXT:    s_subb_u32 s15, 0, s3
-; GCN-NEXT:    s_ashr_i32 s8, s11, 31
-; GCN-NEXT:    v_rcp_f32_e32 v2, v2
-; GCN-NEXT:    s_mov_b32 s9, s8
-; GCN-NEXT:    s_add_u32 s10, s10, s8
-; GCN-NEXT:    v_mul_f32_e32 v2, 0x5f7ffffc, v2
-; GCN-NEXT:    s_addc_u32 s11, s11, s8
-; GCN-NEXT:    s_xor_b64 s[12:13], s[8:9], s[0:1]
-; GCN-NEXT:    v_mul_f32_e32 v3, 0x2f800000, v2
-; GCN-NEXT:    s_xor_b64 s[8:9], s[10:11], s[8:9]
-; GCN-NEXT:    v_mov_b32_e32 v5, s13
-; GCN-NEXT:    v_trunc_f32_e32 v3, v3
-; GCN-NEXT:    v_mov_b32_e32 v6, s9
-; GCN-NEXT:    v_mac_f32_e32 v2, 0xcf800000, v3
-; GCN-NEXT:    v_cvt_u32_f32_e32 v3, v3
+; GCN-NEXT:    s_mov_b32 s15, s14
+; GCN-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v0
+; GCN-NEXT:    v_mul_f32_e32 v2, 0x2f800000, v0
+; GCN-NEXT:    v_trunc_f32_e32 v2, v2
+; GCN-NEXT:    v_mac_f32_e32 v0, 0xcf800000, v2
+; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v0
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v2, v2
-; GCN-NEXT:    v_mul_lo_u32 v7, s14, v3
-; GCN-NEXT:    v_mul_lo_u32 v8, s15, v2
-; GCN-NEXT:    v_mul_hi_u32 v9, s14, v2
-; GCN-NEXT:    v_mul_lo_u32 v10, s14, v2
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v9, v7
-; GCN-NEXT:    v_mul_hi_u32 v9, v2, v10
-; GCN-NEXT:    v_mul_hi_u32 v11, v3, v10
-; GCN-NEXT:    v_mul_lo_u32 v10, v3, v10
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v7, v8
-; GCN-NEXT:    v_mul_hi_u32 v8, v2, v7
-; GCN-NEXT:    v_mul_lo_u32 v12, v2, v7
-; GCN-NEXT:    v_mul_hi_u32 v13, v3, v7
-; GCN-NEXT:    v_mul_lo_u32 v7, v3, v7
-; GCN-NEXT:    v_add_i32_e32 v9, vcc, v9, v12
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v1, v8, vcc
-; GCN-NEXT:    v_add_i32_e32 v9, vcc, v10, v9
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v8, v11, vcc
-; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v13, v0, vcc
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v8, v7
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v1, v9, vcc
-; GCN-NEXT:    v_add_i32_e64 v2, s[0:1], v2, v7
-; GCN-NEXT:    v_addc_u32_e64 v7, vcc, v3, v8, s[0:1]
-; GCN-NEXT:    v_mul_hi_u32 v9, s14, v2
-; GCN-NEXT:    v_mul_lo_u32 v10, s15, v2
-; GCN-NEXT:    v_mul_lo_u32 v11, s14, v2
-; GCN-NEXT:    v_mul_lo_u32 v12, s14, v7
-; GCN-NEXT:    v_mul_hi_u32 v13, v7, v11
-; GCN-NEXT:    v_mul_lo_u32 v14, v7, v11
-; GCN-NEXT:    v_mul_hi_u32 v11, v2, v11
-; GCN-NEXT:    v_add_i32_e32 v9, vcc, v9, v12
-; GCN-NEXT:    v_add_i32_e32 v9, vcc, v10, v9
-; GCN-NEXT:    v_mul_hi_u32 v10, v7, v9
-; GCN-NEXT:    v_mul_hi_u32 v12, v2, v9
-; GCN-NEXT:    v_mul_lo_u32 v15, v2, v9
-; GCN-NEXT:    v_mul_lo_u32 v7, v7, v9
-; GCN-NEXT:    v_add_i32_e32 v9, vcc, v11, v15
-; GCN-NEXT:    v_addc_u32_e32 v11, vcc, v1, v12, vcc
-; GCN-NEXT:    v_add_i32_e32 v9, vcc, v14, v9
-; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v11, v13, vcc
-; GCN-NEXT:    v_addc_u32_e32 v10, vcc, v10, v0, vcc
-; GCN-NEXT:    v_add_i32_e32 v3, vcc, v3, v8
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v9, v7
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v1, v10, vcc
-; GCN-NEXT:    v_addc_u32_e64 v3, vcc, v3, v8, s[0:1]
-; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v7
-; GCN-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; GCN-NEXT:    v_mul_hi_u32 v7, s8, v2
-; GCN-NEXT:    v_mul_hi_u32 v8, s9, v2
-; GCN-NEXT:    v_mul_lo_u32 v2, s9, v2
-; GCN-NEXT:    v_mul_hi_u32 v9, s8, v3
-; GCN-NEXT:    v_mul_lo_u32 v10, s8, v3
-; GCN-NEXT:    v_mul_hi_u32 v11, s9, v3
-; GCN-NEXT:    v_mul_lo_u32 v3, s9, v3
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v7, v10
-; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v1, v9, vcc
-; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v7
-; GCN-NEXT:    v_addc_u32_e32 v2, vcc, v9, v8, vcc
-; GCN-NEXT:    v_addc_u32_e32 v0, vcc, v11, v0, vcc
-; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v3
-; GCN-NEXT:    v_addc_u32_e32 v0, vcc, v1, v0, vcc
-; GCN-NEXT:    v_mul_hi_u32 v1, s2, v2
-; GCN-NEXT:    v_mul_lo_u32 v3, s3, v2
-; GCN-NEXT:    v_mul_lo_u32 v7, s2, v2
-; GCN-NEXT:    v_mul_lo_u32 v8, s2, v0
-; GCN-NEXT:    v_add_i32_e32 v9, vcc, 2, v2
-; GCN-NEXT:    v_addc_u32_e32 v10, vcc, 0, v0, vcc
-; GCN-NEXT:    v_add_i32_e32 v11, vcc, 1, v2
-; GCN-NEXT:    v_addc_u32_e32 v12, vcc, 0, v0, vcc
-; GCN-NEXT:    v_add_i32_e32 v1, vcc, v1, v8
-; GCN-NEXT:    v_add_i32_e32 v1, vcc, v1, v3
-; GCN-NEXT:    v_sub_i32_e32 v3, vcc, s9, v1
-; GCN-NEXT:    v_sub_i32_e32 v7, vcc, s8, v7
-; GCN-NEXT:    v_subb_u32_e64 v3, s[0:1], v3, v4, vcc
-; GCN-NEXT:    v_subrev_i32_e64 v4, s[0:1], s2, v7
-; GCN-NEXT:    v_subb_u32_e32 v1, vcc, v6, v1, vcc
-; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s2, v7
+; GCN-NEXT:    v_mul_hi_u32 v4, s4, v0
+; GCN-NEXT:    v_mul_lo_u32 v3, s4, v2
+; GCN-NEXT:    v_mul_lo_u32 v6, s5, v0
+; GCN-NEXT:    v_mul_lo_u32 v5, s4, v0
+; GCN-NEXT:    v_add_i32_e32 v3, vcc, v4, v3
+; GCN-NEXT:    v_add_i32_e32 v3, vcc, v3, v6
+; GCN-NEXT:    v_mul_hi_u32 v4, v0, v5
+; GCN-NEXT:    v_mul_lo_u32 v6, v0, v3
+; GCN-NEXT:    v_mul_hi_u32 v8, v0, v3
+; GCN-NEXT:    v_mul_hi_u32 v9, v2, v3
+; GCN-NEXT:    v_mul_lo_u32 v3, v2, v3
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v4, v6
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v7, v8, vcc
+; GCN-NEXT:    v_mul_lo_u32 v8, v2, v5
+; GCN-NEXT:    v_mul_hi_u32 v5, v2, v5
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v8, v4
+; GCN-NEXT:    v_addc_u32_e32 v4, vcc, v6, v5, vcc
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v9, v1, vcc
+; GCN-NEXT:    v_add_i32_e32 v3, vcc, v4, v3
+; GCN-NEXT:    v_add_i32_e64 v0, s[0:1], v0, v3
+; GCN-NEXT:    v_addc_u32_e32 v4, vcc, v7, v5, vcc
+; GCN-NEXT:    v_addc_u32_e64 v3, vcc, v2, v4, s[0:1]
+; GCN-NEXT:    v_mul_lo_u32 v5, s4, v3
+; GCN-NEXT:    v_mul_hi_u32 v6, s4, v0
+; GCN-NEXT:    v_mul_lo_u32 v8, s5, v0
+; GCN-NEXT:    s_mov_b32 s5, s9
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v6, v5
+; GCN-NEXT:    v_mul_lo_u32 v6, s4, v0
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v8, v5
+; GCN-NEXT:    v_mul_lo_u32 v10, v0, v5
+; GCN-NEXT:    v_mul_hi_u32 v12, v0, v5
+; GCN-NEXT:    v_mul_hi_u32 v11, v0, v6
+; GCN-NEXT:    v_mul_hi_u32 v9, v3, v6
+; GCN-NEXT:    v_mul_lo_u32 v6, v3, v6
+; GCN-NEXT:    v_mul_hi_u32 v8, v3, v5
+; GCN-NEXT:    v_add_i32_e32 v10, vcc, v11, v10
+; GCN-NEXT:    v_addc_u32_e32 v11, vcc, v7, v12, vcc
+; GCN-NEXT:    v_mul_lo_u32 v3, v3, v5
+; GCN-NEXT:    v_add_i32_e32 v6, vcc, v6, v10
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v11, v9, vcc
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v8, v1, vcc
+; GCN-NEXT:    v_add_i32_e32 v3, vcc, v6, v3
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v7, v5, vcc
+; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v4
+; GCN-NEXT:    v_addc_u32_e64 v2, vcc, v2, v5, s[0:1]
+; GCN-NEXT:    s_add_u32 s0, s10, s14
+; GCN-NEXT:    s_addc_u32 s1, s11, s14
+; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v3
+; GCN-NEXT:    s_xor_b64 s[10:11], s[0:1], s[14:15]
+; GCN-NEXT:    v_addc_u32_e32 v2, vcc, 0, v2, vcc
+; GCN-NEXT:    v_mul_lo_u32 v3, s10, v2
+; GCN-NEXT:    v_mul_hi_u32 v4, s10, v0
+; GCN-NEXT:    v_mul_hi_u32 v5, s10, v2
+; GCN-NEXT:    v_mul_hi_u32 v6, s11, v2
+; GCN-NEXT:    v_mul_lo_u32 v2, s11, v2
+; GCN-NEXT:    v_add_i32_e32 v3, vcc, v4, v3
+; GCN-NEXT:    v_addc_u32_e32 v4, vcc, v7, v5, vcc
+; GCN-NEXT:    v_mul_lo_u32 v5, s11, v0
+; GCN-NEXT:    v_mul_hi_u32 v0, s11, v0
+; GCN-NEXT:    s_mov_b32 s4, s8
+; GCN-NEXT:    v_add_i32_e32 v3, vcc, v5, v3
+; GCN-NEXT:    v_addc_u32_e32 v0, vcc, v4, v0, vcc
+; GCN-NEXT:    v_addc_u32_e32 v1, vcc, v6, v1, vcc
+; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v2
+; GCN-NEXT:    v_addc_u32_e32 v1, vcc, v7, v1, vcc
+; GCN-NEXT:    v_mul_lo_u32 v2, s2, v1
+; GCN-NEXT:    v_mul_hi_u32 v3, s2, v0
+; GCN-NEXT:    v_mul_lo_u32 v4, s3, v0
+; GCN-NEXT:    v_mov_b32_e32 v5, s3
+; GCN-NEXT:    v_add_i32_e32 v2, vcc, v3, v2
+; GCN-NEXT:    v_mul_lo_u32 v3, s2, v0
+; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v4
+; GCN-NEXT:    v_sub_i32_e32 v4, vcc, s11, v2
+; GCN-NEXT:    v_sub_i32_e32 v3, vcc, s10, v3
+; GCN-NEXT:    v_subb_u32_e64 v4, s[0:1], v4, v5, vcc
+; GCN-NEXT:    v_subrev_i32_e64 v5, s[0:1], s2, v3
+; GCN-NEXT:    v_subbrev_u32_e64 v4, s[0:1], 0, v4, s[0:1]
+; GCN-NEXT:    v_cmp_le_u32_e64 s[0:1], s3, v4
+; GCN-NEXT:    v_cndmask_b32_e64 v6, 0, -1, s[0:1]
+; GCN-NEXT:    v_cmp_le_u32_e64 s[0:1], s2, v5
+; GCN-NEXT:    v_cndmask_b32_e64 v5, 0, -1, s[0:1]
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], s3, v4
+; GCN-NEXT:    v_cndmask_b32_e64 v4, v6, v5, s[0:1]
+; GCN-NEXT:    v_add_i32_e64 v5, s[0:1], 2, v0
+; GCN-NEXT:    v_addc_u32_e64 v6, s[0:1], 0, v1, s[0:1]
+; GCN-NEXT:    v_add_i32_e64 v7, s[0:1], 1, v0
+; GCN-NEXT:    v_addc_u32_e64 v8, s[0:1], 0, v1, s[0:1]
+; GCN-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v4
+; GCN-NEXT:    v_cndmask_b32_e64 v4, v8, v6, s[0:1]
+; GCN-NEXT:    v_mov_b32_e32 v6, s11
+; GCN-NEXT:    v_subb_u32_e32 v2, vcc, v6, v2, vcc
+; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s3, v2
 ; GCN-NEXT:    v_cndmask_b32_e64 v6, 0, -1, vcc
-; GCN-NEXT:    v_subbrev_u32_e64 v3, vcc, 0, v3, s[0:1]
-; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s2, v4
-; GCN-NEXT:    v_cndmask_b32_e64 v4, 0, -1, vcc
-; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s3, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v7, 0, -1, vcc
-; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s3, v3
-; GCN-NEXT:    v_cndmask_b32_e64 v8, 0, -1, vcc
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, s3, v1
-; GCN-NEXT:    v_cndmask_b32_e32 v1, v7, v6, vcc
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, s3, v3
-; GCN-NEXT:    v_cndmask_b32_e32 v3, v8, v4, vcc
-; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v3
-; GCN-NEXT:    v_cndmask_b32_e32 v3, v12, v10, vcc
-; GCN-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v0, v0, v3, s[0:1]
-; GCN-NEXT:    v_cndmask_b32_e32 v1, v11, v9, vcc
-; GCN-NEXT:    v_cndmask_b32_e64 v1, v2, v1, s[0:1]
-; GCN-NEXT:    v_xor_b32_e32 v2, s13, v0
-; GCN-NEXT:    v_xor_b32_e32 v0, s12, v1
-; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, s12, v0
-; GCN-NEXT:    v_subb_u32_e32 v1, vcc, v2, v5, vcc
+; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s2, v3
+; GCN-NEXT:    v_cndmask_b32_e64 v3, 0, -1, vcc
+; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, s3, v2
+; GCN-NEXT:    v_cndmask_b32_e32 v2, v6, v3, vcc
+; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v2, v7, v5, s[0:1]
+; GCN-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
+; GCN-NEXT:    s_xor_b64 s[0:1], s[14:15], s[12:13]
+; GCN-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; GCN-NEXT:    v_xor_b32_e32 v0, s0, v0
+; GCN-NEXT:    v_xor_b32_e32 v1, s1, v1
+; GCN-NEXT:    v_mov_b32_e32 v2, s1
+; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, s0, v0
+; GCN-NEXT:    v_subb_u32_e32 v1, vcc, v1, v2, vcc
 ; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
 ; GCN-NEXT:    s_endpgm
   %result = sdiv i64 %x, %y
   store i64 %result, i64 addrspace(1)* %out
   ret void
 }
 
 define i64 @v_test_sdiv(i64 %x, i64 %y) {
 ; GCN-LABEL: v_test_sdiv:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GCN-NEXT:    v_ashrrev_i32_e32 v4, 31, v3
-; GCN-NEXT:    v_mov_b32_e32 v5, 0
-; GCN-NEXT:    v_mov_b32_e32 v6, 0
-; GCN-NEXT:    v_ashrrev_i32_e32 v7, 31, v1
 ; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v4
 ; GCN-NEXT:    v_addc_u32_e32 v3, vcc, v3, v4, vcc
-; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v7
-; GCN-NEXT:    v_addc_u32_e32 v1, vcc, v1, v7, vcc
-; GCN-NEXT:    v_xor_b32_e32 v8, v7, v4
-; GCN-NEXT:    v_xor_b32_e32 v3, v3, v4
 ; GCN-NEXT:    v_xor_b32_e32 v2, v2, v4
-; GCN-NEXT:    v_xor_b32_e32 v1, v1, v7
-; GCN-NEXT:    v_xor_b32_e32 v0, v0, v7
-; GCN-NEXT:    v_cvt_f32_u32_e32 v4, v2
-; GCN-NEXT:    v_cvt_f32_u32_e32 v7, v3
-; GCN-NEXT:    v_sub_i32_e32 v9, vcc, 0, v2
-; GCN-NEXT:    v_subb_u32_e32 v10, vcc, 0, v3, vcc
-; GCN-NEXT:    v_mac_f32_e32 v4, 0x4f800000, v7
-; GCN-NEXT:    v_rcp_f32_e32 v4, v4
-; GCN-NEXT:    v_mul_f32_e32 v4, 0x5f7ffffc, v4
-; GCN-NEXT:    v_mul_f32_e32 v7, 0x2f800000, v4
-; GCN-NEXT:    v_trunc_f32_e32 v7, v7
-; GCN-NEXT:    v_mac_f32_e32 v4, 0xcf800000, v7
-; GCN-NEXT:    v_cvt_u32_f32_e32 v7, v7
-; GCN-NEXT:    v_cvt_u32_f32_e32 v4, v4
-; GCN-NEXT:    v_mul_lo_u32 v11, v9, v7
-; GCN-NEXT:    v_mul_lo_u32 v12, v10, v4
-; GCN-NEXT:    v_mul_hi_u32 v13, v9, v4
-; GCN-NEXT:    v_mul_lo_u32 v14, v9, v4
-; GCN-NEXT:    v_add_i32_e32 v11, vcc, v13, v11
-; GCN-NEXT:    v_mul_hi_u32 v13, v4, v14
-; GCN-NEXT:    v_mul_hi_u32 v15, v7, v14
-; GCN-NEXT:    v_mul_lo_u32 v14, v7, v14
-; GCN-NEXT:    v_add_i32_e32 v11, vcc, v11, v12
-; GCN-NEXT:    v_mul_hi_u32 v12, v4, v11
-; GCN-NEXT:    v_mul_lo_u32 v16, v4, v11
-; GCN-NEXT:    v_mul_hi_u32 v17, v7, v11
-; GCN-NEXT:    v_mul_lo_u32 v11, v7, v11
-; GCN-NEXT:    v_add_i32_e32 v13, vcc, v13, v16
-; GCN-NEXT:    v_addc_u32_e32 v12, vcc, v6, v12, vcc
-; GCN-NEXT:    v_add_i32_e32 v13, vcc, v14, v13
-; GCN-NEXT:    v_addc_u32_e32 v12, vcc, v12, v15, vcc
-; GCN-NEXT:    v_addc_u32_e32 v13, vcc, v17, v5, vcc
-; GCN-NEXT:    v_add_i32_e32 v11, vcc, v12, v11
-; GCN-NEXT:    v_addc_u32_e32 v12, vcc, v6, v13, vcc
-; GCN-NEXT:    v_add_i32_e64 v4, s[4:5], v4, v11
-; GCN-NEXT:    v_addc_u32_e64 v11, vcc, v7, v12, s[4:5]
-; GCN-NEXT:    v_mul_hi_u32 v13, v9, v4
-; GCN-NEXT:    v_mul_lo_u32 v10, v10, v4
-; GCN-NEXT:    v_mul_lo_u32 v14, v9, v4
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v7, v12
-; GCN-NEXT:    v_mul_lo_u32 v9, v9, v11
-; GCN-NEXT:    v_mul_hi_u32 v12, v11, v14
-; GCN-NEXT:    v_mul_lo_u32 v15, v11, v14
-; GCN-NEXT:    v_mul_hi_u32 v14, v4, v14
-; GCN-NEXT:    v_add_i32_e32 v9, vcc, v13, v9
+; GCN-NEXT:    v_xor_b32_e32 v3, v3, v4
+; GCN-NEXT:    v_cvt_f32_u32_e32 v5, v2
+; GCN-NEXT:    v_cvt_f32_u32_e32 v6, v3
+; GCN-NEXT:    v_sub_i32_e32 v7, vcc, 0, v2
+; GCN-NEXT:    v_subb_u32_e32 v8, vcc, 0, v3, vcc
+; GCN-NEXT:    v_mov_b32_e32 v15, 0
+; GCN-NEXT:    v_mac_f32_e32 v5, 0x4f800000, v6
+; GCN-NEXT:    v_rcp_f32_e32 v5, v5
+; GCN-NEXT:    v_mov_b32_e32 v14, 0
+; GCN-NEXT:    v_mul_f32_e32 v5, 0x5f7ffffc, v5
+; GCN-NEXT:    v_mul_f32_e32 v6, 0x2f800000, v5
+; GCN-NEXT:    v_trunc_f32_e32 v6, v6
+; GCN-NEXT:    v_mac_f32_e32 v5, 0xcf800000, v6
+; GCN-NEXT:    v_cvt_u32_f32_e32 v5, v5
+; GCN-NEXT:    v_cvt_u32_f32_e32 v6, v6
+; GCN-NEXT:    v_mul_hi_u32 v9, v7, v5
+; GCN-NEXT:    v_mul_lo_u32 v10, v7, v6
+; GCN-NEXT:    v_mul_lo_u32 v11, v8, v5
 ; GCN-NEXT:    v_add_i32_e32 v9, vcc, v9, v10
-; GCN-NEXT:    v_mul_hi_u32 v10, v11, v9
-; GCN-NEXT:    v_mul_hi_u32 v13, v4, v9
-; GCN-NEXT:    v_mul_lo_u32 v16, v4, v9
-; GCN-NEXT:    v_mul_lo_u32 v9, v11, v9
-; GCN-NEXT:    v_add_i32_e32 v11, vcc, v14, v16
-; GCN-NEXT:    v_addc_u32_e32 v13, vcc, v6, v13, vcc
-; GCN-NEXT:    v_add_i32_e32 v11, vcc, v15, v11
-; GCN-NEXT:    v_addc_u32_e32 v11, vcc, v13, v12, vcc
-; GCN-NEXT:    v_addc_u32_e32 v10, vcc, v10, v5, vcc
-; GCN-NEXT:    v_add_i32_e32 v9, vcc, v11, v9
-; GCN-NEXT:    v_addc_u32_e32 v10, vcc, v6, v10, vcc
-; GCN-NEXT:    v_addc_u32_e64 v7, vcc, v7, v10, s[4:5]
-; GCN-NEXT:    v_add_i32_e32 v4, vcc, v4, v9
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
-; GCN-NEXT:    v_mul_hi_u32 v9, v0, v4
-; GCN-NEXT:    v_mul_hi_u32 v10, v1, v4
-; GCN-NEXT:    v_mul_lo_u32 v4, v1, v4
-; GCN-NEXT:    v_mul_hi_u32 v11, v0, v7
-; GCN-NEXT:    v_mul_lo_u32 v12, v0, v7
-; GCN-NEXT:    v_mul_hi_u32 v13, v1, v7
-; GCN-NEXT:    v_mul_lo_u32 v7, v1, v7
-; GCN-NEXT:    v_add_i32_e32 v9, vcc, v9, v12
-; GCN-NEXT:    v_addc_u32_e32 v11, vcc, v6, v11, vcc
-; GCN-NEXT:    v_add_i32_e32 v4, vcc, v4, v9
-; GCN-NEXT:    v_addc_u32_e32 v4, vcc, v11, v10, vcc
-; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v13, v5, vcc
-; GCN-NEXT:    v_add_i32_e32 v4, vcc, v4, v7
-; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v6, v5, vcc
-; GCN-NEXT:    v_mul_hi_u32 v6, v2, v4
-; GCN-NEXT:    v_mul_lo_u32 v7, v3, v4
-; GCN-NEXT:    v_mul_lo_u32 v9, v2, v4
-; GCN-NEXT:    v_mul_lo_u32 v10, v2, v5
-; GCN-NEXT:    v_add_i32_e32 v11, vcc, 2, v4
-; GCN-NEXT:    v_addc_u32_e32 v12, vcc, 0, v5, vcc
-; GCN-NEXT:    v_add_i32_e32 v13, vcc, 1, v4
-; GCN-NEXT:    v_addc_u32_e32 v14, vcc, 0, v5, vcc
+; GCN-NEXT:    v_mul_lo_u32 v10, v7, v5
+; GCN-NEXT:    v_add_i32_e32 v9, vcc, v9, v11
+; GCN-NEXT:    v_mul_lo_u32 v12, v5, v9
+; GCN-NEXT:    v_mul_hi_u32 v11, v5, v9
+; GCN-NEXT:    v_mul_hi_u32 v13, v5, v10
+; GCN-NEXT:    v_mul_hi_u32 v16, v6, v9
+; GCN-NEXT:    v_mul_lo_u32 v9, v6, v9
+; GCN-NEXT:    v_add_i32_e32 v12, vcc, v13, v12
+; GCN-NEXT:    v_mul_lo_u32 v13, v6, v10
+; GCN-NEXT:    v_mul_hi_u32 v10, v6, v10
+; GCN-NEXT:    v_addc_u32_e32 v11, vcc, v15, v11, vcc
+; GCN-NEXT:    v_add_i32_e32 v12, vcc, v13, v12
+; GCN-NEXT:    v_addc_u32_e32 v10, vcc, v11, v10, vcc
+; GCN-NEXT:    v_addc_u32_e32 v11, vcc, v16, v14, vcc
+; GCN-NEXT:    v_add_i32_e32 v9, vcc, v10, v9
+; GCN-NEXT:    v_add_i32_e64 v5, s[4:5], v5, v9
+; GCN-NEXT:    v_addc_u32_e32 v10, vcc, v15, v11, vcc
+; GCN-NEXT:    v_addc_u32_e64 v9, vcc, v6, v10, s[4:5]
+; GCN-NEXT:    v_mul_lo_u32 v11, v7, v9
+; GCN-NEXT:    v_mul_hi_u32 v12, v7, v5
+; GCN-NEXT:    v_mul_lo_u32 v8, v8, v5
+; GCN-NEXT:    v_mul_lo_u32 v7, v7, v5
+; GCN-NEXT:    v_add_i32_e32 v11, vcc, v12, v11
+; GCN-NEXT:    v_add_i32_e32 v8, vcc, v11, v8
+; GCN-NEXT:    v_mul_lo_u32 v13, v5, v8
+; GCN-NEXT:    v_mul_hi_u32 v16, v5, v7
+; GCN-NEXT:    v_mul_hi_u32 v17, v5, v8
+; GCN-NEXT:    v_mul_hi_u32 v12, v9, v7
+; GCN-NEXT:    v_mul_lo_u32 v7, v9, v7
+; GCN-NEXT:    v_add_i32_e32 v13, vcc, v16, v13
+; GCN-NEXT:    v_mul_hi_u32 v11, v9, v8
+; GCN-NEXT:    v_addc_u32_e32 v16, vcc, v15, v17, vcc
+; GCN-NEXT:    v_mul_lo_u32 v8, v9, v8
+; GCN-NEXT:    v_add_i32_e32 v7, vcc, v7, v13
+; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v16, v12, vcc
+; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v11, v14, vcc
+; GCN-NEXT:    v_add_i32_e32 v7, vcc, v7, v8
+; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v15, v9, vcc
 ; GCN-NEXT:    v_add_i32_e32 v6, vcc, v6, v10
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, v6, v7
-; GCN-NEXT:    v_sub_i32_e32 v7, vcc, v1, v6
+; GCN-NEXT:    v_addc_u32_e64 v6, vcc, v6, v8, s[4:5]
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v7
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, 0, v6, vcc
+; GCN-NEXT:    v_ashrrev_i32_e32 v7, 31, v1
+; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v7
+; GCN-NEXT:    v_xor_b32_e32 v0, v0, v7
+; GCN-NEXT:    v_mul_lo_u32 v8, v0, v6
+; GCN-NEXT:    v_mul_hi_u32 v9, v0, v5
+; GCN-NEXT:    v_mul_hi_u32 v10, v0, v6
+; GCN-NEXT:    v_addc_u32_e32 v1, vcc, v1, v7, vcc
+; GCN-NEXT:    v_xor_b32_e32 v1, v1, v7
+; GCN-NEXT:    v_add_i32_e32 v8, vcc, v9, v8
+; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v15, v10, vcc
+; GCN-NEXT:    v_mul_lo_u32 v10, v1, v5
+; GCN-NEXT:    v_mul_hi_u32 v5, v1, v5
+; GCN-NEXT:    v_mul_hi_u32 v11, v1, v6
+; GCN-NEXT:    v_mul_lo_u32 v6, v1, v6
+; GCN-NEXT:    v_add_i32_e32 v8, vcc, v10, v8
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v9, v5, vcc
+; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v11, v14, vcc
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v6
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v15, v8, vcc
+; GCN-NEXT:    v_mul_lo_u32 v8, v2, v6
+; GCN-NEXT:    v_mul_hi_u32 v9, v2, v5
+; GCN-NEXT:    v_mul_lo_u32 v10, v3, v5
+; GCN-NEXT:    v_add_i32_e32 v8, vcc, v9, v8
+; GCN-NEXT:    v_mul_lo_u32 v9, v2, v5
+; GCN-NEXT:    v_add_i32_e32 v8, vcc, v8, v10
+; GCN-NEXT:    v_sub_i32_e32 v10, vcc, v1, v8
 ; GCN-NEXT:    v_sub_i32_e32 v0, vcc, v0, v9
-; GCN-NEXT:    v_subb_u32_e64 v7, s[4:5], v7, v3, vcc
-; GCN-NEXT:    v_sub_i32_e64 v9, s[4:5], v0, v2
-; GCN-NEXT:    v_subb_u32_e32 v1, vcc, v1, v6, vcc
+; GCN-NEXT:    v_subb_u32_e64 v9, s[4:5], v10, v3, vcc
+; GCN-NEXT:    v_sub_i32_e64 v10, s[4:5], v0, v2
+; GCN-NEXT:    v_subbrev_u32_e64 v9, s[4:5], 0, v9, s[4:5]
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[4:5], v9, v3
+; GCN-NEXT:    v_cndmask_b32_e64 v11, 0, -1, s[4:5]
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[4:5], v10, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v10, 0, -1, s[4:5]
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[4:5], v9, v3
+; GCN-NEXT:    v_subb_u32_e32 v1, vcc, v1, v8, vcc
+; GCN-NEXT:    v_cndmask_b32_e64 v9, v11, v10, s[4:5]
+; GCN-NEXT:    v_add_i32_e64 v10, s[4:5], 2, v5
+; GCN-NEXT:    v_addc_u32_e64 v11, s[4:5], 0, v6, s[4:5]
+; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v1, v3
+; GCN-NEXT:    v_add_i32_e64 v12, s[4:5], 1, v5
+; GCN-NEXT:    v_cndmask_b32_e64 v8, 0, -1, vcc
 ; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v0, v2
+; GCN-NEXT:    v_addc_u32_e64 v13, s[4:5], 0, v6, s[4:5]
 ; GCN-NEXT:    v_cndmask_b32_e64 v0, 0, -1, vcc
-; GCN-NEXT:    v_subbrev_u32_e64 v6, vcc, 0, v7, s[4:5]
-; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v9, v2
-; GCN-NEXT:    v_cndmask_b32_e64 v2, 0, -1, vcc
-; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v1, v3
-; GCN-NEXT:    v_cndmask_b32_e64 v7, 0, -1, vcc
-; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v6, v3
-; GCN-NEXT:    v_cndmask_b32_e64 v9, 0, -1, vcc
 ; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, v1, v3
-; GCN-NEXT:    v_cndmask_b32_e32 v0, v7, v0, vcc
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, v6, v3
-; GCN-NEXT:    v_cndmask_b32_e32 v1, v9, v2, vcc
-; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v1
-; GCN-NEXT:    v_cndmask_b32_e32 v1, v14, v12, vcc
-; GCN-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, v0
-; GCN-NEXT:    v_cndmask_b32_e64 v0, v5, v1, s[4:5]
-; GCN-NEXT:    v_cndmask_b32_e32 v1, v13, v11, vcc
-; GCN-NEXT:    v_cndmask_b32_e64 v1, v4, v1, s[4:5]
-; GCN-NEXT:    v_xor_b32_e32 v2, v0, v8
-; GCN-NEXT:    v_xor_b32_e32 v0, v1, v8
-; GCN-NEXT:    v_sub_i32_e32 v0, vcc, v0, v8
-; GCN-NEXT:    v_subb_u32_e32 v1, vcc, v2, v8, vcc
+; GCN-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, v9
+; GCN-NEXT:    v_cndmask_b32_e32 v0, v8, v0, vcc
+; GCN-NEXT:    v_cndmask_b32_e64 v9, v13, v11, s[4:5]
+; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
+; GCN-NEXT:    v_cndmask_b32_e64 v1, v12, v10, s[4:5]
+; GCN-NEXT:    v_cndmask_b32_e32 v0, v6, v9, vcc
+; GCN-NEXT:    v_xor_b32_e32 v2, v7, v4
+; GCN-NEXT:    v_cndmask_b32_e32 v1, v5, v1, vcc
+; GCN-NEXT:    v_xor_b32_e32 v3, v0, v2
+; GCN-NEXT:    v_xor_b32_e32 v0, v1, v2
+; GCN-NEXT:    v_sub_i32_e32 v0, vcc, v0, v2
+; GCN-NEXT:    v_subb_u32_e32 v1, vcc, v3, v2, vcc
 ; GCN-NEXT:    s_setpc_b64 s[30:31]
   %result = sdiv i64 %x, %y
   ret i64 %result
 }
 
 define amdgpu_kernel void @s_test_sdiv24_64(i64 addrspace(1)* %out, i64 %x, i64 %y) {
 ; GCN-LABEL: s_test_sdiv24_64:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; GCN-NEXT:    s_load_dword s9, s[0:1], 0xe
+; GCN-NEXT:    s_load_dword s1, s[0:1], 0xe
 ; GCN-NEXT:    s_mov_b32 s3, 0xf000
 ; GCN-NEXT:    s_mov_b32 s2, -1
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN-NEXT:    s_mov_b32 s0, s4
+; GCN-NEXT:    s_ashr_i64 s[8:9], s[0:1], 40
+; GCN-NEXT:    v_cvt_f32_i32_e32 v0, s8
 ; GCN-NEXT:    s_mov_b32 s1, s5
 ; GCN-NEXT:    s_ashr_i64 s[4:5], s[6:7], 40
-; GCN-NEXT:    s_ashr_i64 s[6:7], s[8:9], 40
-; GCN-NEXT:    s_xor_b32 s5, s4, s6
-; GCN-NEXT:    v_cvt_f32_i32_e32 v0, s6
 ; GCN-NEXT:    v_cvt_f32_i32_e32 v1, s4
-; GCN-NEXT:    s_ashr_i32 s4, s5, 30
 ; GCN-NEXT:    v_rcp_iflag_f32_e32 v2, v0
+; GCN-NEXT:    s_xor_b32 s4, s4, s8
+; GCN-NEXT:    s_ashr_i32 s4, s4, 30
 ; GCN-NEXT:    s_or_b32 s4, s4, 1
 ; GCN-NEXT:    v_mul_f32_e32 v2, v1, v2
 ; GCN-NEXT:    v_trunc_f32_e32 v2, v2
-; GCN-NEXT:    v_mov_b32_e32 v3, s4
 ; GCN-NEXT:    v_mad_f32 v1, -v2, v0, v1
+; GCN-NEXT:    v_cvt_i32_f32_e32 v2, v2
+; GCN-NEXT:    v_mov_b32_e32 v3, s4
 ; GCN-NEXT:    v_cmp_ge_f32_e64 vcc, |v1|, |v0|
 ; GCN-NEXT:    v_cndmask_b32_e32 v0, 0, v3, vcc
-; GCN-NEXT:    v_cvt_i32_f32_e32 v1, v2
-; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v1
+; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v2
 ; GCN-NEXT:    v_bfe_i32 v0, v0, 0, 24
 ; GCN-NEXT:    v_ashrrev_i32_e32 v1, 31, v0
 ; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 ; GCN-NEXT:    s_endpgm
   %1 = ashr i64 %x, 40
   %2 = ashr i64 %y, 40
   %result = sdiv i64 %1, %2
   store i64 %result, i64 addrspace(1)* %out
   ret void
 }
 
 define i64 @v_test_sdiv24_64(i64 %x, i64 %y) {
 ; GCN-LABEL: v_test_sdiv24_64:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT:    v_lshrrev_b32_e32 v0, 8, v1
-; GCN-NEXT:    v_lshrrev_b32_e32 v1, 8, v3
-; GCN-NEXT:    v_cvt_f32_u32_e32 v2, v1
+; GCN-NEXT:    v_lshrrev_b32_e32 v0, 8, v3
+; GCN-NEXT:    v_cvt_f32_u32_e32 v2, v0
+; GCN-NEXT:    v_lshrrev_b32_e32 v1, 8, v1
 ; GCN-NEXT:    v_rcp_iflag_f32_e32 v2, v2
 ; GCN-NEXT:    v_mul_f32_e32 v2, 0x4f800000, v2
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v2, v2
-; GCN-NEXT:    v_mul_hi_u32 v3, v2, v1
-; GCN-NEXT:    v_mul_lo_u32 v4, v2, v1
-; GCN-NEXT:    v_sub_i32_e32 v5, vcc, 0, v4
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v3
-; GCN-NEXT:    v_cndmask_b32_e32 v3, v4, v5, vcc
+; GCN-NEXT:    v_mul_lo_u32 v3, v2, v0
+; GCN-NEXT:    v_mul_hi_u32 v4, v2, v0
+; GCN-NEXT:    v_sub_i32_e32 v5, vcc, 0, v3
+; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v4
+; GCN-NEXT:    v_cndmask_b32_e32 v3, v3, v5, vcc
 ; GCN-NEXT:    v_mul_hi_u32 v3, v3, v2
 ; GCN-NEXT:    v_add_i32_e64 v4, s[4:5], v2, v3
 ; GCN-NEXT:    v_sub_i32_e64 v2, s[4:5], v2, v3
 ; GCN-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
-; GCN-NEXT:    v_mul_hi_u32 v2, v2, v0
-; GCN-NEXT:    v_mul_lo_u32 v3, v2, v1
+; GCN-NEXT:    v_mul_hi_u32 v2, v2, v1
+; GCN-NEXT:    v_mul_lo_u32 v3, v2, v0
 ; GCN-NEXT:    v_add_i32_e32 v4, vcc, 1, v2
 ; GCN-NEXT:    v_add_i32_e32 v5, vcc, -1, v2
-; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v0, v3
-; GCN-NEXT:    v_sub_i32_e64 v0, s[4:5], v0, v3
-; GCN-NEXT:    v_cmp_ge_u32_e64 s[4:5], v0, v1
+; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v1, v3
+; GCN-NEXT:    v_sub_i32_e64 v1, s[4:5], v1, v3
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[4:5], v1, v0
 ; GCN-NEXT:    s_and_b64 s[4:5], s[4:5], vcc
 ; GCN-NEXT:    v_cndmask_b32_e64 v0, v2, v4, s[4:5]
 ; GCN-NEXT:    v_cndmask_b32_e32 v0, v5, v0, vcc
 ; GCN-NEXT:    v_mov_b32_e32 v1, 0
 ; GCN-NEXT:    s_setpc_b64 s[30:31]
   %1 = lshr i64 %x, 40
   %2 = lshr i64 %y, 40
   %result = sdiv i64 %1, %2
   ret i64 %result
 }
 
 define amdgpu_kernel void @s_test_sdiv32_64(i64 addrspace(1)* %out, i64 %x, i64 %y) {
 ; GCN-LABEL: s_test_sdiv32_64:
 ; GCN:       ; %bb.0:
-; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GCN-NEXT:    s_load_dword s2, s[0:1], 0xe
+; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GCN-NEXT:    s_mov_b32 s11, 0xf000
 ; GCN-NEXT:    s_mov_b32 s10, -1
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_mov_b32 s8, s4
-; GCN-NEXT:    s_mov_b32 s9, s5
 ; GCN-NEXT:    s_ashr_i32 s3, s2, 31
-; GCN-NEXT:    s_ashr_i32 s4, s7, 31
 ; GCN-NEXT:    s_add_i32 s2, s2, s3
-; GCN-NEXT:    s_add_i32 s0, s7, s4
-; GCN-NEXT:    s_xor_b32 s5, s2, s3
-; GCN-NEXT:    s_xor_b32 s2, s0, s4
-; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s5
+; GCN-NEXT:    s_xor_b32 s12, s2, s3
+; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s12
+; GCN-NEXT:    s_ashr_i32 s2, s7, 31
+; GCN-NEXT:    s_add_i32 s6, s7, s2
+; GCN-NEXT:    s_xor_b32 s6, s6, s2
 ; GCN-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GCN-NEXT:    s_xor_b32 s2, s2, s3
+; GCN-NEXT:    s_mov_b32 s8, s4
+; GCN-NEXT:    s_mov_b32 s9, s5
 ; GCN-NEXT:    v_mul_f32_e32 v0, 0x4f800000, v0
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v0
-; GCN-NEXT:    v_mul_hi_u32 v1, v0, s5
-; GCN-NEXT:    v_mul_lo_u32 v2, v0, s5
-; GCN-NEXT:    v_sub_i32_e32 v3, vcc, 0, v2
-; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v1, v2, v3, s[0:1]
+; GCN-NEXT:    v_mul_lo_u32 v1, v0, s12
+; GCN-NEXT:    v_mul_hi_u32 v2, v0, s12
+; GCN-NEXT:    v_sub_i32_e32 v3, vcc, 0, v1
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v1, v1, v3, s[0:1]
 ; GCN-NEXT:    v_mul_hi_u32 v1, v1, v0
 ; GCN-NEXT:    v_add_i32_e32 v2, vcc, v1, v0
 ; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, v1, v0
 ; GCN-NEXT:    v_cndmask_b32_e64 v0, v0, v2, s[0:1]
-; GCN-NEXT:    v_mul_hi_u32 v0, v0, s2
-; GCN-NEXT:    v_mul_lo_u32 v1, v0, s5
+; GCN-NEXT:    v_mul_hi_u32 v0, v0, s6
+; GCN-NEXT:    v_mul_lo_u32 v1, v0, s12
 ; GCN-NEXT:    v_add_i32_e32 v2, vcc, 1, v0
 ; GCN-NEXT:    v_add_i32_e32 v3, vcc, -1, v0
-; GCN-NEXT:    v_cmp_ge_u32_e64 s[0:1], s2, v1
-; GCN-NEXT:    v_sub_i32_e32 v1, vcc, s2, v1
-; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s5, v1
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[0:1], s6, v1
+; GCN-NEXT:    v_sub_i32_e32 v1, vcc, s6, v1
+; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s12, v1
 ; GCN-NEXT:    s_and_b64 vcc, vcc, s[0:1]
 ; GCN-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
 ; GCN-NEXT:    v_cndmask_b32_e64 v0, v3, v0, s[0:1]
-; GCN-NEXT:    s_xor_b32 s0, s4, s3
-; GCN-NEXT:    v_xor_b32_e32 v0, s0, v0
-; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, s0, v0
+; GCN-NEXT:    v_xor_b32_e32 v0, s2, v0
+; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, s2, v0
 ; GCN-NEXT:    v_ashrrev_i32_e32 v1, 31, v0
 ; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[8:11], 0
 ; GCN-NEXT:    s_endpgm
   %1 = ashr i64 %x, 32
   %2 = ashr i64 %y, 32
   %result = sdiv i64 %1, %2
   store i64 %result, i64 addrspace(1)* %out
   ret void
 }
 
 define amdgpu_kernel void @s_test_sdiv31_64(i64 addrspace(1)* %out, i64 %x, i64 %y) {
 ; GCN-LABEL: s_test_sdiv31_64:
 ; GCN:       ; %bb.0:
-; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GCN-NEXT:    s_load_dword s3, s[0:1], 0xe
+; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GCN-NEXT:    s_mov_b32 s11, 0xf000
 ; GCN-NEXT:    s_mov_b32 s10, -1
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_mov_b32 s8, s4
-; GCN-NEXT:    s_mov_b32 s9, s5
-; GCN-NEXT:    s_ashr_i64 s[0:1], s[6:7], 33
 ; GCN-NEXT:    s_ashr_i64 s[2:3], s[2:3], 33
+; GCN-NEXT:    s_ashr_i32 s12, s2, 31
+; GCN-NEXT:    s_add_i32 s2, s2, s12
+; GCN-NEXT:    s_xor_b32 s13, s2, s12
+; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s13
+; GCN-NEXT:    s_ashr_i64 s[2:3], s[6:7], 33
 ; GCN-NEXT:    s_ashr_i32 s3, s2, 31
-; GCN-NEXT:    s_ashr_i32 s4, s0, 31
-; GCN-NEXT:    s_add_i32 s1, s2, s3
-; GCN-NEXT:    s_add_i32 s0, s0, s4
-; GCN-NEXT:    s_xor_b32 s5, s1, s3
-; GCN-NEXT:    s_xor_b32 s2, s0, s4
-; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s5
+; GCN-NEXT:    s_add_i32 s2, s2, s3
 ; GCN-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GCN-NEXT:    s_xor_b32 s2, s2, s3
+; GCN-NEXT:    s_xor_b32 s3, s3, s12
+; GCN-NEXT:    s_mov_b32 s8, s4
 ; GCN-NEXT:    v_mul_f32_e32 v0, 0x4f800000, v0
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v0
-; GCN-NEXT:    v_mul_hi_u32 v1, v0, s5
-; GCN-NEXT:    v_mul_lo_u32 v2, v0, s5
-; GCN-NEXT:    v_sub_i32_e32 v3, vcc, 0, v2
-; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v1, v2, v3, s[0:1]
+; GCN-NEXT:    s_mov_b32 s9, s5
+; GCN-NEXT:    v_mul_lo_u32 v1, v0, s13
+; GCN-NEXT:    v_mul_hi_u32 v2, v0, s13
+; GCN-NEXT:    v_sub_i32_e32 v3, vcc, 0, v1
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v1, v1, v3, s[0:1]
 ; GCN-NEXT:    v_mul_hi_u32 v1, v1, v0
 ; GCN-NEXT:    v_add_i32_e32 v2, vcc, v1, v0
 ; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, v1, v0
 ; GCN-NEXT:    v_cndmask_b32_e64 v0, v0, v2, s[0:1]
 ; GCN-NEXT:    v_mul_hi_u32 v0, v0, s2
-; GCN-NEXT:    v_mul_lo_u32 v1, v0, s5
+; GCN-NEXT:    v_mul_lo_u32 v1, v0, s13
 ; GCN-NEXT:    v_add_i32_e32 v2, vcc, 1, v0
 ; GCN-NEXT:    v_add_i32_e32 v3, vcc, -1, v0
 ; GCN-NEXT:    v_cmp_ge_u32_e64 s[0:1], s2, v1
 ; GCN-NEXT:    v_sub_i32_e32 v1, vcc, s2, v1
-; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s5, v1
+; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s13, v1
 ; GCN-NEXT:    s_and_b64 vcc, vcc, s[0:1]
 ; GCN-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
 ; GCN-NEXT:    v_cndmask_b32_e64 v0, v3, v0, s[0:1]
-; GCN-NEXT:    s_xor_b32 s0, s4, s3
-; GCN-NEXT:    v_xor_b32_e32 v0, s0, v0
-; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, s0, v0
+; GCN-NEXT:    v_xor_b32_e32 v0, s3, v0
+; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, s3, v0
 ; GCN-NEXT:    v_ashrrev_i32_e32 v1, 31, v0
 ; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[8:11], 0
 ; GCN-NEXT:    s_endpgm
   %1 = ashr i64 %x, 33
   %2 = ashr i64 %y, 33
   %result = sdiv i64 %1, %2
   store i64 %result, i64 addrspace(1)* %out
   ret void
 }
 
 define amdgpu_kernel void @s_test_sdiv23_64(i64 addrspace(1)* %out, i64 %x, i64 %y) {
 ; GCN-LABEL: s_test_sdiv23_64:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; GCN-NEXT:    s_load_dword s9, s[0:1], 0xe
+; GCN-NEXT:    s_load_dword s1, s[0:1], 0xe
 ; GCN-NEXT:    s_mov_b32 s3, 0xf000
 ; GCN-NEXT:    s_mov_b32 s2, -1
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN-NEXT:    s_mov_b32 s0, s4
+; GCN-NEXT:    s_ashr_i64 s[8:9], s[0:1], 41
+; GCN-NEXT:    v_cvt_f32_i32_e32 v0, s8
 ; GCN-NEXT:    s_mov_b32 s1, s5
 ; GCN-NEXT:    s_ashr_i64 s[4:5], s[6:7], 41
-; GCN-NEXT:    s_ashr_i64 s[6:7], s[8:9], 41
-; GCN-NEXT:    s_xor_b32 s5, s4, s6
-; GCN-NEXT:    v_cvt_f32_i32_e32 v0, s6
 ; GCN-NEXT:    v_cvt_f32_i32_e32 v1, s4
-; GCN-NEXT:    s_ashr_i32 s4, s5, 30
 ; GCN-NEXT:    v_rcp_iflag_f32_e32 v2, v0
+; GCN-NEXT:    s_xor_b32 s4, s4, s8
+; GCN-NEXT:    s_ashr_i32 s4, s4, 30
 ; GCN-NEXT:    s_or_b32 s4, s4, 1
 ; GCN-NEXT:    v_mul_f32_e32 v2, v1, v2
 ; GCN-NEXT:    v_trunc_f32_e32 v2, v2
-; GCN-NEXT:    v_mov_b32_e32 v3, s4
 ; GCN-NEXT:    v_mad_f32 v1, -v2, v0, v1
+; GCN-NEXT:    v_cvt_i32_f32_e32 v2, v2
+; GCN-NEXT:    v_mov_b32_e32 v3, s4
 ; GCN-NEXT:    v_cmp_ge_f32_e64 vcc, |v1|, |v0|
 ; GCN-NEXT:    v_cndmask_b32_e32 v0, 0, v3, vcc
-; GCN-NEXT:    v_cvt_i32_f32_e32 v1, v2
-; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v1
+; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v2
 ; GCN-NEXT:    v_bfe_i32 v0, v0, 0, 23
 ; GCN-NEXT:    v_ashrrev_i32_e32 v1, 31, v0
 ; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 ; GCN-NEXT:    s_endpgm
   %1 = ashr i64 %x, 41
   %2 = ashr i64 %y, 41
   %result = sdiv i64 %1, %2
   store i64 %result, i64 addrspace(1)* %out
   ret void
 }
 
 define amdgpu_kernel void @s_test_sdiv25_64(i64 addrspace(1)* %out, i64 %x, i64 %y) {
 ; GCN-LABEL: s_test_sdiv25_64:
 ; GCN:       ; %bb.0:
-; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GCN-NEXT:    s_load_dword s3, s[0:1], 0xe
+; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GCN-NEXT:    s_mov_b32 s11, 0xf000
 ; GCN-NEXT:    s_mov_b32 s10, -1
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_mov_b32 s8, s4
-; GCN-NEXT:    s_mov_b32 s9, s5
-; GCN-NEXT:    s_ashr_i64 s[0:1], s[6:7], 39
 ; GCN-NEXT:    s_ashr_i64 s[2:3], s[2:3], 39
+; GCN-NEXT:    s_ashr_i32 s12, s2, 31
+; GCN-NEXT:    s_add_i32 s2, s2, s12
+; GCN-NEXT:    s_xor_b32 s13, s2, s12
+; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s13
+; GCN-NEXT:    s_ashr_i64 s[2:3], s[6:7], 39
 ; GCN-NEXT:    s_ashr_i32 s3, s2, 31
-; GCN-NEXT:    s_ashr_i32 s4, s0, 31
-; GCN-NEXT:    s_add_i32 s1, s2, s3
-; GCN-NEXT:    s_add_i32 s0, s0, s4
-; GCN-NEXT:    s_xor_b32 s5, s1, s3
-; GCN-NEXT:    s_xor_b32 s2, s0, s4
-; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s5
+; GCN-NEXT:    s_add_i32 s2, s2, s3
 ; GCN-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GCN-NEXT:    s_xor_b32 s2, s2, s3
+; GCN-NEXT:    s_xor_b32 s3, s3, s12
+; GCN-NEXT:    s_mov_b32 s8, s4
 ; GCN-NEXT:    v_mul_f32_e32 v0, 0x4f800000, v0
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v0
-; GCN-NEXT:    v_mul_hi_u32 v1, v0, s5
-; GCN-NEXT:    v_mul_lo_u32 v2, v0, s5
-; GCN-NEXT:    v_sub_i32_e32 v3, vcc, 0, v2
-; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v1, v2, v3, s[0:1]
+; GCN-NEXT:    s_mov_b32 s9, s5
+; GCN-NEXT:    v_mul_lo_u32 v1, v0, s13
+; GCN-NEXT:    v_mul_hi_u32 v2, v0, s13
+; GCN-NEXT:    v_sub_i32_e32 v3, vcc, 0, v1
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v1, v1, v3, s[0:1]
 ; GCN-NEXT:    v_mul_hi_u32 v1, v1, v0
 ; GCN-NEXT:    v_add_i32_e32 v2, vcc, v1, v0
 ; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, v1, v0
 ; GCN-NEXT:    v_cndmask_b32_e64 v0, v0, v2, s[0:1]
 ; GCN-NEXT:    v_mul_hi_u32 v0, v0, s2
-; GCN-NEXT:    v_mul_lo_u32 v1, v0, s5
+; GCN-NEXT:    v_mul_lo_u32 v1, v0, s13
 ; GCN-NEXT:    v_add_i32_e32 v2, vcc, 1, v0
 ; GCN-NEXT:    v_add_i32_e32 v3, vcc, -1, v0
 ; GCN-NEXT:    v_cmp_ge_u32_e64 s[0:1], s2, v1
 ; GCN-NEXT:    v_sub_i32_e32 v1, vcc, s2, v1
-; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s5, v1
+; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s13, v1
 ; GCN-NEXT:    s_and_b64 vcc, vcc, s[0:1]
 ; GCN-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
 ; GCN-NEXT:    v_cndmask_b32_e64 v0, v3, v0, s[0:1]
-; GCN-NEXT:    s_xor_b32 s0, s4, s3
-; GCN-NEXT:    v_xor_b32_e32 v0, s0, v0
-; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, s0, v0
+; GCN-NEXT:    v_xor_b32_e32 v0, s3, v0
+; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, s3, v0
 ; GCN-NEXT:    v_ashrrev_i32_e32 v1, 31, v0
 ; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[8:11], 0
 ; GCN-NEXT:    s_endpgm
   %1 = ashr i64 %x, 39
   %2 = ashr i64 %y, 39
   %result = sdiv i64 %1, %2
   store i64 %result, i64 addrspace(1)* %out
   ret void
 }
 
 define amdgpu_kernel void @s_test_sdiv24_v2i64(<2 x i64> addrspace(1)* %out, <2 x i64> %x, <2 x i64> %y) {
 ; GCN-LABEL: s_test_sdiv24_v2i64:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
 ; GCN-NEXT:    s_load_dwordx4 s[8:11], s[0:1], 0xd
-; GCN-NEXT:    s_load_dwordx4 s[12:15], s[0:1], 0x11
+; GCN-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x11
 ; GCN-NEXT:    s_mov_b32 s7, 0xf000
 ; GCN-NEXT:    s_mov_b32 s6, -1
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_ashr_i64 s[0:1], s[8:9], 40
-; GCN-NEXT:    s_ashr_i64 s[2:3], s[10:11], 40
-; GCN-NEXT:    s_ashr_i64 s[8:9], s[12:13], 40
-; GCN-NEXT:    s_ashr_i64 s[10:11], s[14:15], 40
-; GCN-NEXT:    s_xor_b32 s1, s2, s10
+; GCN-NEXT:    s_ashr_i64 s[10:11], s[10:11], 40
+; GCN-NEXT:    s_ashr_i64 s[2:3], s[2:3], 40
 ; GCN-NEXT:    v_cvt_f32_i32_e32 v0, s2
 ; GCN-NEXT:    v_cvt_f32_i32_e32 v1, s10
-; GCN-NEXT:    s_xor_b32 s2, s0, s8
-; GCN-NEXT:    v_cvt_f32_i32_e32 v2, s0
-; GCN-NEXT:    v_cvt_f32_i32_e32 v3, s8
-; GCN-NEXT:    s_ashr_i32 s0, s1, 30
+; GCN-NEXT:    s_ashr_i64 s[0:1], s[0:1], 40
+; GCN-NEXT:    s_xor_b32 s1, s10, s2
+; GCN-NEXT:    v_rcp_iflag_f32_e32 v2, v0
+; GCN-NEXT:    s_ashr_i32 s1, s1, 30
+; GCN-NEXT:    s_or_b32 s1, s1, 1
+; GCN-NEXT:    v_mov_b32_e32 v3, s1
+; GCN-NEXT:    v_mul_f32_e32 v2, v1, v2
+; GCN-NEXT:    v_trunc_f32_e32 v2, v2
+; GCN-NEXT:    v_mad_f32 v1, -v2, v0, v1
+; GCN-NEXT:    v_cvt_i32_f32_e32 v2, v2
+; GCN-NEXT:    v_cmp_ge_f32_e64 vcc, |v1|, |v0|
+; GCN-NEXT:    v_cvt_f32_i32_e32 v1, s0
+; GCN-NEXT:    v_cndmask_b32_e32 v0, 0, v3, vcc
+; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v2
+; GCN-NEXT:    s_ashr_i64 s[8:9], s[8:9], 40
+; GCN-NEXT:    v_bfe_i32 v2, v0, 0, 24
+; GCN-NEXT:    v_cvt_f32_i32_e32 v0, s8
 ; GCN-NEXT:    v_rcp_iflag_f32_e32 v4, v1
-; GCN-NEXT:    s_ashr_i32 s1, s2, 30
-; GCN-NEXT:    v_rcp_iflag_f32_e32 v5, v3
+; GCN-NEXT:    s_xor_b32 s0, s8, s0
+; GCN-NEXT:    s_ashr_i32 s0, s0, 30
 ; GCN-NEXT:    s_or_b32 s0, s0, 1
 ; GCN-NEXT:    v_mul_f32_e32 v4, v0, v4
-; GCN-NEXT:    s_or_b32 s1, s1, 1
-; GCN-NEXT:    v_mul_f32_e32 v5, v2, v5
 ; GCN-NEXT:    v_trunc_f32_e32 v4, v4
-; GCN-NEXT:    v_mov_b32_e32 v6, s0
-; GCN-NEXT:    v_trunc_f32_e32 v5, v5
-; GCN-NEXT:    v_mov_b32_e32 v7, s1
 ; GCN-NEXT:    v_mad_f32 v0, -v4, v1, v0
 ; GCN-NEXT:    v_cvt_i32_f32_e32 v4, v4
-; GCN-NEXT:    v_mad_f32 v2, -v5, v3, v2
-; GCN-NEXT:    v_cvt_i32_f32_e32 v5, v5
+; GCN-NEXT:    v_mov_b32_e32 v5, s0
 ; GCN-NEXT:    v_cmp_ge_f32_e64 vcc, |v0|, |v1|
-; GCN-NEXT:    v_cndmask_b32_e32 v0, 0, v6, vcc
-; GCN-NEXT:    v_cmp_ge_f32_e64 vcc, |v2|, |v3|
-; GCN-NEXT:    v_cndmask_b32_e32 v1, 0, v7, vcc
+; GCN-NEXT:    v_cndmask_b32_e32 v0, 0, v5, vcc
 ; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v4
-; GCN-NEXT:    v_add_i32_e32 v1, vcc, v1, v5
-; GCN-NEXT:    v_bfe_i32 v2, v0, 0, 24
-; GCN-NEXT:    v_bfe_i32 v0, v1, 0, 24
+; GCN-NEXT:    v_bfe_i32 v0, v0, 0, 24
 ; GCN-NEXT:    v_ashrrev_i32_e32 v3, 31, v2
 ; GCN-NEXT:    v_ashrrev_i32_e32 v1, 31, v0
 ; GCN-NEXT:    buffer_store_dwordx4 v[0:3], off, s[4:7], 0
 ; GCN-NEXT:    s_endpgm
   %1 = ashr <2 x i64> %x, <i64 40, i64 40>
   %2 = ashr <2 x i64> %y, <i64 40, i64 40>
   %result = sdiv <2 x i64> %1, %2
   store <2 x i64> %result, <2 x i64> addrspace(1)* %out
   ret void
 }
 
 define amdgpu_kernel void @s_test_sdiv24_48(i48 addrspace(1)* %out, i48 %x, i48 %y) {
 ; GCN-LABEL: s_test_sdiv24_48:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
 ; GCN-NEXT:    s_load_dword s2, s[0:1], 0xb
 ; GCN-NEXT:    s_load_dword s3, s[0:1], 0xc
 ; GCN-NEXT:    s_load_dword s8, s[0:1], 0xd
-; GCN-NEXT:    s_load_dword s9, s[0:1], 0xe
+; GCN-NEXT:    s_load_dword s0, s[0:1], 0xe
 ; GCN-NEXT:    s_mov_b32 s7, 0xf000
-; GCN-NEXT:    s_mov_b32 s6, -1
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_sext_i32_i16 s0, s3
-; GCN-NEXT:    s_sext_i32_i16 s1, s9
+; GCN-NEXT:    v_mov_b32_e32 v2, s2
+; GCN-NEXT:    s_sext_i32_i16 s1, s3
 ; GCN-NEXT:    v_mov_b32_e32 v0, s8
-; GCN-NEXT:    v_mov_b32_e32 v1, s2
-; GCN-NEXT:    v_alignbit_b32 v0, s1, v0, 24
-; GCN-NEXT:    v_alignbit_b32 v1, s0, v1, 24
-; GCN-NEXT:    v_xor_b32_e32 v2, v1, v0
-; GCN-NEXT:    v_cvt_f32_i32_e32 v1, v1
-; GCN-NEXT:    v_cvt_f32_i32_e32 v0, v0
-; GCN-NEXT:    v_ashrrev_i32_e32 v2, 30, v2
-; GCN-NEXT:    v_rcp_iflag_f32_e32 v3, v0
-; GCN-NEXT:    v_or_b32_e32 v2, 1, v2
-; GCN-NEXT:    v_mul_f32_e32 v3, v1, v3
-; GCN-NEXT:    v_trunc_f32_e32 v3, v3
-; GCN-NEXT:    v_mad_f32 v1, -v3, v0, v1
-; GCN-NEXT:    v_cvt_i32_f32_e32 v3, v3
-; GCN-NEXT:    v_cmp_ge_f32_e64 vcc, |v1|, |v0|
-; GCN-NEXT:    v_cndmask_b32_e32 v0, 0, v2, vcc
-; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v3
+; GCN-NEXT:    s_sext_i32_i16 s0, s0
+; GCN-NEXT:    v_alignbit_b32 v0, s0, v0, 24
+; GCN-NEXT:    v_cvt_f32_i32_e32 v1, v0
+; GCN-NEXT:    v_alignbit_b32 v2, s1, v2, 24
+; GCN-NEXT:    v_cvt_f32_i32_e32 v3, v2
+; GCN-NEXT:    v_xor_b32_e32 v0, v2, v0
+; GCN-NEXT:    v_rcp_iflag_f32_e32 v4, v1
+; GCN-NEXT:    v_ashrrev_i32_e32 v0, 30, v0
+; GCN-NEXT:    v_or_b32_e32 v0, 1, v0
+; GCN-NEXT:    s_mov_b32 s6, -1
+; GCN-NEXT:    v_mul_f32_e32 v2, v3, v4
+; GCN-NEXT:    v_trunc_f32_e32 v2, v2
+; GCN-NEXT:    v_mad_f32 v3, -v2, v1, v3
+; GCN-NEXT:    v_cvt_i32_f32_e32 v2, v2
+; GCN-NEXT:    v_cmp_ge_f32_e64 vcc, |v3|, |v1|
+; GCN-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
+; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v2
 ; GCN-NEXT:    v_bfe_i32 v0, v0, 0, 24
 ; GCN-NEXT:    v_ashrrev_i32_e32 v1, 31, v0
 ; GCN-NEXT:    buffer_store_dword v0, off, s[4:7], 0
 ; GCN-NEXT:    buffer_store_short v1, off, s[4:7], 0 offset:4
 ; GCN-NEXT:    s_endpgm
   %1 = ashr i48 %x, 24
   %2 = ashr i48 %y, 24
   %result = sdiv i48 %1, %2
   store i48 %result, i48 addrspace(1)* %out
   ret void
 }
 
 define amdgpu_kernel void @s_test_sdiv_k_num_i64(i64 addrspace(1)* %out, i64 %x) {
 ; GCN-LABEL: s_test_sdiv_k_num_i64:
 ; GCN:       ; %bb.0:
-; GCN-NEXT:    s_load_dwordx4 s[8:11], s[0:1], 0x9
-; GCN-NEXT:    s_mov_b32 s7, 0xf000
-; GCN-NEXT:    s_mov_b32 s6, -1
-; GCN-NEXT:    v_mov_b32_e32 v0, 0
+; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN-NEXT:    v_mov_b32_e32 v2, 0
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_ashr_i32 s2, s11, 31
-; GCN-NEXT:    v_mov_b32_e32 v1, 0
-; GCN-NEXT:    s_mov_b32 s4, s8
-; GCN-NEXT:    s_mov_b32 s5, s9
+; GCN-NEXT:    s_ashr_i32 s2, s7, 31
+; GCN-NEXT:    s_add_u32 s0, s6, s2
+; GCN-NEXT:    s_addc_u32 s1, s7, s2
 ; GCN-NEXT:    s_mov_b32 s3, s2
-; GCN-NEXT:    s_add_u32 s0, s10, s2
-; GCN-NEXT:    v_mov_b32_e32 v2, s2
-; GCN-NEXT:    s_addc_u32 s1, s11, s2
 ; GCN-NEXT:    s_xor_b64 s[8:9], s[0:1], s[2:3]
-; GCN-NEXT:    v_cvt_f32_u32_e32 v3, s8
-; GCN-NEXT:    v_cvt_f32_u32_e32 v4, s9
+; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s8
+; GCN-NEXT:    v_cvt_f32_u32_e32 v1, s9
 ; GCN-NEXT:    s_sub_u32 s3, 0, s8
-; GCN-NEXT:    v_mov_b32_e32 v5, s9
-; GCN-NEXT:    v_mac_f32_e32 v3, 0x4f800000, v4
 ; GCN-NEXT:    s_subb_u32 s10, 0, s9
-; GCN-NEXT:    v_rcp_f32_e32 v3, v3
-; GCN-NEXT:    v_mul_f32_e32 v3, 0x5f7ffffc, v3
-; GCN-NEXT:    v_mul_f32_e32 v4, 0x2f800000, v3
-; GCN-NEXT:    v_trunc_f32_e32 v4, v4
-; GCN-NEXT:    v_mac_f32_e32 v3, 0xcf800000, v4
-; GCN-NEXT:    v_cvt_u32_f32_e32 v4, v4
+; GCN-NEXT:    s_mov_b32 s7, 0xf000
+; GCN-NEXT:    v_mac_f32_e32 v0, 0x4f800000, v1
+; GCN-NEXT:    v_rcp_f32_e32 v0, v0
+; GCN-NEXT:    v_mov_b32_e32 v1, 0
+; GCN-NEXT:    s_mov_b32 s6, -1
+; GCN-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v0
+; GCN-NEXT:    v_mul_f32_e32 v3, 0x2f800000, v0
+; GCN-NEXT:    v_trunc_f32_e32 v3, v3
+; GCN-NEXT:    v_mac_f32_e32 v0, 0xcf800000, v3
+; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v0
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v3, v3
+; GCN-NEXT:    v_mul_hi_u32 v5, s3, v0
+; GCN-NEXT:    v_mul_lo_u32 v4, s3, v3
+; GCN-NEXT:    v_mul_lo_u32 v7, s10, v0
+; GCN-NEXT:    v_mul_lo_u32 v6, s3, v0
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v5, v4
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v4, v7
+; GCN-NEXT:    v_mul_hi_u32 v5, v0, v6
+; GCN-NEXT:    v_mul_lo_u32 v7, v0, v4
+; GCN-NEXT:    v_mul_hi_u32 v9, v0, v4
+; GCN-NEXT:    v_mul_lo_u32 v8, v3, v6
+; GCN-NEXT:    v_mul_hi_u32 v6, v3, v6
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v7
+; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v2, v9, vcc
+; GCN-NEXT:    v_mul_hi_u32 v9, v3, v4
+; GCN-NEXT:    v_mul_lo_u32 v4, v3, v4
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v8, v5
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v7, v6, vcc
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v9, v1, vcc
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v5, v4
+; GCN-NEXT:    v_add_i32_e64 v0, s[0:1], v0, v4
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v2, v6, vcc
+; GCN-NEXT:    v_addc_u32_e64 v4, vcc, v3, v5, s[0:1]
 ; GCN-NEXT:    v_mul_lo_u32 v6, s3, v4
-; GCN-NEXT:    v_mul_lo_u32 v7, s10, v3
-; GCN-NEXT:    v_mul_hi_u32 v8, s3, v3
-; GCN-NEXT:    v_mul_lo_u32 v9, s3, v3
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, v8, v6
-; GCN-NEXT:    v_mul_hi_u32 v8, v3, v9
-; GCN-NEXT:    v_mul_hi_u32 v10, v4, v9
-; GCN-NEXT:    v_mul_lo_u32 v9, v4, v9
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, v6, v7
-; GCN-NEXT:    v_mul_hi_u32 v7, v3, v6
-; GCN-NEXT:    v_mul_lo_u32 v11, v3, v6
-; GCN-NEXT:    v_mul_hi_u32 v12, v4, v6
-; GCN-NEXT:    v_mul_lo_u32 v6, v4, v6
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v8, v11
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v1, v7, vcc
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v9, v8
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v7, v10, vcc
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v12, v0, vcc
+; GCN-NEXT:    v_mul_hi_u32 v7, s3, v0
+; GCN-NEXT:    v_mul_lo_u32 v8, s10, v0
 ; GCN-NEXT:    v_add_i32_e32 v6, vcc, v7, v6
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v1, v8, vcc
-; GCN-NEXT:    v_add_i32_e64 v3, s[0:1], v3, v6
-; GCN-NEXT:    v_addc_u32_e64 v6, vcc, v4, v7, s[0:1]
-; GCN-NEXT:    v_mul_hi_u32 v8, s3, v3
-; GCN-NEXT:    v_mul_lo_u32 v9, s10, v3
-; GCN-NEXT:    v_mul_lo_u32 v10, s3, v3
-; GCN-NEXT:    v_mul_lo_u32 v11, s3, v6
-; GCN-NEXT:    v_mul_hi_u32 v12, v6, v10
-; GCN-NEXT:    v_mul_lo_u32 v13, v6, v10
-; GCN-NEXT:    v_mul_hi_u32 v10, v3, v10
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v8, v11
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v9, v8
-; GCN-NEXT:    v_mul_hi_u32 v9, v6, v8
-; GCN-NEXT:    v_mul_hi_u32 v11, v3, v8
-; GCN-NEXT:    v_mul_lo_u32 v14, v3, v8
-; GCN-NEXT:    v_mul_lo_u32 v6, v6, v8
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v10, v14
-; GCN-NEXT:    v_addc_u32_e32 v10, vcc, v1, v11, vcc
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v13, v8
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v10, v12, vcc
-; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v9, v0, vcc
-; GCN-NEXT:    v_add_i32_e32 v4, vcc, v4, v7
+; GCN-NEXT:    v_mul_lo_u32 v7, s3, v0
 ; GCN-NEXT:    v_add_i32_e32 v6, vcc, v8, v6
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v1, v9, vcc
-; GCN-NEXT:    v_addc_u32_e64 v4, vcc, v4, v7, s[0:1]
-; GCN-NEXT:    v_add_i32_e32 v3, vcc, v3, v6
-; GCN-NEXT:    v_addc_u32_e32 v4, vcc, 0, v4, vcc
+; GCN-NEXT:    v_mul_lo_u32 v10, v0, v6
+; GCN-NEXT:    v_mul_hi_u32 v12, v0, v6
+; GCN-NEXT:    v_mul_hi_u32 v11, v0, v7
+; GCN-NEXT:    v_mul_hi_u32 v9, v4, v7
+; GCN-NEXT:    v_mul_lo_u32 v7, v4, v7
+; GCN-NEXT:    v_mul_hi_u32 v8, v4, v6
+; GCN-NEXT:    v_add_i32_e32 v10, vcc, v11, v10
+; GCN-NEXT:    v_addc_u32_e32 v11, vcc, v2, v12, vcc
+; GCN-NEXT:    v_mul_lo_u32 v4, v4, v6
+; GCN-NEXT:    v_add_i32_e32 v7, vcc, v7, v10
+; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v11, v9, vcc
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v8, v1, vcc
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v7, v4
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v2, v6, vcc
+; GCN-NEXT:    v_add_i32_e32 v3, vcc, v3, v5
+; GCN-NEXT:    v_addc_u32_e64 v3, vcc, v3, v6, s[0:1]
+; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v4
+; GCN-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; GCN-NEXT:    v_mul_hi_u32 v5, 24, v0
+; GCN-NEXT:    v_mul_lo_u32 v4, v3, 24
 ; GCN-NEXT:    v_mul_hi_u32 v6, 24, v3
+; GCN-NEXT:    v_mul_hi_u32 v0, 0, v0
 ; GCN-NEXT:    v_mul_hi_u32 v3, 0, v3
-; GCN-NEXT:    v_mul_hi_u32 v7, 24, v4
-; GCN-NEXT:    v_mul_lo_u32 v8, v4, 24
-; GCN-NEXT:    v_mul_hi_u32 v4, 0, v4
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, v6, v8
-; GCN-NEXT:    v_addc_u32_e32 v1, vcc, v1, v7, vcc
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, 0, v6
-; GCN-NEXT:    v_addc_u32_e32 v1, vcc, v1, v3, vcc
-; GCN-NEXT:    v_addc_u32_e32 v0, vcc, v4, v0, vcc
-; GCN-NEXT:    v_mul_hi_u32 v3, s8, v1
-; GCN-NEXT:    v_mul_lo_u32 v4, s9, v1
-; GCN-NEXT:    v_mul_lo_u32 v6, s8, v1
-; GCN-NEXT:    v_mul_lo_u32 v7, s8, v0
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, 2, v1
-; GCN-NEXT:    v_addc_u32_e32 v9, vcc, 0, v0, vcc
-; GCN-NEXT:    v_add_i32_e32 v10, vcc, 1, v1
-; GCN-NEXT:    v_addc_u32_e32 v11, vcc, 0, v0, vcc
-; GCN-NEXT:    v_add_i32_e32 v3, vcc, v3, v7
-; GCN-NEXT:    v_add_i32_e32 v3, vcc, v3, v4
-; GCN-NEXT:    v_sub_i32_e32 v4, vcc, 0, v3
-; GCN-NEXT:    v_sub_i32_e32 v6, vcc, 24, v6
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v5, v4
+; GCN-NEXT:    v_addc_u32_e32 v2, vcc, v2, v6, vcc
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, 0, v4
+; GCN-NEXT:    v_addc_u32_e32 v0, vcc, v2, v0, vcc
+; GCN-NEXT:    v_addc_u32_e32 v1, vcc, v3, v1, vcc
+; GCN-NEXT:    v_mul_lo_u32 v2, s8, v1
+; GCN-NEXT:    v_mul_hi_u32 v3, s8, v0
+; GCN-NEXT:    v_mul_lo_u32 v4, s9, v0
+; GCN-NEXT:    v_mov_b32_e32 v5, s9
+; GCN-NEXT:    v_add_i32_e32 v2, vcc, v3, v2
+; GCN-NEXT:    v_mul_lo_u32 v3, s8, v0
+; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v4
+; GCN-NEXT:    v_sub_i32_e32 v4, vcc, 0, v2
+; GCN-NEXT:    v_sub_i32_e32 v3, vcc, 24, v3
 ; GCN-NEXT:    v_subb_u32_e64 v4, s[0:1], v4, v5, vcc
-; GCN-NEXT:    v_subrev_i32_e64 v5, s[0:1], s8, v6
-; GCN-NEXT:    v_subb_u32_e32 v3, vcc, 0, v3, vcc
-; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s8, v6
+; GCN-NEXT:    v_subrev_i32_e64 v5, s[0:1], s8, v3
+; GCN-NEXT:    v_subbrev_u32_e64 v4, s[0:1], 0, v4, s[0:1]
+; GCN-NEXT:    v_cmp_le_u32_e64 s[0:1], s9, v4
+; GCN-NEXT:    v_cndmask_b32_e64 v6, 0, -1, s[0:1]
+; GCN-NEXT:    v_cmp_le_u32_e64 s[0:1], s8, v5
+; GCN-NEXT:    v_cndmask_b32_e64 v5, 0, -1, s[0:1]
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], s9, v4
+; GCN-NEXT:    v_cndmask_b32_e64 v4, v6, v5, s[0:1]
+; GCN-NEXT:    v_add_i32_e64 v5, s[0:1], 2, v0
+; GCN-NEXT:    v_addc_u32_e64 v6, s[0:1], 0, v1, s[0:1]
+; GCN-NEXT:    v_add_i32_e64 v7, s[0:1], 1, v0
+; GCN-NEXT:    v_subb_u32_e32 v2, vcc, 0, v2, vcc
+; GCN-NEXT:    v_addc_u32_e64 v8, s[0:1], 0, v1, s[0:1]
+; GCN-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v4
+; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s9, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v4, v8, v6, s[0:1]
 ; GCN-NEXT:    v_cndmask_b32_e64 v6, 0, -1, vcc
-; GCN-NEXT:    v_subbrev_u32_e64 v4, vcc, 0, v4, s[0:1]
-; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s8, v5
-; GCN-NEXT:    v_cndmask_b32_e64 v5, 0, -1, vcc
-; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s9, v3
-; GCN-NEXT:    v_cndmask_b32_e64 v7, 0, -1, vcc
-; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s9, v4
-; GCN-NEXT:    v_cndmask_b32_e64 v12, 0, -1, vcc
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, s9, v3
-; GCN-NEXT:    v_cndmask_b32_e32 v3, v7, v6, vcc
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, s9, v4
-; GCN-NEXT:    v_cndmask_b32_e32 v4, v12, v5, vcc
-; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v4
-; GCN-NEXT:    v_cndmask_b32_e32 v4, v11, v9, vcc
-; GCN-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v3
-; GCN-NEXT:    v_cndmask_b32_e64 v0, v0, v4, s[0:1]
-; GCN-NEXT:    v_cndmask_b32_e32 v3, v10, v8, vcc
-; GCN-NEXT:    v_cndmask_b32_e64 v1, v1, v3, s[0:1]
-; GCN-NEXT:    v_xor_b32_e32 v3, s2, v0
-; GCN-NEXT:    v_xor_b32_e32 v0, s2, v1
+; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s8, v3
+; GCN-NEXT:    v_cndmask_b32_e64 v3, 0, -1, vcc
+; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, s9, v2
+; GCN-NEXT:    v_cndmask_b32_e32 v2, v6, v3, vcc
+; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v2, v7, v5, s[0:1]
+; GCN-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
+; GCN-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; GCN-NEXT:    v_xor_b32_e32 v0, s2, v0
+; GCN-NEXT:    v_xor_b32_e32 v1, s2, v1
+; GCN-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, s2, v0
-; GCN-NEXT:    v_subb_u32_e32 v1, vcc, v3, v2, vcc
+; GCN-NEXT:    v_subb_u32_e32 v1, vcc, v1, v2, vcc
 ; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
 ; GCN-NEXT:    s_endpgm
   %result = sdiv i64 24, %x
   store i64 %result, i64 addrspace(1)* %out
   ret void
 }
 
 define i64 @v_test_sdiv_k_num_i64(i64 %x) {
 ; GCN-LABEL: v_test_sdiv_k_num_i64:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GCN-NEXT:    v_ashrrev_i32_e32 v2, 31, v1
-; GCN-NEXT:    v_mov_b32_e32 v3, 0
-; GCN-NEXT:    v_mov_b32_e32 v4, 0
 ; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v2
 ; GCN-NEXT:    v_addc_u32_e32 v1, vcc, v1, v2, vcc
-; GCN-NEXT:    v_xor_b32_e32 v1, v1, v2
 ; GCN-NEXT:    v_xor_b32_e32 v0, v0, v2
-; GCN-NEXT:    v_cvt_f32_u32_e32 v5, v0
-; GCN-NEXT:    v_cvt_f32_u32_e32 v6, v1
-; GCN-NEXT:    v_sub_i32_e32 v7, vcc, 0, v0
-; GCN-NEXT:    v_subb_u32_e32 v8, vcc, 0, v1, vcc
-; GCN-NEXT:    v_mac_f32_e32 v5, 0x4f800000, v6
-; GCN-NEXT:    v_rcp_f32_e32 v5, v5
-; GCN-NEXT:    v_mul_f32_e32 v5, 0x5f7ffffc, v5
-; GCN-NEXT:    v_mul_f32_e32 v6, 0x2f800000, v5
-; GCN-NEXT:    v_trunc_f32_e32 v6, v6
-; GCN-NEXT:    v_mac_f32_e32 v5, 0xcf800000, v6
-; GCN-NEXT:    v_cvt_u32_f32_e32 v6, v6
-; GCN-NEXT:    v_cvt_u32_f32_e32 v5, v5
-; GCN-NEXT:    v_mul_lo_u32 v9, v7, v6
-; GCN-NEXT:    v_mul_lo_u32 v10, v8, v5
-; GCN-NEXT:    v_mul_hi_u32 v11, v7, v5
-; GCN-NEXT:    v_mul_lo_u32 v12, v7, v5
-; GCN-NEXT:    v_add_i32_e32 v9, vcc, v11, v9
-; GCN-NEXT:    v_mul_hi_u32 v11, v5, v12
-; GCN-NEXT:    v_mul_hi_u32 v13, v6, v12
-; GCN-NEXT:    v_mul_lo_u32 v12, v6, v12
-; GCN-NEXT:    v_add_i32_e32 v9, vcc, v9, v10
-; GCN-NEXT:    v_mul_hi_u32 v10, v5, v9
-; GCN-NEXT:    v_mul_lo_u32 v14, v5, v9
-; GCN-NEXT:    v_mul_hi_u32 v15, v6, v9
-; GCN-NEXT:    v_mul_lo_u32 v9, v6, v9
-; GCN-NEXT:    v_add_i32_e32 v11, vcc, v11, v14
-; GCN-NEXT:    v_addc_u32_e32 v10, vcc, v4, v10, vcc
-; GCN-NEXT:    v_add_i32_e32 v11, vcc, v12, v11
-; GCN-NEXT:    v_addc_u32_e32 v10, vcc, v10, v13, vcc
-; GCN-NEXT:    v_addc_u32_e32 v11, vcc, v15, v3, vcc
-; GCN-NEXT:    v_add_i32_e32 v9, vcc, v10, v9
-; GCN-NEXT:    v_addc_u32_e32 v10, vcc, v4, v11, vcc
-; GCN-NEXT:    v_add_i32_e64 v5, s[4:5], v5, v9
-; GCN-NEXT:    v_addc_u32_e64 v9, vcc, v6, v10, s[4:5]
-; GCN-NEXT:    v_mul_hi_u32 v11, v7, v5
-; GCN-NEXT:    v_mul_lo_u32 v8, v8, v5
-; GCN-NEXT:    v_mul_lo_u32 v12, v7, v5
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, v6, v10
-; GCN-NEXT:    v_mul_lo_u32 v7, v7, v9
-; GCN-NEXT:    v_mul_hi_u32 v10, v9, v12
-; GCN-NEXT:    v_mul_lo_u32 v13, v9, v12
-; GCN-NEXT:    v_mul_hi_u32 v12, v5, v12
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v11, v7
+; GCN-NEXT:    v_xor_b32_e32 v1, v1, v2
+; GCN-NEXT:    v_cvt_f32_u32_e32 v3, v0
+; GCN-NEXT:    v_cvt_f32_u32_e32 v4, v1
+; GCN-NEXT:    v_sub_i32_e32 v5, vcc, 0, v0
+; GCN-NEXT:    v_subb_u32_e32 v6, vcc, 0, v1, vcc
+; GCN-NEXT:    v_mov_b32_e32 v13, 0
+; GCN-NEXT:    v_mac_f32_e32 v3, 0x4f800000, v4
+; GCN-NEXT:    v_rcp_f32_e32 v3, v3
+; GCN-NEXT:    v_mov_b32_e32 v12, 0
+; GCN-NEXT:    v_mul_f32_e32 v3, 0x5f7ffffc, v3
+; GCN-NEXT:    v_mul_f32_e32 v4, 0x2f800000, v3
+; GCN-NEXT:    v_trunc_f32_e32 v4, v4
+; GCN-NEXT:    v_mac_f32_e32 v3, 0xcf800000, v4
+; GCN-NEXT:    v_cvt_u32_f32_e32 v3, v3
+; GCN-NEXT:    v_cvt_u32_f32_e32 v4, v4
+; GCN-NEXT:    v_mul_hi_u32 v7, v5, v3
+; GCN-NEXT:    v_mul_lo_u32 v8, v5, v4
+; GCN-NEXT:    v_mul_lo_u32 v9, v6, v3
 ; GCN-NEXT:    v_add_i32_e32 v7, vcc, v7, v8
-; GCN-NEXT:    v_mul_hi_u32 v8, v9, v7
-; GCN-NEXT:    v_mul_hi_u32 v11, v5, v7
-; GCN-NEXT:    v_mul_lo_u32 v14, v5, v7
-; GCN-NEXT:    v_mul_lo_u32 v7, v9, v7
-; GCN-NEXT:    v_add_i32_e32 v9, vcc, v12, v14
-; GCN-NEXT:    v_addc_u32_e32 v11, vcc, v4, v11, vcc
-; GCN-NEXT:    v_add_i32_e32 v9, vcc, v13, v9
-; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v11, v10, vcc
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v8, v3, vcc
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v9, v7
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v4, v8, vcc
-; GCN-NEXT:    v_addc_u32_e64 v6, vcc, v6, v8, s[4:5]
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v7
-; GCN-NEXT:    v_addc_u32_e32 v6, vcc, 0, v6, vcc
-; GCN-NEXT:    v_mul_hi_u32 v7, 24, v5
-; GCN-NEXT:    v_mul_hi_u32 v5, 0, v5
-; GCN-NEXT:    v_mul_hi_u32 v8, 24, v6
-; GCN-NEXT:    v_mul_lo_u32 v9, v6, 24
-; GCN-NEXT:    v_mul_hi_u32 v6, 0, v6
+; GCN-NEXT:    v_mul_lo_u32 v8, v5, v3
 ; GCN-NEXT:    v_add_i32_e32 v7, vcc, v7, v9
-; GCN-NEXT:    v_addc_u32_e32 v4, vcc, v4, v8, vcc
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, 0, v7
-; GCN-NEXT:    v_addc_u32_e32 v4, vcc, v4, v5, vcc
-; GCN-NEXT:    v_addc_u32_e32 v3, vcc, v6, v3, vcc
-; GCN-NEXT:    v_mul_hi_u32 v5, v0, v4
-; GCN-NEXT:    v_mul_lo_u32 v6, v1, v4
-; GCN-NEXT:    v_mul_lo_u32 v7, v0, v4
-; GCN-NEXT:    v_mul_lo_u32 v8, v0, v3
-; GCN-NEXT:    v_add_i32_e32 v9, vcc, 2, v4
-; GCN-NEXT:    v_addc_u32_e32 v10, vcc, 0, v3, vcc
-; GCN-NEXT:    v_add_i32_e32 v11, vcc, 1, v4
-; GCN-NEXT:    v_addc_u32_e32 v12, vcc, 0, v3, vcc
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v8
+; GCN-NEXT:    v_mul_lo_u32 v10, v3, v7
+; GCN-NEXT:    v_mul_hi_u32 v9, v3, v7
+; GCN-NEXT:    v_mul_hi_u32 v11, v3, v8
+; GCN-NEXT:    v_mul_hi_u32 v14, v4, v7
+; GCN-NEXT:    v_mul_lo_u32 v7, v4, v7
+; GCN-NEXT:    v_add_i32_e32 v10, vcc, v11, v10
+; GCN-NEXT:    v_mul_lo_u32 v11, v4, v8
+; GCN-NEXT:    v_mul_hi_u32 v8, v4, v8
+; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v13, v9, vcc
+; GCN-NEXT:    v_add_i32_e32 v10, vcc, v11, v10
+; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v9, v8, vcc
+; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v14, v12, vcc
+; GCN-NEXT:    v_add_i32_e32 v7, vcc, v8, v7
+; GCN-NEXT:    v_add_i32_e64 v3, s[4:5], v3, v7
+; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v13, v9, vcc
+; GCN-NEXT:    v_addc_u32_e64 v7, vcc, v4, v8, s[4:5]
+; GCN-NEXT:    v_mul_lo_u32 v9, v5, v7
+; GCN-NEXT:    v_mul_hi_u32 v10, v5, v3
+; GCN-NEXT:    v_mul_lo_u32 v6, v6, v3
+; GCN-NEXT:    v_mul_lo_u32 v5, v5, v3
+; GCN-NEXT:    v_add_i32_e32 v9, vcc, v10, v9
+; GCN-NEXT:    v_add_i32_e32 v6, vcc, v9, v6
+; GCN-NEXT:    v_mul_lo_u32 v11, v3, v6
+; GCN-NEXT:    v_mul_hi_u32 v14, v3, v5
+; GCN-NEXT:    v_mul_hi_u32 v15, v3, v6
+; GCN-NEXT:    v_mul_hi_u32 v10, v7, v5
+; GCN-NEXT:    v_mul_lo_u32 v5, v7, v5
+; GCN-NEXT:    v_add_i32_e32 v11, vcc, v14, v11
+; GCN-NEXT:    v_mul_hi_u32 v9, v7, v6
+; GCN-NEXT:    v_addc_u32_e32 v14, vcc, v13, v15, vcc
+; GCN-NEXT:    v_mul_lo_u32 v6, v7, v6
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v11
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v14, v10, vcc
+; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v9, v12, vcc
 ; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v6
-; GCN-NEXT:    v_sub_i32_e32 v6, vcc, 0, v5
-; GCN-NEXT:    v_sub_i32_e32 v7, vcc, 24, v7
-; GCN-NEXT:    v_subb_u32_e64 v6, s[4:5], v6, v1, vcc
-; GCN-NEXT:    v_sub_i32_e64 v8, s[4:5], v7, v0
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v13, v7, vcc
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v4, v8
+; GCN-NEXT:    v_addc_u32_e64 v4, vcc, v4, v6, s[4:5]
+; GCN-NEXT:    v_add_i32_e32 v3, vcc, v3, v5
+; GCN-NEXT:    v_addc_u32_e32 v4, vcc, 0, v4, vcc
+; GCN-NEXT:    v_mul_hi_u32 v6, 24, v3
+; GCN-NEXT:    v_mul_lo_u32 v5, v4, 24
+; GCN-NEXT:    v_mul_hi_u32 v7, 24, v4
+; GCN-NEXT:    v_mul_hi_u32 v3, 0, v3
+; GCN-NEXT:    v_mul_hi_u32 v4, 0, v4
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v6, v5
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v13, v7, vcc
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, 0, v5
+; GCN-NEXT:    v_addc_u32_e32 v3, vcc, v6, v3, vcc
+; GCN-NEXT:    v_addc_u32_e32 v4, vcc, v4, v12, vcc
+; GCN-NEXT:    v_mul_lo_u32 v5, v0, v4
+; GCN-NEXT:    v_mul_hi_u32 v6, v0, v3
+; GCN-NEXT:    v_mul_lo_u32 v7, v1, v3
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v6, v5
+; GCN-NEXT:    v_mul_lo_u32 v6, v0, v3
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v7
+; GCN-NEXT:    v_sub_i32_e32 v7, vcc, 0, v5
+; GCN-NEXT:    v_sub_i32_e32 v6, vcc, 24, v6
+; GCN-NEXT:    v_subb_u32_e64 v7, s[4:5], v7, v1, vcc
+; GCN-NEXT:    v_sub_i32_e64 v8, s[4:5], v6, v0
+; GCN-NEXT:    v_subbrev_u32_e64 v7, s[4:5], 0, v7, s[4:5]
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[4:5], v7, v1
+; GCN-NEXT:    v_cndmask_b32_e64 v9, 0, -1, s[4:5]
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[4:5], v8, v0
+; GCN-NEXT:    v_cndmask_b32_e64 v8, 0, -1, s[4:5]
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[4:5], v7, v1
+; GCN-NEXT:    v_cndmask_b32_e64 v7, v9, v8, s[4:5]
+; GCN-NEXT:    v_add_i32_e64 v8, s[4:5], 2, v3
+; GCN-NEXT:    v_addc_u32_e64 v9, s[4:5], 0, v4, s[4:5]
+; GCN-NEXT:    v_add_i32_e64 v10, s[4:5], 1, v3
 ; GCN-NEXT:    v_subb_u32_e32 v5, vcc, 0, v5, vcc
-; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v7, v0
-; GCN-NEXT:    v_cndmask_b32_e64 v7, 0, -1, vcc
-; GCN-NEXT:    v_subbrev_u32_e64 v6, vcc, 0, v6, s[4:5]
-; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v8, v0
-; GCN-NEXT:    v_cndmask_b32_e64 v0, 0, -1, vcc
+; GCN-NEXT:    v_addc_u32_e64 v11, s[4:5], 0, v4, s[4:5]
+; GCN-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, v7
 ; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v5, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v8, 0, -1, vcc
-; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v6, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v13, 0, -1, vcc
+; GCN-NEXT:    v_cndmask_b32_e64 v7, v11, v9, s[4:5]
+; GCN-NEXT:    v_cndmask_b32_e64 v9, 0, -1, vcc
+; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v6, v0
+; GCN-NEXT:    v_cndmask_b32_e64 v0, 0, -1, vcc
 ; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, v5, v1
-; GCN-NEXT:    v_cndmask_b32_e32 v5, v8, v7, vcc
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, v6, v1
-; GCN-NEXT:    v_cndmask_b32_e32 v0, v13, v0, vcc
+; GCN-NEXT:    v_cndmask_b32_e32 v0, v9, v0, vcc
 ; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
-; GCN-NEXT:    v_cndmask_b32_e32 v0, v12, v10, vcc
-; GCN-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, v5
-; GCN-NEXT:    v_cndmask_b32_e64 v0, v3, v0, s[4:5]
-; GCN-NEXT:    v_cndmask_b32_e32 v1, v11, v9, vcc
-; GCN-NEXT:    v_cndmask_b32_e64 v1, v4, v1, s[4:5]
+; GCN-NEXT:    v_cndmask_b32_e64 v1, v10, v8, s[4:5]
+; GCN-NEXT:    v_cndmask_b32_e32 v0, v4, v7, vcc
+; GCN-NEXT:    v_cndmask_b32_e32 v1, v3, v1, vcc
 ; GCN-NEXT:    v_xor_b32_e32 v3, v0, v2
 ; GCN-NEXT:    v_xor_b32_e32 v0, v1, v2
 ; GCN-NEXT:    v_sub_i32_e32 v0, vcc, v0, v2
 ; GCN-NEXT:    v_subb_u32_e32 v1, vcc, v3, v2, vcc
 ; GCN-NEXT:    s_setpc_b64 s[30:31]
   %result = sdiv i64 24, %x
   ret i64 %result
 }
 
 define i64 @v_test_sdiv_pow2_k_num_i64(i64 %x) {
 ; GCN-LABEL: v_test_sdiv_pow2_k_num_i64:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GCN-NEXT:    v_ashrrev_i32_e32 v2, 31, v1
-; GCN-NEXT:    v_mov_b32_e32 v3, 0
-; GCN-NEXT:    v_mov_b32_e32 v4, 0
-; GCN-NEXT:    s_mov_b32 s6, 0x8000
 ; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v2
 ; GCN-NEXT:    v_addc_u32_e32 v1, vcc, v1, v2, vcc
-; GCN-NEXT:    v_xor_b32_e32 v1, v1, v2
 ; GCN-NEXT:    v_xor_b32_e32 v0, v0, v2
-; GCN-NEXT:    v_cvt_f32_u32_e32 v5, v0
-; GCN-NEXT:    v_cvt_f32_u32_e32 v6, v1
-; GCN-NEXT:    v_sub_i32_e32 v7, vcc, 0, v0
-; GCN-NEXT:    v_subb_u32_e32 v8, vcc, 0, v1, vcc
-; GCN-NEXT:    v_mac_f32_e32 v5, 0x4f800000, v6
-; GCN-NEXT:    v_rcp_f32_e32 v5, v5
-; GCN-NEXT:    v_mul_f32_e32 v5, 0x5f7ffffc, v5
-; GCN-NEXT:    v_mul_f32_e32 v6, 0x2f800000, v5
-; GCN-NEXT:    v_trunc_f32_e32 v6, v6
-; GCN-NEXT:    v_mac_f32_e32 v5, 0xcf800000, v6
-; GCN-NEXT:    v_cvt_u32_f32_e32 v6, v6
-; GCN-NEXT:    v_cvt_u32_f32_e32 v5, v5
-; GCN-NEXT:    v_mul_lo_u32 v9, v7, v6
-; GCN-NEXT:    v_mul_lo_u32 v10, v8, v5
-; GCN-NEXT:    v_mul_hi_u32 v11, v7, v5
-; GCN-NEXT:    v_mul_lo_u32 v12, v7, v5
-; GCN-NEXT:    v_add_i32_e32 v9, vcc, v11, v9
-; GCN-NEXT:    v_mul_hi_u32 v11, v5, v12
-; GCN-NEXT:    v_mul_hi_u32 v13, v6, v12
-; GCN-NEXT:    v_mul_lo_u32 v12, v6, v12
-; GCN-NEXT:    v_add_i32_e32 v9, vcc, v9, v10
-; GCN-NEXT:    v_mul_hi_u32 v10, v5, v9
-; GCN-NEXT:    v_mul_lo_u32 v14, v5, v9
-; GCN-NEXT:    v_mul_hi_u32 v15, v6, v9
-; GCN-NEXT:    v_mul_lo_u32 v9, v6, v9
-; GCN-NEXT:    v_add_i32_e32 v11, vcc, v11, v14
-; GCN-NEXT:    v_addc_u32_e32 v10, vcc, v4, v10, vcc
-; GCN-NEXT:    v_add_i32_e32 v11, vcc, v12, v11
-; GCN-NEXT:    v_addc_u32_e32 v10, vcc, v10, v13, vcc
-; GCN-NEXT:    v_addc_u32_e32 v11, vcc, v15, v3, vcc
-; GCN-NEXT:    v_add_i32_e32 v9, vcc, v10, v9
-; GCN-NEXT:    v_addc_u32_e32 v10, vcc, v4, v11, vcc
-; GCN-NEXT:    v_add_i32_e64 v5, s[4:5], v5, v9
-; GCN-NEXT:    v_addc_u32_e64 v9, vcc, v6, v10, s[4:5]
-; GCN-NEXT:    v_mul_hi_u32 v11, v7, v5
-; GCN-NEXT:    v_mul_lo_u32 v8, v8, v5
-; GCN-NEXT:    v_mul_lo_u32 v12, v7, v5
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, v6, v10
-; GCN-NEXT:    v_mul_lo_u32 v7, v7, v9
-; GCN-NEXT:    v_mul_hi_u32 v10, v9, v12
-; GCN-NEXT:    v_mul_lo_u32 v13, v9, v12
-; GCN-NEXT:    v_mul_hi_u32 v12, v5, v12
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v11, v7
+; GCN-NEXT:    v_xor_b32_e32 v1, v1, v2
+; GCN-NEXT:    v_cvt_f32_u32_e32 v3, v0
+; GCN-NEXT:    v_cvt_f32_u32_e32 v4, v1
+; GCN-NEXT:    v_sub_i32_e32 v5, vcc, 0, v0
+; GCN-NEXT:    v_subb_u32_e32 v6, vcc, 0, v1, vcc
+; GCN-NEXT:    v_mov_b32_e32 v13, 0
+; GCN-NEXT:    v_mac_f32_e32 v3, 0x4f800000, v4
+; GCN-NEXT:    v_rcp_f32_e32 v3, v3
+; GCN-NEXT:    v_mov_b32_e32 v12, 0
+; GCN-NEXT:    v_mul_f32_e32 v3, 0x5f7ffffc, v3
+; GCN-NEXT:    v_mul_f32_e32 v4, 0x2f800000, v3
+; GCN-NEXT:    v_trunc_f32_e32 v4, v4
+; GCN-NEXT:    v_mac_f32_e32 v3, 0xcf800000, v4
+; GCN-NEXT:    v_cvt_u32_f32_e32 v3, v3
+; GCN-NEXT:    v_cvt_u32_f32_e32 v4, v4
+; GCN-NEXT:    v_mul_hi_u32 v7, v5, v3
+; GCN-NEXT:    v_mul_lo_u32 v8, v5, v4
+; GCN-NEXT:    v_mul_lo_u32 v9, v6, v3
 ; GCN-NEXT:    v_add_i32_e32 v7, vcc, v7, v8
-; GCN-NEXT:    v_mul_hi_u32 v8, v9, v7
-; GCN-NEXT:    v_mul_hi_u32 v11, v5, v7
-; GCN-NEXT:    v_mul_lo_u32 v14, v5, v7
-; GCN-NEXT:    v_mul_lo_u32 v7, v9, v7
-; GCN-NEXT:    v_add_i32_e32 v9, vcc, v12, v14
-; GCN-NEXT:    v_addc_u32_e32 v11, vcc, v4, v11, vcc
-; GCN-NEXT:    v_add_i32_e32 v9, vcc, v13, v9
-; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v11, v10, vcc
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v8, v3, vcc
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v9, v7
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v4, v8, vcc
-; GCN-NEXT:    v_addc_u32_e64 v6, vcc, v6, v8, s[4:5]
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v7
-; GCN-NEXT:    v_addc_u32_e32 v6, vcc, 0, v6, vcc
-; GCN-NEXT:    v_mul_hi_u32 v7, s6, v5
-; GCN-NEXT:    v_mul_hi_u32 v5, 0, v5
-; GCN-NEXT:    v_mul_hi_u32 v8, s6, v6
-; GCN-NEXT:    v_lshlrev_b32_e32 v9, 15, v6
-; GCN-NEXT:    v_mul_hi_u32 v6, 0, v6
+; GCN-NEXT:    v_mul_lo_u32 v8, v5, v3
 ; GCN-NEXT:    v_add_i32_e32 v7, vcc, v7, v9
-; GCN-NEXT:    v_addc_u32_e32 v4, vcc, v4, v8, vcc
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, 0, v7
-; GCN-NEXT:    v_addc_u32_e32 v4, vcc, v4, v5, vcc
-; GCN-NEXT:    v_addc_u32_e32 v3, vcc, v6, v3, vcc
-; GCN-NEXT:    v_mul_hi_u32 v5, v0, v4
-; GCN-NEXT:    v_mul_lo_u32 v6, v1, v4
-; GCN-NEXT:    v_mul_lo_u32 v7, v0, v4
-; GCN-NEXT:    v_mul_lo_u32 v8, v0, v3
-; GCN-NEXT:    v_add_i32_e32 v9, vcc, 2, v4
-; GCN-NEXT:    v_addc_u32_e32 v10, vcc, 0, v3, vcc
-; GCN-NEXT:    v_add_i32_e32 v11, vcc, 1, v4
-; GCN-NEXT:    v_addc_u32_e32 v12, vcc, 0, v3, vcc
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v8
+; GCN-NEXT:    v_mul_lo_u32 v10, v3, v7
+; GCN-NEXT:    v_mul_hi_u32 v9, v3, v7
+; GCN-NEXT:    v_mul_hi_u32 v11, v3, v8
+; GCN-NEXT:    v_mul_hi_u32 v14, v4, v7
+; GCN-NEXT:    v_mul_lo_u32 v7, v4, v7
+; GCN-NEXT:    v_add_i32_e32 v10, vcc, v11, v10
+; GCN-NEXT:    v_mul_lo_u32 v11, v4, v8
+; GCN-NEXT:    v_mul_hi_u32 v8, v4, v8
+; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v13, v9, vcc
+; GCN-NEXT:    v_add_i32_e32 v10, vcc, v11, v10
+; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v9, v8, vcc
+; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v14, v12, vcc
+; GCN-NEXT:    v_add_i32_e32 v7, vcc, v8, v7
+; GCN-NEXT:    v_add_i32_e64 v3, s[4:5], v3, v7
+; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v13, v9, vcc
+; GCN-NEXT:    v_addc_u32_e64 v7, vcc, v4, v8, s[4:5]
+; GCN-NEXT:    v_mul_lo_u32 v9, v5, v7
+; GCN-NEXT:    v_mul_hi_u32 v10, v5, v3
+; GCN-NEXT:    v_mul_lo_u32 v6, v6, v3
+; GCN-NEXT:    v_mul_lo_u32 v5, v5, v3
+; GCN-NEXT:    v_add_i32_e32 v9, vcc, v10, v9
+; GCN-NEXT:    v_add_i32_e32 v6, vcc, v9, v6
+; GCN-NEXT:    v_mul_lo_u32 v11, v3, v6
+; GCN-NEXT:    v_mul_hi_u32 v14, v3, v5
+; GCN-NEXT:    v_mul_hi_u32 v15, v3, v6
+; GCN-NEXT:    v_mul_hi_u32 v10, v7, v5
+; GCN-NEXT:    v_mul_lo_u32 v5, v7, v5
+; GCN-NEXT:    v_add_i32_e32 v11, vcc, v14, v11
+; GCN-NEXT:    v_mul_hi_u32 v9, v7, v6
+; GCN-NEXT:    v_addc_u32_e32 v14, vcc, v13, v15, vcc
+; GCN-NEXT:    v_mul_lo_u32 v6, v7, v6
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v11
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v14, v10, vcc
+; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v9, v12, vcc
 ; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v6
-; GCN-NEXT:    v_sub_i32_e32 v6, vcc, 0, v5
-; GCN-NEXT:    v_sub_i32_e32 v7, vcc, s6, v7
-; GCN-NEXT:    v_subb_u32_e64 v6, s[4:5], v6, v1, vcc
-; GCN-NEXT:    v_sub_i32_e64 v8, s[4:5], v7, v0
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v13, v7, vcc
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v4, v8
+; GCN-NEXT:    v_addc_u32_e64 v4, vcc, v4, v6, s[4:5]
+; GCN-NEXT:    v_add_i32_e32 v3, vcc, v3, v5
+; GCN-NEXT:    s_mov_b32 s4, 0x8000
+; GCN-NEXT:    v_addc_u32_e32 v4, vcc, 0, v4, vcc
+; GCN-NEXT:    v_mul_hi_u32 v5, s4, v3
+; GCN-NEXT:    v_mul_hi_u32 v6, s4, v4
+; GCN-NEXT:    v_lshlrev_b32_e32 v7, 15, v4
+; GCN-NEXT:    v_mul_hi_u32 v3, 0, v3
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v7
+; GCN-NEXT:    v_mul_hi_u32 v4, 0, v4
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v13, v6, vcc
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, 0, v5
+; GCN-NEXT:    v_addc_u32_e32 v3, vcc, v6, v3, vcc
+; GCN-NEXT:    v_addc_u32_e32 v4, vcc, v4, v12, vcc
+; GCN-NEXT:    v_mul_lo_u32 v5, v0, v4
+; GCN-NEXT:    v_mul_hi_u32 v6, v0, v3
+; GCN-NEXT:    v_mul_lo_u32 v7, v1, v3
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v6, v5
+; GCN-NEXT:    v_mul_lo_u32 v6, v0, v3
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v7
+; GCN-NEXT:    v_sub_i32_e32 v7, vcc, 0, v5
+; GCN-NEXT:    v_sub_i32_e32 v6, vcc, s4, v6
+; GCN-NEXT:    v_subb_u32_e64 v7, s[4:5], v7, v1, vcc
+; GCN-NEXT:    v_sub_i32_e64 v8, s[4:5], v6, v0
+; GCN-NEXT:    v_subbrev_u32_e64 v7, s[4:5], 0, v7, s[4:5]
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[4:5], v7, v1
+; GCN-NEXT:    v_cndmask_b32_e64 v9, 0, -1, s[4:5]
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[4:5], v8, v0
+; GCN-NEXT:    v_cndmask_b32_e64 v8, 0, -1, s[4:5]
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[4:5], v7, v1
+; GCN-NEXT:    v_cndmask_b32_e64 v7, v9, v8, s[4:5]
+; GCN-NEXT:    v_add_i32_e64 v8, s[4:5], 2, v3
+; GCN-NEXT:    v_addc_u32_e64 v9, s[4:5], 0, v4, s[4:5]
+; GCN-NEXT:    v_add_i32_e64 v10, s[4:5], 1, v3
 ; GCN-NEXT:    v_subb_u32_e32 v5, vcc, 0, v5, vcc
-; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v7, v0
-; GCN-NEXT:    v_cndmask_b32_e64 v7, 0, -1, vcc
-; GCN-NEXT:    v_subbrev_u32_e64 v6, vcc, 0, v6, s[4:5]
-; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v8, v0
-; GCN-NEXT:    v_cndmask_b32_e64 v0, 0, -1, vcc
+; GCN-NEXT:    v_addc_u32_e64 v11, s[4:5], 0, v4, s[4:5]
+; GCN-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, v7
 ; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v5, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v8, 0, -1, vcc
-; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v6, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v13, 0, -1, vcc
+; GCN-NEXT:    v_cndmask_b32_e64 v7, v11, v9, s[4:5]
+; GCN-NEXT:    v_cndmask_b32_e64 v9, 0, -1, vcc
+; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v6, v0
+; GCN-NEXT:    v_cndmask_b32_e64 v0, 0, -1, vcc
 ; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, v5, v1
-; GCN-NEXT:    v_cndmask_b32_e32 v5, v8, v7, vcc
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, v6, v1
-; GCN-NEXT:    v_cndmask_b32_e32 v0, v13, v0, vcc
+; GCN-NEXT:    v_cndmask_b32_e32 v0, v9, v0, vcc
 ; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
-; GCN-NEXT:    v_cndmask_b32_e32 v0, v12, v10, vcc
-; GCN-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, v5
-; GCN-NEXT:    v_cndmask_b32_e64 v0, v3, v0, s[4:5]
-; GCN-NEXT:    v_cndmask_b32_e32 v1, v11, v9, vcc
-; GCN-NEXT:    v_cndmask_b32_e64 v1, v4, v1, s[4:5]
+; GCN-NEXT:    v_cndmask_b32_e64 v1, v10, v8, s[4:5]
+; GCN-NEXT:    v_cndmask_b32_e32 v0, v4, v7, vcc
+; GCN-NEXT:    v_cndmask_b32_e32 v1, v3, v1, vcc
 ; GCN-NEXT:    v_xor_b32_e32 v3, v0, v2
 ; GCN-NEXT:    v_xor_b32_e32 v0, v1, v2
 ; GCN-NEXT:    v_sub_i32_e32 v0, vcc, v0, v2
 ; GCN-NEXT:    v_subb_u32_e32 v1, vcc, v3, v2, vcc
 ; GCN-NEXT:    s_setpc_b64 s[30:31]
   %result = sdiv i64 32768, %x
   ret i64 %result
 }

 ; GCN-NEXT:    s_setpc_b64 s[30:31]
   %result = sdiv i64 %x, 32768
   ret i64 %result
 }
 
 define amdgpu_kernel void @s_test_sdiv24_k_num_i64(i64 addrspace(1)* %out, i64 %x) {
 ; GCN-LABEL: s_test_sdiv24_k_num_i64:
 ; GCN:       ; %bb.0:
-; GCN-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
-; GCN-NEXT:    s_mov_b32 s7, 0xf000
-; GCN-NEXT:    s_mov_b32 s6, -1
+; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN-NEXT:    s_mov_b32 s3, 0xf000
+; GCN-NEXT:    s_mov_b32 s2, -1
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_mov_b32 s2, 0x41c00000
-; GCN-NEXT:    s_mov_b32 s4, s0
-; GCN-NEXT:    s_mov_b32 s5, s1
-; GCN-NEXT:    s_ashr_i64 s[0:1], s[2:3], 40
-; GCN-NEXT:    v_cvt_f32_i32_e32 v0, s0
-; GCN-NEXT:    s_ashr_i32 s0, s0, 30
+; GCN-NEXT:    s_ashr_i64 s[6:7], s[6:7], 40
+; GCN-NEXT:    v_cvt_f32_i32_e32 v0, s6
+; GCN-NEXT:    s_mov_b32 s7, 0x41c00000
+; GCN-NEXT:    s_mov_b32 s0, s4
+; GCN-NEXT:    s_ashr_i32 s4, s6, 30
 ; GCN-NEXT:    v_rcp_iflag_f32_e32 v1, v0
-; GCN-NEXT:    s_or_b32 s0, s0, 1
-; GCN-NEXT:    v_mul_f32_e32 v1, s2, v1
-; GCN-NEXT:    v_mov_b32_e32 v2, s0
+; GCN-NEXT:    s_or_b32 s4, s4, 1
+; GCN-NEXT:    v_mov_b32_e32 v2, s4
+; GCN-NEXT:    s_mov_b32 s1, s5
+; GCN-NEXT:    v_mul_f32_e32 v1, s7, v1
 ; GCN-NEXT:    v_trunc_f32_e32 v1, v1
-; GCN-NEXT:    v_mad_f32 v3, -v1, v0, s2
+; GCN-NEXT:    v_mad_f32 v3, -v1, v0, s7
+; GCN-NEXT:    v_cvt_i32_f32_e32 v1, v1
 ; GCN-NEXT:    v_cmp_ge_f32_e64 vcc, |v3|, |v0|
 ; GCN-NEXT:    v_cndmask_b32_e32 v0, 0, v2, vcc
-; GCN-NEXT:    v_cvt_i32_f32_e32 v1, v1
 ; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v1
 ; GCN-NEXT:    v_bfe_i32 v0, v0, 0, 24
 ; GCN-NEXT:    v_ashrrev_i32_e32 v1, 31, v0
-; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
+; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 ; GCN-NEXT:    s_endpgm
   %x.shr = ashr i64 %x, 40
   %result = sdiv i64 24, %x.shr
   store i64 %result, i64 addrspace(1)* %out
   ret void
 }
 
 define amdgpu_kernel void @s_test_sdiv24_k_den_i64(i64 addrspace(1)* %out, i64 %x) {
 ; GCN-LABEL: s_test_sdiv24_k_den_i64:
 ; GCN:       ; %bb.0:
-; GCN-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
-; GCN-NEXT:    s_mov_b32 s7, 0xf000
-; GCN-NEXT:    s_mov_b32 s6, -1
+; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN-NEXT:    s_mov_b32 s8, 0x46b6fe00
+; GCN-NEXT:    s_mov_b32 s3, 0xf000
+; GCN-NEXT:    s_mov_b32 s2, -1
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_mov_b32 s2, 0x46b6fe00
-; GCN-NEXT:    s_mov_b32 s4, s0
-; GCN-NEXT:    s_mov_b32 s5, s1
-; GCN-NEXT:    s_ashr_i64 s[0:1], s[2:3], 40
-; GCN-NEXT:    v_cvt_f32_i32_e32 v0, s0
-; GCN-NEXT:    s_ashr_i32 s0, s0, 30
+; GCN-NEXT:    s_ashr_i64 s[6:7], s[6:7], 40
+; GCN-NEXT:    v_cvt_f32_i32_e32 v0, s6
+; GCN-NEXT:    s_mov_b32 s0, s4
+; GCN-NEXT:    s_ashr_i32 s4, s6, 30
+; GCN-NEXT:    s_or_b32 s4, s4, 1
 ; GCN-NEXT:    v_mul_f32_e32 v1, 0x38331158, v0
-; GCN-NEXT:    s_or_b32 s0, s0, 1
 ; GCN-NEXT:    v_trunc_f32_e32 v1, v1
-; GCN-NEXT:    v_mov_b32_e32 v2, s0
-; GCN-NEXT:    v_mad_f32 v0, -v1, s2, v0
-; GCN-NEXT:    v_cmp_ge_f32_e64 vcc, |v0|, s2
-; GCN-NEXT:    v_cndmask_b32_e32 v0, 0, v2, vcc
+; GCN-NEXT:    v_mad_f32 v0, -v1, s8, v0
 ; GCN-NEXT:    v_cvt_i32_f32_e32 v1, v1
+; GCN-NEXT:    v_mov_b32_e32 v2, s4
+; GCN-NEXT:    v_cmp_ge_f32_e64 vcc, |v0|, s8
+; GCN-NEXT:    v_cndmask_b32_e32 v0, 0, v2, vcc
 ; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v1
 ; GCN-NEXT:    v_bfe_i32 v0, v0, 0, 24
+; GCN-NEXT:    s_mov_b32 s1, s5
 ; GCN-NEXT:    v_ashrrev_i32_e32 v1, 31, v0
-; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
+; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 ; GCN-NEXT:    s_endpgm
   %x.shr = ashr i64 %x, 40
   %result = sdiv i64 %x.shr, 23423
   store i64 %result, i64 addrspace(1)* %out
   ret void
 }
 
 define i64 @v_test_sdiv24_k_num_i64(i64 %x) {
 ; GCN-LABEL: v_test_sdiv24_k_num_i64:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GCN-NEXT:    v_ashr_i64 v[0:1], v[0:1], 40
 ; GCN-NEXT:    s_mov_b32 s4, 0x41c00000
 ; GCN-NEXT:    v_cvt_f32_i32_e32 v1, v0
 ; GCN-NEXT:    v_ashrrev_i32_e32 v0, 30, v0
-; GCN-NEXT:    v_rcp_iflag_f32_e32 v2, v1
 ; GCN-NEXT:    v_or_b32_e32 v0, 1, v0
+; GCN-NEXT:    v_rcp_iflag_f32_e32 v2, v1
 ; GCN-NEXT:    v_mul_f32_e32 v2, s4, v2
 ; GCN-NEXT:    v_trunc_f32_e32 v2, v2
 ; GCN-NEXT:    v_mad_f32 v3, -v2, v1, s4
+; GCN-NEXT:    v_cvt_i32_f32_e32 v2, v2
 ; GCN-NEXT:    v_cmp_ge_f32_e64 vcc, |v3|, |v1|
 ; GCN-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
-; GCN-NEXT:    v_cvt_i32_f32_e32 v1, v2
-; GCN-NEXT:    v_add_i32_e32 v0, vcc, v1, v0
+; GCN-NEXT:    v_add_i32_e32 v0, vcc, v2, v0
 ; GCN-NEXT:    v_bfe_i32 v0, v0, 0, 24
 ; GCN-NEXT:    v_ashrrev_i32_e32 v1, 31, v0
 ; GCN-NEXT:    s_setpc_b64 s[30:31]
   %x.shr = ashr i64 %x, 40
   %result = sdiv i64 24, %x.shr
   ret i64 %result
 }
 
 define i64 @v_test_sdiv24_pow2_k_num_i64(i64 %x) {
 ; GCN-LABEL: v_test_sdiv24_pow2_k_num_i64:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GCN-NEXT:    v_ashr_i64 v[0:1], v[0:1], 40
 ; GCN-NEXT:    s_mov_b32 s4, 0x47000000
 ; GCN-NEXT:    v_cvt_f32_i32_e32 v1, v0
 ; GCN-NEXT:    v_ashrrev_i32_e32 v0, 30, v0
-; GCN-NEXT:    v_rcp_iflag_f32_e32 v2, v1
 ; GCN-NEXT:    v_or_b32_e32 v0, 1, v0
+; GCN-NEXT:    v_rcp_iflag_f32_e32 v2, v1
 ; GCN-NEXT:    v_mul_f32_e32 v2, s4, v2
 ; GCN-NEXT:    v_trunc_f32_e32 v2, v2
 ; GCN-NEXT:    v_mad_f32 v3, -v2, v1, s4
+; GCN-NEXT:    v_cvt_i32_f32_e32 v2, v2
 ; GCN-NEXT:    v_cmp_ge_f32_e64 vcc, |v3|, |v1|
 ; GCN-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
-; GCN-NEXT:    v_cvt_i32_f32_e32 v1, v2
-; GCN-NEXT:    v_add_i32_e32 v0, vcc, v1, v0
+; GCN-NEXT:    v_add_i32_e32 v0, vcc, v2, v0
 ; GCN-NEXT:    v_bfe_i32 v0, v0, 0, 24
 ; GCN-NEXT:    v_ashrrev_i32_e32 v1, 31, v0
 ; GCN-NEXT:    s_setpc_b64 s[30:31]
   %x.shr = ashr i64 %x, 40
   %result = sdiv i64 32768, %x.shr
   ret i64 %result
 }
 

llvm/test/CodeGen/AMDGPU/srem64.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
+; RUN: llc -march=amdgcn -mcpu=gfx600 -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
 
 define amdgpu_kernel void @s_test_srem(i64 addrspace(1)* %out, i64 %x, i64 %y) {
 ; GCN-LABEL: s_test_srem:
 ; GCN:       ; %bb.0:
+; GCN-NEXT:    s_load_dwordx2 s[12:13], s[0:1], 0xd
+; GCN-NEXT:    v_mov_b32_e32 v2, 0
 ; GCN-NEXT:    s_load_dwordx4 s[8:11], s[0:1], 0x9
-; GCN-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xd
 ; GCN-NEXT:    s_mov_b32 s7, 0xf000
 ; GCN-NEXT:    s_mov_b32 s6, -1
-; GCN-NEXT:    v_mov_b32_e32 v0, 0
-; GCN-NEXT:    v_mov_b32_e32 v1, 0
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s12
+; GCN-NEXT:    v_cvt_f32_u32_e32 v1, s13
+; GCN-NEXT:    s_sub_u32 s2, 0, s12
+; GCN-NEXT:    s_subb_u32 s3, 0, s13
 ; GCN-NEXT:    s_mov_b32 s4, s8
+; GCN-NEXT:    v_mac_f32_e32 v0, 0x4f800000, v1
+; GCN-NEXT:    v_rcp_f32_e32 v0, v0
+; GCN-NEXT:    v_mov_b32_e32 v1, 0
 ; GCN-NEXT:    s_mov_b32 s5, s9
-; GCN-NEXT:    v_cvt_f32_u32_e32 v2, s2
-; GCN-NEXT:    v_cvt_f32_u32_e32 v3, s3
-; GCN-NEXT:    s_sub_u32 s8, 0, s2
-; GCN-NEXT:    v_mov_b32_e32 v4, s3
-; GCN-NEXT:    v_mov_b32_e32 v5, s11
-; GCN-NEXT:    v_mac_f32_e32 v2, 0x4f800000, v3
-; GCN-NEXT:    s_subb_u32 s9, 0, s3
-; GCN-NEXT:    v_rcp_f32_e32 v2, v2
-; GCN-NEXT:    v_mul_f32_e32 v2, 0x5f7ffffc, v2
-; GCN-NEXT:    v_mul_f32_e32 v3, 0x2f800000, v2
+; GCN-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v0
+; GCN-NEXT:    v_mul_f32_e32 v3, 0x2f800000, v0
 ; GCN-NEXT:    v_trunc_f32_e32 v3, v3
-; GCN-NEXT:    v_mac_f32_e32 v2, 0xcf800000, v3
+; GCN-NEXT:    v_mac_f32_e32 v0, 0xcf800000, v3
+; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v0
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v3, v3
-; GCN-NEXT:    v_cvt_u32_f32_e32 v2, v2
-; GCN-NEXT:    v_mul_lo_u32 v6, s8, v3
-; GCN-NEXT:    v_mul_lo_u32 v7, s9, v2
-; GCN-NEXT:    v_mul_hi_u32 v8, s8, v2
-; GCN-NEXT:    v_mul_lo_u32 v9, s8, v2
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, v8, v6
-; GCN-NEXT:    v_mul_hi_u32 v8, v2, v9
-; GCN-NEXT:    v_mul_hi_u32 v10, v3, v9
-; GCN-NEXT:    v_mul_lo_u32 v9, v3, v9
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, v6, v7
-; GCN-NEXT:    v_mul_hi_u32 v7, v2, v6
-; GCN-NEXT:    v_mul_lo_u32 v11, v2, v6
-; GCN-NEXT:    v_mul_hi_u32 v12, v3, v6
-; GCN-NEXT:    v_mul_lo_u32 v6, v3, v6
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v8, v11
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v1, v7, vcc
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v9, v8
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v7, v10, vcc
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v12, v0, vcc
+; GCN-NEXT:    v_mul_hi_u32 v5, s2, v0
+; GCN-NEXT:    v_mul_lo_u32 v4, s2, v3
+; GCN-NEXT:    v_mul_lo_u32 v7, s3, v0
+; GCN-NEXT:    v_mul_lo_u32 v6, s2, v0
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v5, v4
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v4, v7
+; GCN-NEXT:    v_mul_hi_u32 v5, v0, v6
+; GCN-NEXT:    v_mul_lo_u32 v7, v0, v4
+; GCN-NEXT:    v_mul_hi_u32 v9, v0, v4
+; GCN-NEXT:    v_mul_lo_u32 v8, v3, v6
+; GCN-NEXT:    v_mul_hi_u32 v6, v3, v6
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v7
+; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v2, v9, vcc
+; GCN-NEXT:    v_mul_hi_u32 v9, v3, v4
+; GCN-NEXT:    v_mul_lo_u32 v4, v3, v4
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v8, v5
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v7, v6, vcc
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v9, v1, vcc
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v5, v4
+; GCN-NEXT:    v_add_i32_e64 v0, s[0:1], v0, v4
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v2, v6, vcc
+; GCN-NEXT:    v_addc_u32_e64 v4, vcc, v3, v5, s[0:1]
+; GCN-NEXT:    v_mul_lo_u32 v6, s2, v4
+; GCN-NEXT:    v_mul_hi_u32 v7, s2, v0
+; GCN-NEXT:    v_mul_lo_u32 v8, s3, v0
 ; GCN-NEXT:    v_add_i32_e32 v6, vcc, v7, v6
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v1, v8, vcc
-; GCN-NEXT:    v_add_i32_e64 v2, s[0:1], v2, v6
-; GCN-NEXT:    v_addc_u32_e64 v6, vcc, v3, v7, s[0:1]
-; GCN-NEXT:    v_mul_hi_u32 v8, s8, v2
-; GCN-NEXT:    v_mul_lo_u32 v9, s9, v2
-; GCN-NEXT:    v_mul_lo_u32 v10, s8, v2
-; GCN-NEXT:    v_mul_lo_u32 v11, s8, v6
-; GCN-NEXT:    v_mul_hi_u32 v12, v6, v10
-; GCN-NEXT:    v_mul_lo_u32 v13, v6, v10
-; GCN-NEXT:    v_mul_hi_u32 v10, v2, v10
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v8, v11
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v9, v8
-; GCN-NEXT:    v_mul_hi_u32 v9, v6, v8
-; GCN-NEXT:    v_mul_hi_u32 v11, v2, v8
-; GCN-NEXT:    v_mul_lo_u32 v14, v2, v8
-; GCN-NEXT:    v_mul_lo_u32 v6, v6, v8
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v10, v14
-; GCN-NEXT:    v_addc_u32_e32 v10, vcc, v1, v11, vcc
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v13, v8
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v10, v12, vcc
-; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v9, v0, vcc
-; GCN-NEXT:    v_add_i32_e32 v3, vcc, v3, v7
+; GCN-NEXT:    v_mul_lo_u32 v7, s2, v0
 ; GCN-NEXT:    v_add_i32_e32 v6, vcc, v8, v6
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v1, v9, vcc
-; GCN-NEXT:    v_addc_u32_e64 v3, vcc, v3, v7, s[0:1]
-; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v6
+; GCN-NEXT:    v_mul_lo_u32 v10, v0, v6
+; GCN-NEXT:    v_mul_hi_u32 v12, v0, v6
+; GCN-NEXT:    v_mul_hi_u32 v11, v0, v7
+; GCN-NEXT:    v_mul_hi_u32 v9, v4, v7
+; GCN-NEXT:    v_mul_lo_u32 v7, v4, v7
+; GCN-NEXT:    v_mul_hi_u32 v8, v4, v6
+; GCN-NEXT:    v_add_i32_e32 v10, vcc, v11, v10
+; GCN-NEXT:    v_addc_u32_e32 v11, vcc, v2, v12, vcc
+; GCN-NEXT:    v_mul_lo_u32 v4, v4, v6
+; GCN-NEXT:    v_add_i32_e32 v7, vcc, v7, v10
+; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v11, v9, vcc
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v8, v1, vcc
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v7, v4
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v2, v6, vcc
+; GCN-NEXT:    v_add_i32_e32 v3, vcc, v3, v5
+; GCN-NEXT:    v_addc_u32_e64 v3, vcc, v3, v6, s[0:1]
+; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v4
 ; GCN-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; GCN-NEXT:    v_mul_hi_u32 v6, s10, v2
-; GCN-NEXT:    v_mul_hi_u32 v7, s11, v2
-; GCN-NEXT:    v_mul_lo_u32 v2, s11, v2
-; GCN-NEXT:    v_mul_hi_u32 v8, s10, v3
-; GCN-NEXT:    v_mul_lo_u32 v9, s10, v3
-; GCN-NEXT:    v_mul_hi_u32 v10, s11, v3
+; GCN-NEXT:    v_mul_lo_u32 v4, s10, v3
+; GCN-NEXT:    v_mul_hi_u32 v5, s10, v0
+; GCN-NEXT:    v_mul_hi_u32 v6, s10, v3
+; GCN-NEXT:    v_mul_hi_u32 v7, s11, v3
 ; GCN-NEXT:    v_mul_lo_u32 v3, s11, v3
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, v6, v9
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v1, v8, vcc
-; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v6
-; GCN-NEXT:    v_addc_u32_e32 v2, vcc, v8, v7, vcc
-; GCN-NEXT:    v_addc_u32_e32 v0, vcc, v10, v0, vcc
-; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v3
-; GCN-NEXT:    v_addc_u32_e32 v0, vcc, v1, v0, vcc
-; GCN-NEXT:    v_mul_hi_u32 v1, s2, v2
-; GCN-NEXT:    v_mul_lo_u32 v3, s3, v2
-; GCN-NEXT:    v_mul_lo_u32 v2, s2, v2
-; GCN-NEXT:    v_mul_lo_u32 v0, s2, v0
-; GCN-NEXT:    v_add_i32_e32 v0, vcc, v1, v0
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v5, v4
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v2, v6, vcc
+; GCN-NEXT:    v_mul_lo_u32 v6, s11, v0
+; GCN-NEXT:    v_mul_hi_u32 v0, s11, v0
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v6, v4
+; GCN-NEXT:    v_addc_u32_e32 v0, vcc, v5, v0, vcc
+; GCN-NEXT:    v_addc_u32_e32 v1, vcc, v7, v1, vcc
 ; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v3
-; GCN-NEXT:    v_sub_i32_e32 v1, vcc, s11, v0
-; GCN-NEXT:    v_sub_i32_e32 v2, vcc, s10, v2
-; GCN-NEXT:    v_subb_u32_e64 v1, s[0:1], v1, v4, vcc
-; GCN-NEXT:    v_subb_u32_e32 v0, vcc, v5, v0, vcc
-; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s2, v2
+; GCN-NEXT:    v_addc_u32_e32 v1, vcc, v2, v1, vcc
+; GCN-NEXT:    v_mul_lo_u32 v1, s12, v1
+; GCN-NEXT:    v_mul_hi_u32 v2, s12, v0
+; GCN-NEXT:    v_mul_lo_u32 v3, s13, v0
+; GCN-NEXT:    v_mul_lo_u32 v0, s12, v0
+; GCN-NEXT:    v_add_i32_e32 v1, vcc, v2, v1
+; GCN-NEXT:    v_add_i32_e32 v1, vcc, v1, v3
+; GCN-NEXT:    v_sub_i32_e32 v2, vcc, s11, v1
+; GCN-NEXT:    v_mov_b32_e32 v3, s13
+; GCN-NEXT:    v_sub_i32_e32 v0, vcc, s10, v0
+; GCN-NEXT:    v_subb_u32_e64 v2, s[0:1], v2, v3, vcc
+; GCN-NEXT:    v_subrev_i32_e64 v4, s[0:1], s12, v0
+; GCN-NEXT:    v_subb_u32_e64 v3, s[2:3], v2, v3, s[0:1]
+; GCN-NEXT:    v_subbrev_u32_e64 v2, s[0:1], 0, v2, s[0:1]
+; GCN-NEXT:    v_cmp_le_u32_e64 s[0:1], s13, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v5, 0, -1, s[0:1]
+; GCN-NEXT:    v_cmp_le_u32_e64 s[0:1], s12, v4
+; GCN-NEXT:    v_cndmask_b32_e64 v6, 0, -1, s[0:1]
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], s13, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v5, v5, v6, s[0:1]
+; GCN-NEXT:    v_subrev_i32_e64 v6, s[0:1], s12, v4
+; GCN-NEXT:    v_subbrev_u32_e64 v3, s[0:1], 0, v3, s[0:1]
+; GCN-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v5
+; GCN-NEXT:    v_cndmask_b32_e64 v2, v2, v3, s[0:1]
+; GCN-NEXT:    v_mov_b32_e32 v3, s11
+; GCN-NEXT:    v_subb_u32_e32 v1, vcc, v3, v1, vcc
+; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s13, v1
 ; GCN-NEXT:    v_cndmask_b32_e64 v3, 0, -1, vcc
-; GCN-NEXT:    v_subrev_i32_e32 v5, vcc, s2, v2
-; GCN-NEXT:    v_subb_u32_e64 v4, s[0:1], v1, v4, vcc
-; GCN-NEXT:    v_subbrev_u32_e32 v1, vcc, 0, v1, vcc
-; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s2, v5
-; GCN-NEXT:    v_cndmask_b32_e64 v6, 0, -1, vcc
-; GCN-NEXT:    v_subrev_i32_e32 v7, vcc, s2, v5
-; GCN-NEXT:    v_cmp_le_u32_e64 s[0:1], s3, v0
-; GCN-NEXT:    v_cndmask_b32_e64 v8, 0, -1, s[0:1]
-; GCN-NEXT:    v_cmp_le_u32_e64 s[0:1], s3, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v9, 0, -1, s[0:1]
-; GCN-NEXT:    v_subbrev_u32_e32 v4, vcc, 0, v4, vcc
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, s3, v0
-; GCN-NEXT:    v_cndmask_b32_e32 v3, v8, v3, vcc
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, s3, v1
-; GCN-NEXT:    v_cndmask_b32_e32 v6, v9, v6, vcc
-; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v6
-; GCN-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
-; GCN-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v3
-; GCN-NEXT:    v_cndmask_b32_e64 v1, v0, v1, s[0:1]
-; GCN-NEXT:    v_cndmask_b32_e32 v0, v5, v7, vcc
-; GCN-NEXT:    v_cndmask_b32_e64 v0, v2, v0, s[0:1]
+; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s12, v0
+; GCN-NEXT:    v_cndmask_b32_e64 v5, 0, -1, vcc
+; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, s13, v1
+; GCN-NEXT:    v_cndmask_b32_e32 v3, v3, v5, vcc
+; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v3
+; GCN-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc
+; GCN-NEXT:    v_cndmask_b32_e64 v2, v4, v6, s[0:1]
+; GCN-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
 ; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
 ; GCN-NEXT:    s_endpgm
   %result = urem i64 %x, %y
   store i64 %result, i64 addrspace(1)* %out
   ret void
 }
 
 define i64 @v_test_srem(i64 %x, i64 %y) {
 ; GCN-LABEL: v_test_srem:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GCN-NEXT:    v_ashrrev_i32_e32 v4, 31, v3
-; GCN-NEXT:    v_mov_b32_e32 v5, 0
-; GCN-NEXT:    v_mov_b32_e32 v6, 0
-; GCN-NEXT:    v_ashrrev_i32_e32 v7, 31, v1
 ; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v4
 ; GCN-NEXT:    v_addc_u32_e32 v3, vcc, v3, v4, vcc
-; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v7
-; GCN-NEXT:    v_addc_u32_e32 v1, vcc, v1, v7, vcc
 ; GCN-NEXT:    v_xor_b32_e32 v3, v3, v4
 ; GCN-NEXT:    v_xor_b32_e32 v2, v2, v4
-; GCN-NEXT:    v_xor_b32_e32 v1, v1, v7
-; GCN-NEXT:    v_xor_b32_e32 v0, v0, v7
 ; GCN-NEXT:    v_cvt_f32_u32_e32 v4, v2
-; GCN-NEXT:    v_cvt_f32_u32_e32 v8, v3
-; GCN-NEXT:    v_sub_i32_e32 v9, vcc, 0, v2
-; GCN-NEXT:    v_subb_u32_e32 v10, vcc, 0, v3, vcc
-; GCN-NEXT:    v_mac_f32_e32 v4, 0x4f800000, v8
+; GCN-NEXT:    v_cvt_f32_u32_e32 v5, v3
+; GCN-NEXT:    v_sub_i32_e32 v6, vcc, 0, v2
+; GCN-NEXT:    v_subb_u32_e32 v7, vcc, 0, v3, vcc
+; GCN-NEXT:    v_mov_b32_e32 v14, 0
+; GCN-NEXT:    v_mac_f32_e32 v4, 0x4f800000, v5
 ; GCN-NEXT:    v_rcp_f32_e32 v4, v4
+; GCN-NEXT:    v_mov_b32_e32 v13, 0
 ; GCN-NEXT:    v_mul_f32_e32 v4, 0x5f7ffffc, v4
-; GCN-NEXT:    v_mul_f32_e32 v8, 0x2f800000, v4
-; GCN-NEXT:    v_trunc_f32_e32 v8, v8
-; GCN-NEXT:    v_mac_f32_e32 v4, 0xcf800000, v8
-; GCN-NEXT:    v_cvt_u32_f32_e32 v8, v8
+; GCN-NEXT:    v_mul_f32_e32 v5, 0x2f800000, v4
+; GCN-NEXT:    v_trunc_f32_e32 v5, v5
+; GCN-NEXT:    v_mac_f32_e32 v4, 0xcf800000, v5
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v4, v4
-; GCN-NEXT:    v_mul_lo_u32 v11, v9, v8
-; GCN-NEXT:    v_mul_lo_u32 v12, v10, v4
-; GCN-NEXT:    v_mul_hi_u32 v13, v9, v4
-; GCN-NEXT:    v_mul_lo_u32 v14, v9, v4
-; GCN-NEXT:    v_add_i32_e32 v11, vcc, v13, v11
-; GCN-NEXT:    v_mul_hi_u32 v13, v4, v14
-; GCN-NEXT:    v_mul_hi_u32 v15, v8, v14
-; GCN-NEXT:    v_mul_lo_u32 v14, v8, v14
-; GCN-NEXT:    v_add_i32_e32 v11, vcc, v11, v12
-; GCN-NEXT:    v_mul_hi_u32 v12, v4, v11
-; GCN-NEXT:    v_mul_lo_u32 v16, v4, v11
-; GCN-NEXT:    v_mul_hi_u32 v17, v8, v11
-; GCN-NEXT:    v_mul_lo_u32 v11, v8, v11
-; GCN-NEXT:    v_add_i32_e32 v13, vcc, v13, v16
-; GCN-NEXT:    v_addc_u32_e32 v12, vcc, v6, v12, vcc
-; GCN-NEXT:    v_add_i32_e32 v13, vcc, v14, v13
-; GCN-NEXT:    v_addc_u32_e32 v12, vcc, v12, v15, vcc
-; GCN-NEXT:    v_addc_u32_e32 v13, vcc, v17, v5, vcc
+; GCN-NEXT:    v_cvt_u32_f32_e32 v5, v5
+; GCN-NEXT:    v_mul_hi_u32 v8, v6, v4
+; GCN-NEXT:    v_mul_lo_u32 v9, v6, v5
+; GCN-NEXT:    v_mul_lo_u32 v10, v7, v4
+; GCN-NEXT:    v_add_i32_e32 v8, vcc, v8, v9
+; GCN-NEXT:    v_mul_lo_u32 v9, v6, v4
+; GCN-NEXT:    v_add_i32_e32 v8, vcc, v8, v10
+; GCN-NEXT:    v_mul_lo_u32 v11, v4, v8
+; GCN-NEXT:    v_mul_hi_u32 v10, v4, v8
+; GCN-NEXT:    v_mul_hi_u32 v12, v4, v9
+; GCN-NEXT:    v_mul_hi_u32 v15, v5, v8
+; GCN-NEXT:    v_mul_lo_u32 v8, v5, v8
 ; GCN-NEXT:    v_add_i32_e32 v11, vcc, v12, v11
-; GCN-NEXT:    v_addc_u32_e32 v12, vcc, v6, v13, vcc
-; GCN-NEXT:    v_add_i32_e64 v4, s[4:5], v4, v11
-; GCN-NEXT:    v_addc_u32_e64 v11, vcc, v8, v12, s[4:5]
-; GCN-NEXT:    v_mul_hi_u32 v13, v9, v4
-; GCN-NEXT:    v_mul_lo_u32 v10, v10, v4
-; GCN-NEXT:    v_mul_lo_u32 v14, v9, v4
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v8, v12
-; GCN-NEXT:    v_mul_lo_u32 v9, v9, v11
-; GCN-NEXT:    v_mul_hi_u32 v12, v11, v14
-; GCN-NEXT:    v_mul_lo_u32 v15, v11, v14
-; GCN-NEXT:    v_mul_hi_u32 v14, v4, v14
-; GCN-NEXT:    v_add_i32_e32 v9, vcc, v13, v9
-; GCN-NEXT:    v_add_i32_e32 v9, vcc, v9, v10
-; GCN-NEXT:    v_mul_hi_u32 v10, v11, v9
-; GCN-NEXT:    v_mul_hi_u32 v13, v4, v9
-; GCN-NEXT:    v_mul_lo_u32 v16, v4, v9
-; GCN-NEXT:    v_mul_lo_u32 v9, v11, v9
-; GCN-NEXT:    v_add_i32_e32 v11, vcc, v14, v16
-; GCN-NEXT:    v_addc_u32_e32 v13, vcc, v6, v13, vcc
-; GCN-NEXT:    v_add_i32_e32 v11, vcc, v15, v11
-; GCN-NEXT:    v_addc_u32_e32 v11, vcc, v13, v12, vcc
-; GCN-NEXT:    v_addc_u32_e32 v10, vcc, v10, v5, vcc
-; GCN-NEXT:    v_add_i32_e32 v9, vcc, v11, v9
-; GCN-NEXT:    v_addc_u32_e32 v10, vcc, v6, v10, vcc
-; GCN-NEXT:    v_addc_u32_e64 v8, vcc, v8, v10, s[4:5]
-; GCN-NEXT:    v_add_i32_e32 v4, vcc, v4, v9
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, 0, v8, vcc
-; GCN-NEXT:    v_mul_hi_u32 v9, v0, v4
-; GCN-NEXT:    v_mul_hi_u32 v10, v1, v4
-; GCN-NEXT:    v_mul_lo_u32 v4, v1, v4
-; GCN-NEXT:    v_mul_hi_u32 v11, v0, v8
-; GCN-NEXT:    v_mul_lo_u32 v12, v0, v8
-; GCN-NEXT:    v_mul_hi_u32 v13, v1, v8
-; GCN-NEXT:    v_mul_lo_u32 v8, v1, v8
-; GCN-NEXT:    v_add_i32_e32 v9, vcc, v9, v12
-; GCN-NEXT:    v_addc_u32_e32 v11, vcc, v6, v11, vcc
-; GCN-NEXT:    v_add_i32_e32 v4, vcc, v4, v9
-; GCN-NEXT:    v_addc_u32_e32 v4, vcc, v11, v10, vcc
-; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v13, v5, vcc
-; GCN-NEXT:    v_add_i32_e32 v4, vcc, v4, v8
-; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v6, v5, vcc
-; GCN-NEXT:    v_mul_hi_u32 v6, v2, v4
+; GCN-NEXT:    v_mul_lo_u32 v12, v5, v9
+; GCN-NEXT:    v_mul_hi_u32 v9, v5, v9
+; GCN-NEXT:    v_addc_u32_e32 v10, vcc, v14, v10, vcc
+; GCN-NEXT:    v_add_i32_e32 v11, vcc, v12, v11
+; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v10, v9, vcc
+; GCN-NEXT:    v_addc_u32_e32 v10, vcc, v15, v13, vcc
+; GCN-NEXT:    v_add_i32_e32 v8, vcc, v9, v8
+; GCN-NEXT:    v_add_i32_e64 v4, s[4:5], v4, v8
+; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v14, v10, vcc
+; GCN-NEXT:    v_addc_u32_e64 v8, vcc, v5, v9, s[4:5]
+; GCN-NEXT:    v_mul_lo_u32 v10, v6, v8
+; GCN-NEXT:    v_mul_hi_u32 v11, v6, v4
+; GCN-NEXT:    v_mul_lo_u32 v7, v7, v4
+; GCN-NEXT:    v_mul_lo_u32 v6, v6, v4
+; GCN-NEXT:    v_add_i32_e32 v10, vcc, v11, v10
+; GCN-NEXT:    v_add_i32_e32 v7, vcc, v10, v7
+; GCN-NEXT:    v_mul_lo_u32 v12, v4, v7
+; GCN-NEXT:    v_mul_hi_u32 v15, v4, v6
+; GCN-NEXT:    v_mul_hi_u32 v16, v4, v7
+; GCN-NEXT:    v_mul_hi_u32 v11, v8, v6
+; GCN-NEXT:    v_mul_lo_u32 v6, v8, v6
+; GCN-NEXT:    v_add_i32_e32 v12, vcc, v15, v12
+; GCN-NEXT:    v_mul_hi_u32 v10, v8, v7
+; GCN-NEXT:    v_addc_u32_e32 v15, vcc, v14, v16, vcc
+; GCN-NEXT:    v_mul_lo_u32 v7, v8, v7
+; GCN-NEXT:    v_add_i32_e32 v6, vcc, v6, v12
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v15, v11, vcc
+; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v10, v13, vcc
+; GCN-NEXT:    v_add_i32_e32 v6, vcc, v6, v7
+; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v14, v8, vcc
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v9
+; GCN-NEXT:    v_addc_u32_e64 v5, vcc, v5, v7, s[4:5]
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v4, v6
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
+; GCN-NEXT:    v_ashrrev_i32_e32 v6, 31, v1
+; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v6
+; GCN-NEXT:    v_xor_b32_e32 v0, v0, v6
+; GCN-NEXT:    v_mul_lo_u32 v7, v0, v5
+; GCN-NEXT:    v_mul_hi_u32 v8, v0, v4
+; GCN-NEXT:    v_mul_hi_u32 v9, v0, v5
+; GCN-NEXT:    v_addc_u32_e32 v1, vcc, v1, v6, vcc
+; GCN-NEXT:    v_xor_b32_e32 v1, v1, v6
+; GCN-NEXT:    v_add_i32_e32 v7, vcc, v8, v7
+; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v14, v9, vcc
+; GCN-NEXT:    v_mul_lo_u32 v9, v1, v4
+; GCN-NEXT:    v_mul_hi_u32 v4, v1, v4
+; GCN-NEXT:    v_mul_hi_u32 v10, v1, v5
+; GCN-NEXT:    v_mul_lo_u32 v5, v1, v5
+; GCN-NEXT:    v_add_i32_e32 v7, vcc, v9, v7
+; GCN-NEXT:    v_addc_u32_e32 v4, vcc, v8, v4, vcc
+; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v10, v13, vcc
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v4, v5
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v14, v7, vcc
+; GCN-NEXT:    v_mul_lo_u32 v5, v2, v5
+; GCN-NEXT:    v_mul_hi_u32 v7, v2, v4
 ; GCN-NEXT:    v_mul_lo_u32 v8, v3, v4
 ; GCN-NEXT:    v_mul_lo_u32 v4, v2, v4
-; GCN-NEXT:    v_mul_lo_u32 v5, v2, v5
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, v6, v5
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v7, v5
 ; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v8
-; GCN-NEXT:    v_sub_i32_e32 v6, vcc, v1, v5
+; GCN-NEXT:    v_sub_i32_e32 v7, vcc, v1, v5
 ; GCN-NEXT:    v_sub_i32_e32 v0, vcc, v0, v4
-; GCN-NEXT:    v_subb_u32_e64 v4, s[4:5], v6, v3, vcc
+; GCN-NEXT:    v_subb_u32_e64 v4, s[4:5], v7, v3, vcc
+; GCN-NEXT:    v_sub_i32_e64 v7, s[4:5], v0, v2
+; GCN-NEXT:    v_subb_u32_e64 v8, s[6:7], v4, v3, s[4:5]
+; GCN-NEXT:    v_subbrev_u32_e64 v4, s[4:5], 0, v4, s[4:5]
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[4:5], v4, v3
 ; GCN-NEXT:    v_subb_u32_e32 v1, vcc, v1, v5, vcc
-; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v0, v2
-; GCN-NEXT:    v_cndmask_b32_e64 v5, 0, -1, vcc
-; GCN-NEXT:    v_sub_i32_e32 v6, vcc, v0, v2
-; GCN-NEXT:    v_subb_u32_e64 v8, s[4:5], v4, v3, vcc
-; GCN-NEXT:    v_subbrev_u32_e32 v4, vcc, 0, v4, vcc
-; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v6, v2
-; GCN-NEXT:    v_cndmask_b32_e64 v9, 0, -1, vcc
-; GCN-NEXT:    v_sub_i32_e32 v2, vcc, v6, v2
-; GCN-NEXT:    v_cmp_ge_u32_e64 s[4:5], v1, v3
+; GCN-NEXT:    v_cndmask_b32_e64 v9, 0, -1, s[4:5]
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[4:5], v7, v2
 ; GCN-NEXT:    v_cndmask_b32_e64 v10, 0, -1, s[4:5]
-; GCN-NEXT:    v_cmp_ge_u32_e64 s[4:5], v4, v3
-; GCN-NEXT:    v_cndmask_b32_e64 v11, 0, -1, s[4:5]
-; GCN-NEXT:    v_subbrev_u32_e32 v8, vcc, 0, v8, vcc
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[4:5], v4, v3
+; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v1, v3
+; GCN-NEXT:    v_cndmask_b32_e64 v9, v9, v10, s[4:5]
+; GCN-NEXT:    v_sub_i32_e64 v10, s[4:5], v7, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v5, 0, -1, vcc
+; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v0, v2
+; GCN-NEXT:    v_subbrev_u32_e64 v8, s[4:5], 0, v8, s[4:5]
+; GCN-NEXT:    v_cndmask_b32_e64 v2, 0, -1, vcc
 ; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, v1, v3
-; GCN-NEXT:    v_cndmask_b32_e32 v5, v10, v5, vcc
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
-; GCN-NEXT:    v_cndmask_b32_e32 v3, v11, v9, vcc
-; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v3
-; GCN-NEXT:    v_cndmask_b32_e32 v3, v4, v8, vcc
-; GCN-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, v5
-; GCN-NEXT:    v_cndmask_b32_e64 v1, v1, v3, s[4:5]
-; GCN-NEXT:    v_cndmask_b32_e32 v2, v6, v2, vcc
-; GCN-NEXT:    v_cndmask_b32_e64 v0, v0, v2, s[4:5]
-; GCN-NEXT:    v_xor_b32_e32 v1, v1, v7
-; GCN-NEXT:    v_xor_b32_e32 v0, v0, v7
-; GCN-NEXT:    v_sub_i32_e32 v0, vcc, v0, v7
-; GCN-NEXT:    v_subb_u32_e32 v1, vcc, v1, v7, vcc
+; GCN-NEXT:    v_cndmask_b32_e32 v2, v5, v2, vcc
+; GCN-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, v9
+; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v2, v7, v10, s[4:5]
+; GCN-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
+; GCN-NEXT:    v_cndmask_b32_e64 v4, v4, v8, s[4:5]
+; GCN-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; GCN-NEXT:    v_xor_b32_e32 v0, v0, v6
+; GCN-NEXT:    v_xor_b32_e32 v1, v1, v6
+; GCN-NEXT:    v_sub_i32_e32 v0, vcc, v0, v6
+; GCN-NEXT:    v_subb_u32_e32 v1, vcc, v1, v6, vcc
 ; GCN-NEXT:    s_setpc_b64 s[30:31]
   %result = srem i64 %x, %y
   ret i64 %result
 }
 
 define amdgpu_kernel void @s_test_srem23_64(i64 addrspace(1)* %out, i64 %x, i64 %y) {
 ; GCN-LABEL: s_test_srem23_64:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; GCN-NEXT:    s_load_dword s9, s[0:1], 0xe
+; GCN-NEXT:    s_load_dword s1, s[0:1], 0xe
 ; GCN-NEXT:    s_mov_b32 s3, 0xf000
 ; GCN-NEXT:    s_mov_b32 s2, -1
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_mov_b32 s0, s4
-; GCN-NEXT:    s_mov_b32 s1, s5
-; GCN-NEXT:    s_ashr_i64 s[4:5], s[6:7], 41
-; GCN-NEXT:    s_ashr_i64 s[6:7], s[8:9], 41
-; GCN-NEXT:    s_xor_b32 s5, s4, s6
-; GCN-NEXT:    v_cvt_f32_i32_e32 v0, s6
-; GCN-NEXT:    v_cvt_f32_i32_e32 v1, s4
-; GCN-NEXT:    s_ashr_i32 s5, s5, 30
+; GCN-NEXT:    s_ashr_i64 s[6:7], s[6:7], 41
+; GCN-NEXT:    s_ashr_i64 s[0:1], s[0:1], 41
+; GCN-NEXT:    v_cvt_f32_i32_e32 v0, s0
+; GCN-NEXT:    v_cvt_f32_i32_e32 v1, s6
+; GCN-NEXT:    s_xor_b32 s1, s6, s0
+; GCN-NEXT:    s_ashr_i32 s1, s1, 30
 ; GCN-NEXT:    v_rcp_iflag_f32_e32 v2, v0
-; GCN-NEXT:    s_or_b32 s5, s5, 1
+; GCN-NEXT:    s_or_b32 s1, s1, 1
+; GCN-NEXT:    v_mov_b32_e32 v3, s1
+; GCN-NEXT:    s_mov_b32 s1, s5
 ; GCN-NEXT:    v_mul_f32_e32 v2, v1, v2
 ; GCN-NEXT:    v_trunc_f32_e32 v2, v2
-; GCN-NEXT:    v_mov_b32_e32 v3, s5
 ; GCN-NEXT:    v_mad_f32 v1, -v2, v0, v1
+; GCN-NEXT:    v_cvt_i32_f32_e32 v2, v2
 ; GCN-NEXT:    v_cmp_ge_f32_e64 vcc, |v1|, |v0|
 ; GCN-NEXT:    v_cndmask_b32_e32 v0, 0, v3, vcc
-; GCN-NEXT:    v_cvt_i32_f32_e32 v1, v2
-; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v1
-; GCN-NEXT:    v_mul_lo_u32 v0, v0, s6
-; GCN-NEXT:    v_sub_i32_e32 v0, vcc, s4, v0
+; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v2
+; GCN-NEXT:    v_mul_lo_u32 v0, v0, s0
+; GCN-NEXT:    s_mov_b32 s0, s4
+; GCN-NEXT:    v_sub_i32_e32 v0, vcc, s6, v0
 ; GCN-NEXT:    v_bfe_i32 v0, v0, 0, 23
 ; GCN-NEXT:    v_ashrrev_i32_e32 v1, 31, v0
 ; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 ; GCN-NEXT:    s_endpgm
   %1 = ashr i64 %x, 41
   %2 = ashr i64 %y, 41
   %result = srem i64 %1, %2
   store i64 %result, i64 addrspace(1)* %out
   ret void
 }
 
 define amdgpu_kernel void @s_test_srem24_64(i64 addrspace(1)* %out, i64 %x, i64 %y) {
 ; GCN-LABEL: s_test_srem24_64:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; GCN-NEXT:    s_load_dword s9, s[0:1], 0xe
+; GCN-NEXT:    s_load_dword s1, s[0:1], 0xe
 ; GCN-NEXT:    s_mov_b32 s3, 0xf000
 ; GCN-NEXT:    s_mov_b32 s2, -1
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_mov_b32 s0, s4
-; GCN-NEXT:    s_mov_b32 s1, s5
-; GCN-NEXT:    s_ashr_i64 s[4:5], s[6:7], 40
-; GCN-NEXT:    s_ashr_i64 s[6:7], s[8:9], 40
-; GCN-NEXT:    s_xor_b32 s5, s4, s6
-; GCN-NEXT:    v_cvt_f32_i32_e32 v0, s6
-; GCN-NEXT:    v_cvt_f32_i32_e32 v1, s4
-; GCN-NEXT:    s_ashr_i32 s5, s5, 30
+; GCN-NEXT:    s_ashr_i64 s[6:7], s[6:7], 40
+; GCN-NEXT:    s_ashr_i64 s[0:1], s[0:1], 40
+; GCN-NEXT:    v_cvt_f32_i32_e32 v0, s0
+; GCN-NEXT:    v_cvt_f32_i32_e32 v1, s6
+; GCN-NEXT:    s_xor_b32 s1, s6, s0
+; GCN-NEXT:    s_ashr_i32 s1, s1, 30
 ; GCN-NEXT:    v_rcp_iflag_f32_e32 v2, v0
-; GCN-NEXT:    s_or_b32 s5, s5, 1
+; GCN-NEXT:    s_or_b32 s1, s1, 1
+; GCN-NEXT:    v_mov_b32_e32 v3, s1
+; GCN-NEXT:    s_mov_b32 s1, s5
 ; GCN-NEXT:    v_mul_f32_e32 v2, v1, v2
 ; GCN-NEXT:    v_trunc_f32_e32 v2, v2
-; GCN-NEXT:    v_mov_b32_e32 v3, s5
 ; GCN-NEXT:    v_mad_f32 v1, -v2, v0, v1
+; GCN-NEXT:    v_cvt_i32_f32_e32 v2, v2
 ; GCN-NEXT:    v_cmp_ge_f32_e64 vcc, |v1|, |v0|
 ; GCN-NEXT:    v_cndmask_b32_e32 v0, 0, v3, vcc
-; GCN-NEXT:    v_cvt_i32_f32_e32 v1, v2
-; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v1
-; GCN-NEXT:    v_mul_lo_u32 v0, v0, s6
-; GCN-NEXT:    v_sub_i32_e32 v0, vcc, s4, v0
+; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v2
+; GCN-NEXT:    v_mul_lo_u32 v0, v0, s0
+; GCN-NEXT:    s_mov_b32 s0, s4
+; GCN-NEXT:    v_sub_i32_e32 v0, vcc, s6, v0
 ; GCN-NEXT:    v_bfe_i32 v0, v0, 0, 24
 ; GCN-NEXT:    v_ashrrev_i32_e32 v1, 31, v0
 ; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 ; GCN-NEXT:    s_endpgm
   %1 = ashr i64 %x, 40
   %2 = ashr i64 %y, 40
   %result = srem i64 %1, %2
   store i64 %result, i64 addrspace(1)* %out
   ret void
 }
 
 define i64 @v_test_srem24_64(i64 %x, i64 %y) {
 ; GCN-LABEL: v_test_srem24_64:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT:    v_ashr_i64 v[2:3], v[2:3], 40
 ; GCN-NEXT:    v_ashr_i64 v[0:1], v[0:1], 40
-; GCN-NEXT:    v_ashr_i64 v[1:2], v[2:3], 40
-; GCN-NEXT:    v_xor_b32_e32 v2, v0, v1
-; GCN-NEXT:    v_cvt_f32_i32_e32 v3, v0
-; GCN-NEXT:    v_cvt_f32_i32_e32 v4, v1
-; GCN-NEXT:    v_ashrrev_i32_e32 v2, 30, v2
-; GCN-NEXT:    v_rcp_iflag_f32_e32 v5, v4
-; GCN-NEXT:    v_or_b32_e32 v2, 1, v2
-; GCN-NEXT:    v_mul_f32_e32 v5, v3, v5
-; GCN-NEXT:    v_trunc_f32_e32 v5, v5
-; GCN-NEXT:    v_mad_f32 v3, -v5, v4, v3
-; GCN-NEXT:    v_cmp_ge_f32_e64 vcc, |v3|, |v4|
-; GCN-NEXT:    v_cndmask_b32_e32 v2, 0, v2, vcc
-; GCN-NEXT:    v_cvt_i32_f32_e32 v3, v5
-; GCN-NEXT:    v_add_i32_e32 v2, vcc, v3, v2
-; GCN-NEXT:    v_mul_lo_u32 v1, v2, v1
+; GCN-NEXT:    v_cvt_f32_i32_e32 v3, v2
+; GCN-NEXT:    v_cvt_f32_i32_e32 v1, v0
+; GCN-NEXT:    v_xor_b32_e32 v5, v0, v2
+; GCN-NEXT:    v_ashrrev_i32_e32 v5, 30, v5
+; GCN-NEXT:    v_rcp_iflag_f32_e32 v4, v3
+; GCN-NEXT:    v_or_b32_e32 v5, 1, v5
+; GCN-NEXT:    v_mul_f32_e32 v4, v1, v4
+; GCN-NEXT:    v_trunc_f32_e32 v4, v4
+; GCN-NEXT:    v_mad_f32 v1, -v4, v3, v1
+; GCN-NEXT:    v_cvt_i32_f32_e32 v4, v4
+; GCN-NEXT:    v_cmp_ge_f32_e64 vcc, |v1|, |v3|
+; GCN-NEXT:    v_cndmask_b32_e32 v1, 0, v5, vcc
+; GCN-NEXT:    v_add_i32_e32 v1, vcc, v4, v1
+; GCN-NEXT:    v_mul_lo_u32 v1, v1, v2
 ; GCN-NEXT:    v_sub_i32_e32 v0, vcc, v0, v1
 ; GCN-NEXT:    v_bfe_i32 v0, v0, 0, 24
 ; GCN-NEXT:    v_ashrrev_i32_e32 v1, 31, v0
 ; GCN-NEXT:    s_setpc_b64 s[30:31]
   %1 = ashr i64 %x, 40
   %2 = ashr i64 %y, 40
   %result = srem i64 %1, %2
   ret i64 %result
 }
 
 define amdgpu_kernel void @s_test_srem25_64(i64 addrspace(1)* %out, i64 %x, i64 %y) {
 ; GCN-LABEL: s_test_srem25_64:
 ; GCN:       ; %bb.0:
-; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GCN-NEXT:    s_load_dword s3, s[0:1], 0xe
-; GCN-NEXT:    s_mov_b32 s11, 0xf000
-; GCN-NEXT:    s_mov_b32 s10, -1
+; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_mov_b32 s8, s4
-; GCN-NEXT:    s_mov_b32 s9, s5
-; GCN-NEXT:    s_ashr_i64 s[0:1], s[6:7], 39
 ; GCN-NEXT:    s_ashr_i64 s[2:3], s[2:3], 39
-; GCN-NEXT:    s_ashr_i32 s1, s2, 31
-; GCN-NEXT:    s_ashr_i32 s4, s0, 31
-; GCN-NEXT:    s_add_i32 s2, s2, s1
-; GCN-NEXT:    s_add_i32 s0, s0, s4
-; GCN-NEXT:    s_xor_b32 s5, s2, s1
-; GCN-NEXT:    s_xor_b32 s2, s0, s4
-; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s5
+; GCN-NEXT:    s_ashr_i32 s3, s2, 31
+; GCN-NEXT:    s_add_i32 s2, s2, s3
+; GCN-NEXT:    s_xor_b32 s9, s2, s3
+; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s9
+; GCN-NEXT:    s_ashr_i64 s[2:3], s[6:7], 39
+; GCN-NEXT:    s_ashr_i32 s8, s2, 31
+; GCN-NEXT:    s_add_i32 s2, s2, s8
 ; GCN-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GCN-NEXT:    s_xor_b32 s2, s2, s8
+; GCN-NEXT:    s_mov_b32 s7, 0xf000
+; GCN-NEXT:    s_mov_b32 s6, -1
 ; GCN-NEXT:    v_mul_f32_e32 v0, 0x4f800000, v0
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v0
-; GCN-NEXT:    v_mul_hi_u32 v1, v0, s5
-; GCN-NEXT:    v_mul_lo_u32 v2, v0, s5
-; GCN-NEXT:    v_sub_i32_e32 v3, vcc, 0, v2
-; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v1, v2, v3, s[0:1]
+; GCN-NEXT:    v_mul_lo_u32 v1, v0, s9
+; GCN-NEXT:    v_mul_hi_u32 v2, v0, s9
+; GCN-NEXT:    v_sub_i32_e32 v3, vcc, 0, v1
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v1, v1, v3, s[0:1]
 ; GCN-NEXT:    v_mul_hi_u32 v1, v1, v0
 ; GCN-NEXT:    v_add_i32_e32 v2, vcc, v1, v0
 ; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, v1, v0
 ; GCN-NEXT:    v_cndmask_b32_e64 v0, v0, v2, s[0:1]
 ; GCN-NEXT:    v_mul_hi_u32 v0, v0, s2
-; GCN-NEXT:    v_mul_lo_u32 v0, v0, s5
+; GCN-NEXT:    v_mul_lo_u32 v0, v0, s9
 ; GCN-NEXT:    v_sub_i32_e32 v1, vcc, s2, v0
-; GCN-NEXT:    v_add_i32_e32 v2, vcc, s5, v1
 ; GCN-NEXT:    v_cmp_ge_u32_e64 s[0:1], s2, v0
-; GCN-NEXT:    v_cmp_le_u32_e64 s[2:3], s5, v1
-; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, s5, v1
+; GCN-NEXT:    v_add_i32_e32 v2, vcc, s9, v1
+; GCN-NEXT:    v_cmp_le_u32_e64 s[2:3], s9, v1
+; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, s9, v1
 ; GCN-NEXT:    s_and_b64 vcc, s[2:3], s[0:1]
 ; GCN-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc
 ; GCN-NEXT:    v_cndmask_b32_e64 v0, v2, v0, s[0:1]
-; GCN-NEXT:    v_xor_b32_e32 v0, s4, v0
-; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, s4, v0
+; GCN-NEXT:    v_xor_b32_e32 v0, s8, v0
+; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, s8, v0
 ; GCN-NEXT:    v_ashrrev_i32_e32 v1, 31, v0
-; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[8:11], 0
+; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
 ; GCN-NEXT:    s_endpgm
   %1 = ashr i64 %x, 39
   %2 = ashr i64 %y, 39
   %result = srem i64 %1, %2
   store i64 %result, i64 addrspace(1)* %out
   ret void
 }
 
 define amdgpu_kernel void @s_test_srem31_64(i64 addrspace(1)* %out, i64 %x, i64 %y) {
 ; GCN-LABEL: s_test_srem31_64:
 ; GCN:       ; %bb.0:
-; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GCN-NEXT:    s_load_dword s3, s[0:1], 0xe
-; GCN-NEXT:    s_mov_b32 s11, 0xf000
-; GCN-NEXT:    s_mov_b32 s10, -1
+; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_mov_b32 s8, s4
-; GCN-NEXT:    s_mov_b32 s9, s5
-; GCN-NEXT:    s_ashr_i64 s[0:1], s[6:7], 33
 ; GCN-NEXT:    s_ashr_i64 s[2:3], s[2:3], 33
-; GCN-NEXT:    s_ashr_i32 s1, s2, 31
-; GCN-NEXT:    s_ashr_i32 s4, s0, 31
-; GCN-NEXT:    s_add_i32 s2, s2, s1
-; GCN-NEXT:    s_add_i32 s0, s0, s4
-; GCN-NEXT:    s_xor_b32 s5, s2, s1
-; GCN-NEXT:    s_xor_b32 s2, s0, s4
-; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s5
+; GCN-NEXT:    s_ashr_i32 s3, s2, 31
+; GCN-NEXT:    s_add_i32 s2, s2, s3
+; GCN-NEXT:    s_xor_b32 s9, s2, s3
+; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s9
+; GCN-NEXT:    s_ashr_i64 s[2:3], s[6:7], 33
+; GCN-NEXT:    s_ashr_i32 s8, s2, 31
+; GCN-NEXT:    s_add_i32 s2, s2, s8
 ; GCN-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GCN-NEXT:    s_xor_b32 s2, s2, s8
+; GCN-NEXT:    s_mov_b32 s7, 0xf000
+; GCN-NEXT:    s_mov_b32 s6, -1
 ; GCN-NEXT:    v_mul_f32_e32 v0, 0x4f800000, v0
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v0
-; GCN-NEXT:    v_mul_hi_u32 v1, v0, s5
-; GCN-NEXT:    v_mul_lo_u32 v2, v0, s5
-; GCN-NEXT:    v_sub_i32_e32 v3, vcc, 0, v2
-; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v1, v2, v3, s[0:1]
+; GCN-NEXT:    v_mul_lo_u32 v1, v0, s9
+; GCN-NEXT:    v_mul_hi_u32 v2, v0, s9
+; GCN-NEXT:    v_sub_i32_e32 v3, vcc, 0, v1
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v1, v1, v3, s[0:1]
 ; GCN-NEXT:    v_mul_hi_u32 v1, v1, v0
 ; GCN-NEXT:    v_add_i32_e32 v2, vcc, v1, v0
 ; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, v1, v0
 ; GCN-NEXT:    v_cndmask_b32_e64 v0, v0, v2, s[0:1]
 ; GCN-NEXT:    v_mul_hi_u32 v0, v0, s2
-; GCN-NEXT:    v_mul_lo_u32 v0, v0, s5
+; GCN-NEXT:    v_mul_lo_u32 v0, v0, s9
 ; GCN-NEXT:    v_sub_i32_e32 v1, vcc, s2, v0
-; GCN-NEXT:    v_add_i32_e32 v2, vcc, s5, v1
 ; GCN-NEXT:    v_cmp_ge_u32_e64 s[0:1], s2, v0
-; GCN-NEXT:    v_cmp_le_u32_e64 s[2:3], s5, v1
-; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, s5, v1
+; GCN-NEXT:    v_add_i32_e32 v2, vcc, s9, v1
+; GCN-NEXT:    v_cmp_le_u32_e64 s[2:3], s9, v1
+; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, s9, v1
 ; GCN-NEXT:    s_and_b64 vcc, s[2:3], s[0:1]
 ; GCN-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc
 ; GCN-NEXT:    v_cndmask_b32_e64 v0, v2, v0, s[0:1]
-; GCN-NEXT:    v_xor_b32_e32 v0, s4, v0
-; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, s4, v0
+; GCN-NEXT:    v_xor_b32_e32 v0, s8, v0
+; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, s8, v0
 ; GCN-NEXT:    v_ashrrev_i32_e32 v1, 31, v0
-; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[8:11], 0
+; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
 ; GCN-NEXT:    s_endpgm
   %1 = ashr i64 %x, 33
   %2 = ashr i64 %y, 33
   %result = srem i64 %1, %2
   store i64 %result, i64 addrspace(1)* %out
   ret void
 }
 
 ; 32 known sign bits
 define amdgpu_kernel void @s_test_srem32_64(i64 addrspace(1)* %out, i64 %x, i64 %y) {
 ; GCN-LABEL: s_test_srem32_64:
 ; GCN:       ; %bb.0:
-; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GCN-NEXT:    s_load_dword s2, s[0:1], 0xe
-; GCN-NEXT:    s_mov_b32 s11, 0xf000
-; GCN-NEXT:    s_mov_b32 s10, -1
+; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_mov_b32 s8, s4
-; GCN-NEXT:    s_mov_b32 s9, s5
-; GCN-NEXT:    s_ashr_i32 s0, s2, 31
-; GCN-NEXT:    s_ashr_i32 s4, s7, 31
-; GCN-NEXT:    s_add_i32 s2, s2, s0
-; GCN-NEXT:    s_add_i32 s1, s7, s4
-; GCN-NEXT:    s_xor_b32 s5, s2, s0
-; GCN-NEXT:    s_xor_b32 s2, s1, s4
-; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s5
+; GCN-NEXT:    s_mov_b32 s6, -1
+; GCN-NEXT:    s_ashr_i32 s3, s2, 31
+; GCN-NEXT:    s_add_i32 s2, s2, s3
+; GCN-NEXT:    s_xor_b32 s9, s2, s3
+; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s9
+; GCN-NEXT:    s_ashr_i32 s8, s7, 31
+; GCN-NEXT:    s_add_i32 s2, s7, s8
+; GCN-NEXT:    s_xor_b32 s2, s2, s8
 ; GCN-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GCN-NEXT:    s_mov_b32 s7, 0xf000
 ; GCN-NEXT:    v_mul_f32_e32 v0, 0x4f800000, v0
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v0
-; GCN-NEXT:    v_mul_hi_u32 v1, v0, s5
-; GCN-NEXT:    v_mul_lo_u32 v2, v0, s5
-; GCN-NEXT:    v_sub_i32_e32 v3, vcc, 0, v2
-; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v1, v2, v3, s[0:1]
+; GCN-NEXT:    v_mul_lo_u32 v1, v0, s9
+; GCN-NEXT:    v_mul_hi_u32 v2, v0, s9
+; GCN-NEXT:    v_sub_i32_e32 v3, vcc, 0, v1
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v1, v1, v3, s[0:1]
 ; GCN-NEXT:    v_mul_hi_u32 v1, v1, v0
 ; GCN-NEXT:    v_add_i32_e32 v2, vcc, v1, v0
 ; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, v1, v0
 ; GCN-NEXT:    v_cndmask_b32_e64 v0, v0, v2, s[0:1]
 ; GCN-NEXT:    v_mul_hi_u32 v0, v0, s2
-; GCN-NEXT:    v_mul_lo_u32 v0, v0, s5
+; GCN-NEXT:    v_mul_lo_u32 v0, v0, s9
 ; GCN-NEXT:    v_sub_i32_e32 v1, vcc, s2, v0
-; GCN-NEXT:    v_add_i32_e32 v2, vcc, s5, v1
 ; GCN-NEXT:    v_cmp_ge_u32_e64 s[0:1], s2, v0
-; GCN-NEXT:    v_cmp_le_u32_e64 s[2:3], s5, v1
-; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, s5, v1
+; GCN-NEXT:    v_add_i32_e32 v2, vcc, s9, v1
+; GCN-NEXT:    v_cmp_le_u32_e64 s[2:3], s9, v1
+; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, s9, v1
 ; GCN-NEXT:    s_and_b64 vcc, s[2:3], s[0:1]
 ; GCN-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc
 ; GCN-NEXT:    v_cndmask_b32_e64 v0, v2, v0, s[0:1]
-; GCN-NEXT:    v_xor_b32_e32 v0, s4, v0
-; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, s4, v0
+; GCN-NEXT:    v_xor_b32_e32 v0, s8, v0
+; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, s8, v0
 ; GCN-NEXT:    v_ashrrev_i32_e32 v1, 31, v0
-; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[8:11], 0
+; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
 ; GCN-NEXT:    s_endpgm
   %1 = ashr i64 %x, 32
   %2 = ashr i64 %y, 32
   %result = srem i64 %1, %2
   store i64 %result, i64 addrspace(1)* %out
   ret void
 }
 
 ; 33 known sign bits
 define amdgpu_kernel void @s_test_srem33_64(i64 addrspace(1)* %out, i64 %x, i64 %y) {
 ; GCN-LABEL: s_test_srem33_64:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_load_dwordx4 s[8:11], s[0:1], 0x9
-; GCN-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xd
+; GCN-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
+; GCN-NEXT:    v_mov_b32_e32 v7, 0
 ; GCN-NEXT:    s_mov_b32 s7, 0xf000
 ; GCN-NEXT:    s_mov_b32 s6, -1
-; GCN-NEXT:    v_mov_b32_e32 v0, 0
-; GCN-NEXT:    v_mov_b32_e32 v1, 0
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_mov_b32 s4, s8
-; GCN-NEXT:    s_mov_b32 s5, s9
-; GCN-NEXT:    s_ashr_i64 s[0:1], s[10:11], 31
-; GCN-NEXT:    s_ashr_i32 s8, s3, 31
-; GCN-NEXT:    s_ashr_i64 s[2:3], s[2:3], 31
-; GCN-NEXT:    s_mov_b32 s9, s8
-; GCN-NEXT:    s_add_u32 s2, s2, s8
-; GCN-NEXT:    s_addc_u32 s3, s3, s8
-; GCN-NEXT:    s_xor_b64 s[2:3], s[2:3], s[8:9]
-; GCN-NEXT:    v_cvt_f32_u32_e32 v2, s2
-; GCN-NEXT:    v_cvt_f32_u32_e32 v3, s3
-; GCN-NEXT:    s_sub_u32 s12, 0, s2
-; GCN-NEXT:    v_mov_b32_e32 v4, s3
-; GCN-NEXT:    v_mac_f32_e32 v2, 0x4f800000, v3
-; GCN-NEXT:    s_subb_u32 s13, 0, s3
-; GCN-NEXT:    s_ashr_i32 s8, s11, 31
-; GCN-NEXT:    v_rcp_f32_e32 v2, v2
-; GCN-NEXT:    s_mov_b32 s9, s8
-; GCN-NEXT:    s_add_u32 s0, s0, s8
-; GCN-NEXT:    v_mov_b32_e32 v3, s8
-; GCN-NEXT:    v_mul_f32_e32 v2, 0x5f7ffffc, v2
-; GCN-NEXT:    s_addc_u32 s1, s1, s8
-; GCN-NEXT:    v_mul_f32_e32 v5, 0x2f800000, v2
-; GCN-NEXT:    s_xor_b64 s[10:11], s[0:1], s[8:9]
-; GCN-NEXT:    v_trunc_f32_e32 v5, v5
-; GCN-NEXT:    v_mov_b32_e32 v6, s11
-; GCN-NEXT:    v_mac_f32_e32 v2, 0xcf800000, v5
-; GCN-NEXT:    v_cvt_u32_f32_e32 v5, v5
+; GCN-NEXT:    s_ashr_i64 s[2:3], s[10:11], 31
+; GCN-NEXT:    s_ashr_i64 s[4:5], s[0:1], 31
+; GCN-NEXT:    s_ashr_i32 s0, s1, 31
+; GCN-NEXT:    s_add_u32 s4, s4, s0
+; GCN-NEXT:    s_mov_b32 s1, s0
+; GCN-NEXT:    s_addc_u32 s5, s5, s0
+; GCN-NEXT:    s_xor_b64 s[12:13], s[4:5], s[0:1]
+; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s12
+; GCN-NEXT:    v_cvt_f32_u32_e32 v1, s13
+; GCN-NEXT:    s_sub_u32 s4, 0, s12
+; GCN-NEXT:    s_subb_u32 s5, 0, s13
+; GCN-NEXT:    s_ashr_i32 s10, s11, 31
+; GCN-NEXT:    v_mac_f32_e32 v0, 0x4f800000, v1
+; GCN-NEXT:    v_rcp_f32_e32 v0, v0
+; GCN-NEXT:    v_mov_b32_e32 v1, 0
+; GCN-NEXT:    s_mov_b32 s11, s10
+; GCN-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v0
+; GCN-NEXT:    v_mul_f32_e32 v2, 0x2f800000, v0
+; GCN-NEXT:    v_trunc_f32_e32 v2, v2
+; GCN-NEXT:    v_mac_f32_e32 v0, 0xcf800000, v2
+; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v0
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v2, v2
-; GCN-NEXT:    v_mul_lo_u32 v7, s12, v5
-; GCN-NEXT:    v_mul_lo_u32 v8, s13, v2
-; GCN-NEXT:    v_mul_hi_u32 v9, s12, v2
-; GCN-NEXT:    v_mul_lo_u32 v10, s12, v2
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v9, v7
-; GCN-NEXT:    v_mul_hi_u32 v9, v2, v10
-; GCN-NEXT:    v_mul_hi_u32 v11, v5, v10
-; GCN-NEXT:    v_mul_lo_u32 v10, v5, v10
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v7, v8
-; GCN-NEXT:    v_mul_hi_u32 v8, v2, v7
-; GCN-NEXT:    v_mul_lo_u32 v12, v2, v7
-; GCN-NEXT:    v_mul_hi_u32 v13, v5, v7
-; GCN-NEXT:    v_mul_lo_u32 v7, v5, v7
-; GCN-NEXT:    v_add_i32_e32 v9, vcc, v9, v12
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v1, v8, vcc
-; GCN-NEXT:    v_add_i32_e32 v9, vcc, v10, v9
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v8, v11, vcc
-; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v13, v0, vcc
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v8, v7
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v1, v9, vcc
-; GCN-NEXT:    v_add_i32_e64 v2, s[0:1], v2, v7
-; GCN-NEXT:    v_addc_u32_e64 v7, vcc, v5, v8, s[0:1]
-; GCN-NEXT:    v_mul_hi_u32 v9, s12, v2
-; GCN-NEXT:    v_mul_lo_u32 v10, s13, v2
-; GCN-NEXT:    v_mul_lo_u32 v11, s12, v2
-; GCN-NEXT:    v_mul_lo_u32 v12, s12, v7
-; GCN-NEXT:    v_mul_hi_u32 v13, v7, v11
-; GCN-NEXT:    v_mul_lo_u32 v14, v7, v11
-; GCN-NEXT:    v_mul_hi_u32 v11, v2, v11
-; GCN-NEXT:    v_add_i32_e32 v9, vcc, v9, v12
-; GCN-NEXT:    v_add_i32_e32 v9, vcc, v10, v9
-; GCN-NEXT:    v_mul_hi_u32 v10, v7, v9
-; GCN-NEXT:    v_mul_hi_u32 v12, v2, v9
-; GCN-NEXT:    v_mul_lo_u32 v15, v2, v9
-; GCN-NEXT:    v_mul_lo_u32 v7, v7, v9
-; GCN-NEXT:    v_add_i32_e32 v9, vcc, v11, v15
-; GCN-NEXT:    v_addc_u32_e32 v11, vcc, v1, v12, vcc
-; GCN-NEXT:    v_add_i32_e32 v9, vcc, v14, v9
-; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v11, v13, vcc
-; GCN-NEXT:    v_addc_u32_e32 v10, vcc, v10, v0, vcc
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v8
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v9, v7
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v1, v10, vcc
-; GCN-NEXT:    v_addc_u32_e64 v5, vcc, v5, v8, s[0:1]
-; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v7
-; GCN-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; GCN-NEXT:    v_mul_hi_u32 v7, s10, v2
-; GCN-NEXT:    v_mul_hi_u32 v8, s11, v2
-; GCN-NEXT:    v_mul_lo_u32 v2, s11, v2
-; GCN-NEXT:    v_mul_hi_u32 v9, s10, v5
-; GCN-NEXT:    v_mul_lo_u32 v10, s10, v5
-; GCN-NEXT:    v_mul_hi_u32 v11, s11, v5
-; GCN-NEXT:    v_mul_lo_u32 v5, s11, v5
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v7, v10
-; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v1, v9, vcc
-; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v7
-; GCN-NEXT:    v_addc_u32_e32 v2, vcc, v9, v8, vcc
-; GCN-NEXT:    v_addc_u32_e32 v0, vcc, v11, v0, vcc
-; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v5
-; GCN-NEXT:    v_addc_u32_e32 v0, vcc, v1, v0, vcc
-; GCN-NEXT:    v_mul_hi_u32 v1, s2, v2
-; GCN-NEXT:    v_mul_lo_u32 v5, s3, v2
-; GCN-NEXT:    v_mul_lo_u32 v2, s2, v2
-; GCN-NEXT:    v_mul_lo_u32 v0, s2, v0
-; GCN-NEXT:    v_add_i32_e32 v0, vcc, v1, v0
-; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v5
-; GCN-NEXT:    v_sub_i32_e32 v1, vcc, s11, v0
-; GCN-NEXT:    v_sub_i32_e32 v2, vcc, s10, v2
-; GCN-NEXT:    v_subb_u32_e64 v1, s[0:1], v1, v4, vcc
-; GCN-NEXT:    v_subb_u32_e32 v0, vcc, v6, v0, vcc
-; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s2, v2
-; GCN-NEXT:    v_cndmask_b32_e64 v5, 0, -1, vcc
-; GCN-NEXT:    v_subrev_i32_e32 v6, vcc, s2, v2
-; GCN-NEXT:    v_subb_u32_e64 v4, s[0:1], v1, v4, vcc
-; GCN-NEXT:    v_subbrev_u32_e32 v1, vcc, 0, v1, vcc
-; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s2, v6
-; GCN-NEXT:    v_cndmask_b32_e64 v7, 0, -1, vcc
-; GCN-NEXT:    v_subrev_i32_e32 v8, vcc, s2, v6
-; GCN-NEXT:    v_cmp_le_u32_e64 s[0:1], s3, v0
-; GCN-NEXT:    v_cndmask_b32_e64 v9, 0, -1, s[0:1]
-; GCN-NEXT:    v_cmp_le_u32_e64 s[0:1], s3, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v10, 0, -1, s[0:1]
-; GCN-NEXT:    v_subbrev_u32_e32 v4, vcc, 0, v4, vcc
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, s3, v0
-; GCN-NEXT:    v_cndmask_b32_e32 v5, v9, v5, vcc
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, s3, v1
-; GCN-NEXT:    v_cndmask_b32_e32 v7, v10, v7, vcc
-; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v7
-; GCN-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; GCN-NEXT:    v_mul_hi_u32 v4, s4, v0
+; GCN-NEXT:    v_mul_lo_u32 v3, s4, v2
+; GCN-NEXT:    v_mul_lo_u32 v6, s5, v0
+; GCN-NEXT:    v_mul_lo_u32 v5, s4, v0
+; GCN-NEXT:    v_add_i32_e32 v3, vcc, v4, v3
+; GCN-NEXT:    v_add_i32_e32 v3, vcc, v3, v6
+; GCN-NEXT:    v_mul_hi_u32 v4, v0, v5
+; GCN-NEXT:    v_mul_lo_u32 v6, v0, v3
+; GCN-NEXT:    v_mul_hi_u32 v8, v0, v3
+; GCN-NEXT:    v_mul_hi_u32 v9, v2, v3
+; GCN-NEXT:    v_mul_lo_u32 v3, v2, v3
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v4, v6
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v7, v8, vcc
+; GCN-NEXT:    v_mul_lo_u32 v8, v2, v5
+; GCN-NEXT:    v_mul_hi_u32 v5, v2, v5
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v8, v4
+; GCN-NEXT:    v_addc_u32_e32 v4, vcc, v6, v5, vcc
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v9, v1, vcc
+; GCN-NEXT:    v_add_i32_e32 v3, vcc, v4, v3
+; GCN-NEXT:    v_add_i32_e64 v0, s[0:1], v0, v3
+; GCN-NEXT:    v_addc_u32_e32 v4, vcc, v7, v5, vcc
+; GCN-NEXT:    v_addc_u32_e64 v3, vcc, v2, v4, s[0:1]
+; GCN-NEXT:    v_mul_lo_u32 v5, s4, v3
+; GCN-NEXT:    v_mul_hi_u32 v6, s4, v0
+; GCN-NEXT:    v_mul_lo_u32 v8, s5, v0
+; GCN-NEXT:    s_mov_b32 s5, s9
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v6, v5
+; GCN-NEXT:    v_mul_lo_u32 v6, s4, v0
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v8, v5
+; GCN-NEXT:    v_mul_lo_u32 v10, v0, v5
+; GCN-NEXT:    v_mul_hi_u32 v12, v0, v5
+; GCN-NEXT:    v_mul_hi_u32 v11, v0, v6
+; GCN-NEXT:    v_mul_hi_u32 v9, v3, v6
+; GCN-NEXT:    v_mul_lo_u32 v6, v3, v6
+; GCN-NEXT:    v_mul_hi_u32 v8, v3, v5
+; GCN-NEXT:    v_add_i32_e32 v10, vcc, v11, v10
+; GCN-NEXT:    v_addc_u32_e32 v11, vcc, v7, v12, vcc
+; GCN-NEXT:    v_mul_lo_u32 v3, v3, v5
+; GCN-NEXT:    v_add_i32_e32 v6, vcc, v6, v10
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v11, v9, vcc
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v8, v1, vcc
+; GCN-NEXT:    v_add_i32_e32 v3, vcc, v6, v3
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v7, v5, vcc
+; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v4
+; GCN-NEXT:    v_addc_u32_e64 v2, vcc, v2, v5, s[0:1]
+; GCN-NEXT:    s_add_u32 s0, s2, s10
+; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v3
+; GCN-NEXT:    s_addc_u32 s1, s3, s10
+; GCN-NEXT:    s_xor_b64 s[14:15], s[0:1], s[10:11]
+; GCN-NEXT:    v_addc_u32_e32 v2, vcc, 0, v2, vcc
+; GCN-NEXT:    v_mul_lo_u32 v3, s14, v2
+; GCN-NEXT:    v_mul_hi_u32 v4, s14, v0
+; GCN-NEXT:    v_mul_hi_u32 v5, s14, v2
+; GCN-NEXT:    v_mul_hi_u32 v6, s15, v2
+; GCN-NEXT:    v_mul_lo_u32 v2, s15, v2
+; GCN-NEXT:    v_add_i32_e32 v3, vcc, v4, v3
+; GCN-NEXT:    v_addc_u32_e32 v4, vcc, v7, v5, vcc
+; GCN-NEXT:    v_mul_lo_u32 v5, s15, v0
+; GCN-NEXT:    v_mul_hi_u32 v0, s15, v0
+; GCN-NEXT:    s_mov_b32 s4, s8
+; GCN-NEXT:    v_add_i32_e32 v3, vcc, v5, v3
+; GCN-NEXT:    v_addc_u32_e32 v0, vcc, v4, v0, vcc
+; GCN-NEXT:    v_addc_u32_e32 v1, vcc, v6, v1, vcc
+; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v2
+; GCN-NEXT:    v_addc_u32_e32 v1, vcc, v7, v1, vcc
+; GCN-NEXT:    v_mul_lo_u32 v1, s12, v1
+; GCN-NEXT:    v_mul_hi_u32 v2, s12, v0
+; GCN-NEXT:    v_mul_lo_u32 v3, s13, v0
+; GCN-NEXT:    v_mul_lo_u32 v0, s12, v0
+; GCN-NEXT:    v_add_i32_e32 v1, vcc, v2, v1
+; GCN-NEXT:    v_add_i32_e32 v1, vcc, v1, v3
+; GCN-NEXT:    v_sub_i32_e32 v2, vcc, s15, v1
+; GCN-NEXT:    v_mov_b32_e32 v3, s13
+; GCN-NEXT:    v_sub_i32_e32 v0, vcc, s14, v0
+; GCN-NEXT:    v_subb_u32_e64 v2, s[0:1], v2, v3, vcc
+; GCN-NEXT:    v_subrev_i32_e64 v4, s[0:1], s12, v0
+; GCN-NEXT:    v_subb_u32_e64 v3, s[2:3], v2, v3, s[0:1]
+; GCN-NEXT:    v_subbrev_u32_e64 v2, s[0:1], 0, v2, s[0:1]
+; GCN-NEXT:    v_cmp_le_u32_e64 s[0:1], s13, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v5, 0, -1, s[0:1]
+; GCN-NEXT:    v_cmp_le_u32_e64 s[0:1], s12, v4
+; GCN-NEXT:    v_cndmask_b32_e64 v6, 0, -1, s[0:1]
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], s13, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v5, v5, v6, s[0:1]
+; GCN-NEXT:    v_subrev_i32_e64 v6, s[0:1], s12, v4
+; GCN-NEXT:    v_subbrev_u32_e64 v3, s[0:1], 0, v3, s[0:1]
 ; GCN-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v5
-; GCN-NEXT:    v_cndmask_b32_e64 v0, v0, v1, s[0:1]
-; GCN-NEXT:    v_cndmask_b32_e32 v1, v6, v8, vcc
-; GCN-NEXT:    v_cndmask_b32_e64 v1, v2, v1, s[0:1]
-; GCN-NEXT:    v_xor_b32_e32 v2, s8, v0
-; GCN-NEXT:    v_xor_b32_e32 v0, s8, v1
-; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, s8, v0
-; GCN-NEXT:    v_subb_u32_e32 v1, vcc, v2, v3, vcc
+; GCN-NEXT:    v_cndmask_b32_e64 v2, v2, v3, s[0:1]
+; GCN-NEXT:    v_mov_b32_e32 v3, s15
+; GCN-NEXT:    v_subb_u32_e32 v1, vcc, v3, v1, vcc
+; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s13, v1
+; GCN-NEXT:    v_cndmask_b32_e64 v3, 0, -1, vcc
+; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s12, v0
+; GCN-NEXT:    v_cndmask_b32_e64 v5, 0, -1, vcc
+; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, s13, v1
+; GCN-NEXT:    v_cndmask_b32_e32 v3, v3, v5, vcc
+; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v3
+; GCN-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc
+; GCN-NEXT:    v_cndmask_b32_e64 v2, v4, v6, s[0:1]
+; GCN-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
+; GCN-NEXT:    v_xor_b32_e32 v0, s10, v0
+; GCN-NEXT:    v_xor_b32_e32 v1, s10, v1
+; GCN-NEXT:    v_mov_b32_e32 v2, s10
+; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, s10, v0
+; GCN-NEXT:    v_subb_u32_e32 v1, vcc, v1, v2, vcc
 ; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
 ; GCN-NEXT:    s_endpgm
   %1 = ashr i64 %x, 31
   %2 = ashr i64 %y, 31
   %result = srem i64 %1, %2
   store i64 %result, i64 addrspace(1)* %out
   ret void
 }
 
 define amdgpu_kernel void @s_test_srem24_48(i48 addrspace(1)* %out, i48 %x, i48 %y) {
 ; GCN-LABEL: s_test_srem24_48:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
 ; GCN-NEXT:    s_load_dword s2, s[0:1], 0xb
 ; GCN-NEXT:    s_load_dword s3, s[0:1], 0xc
-; GCN-NEXT:    s_load_dword s8, s[0:1], 0xd
-; GCN-NEXT:    s_load_dword s9, s[0:1], 0xe
+; GCN-NEXT:    s_load_dword s6, s[0:1], 0xd
+; GCN-NEXT:    s_load_dword s0, s[0:1], 0xe
 ; GCN-NEXT:    s_mov_b32 s7, 0xf000
-; GCN-NEXT:    s_mov_b32 s6, -1
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_sext_i32_i16 s0, s3
-; GCN-NEXT:    s_sext_i32_i16 s1, s9
-; GCN-NEXT:    v_mov_b32_e32 v0, s8
-; GCN-NEXT:    v_mov_b32_e32 v1, s2
-; GCN-NEXT:    v_alignbit_b32 v0, s1, v0, 24
-; GCN-NEXT:    v_alignbit_b32 v1, s0, v1, 24
-; GCN-NEXT:    v_xor_b32_e32 v2, v1, v0
-; GCN-NEXT:    v_cvt_f32_i32_e32 v3, v1
-; GCN-NEXT:    v_cvt_f32_i32_e32 v4, v0
-; GCN-NEXT:    v_ashrrev_i32_e32 v2, 30, v2
-; GCN-NEXT:    v_rcp_iflag_f32_e32 v5, v4
-; GCN-NEXT:    v_or_b32_e32 v2, 1, v2
-; GCN-NEXT:    v_mul_f32_e32 v5, v3, v5
-; GCN-NEXT:    v_trunc_f32_e32 v5, v5
-; GCN-NEXT:    v_mad_f32 v3, -v5, v4, v3
-; GCN-NEXT:    v_cvt_i32_f32_e32 v5, v5
-; GCN-NEXT:    v_cmp_ge_f32_e64 vcc, |v3|, |v4|
-; GCN-NEXT:    v_cndmask_b32_e32 v2, 0, v2, vcc
-; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v5
-; GCN-NEXT:    v_mul_lo_u32 v0, v2, v0
-; GCN-NEXT:    v_sub_i32_e32 v0, vcc, v1, v0
+; GCN-NEXT:    v_mov_b32_e32 v2, s2
+; GCN-NEXT:    s_sext_i32_i16 s1, s3
+; GCN-NEXT:    v_mov_b32_e32 v0, s6
+; GCN-NEXT:    s_sext_i32_i16 s0, s0
+; GCN-NEXT:    v_alignbit_b32 v0, s0, v0, 24
+; GCN-NEXT:    v_cvt_f32_i32_e32 v1, v0
+; GCN-NEXT:    v_alignbit_b32 v2, s1, v2, 24
+; GCN-NEXT:    v_cvt_f32_i32_e32 v3, v2
+; GCN-NEXT:    v_xor_b32_e32 v5, v2, v0
+; GCN-NEXT:    v_rcp_iflag_f32_e32 v4, v1
+; GCN-NEXT:    v_ashrrev_i32_e32 v5, 30, v5
+; GCN-NEXT:    v_or_b32_e32 v5, 1, v5
+; GCN-NEXT:    s_mov_b32 s6, -1
+; GCN-NEXT:    v_mul_f32_e32 v4, v3, v4
+; GCN-NEXT:    v_trunc_f32_e32 v4, v4
+; GCN-NEXT:    v_mad_f32 v3, -v4, v1, v3
+; GCN-NEXT:    v_cvt_i32_f32_e32 v4, v4
+; GCN-NEXT:    v_cmp_ge_f32_e64 vcc, |v3|, |v1|
+; GCN-NEXT:    v_cndmask_b32_e32 v1, 0, v5, vcc
+; GCN-NEXT:    v_add_i32_e32 v1, vcc, v1, v4
+; GCN-NEXT:    v_mul_lo_u32 v0, v1, v0
+; GCN-NEXT:    v_sub_i32_e32 v0, vcc, v2, v0
 ; GCN-NEXT:    v_bfe_i32 v0, v0, 0, 24
 ; GCN-NEXT:    v_ashrrev_i32_e32 v1, 31, v0
 ; GCN-NEXT:    buffer_store_dword v0, off, s[4:7], 0
 ; GCN-NEXT:    buffer_store_short v1, off, s[4:7], 0 offset:4
 ; GCN-NEXT:    s_endpgm
   %1 = ashr i48 %x, 24
   %2 = ashr i48 %y, 24
   %result = srem i48 %1, %2
   store i48 %result, i48 addrspace(1)* %out
   ret void
 }
 
 define amdgpu_kernel void @s_test_srem_k_num_i64(i64 addrspace(1)* %out, i64 %x) {
 ; GCN-LABEL: s_test_srem_k_num_i64:
 ; GCN:       ; %bb.0:
-; GCN-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
-; GCN-NEXT:    s_mov_b32 s7, 0xf000
-; GCN-NEXT:    s_mov_b32 s6, -1
-; GCN-NEXT:    v_mov_b32_e32 v0, 0
+; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN-NEXT:    v_mov_b32_e32 v2, 0
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_ashr_i32 s8, s3, 31
+; GCN-NEXT:    s_ashr_i32 s0, s7, 31
+; GCN-NEXT:    s_add_u32 s2, s6, s0
+; GCN-NEXT:    s_addc_u32 s3, s7, s0
+; GCN-NEXT:    s_mov_b32 s1, s0
+; GCN-NEXT:    s_xor_b64 s[8:9], s[2:3], s[0:1]
+; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s8
+; GCN-NEXT:    v_cvt_f32_u32_e32 v1, s9
+; GCN-NEXT:    s_sub_u32 s2, 0, s8
+; GCN-NEXT:    s_subb_u32 s3, 0, s9
+; GCN-NEXT:    s_mov_b32 s7, 0xf000
+; GCN-NEXT:    v_mac_f32_e32 v0, 0x4f800000, v1
+; GCN-NEXT:    v_rcp_f32_e32 v0, v0
 ; GCN-NEXT:    v_mov_b32_e32 v1, 0
-; GCN-NEXT:    s_mov_b32 s4, s0
-; GCN-NEXT:    s_mov_b32 s5, s1
-; GCN-NEXT:    s_mov_b32 s9, s8
-; GCN-NEXT:    s_add_u32 s0, s2, s8
-; GCN-NEXT:    s_addc_u32 s1, s3, s8
-; GCN-NEXT:    s_xor_b64 s[2:3], s[0:1], s[8:9]
-; GCN-NEXT:    v_cvt_f32_u32_e32 v2, s2
-; GCN-NEXT:    v_cvt_f32_u32_e32 v3, s3
-; GCN-NEXT:    s_sub_u32 s8, 0, s2
-; GCN-NEXT:    v_mov_b32_e32 v4, s3
-; GCN-NEXT:    v_mac_f32_e32 v2, 0x4f800000, v3
-; GCN-NEXT:    s_subb_u32 s9, 0, s3
-; GCN-NEXT:    v_rcp_f32_e32 v2, v2
-; GCN-NEXT:    v_mul_f32_e32 v2, 0x5f7ffffc, v2
-; GCN-NEXT:    v_mul_f32_e32 v3, 0x2f800000, v2
+; GCN-NEXT:    s_mov_b32 s6, -1
+; GCN-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v0
+; GCN-NEXT:    v_mul_f32_e32 v3, 0x2f800000, v0
 ; GCN-NEXT:    v_trunc_f32_e32 v3, v3
-; GCN-NEXT:    v_mac_f32_e32 v2, 0xcf800000, v3
+; GCN-NEXT:    v_mac_f32_e32 v0, 0xcf800000, v3
+; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v0
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v3, v3
-; GCN-NEXT:    v_cvt_u32_f32_e32 v2, v2
-; GCN-NEXT:    v_mul_lo_u32 v5, s8, v3
-; GCN-NEXT:    v_mul_lo_u32 v6, s9, v2
-; GCN-NEXT:    v_mul_hi_u32 v7, s8, v2
-; GCN-NEXT:    v_mul_lo_u32 v8, s8, v2
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, v7, v5
-; GCN-NEXT:    v_mul_hi_u32 v7, v2, v8
-; GCN-NEXT:    v_mul_hi_u32 v9, v3, v8
-; GCN-NEXT:    v_mul_lo_u32 v8, v3, v8
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v6
-; GCN-NEXT:    v_mul_hi_u32 v6, v2, v5
-; GCN-NEXT:    v_mul_lo_u32 v10, v2, v5
-; GCN-NEXT:    v_mul_hi_u32 v11, v3, v5
-; GCN-NEXT:    v_mul_lo_u32 v5, v3, v5
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v7, v10
-; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v1, v6, vcc
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v8, v7
-; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v6, v9, vcc
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v11, v0, vcc
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, v6, v5
-; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v1, v7, vcc
-; GCN-NEXT:    v_add_i32_e64 v2, s[0:1], v2, v5
-; GCN-NEXT:    v_addc_u32_e64 v5, vcc, v3, v6, s[0:1]
-; GCN-NEXT:    v_mul_hi_u32 v7, s8, v2
-; GCN-NEXT:    v_mul_lo_u32 v8, s9, v2
-; GCN-NEXT:    v_mul_lo_u32 v9, s8, v2
-; GCN-NEXT:    v_mul_lo_u32 v10, s8, v5
-; GCN-NEXT:    v_mul_hi_u32 v11, v5, v9
-; GCN-NEXT:    v_mul_lo_u32 v12, v5, v9
-; GCN-NEXT:    v_mul_hi_u32 v9, v2, v9
+; GCN-NEXT:    v_mul_hi_u32 v5, s2, v0
+; GCN-NEXT:    v_mul_lo_u32 v4, s2, v3
+; GCN-NEXT:    v_mul_lo_u32 v7, s3, v0
+; GCN-NEXT:    v_mul_lo_u32 v6, s2, v0
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v5, v4
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v4, v7
+; GCN-NEXT:    v_mul_hi_u32 v5, v0, v6
+; GCN-NEXT:    v_mul_lo_u32 v7, v0, v4
+; GCN-NEXT:    v_mul_hi_u32 v9, v0, v4
+; GCN-NEXT:    v_mul_lo_u32 v8, v3, v6
+; GCN-NEXT:    v_mul_hi_u32 v6, v3, v6
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v7
+; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v2, v9, vcc
+; GCN-NEXT:    v_mul_hi_u32 v9, v3, v4
+; GCN-NEXT:    v_mul_lo_u32 v4, v3, v4
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v8, v5
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v7, v6, vcc
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v9, v1, vcc
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v5, v4
+; GCN-NEXT:    v_add_i32_e64 v0, s[0:1], v0, v4
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v2, v6, vcc
+; GCN-NEXT:    v_addc_u32_e64 v4, vcc, v3, v5, s[0:1]
+; GCN-NEXT:    v_mul_lo_u32 v6, s2, v4
+; GCN-NEXT:    v_mul_hi_u32 v7, s2, v0
+; GCN-NEXT:    v_mul_lo_u32 v8, s3, v0
+; GCN-NEXT:    v_add_i32_e32 v6, vcc, v7, v6
+; GCN-NEXT:    v_mul_lo_u32 v7, s2, v0
+; GCN-NEXT:    v_add_i32_e32 v6, vcc, v8, v6
+; GCN-NEXT:    v_mul_lo_u32 v10, v0, v6
+; GCN-NEXT:    v_mul_hi_u32 v12, v0, v6
+; GCN-NEXT:    v_mul_hi_u32 v11, v0, v7
+; GCN-NEXT:    v_mul_hi_u32 v9, v4, v7
+; GCN-NEXT:    v_mul_lo_u32 v7, v4, v7
+; GCN-NEXT:    v_mul_hi_u32 v8, v4, v6
+; GCN-NEXT:    v_add_i32_e32 v10, vcc, v11, v10
+; GCN-NEXT:    v_addc_u32_e32 v11, vcc, v2, v12, vcc
+; GCN-NEXT:    v_mul_lo_u32 v4, v4, v6
 ; GCN-NEXT:    v_add_i32_e32 v7, vcc, v7, v10
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v8, v7
-; GCN-NEXT:    v_mul_hi_u32 v8, v5, v7
-; GCN-NEXT:    v_mul_hi_u32 v10, v2, v7
-; GCN-NEXT:    v_mul_lo_u32 v13, v2, v7
-; GCN-NEXT:    v_mul_lo_u32 v5, v5, v7
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v9, v13
-; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v1, v10, vcc
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v12, v7
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v9, v11, vcc
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v8, v0, vcc
-; GCN-NEXT:    v_add_i32_e32 v3, vcc, v3, v6
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, v7, v5
-; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v1, v8, vcc
+; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v11, v9, vcc
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v8, v1, vcc
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v7, v4
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v2, v6, vcc
+; GCN-NEXT:    v_add_i32_e32 v3, vcc, v3, v5
 ; GCN-NEXT:    v_addc_u32_e64 v3, vcc, v3, v6, s[0:1]
-; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v5
+; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v4
 ; GCN-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; GCN-NEXT:    v_mul_hi_u32 v5, 24, v2
-; GCN-NEXT:    v_mul_hi_u32 v2, 0, v2
+; GCN-NEXT:    v_mul_hi_u32 v5, 24, v0
+; GCN-NEXT:    v_mul_lo_u32 v4, v3, 24
 ; GCN-NEXT:    v_mul_hi_u32 v6, 24, v3
-; GCN-NEXT:    v_mul_lo_u32 v7, v3, 24
+; GCN-NEXT:    v_mul_hi_u32 v0, 0, v0
 ; GCN-NEXT:    v_mul_hi_u32 v3, 0, v3
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v7
-; GCN-NEXT:    v_addc_u32_e32 v1, vcc, v1, v6, vcc
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, 0, v5
-; GCN-NEXT:    v_addc_u32_e32 v1, vcc, v1, v2, vcc
-; GCN-NEXT:    v_addc_u32_e32 v0, vcc, v3, v0, vcc
-; GCN-NEXT:    v_mul_hi_u32 v2, s2, v1
-; GCN-NEXT:    v_mul_lo_u32 v3, s3, v1
-; GCN-NEXT:    v_mul_lo_u32 v1, s2, v1
-; GCN-NEXT:    v_mul_lo_u32 v0, s2, v0
-; GCN-NEXT:    v_add_i32_e32 v0, vcc, v2, v0
-; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v3
-; GCN-NEXT:    v_sub_i32_e32 v2, vcc, 0, v0
-; GCN-NEXT:    v_sub_i32_e32 v3, vcc, 24, v1
-; GCN-NEXT:    v_subb_u32_e64 v1, s[0:1], v2, v4, vcc
-; GCN-NEXT:    v_subb_u32_e32 v0, vcc, 0, v0, vcc
-; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s2, v3
-; GCN-NEXT:    v_cndmask_b32_e64 v2, 0, -1, vcc
-; GCN-NEXT:    v_subrev_i32_e32 v5, vcc, s2, v3
-; GCN-NEXT:    v_subb_u32_e64 v4, s[0:1], v1, v4, vcc
-; GCN-NEXT:    v_subbrev_u32_e32 v1, vcc, 0, v1, vcc
-; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s2, v5
-; GCN-NEXT:    v_cndmask_b32_e64 v6, 0, -1, vcc
-; GCN-NEXT:    v_subrev_i32_e32 v7, vcc, s2, v5
-; GCN-NEXT:    v_cmp_le_u32_e64 s[0:1], s3, v0
-; GCN-NEXT:    v_cndmask_b32_e64 v8, 0, -1, s[0:1]
-; GCN-NEXT:    v_cmp_le_u32_e64 s[0:1], s3, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v9, 0, -1, s[0:1]
-; GCN-NEXT:    v_subbrev_u32_e32 v4, vcc, 0, v4, vcc
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, s3, v0
-; GCN-NEXT:    v_cndmask_b32_e32 v2, v8, v2, vcc
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, s3, v1
-; GCN-NEXT:    v_cndmask_b32_e32 v6, v9, v6, vcc
-; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v6
-; GCN-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
-; GCN-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v2
-; GCN-NEXT:    v_cndmask_b32_e64 v1, v0, v1, s[0:1]
-; GCN-NEXT:    v_cndmask_b32_e32 v0, v5, v7, vcc
-; GCN-NEXT:    v_cndmask_b32_e64 v0, v3, v0, s[0:1]
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v5, v4
+; GCN-NEXT:    v_addc_u32_e32 v2, vcc, v2, v6, vcc
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, 0, v4
+; GCN-NEXT:    v_addc_u32_e32 v0, vcc, v2, v0, vcc
+; GCN-NEXT:    v_addc_u32_e32 v1, vcc, v3, v1, vcc
+; GCN-NEXT:    v_mul_lo_u32 v1, s8, v1
+; GCN-NEXT:    v_mul_hi_u32 v2, s8, v0
+; GCN-NEXT:    v_mul_lo_u32 v3, s9, v0
+; GCN-NEXT:    v_mul_lo_u32 v0, s8, v0
+; GCN-NEXT:    v_add_i32_e32 v1, vcc, v2, v1
+; GCN-NEXT:    v_add_i32_e32 v1, vcc, v1, v3
+; GCN-NEXT:    v_sub_i32_e32 v2, vcc, 0, v1
+; GCN-NEXT:    v_mov_b32_e32 v3, s9
+; GCN-NEXT:    v_sub_i32_e32 v0, vcc, 24, v0
+; GCN-NEXT:    v_subb_u32_e64 v2, s[0:1], v2, v3, vcc
+; GCN-NEXT:    v_subrev_i32_e64 v4, s[0:1], s8, v0
+; GCN-NEXT:    v_subb_u32_e64 v3, s[2:3], v2, v3, s[0:1]
+; GCN-NEXT:    v_subbrev_u32_e64 v2, s[0:1], 0, v2, s[0:1]
+; GCN-NEXT:    v_cmp_le_u32_e64 s[0:1], s9, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v5, 0, -1, s[0:1]
+; GCN-NEXT:    v_cmp_le_u32_e64 s[0:1], s8, v4
+; GCN-NEXT:    v_cndmask_b32_e64 v6, 0, -1, s[0:1]
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], s9, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v5, v5, v6, s[0:1]
+; GCN-NEXT:    v_subrev_i32_e64 v6, s[0:1], s8, v4
+; GCN-NEXT:    v_subb_u32_e32 v1, vcc, 0, v1, vcc
+; GCN-NEXT:    v_subbrev_u32_e64 v3, s[0:1], 0, v3, s[0:1]
+; GCN-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v5
+; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s9, v1
+; GCN-NEXT:    v_cndmask_b32_e64 v2, v2, v3, s[0:1]
+; GCN-NEXT:    v_cndmask_b32_e64 v3, 0, -1, vcc
+; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s8, v0
+; GCN-NEXT:    v_cndmask_b32_e64 v5, 0, -1, vcc
+; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, s9, v1
+; GCN-NEXT:    v_cndmask_b32_e32 v3, v3, v5, vcc
+; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v3
+; GCN-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc
+; GCN-NEXT:    v_cndmask_b32_e64 v2, v4, v6, s[0:1]
+; GCN-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
 ; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
 ; GCN-NEXT:    s_endpgm
   %result = srem i64 24, %x
   store i64 %result, i64 addrspace(1)* %out
   ret void
 }
 
 define i64 @v_test_srem_k_num_i64(i64 %x) {
 ; GCN-LABEL: v_test_srem_k_num_i64:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GCN-NEXT:    v_ashrrev_i32_e32 v2, 31, v1
-; GCN-NEXT:    v_mov_b32_e32 v3, 0
-; GCN-NEXT:    v_mov_b32_e32 v4, 0
 ; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v2
 ; GCN-NEXT:    v_addc_u32_e32 v1, vcc, v1, v2, vcc
 ; GCN-NEXT:    v_xor_b32_e32 v1, v1, v2
 ; GCN-NEXT:    v_xor_b32_e32 v0, v0, v2
 ; GCN-NEXT:    v_cvt_f32_u32_e32 v2, v0
-; GCN-NEXT:    v_cvt_f32_u32_e32 v5, v1
-; GCN-NEXT:    v_sub_i32_e32 v6, vcc, 0, v0
-; GCN-NEXT:    v_subb_u32_e32 v7, vcc, 0, v1, vcc
-; GCN-NEXT:    v_mac_f32_e32 v2, 0x4f800000, v5
+; GCN-NEXT:    v_cvt_f32_u32_e32 v3, v1
+; GCN-NEXT:    v_sub_i32_e32 v4, vcc, 0, v0
+; GCN-NEXT:    v_subb_u32_e32 v5, vcc, 0, v1, vcc
+; GCN-NEXT:    v_mov_b32_e32 v12, 0
+; GCN-NEXT:    v_mac_f32_e32 v2, 0x4f800000, v3
 ; GCN-NEXT:    v_rcp_f32_e32 v2, v2
+; GCN-NEXT:    v_mov_b32_e32 v11, 0
 ; GCN-NEXT:    v_mul_f32_e32 v2, 0x5f7ffffc, v2
-; GCN-NEXT:    v_mul_f32_e32 v5, 0x2f800000, v2
-; GCN-NEXT:    v_trunc_f32_e32 v5, v5
-; GCN-NEXT:    v_mac_f32_e32 v2, 0xcf800000, v5
-; GCN-NEXT:    v_cvt_u32_f32_e32 v5, v5
+; GCN-NEXT:    v_mul_f32_e32 v3, 0x2f800000, v2
+; GCN-NEXT:    v_trunc_f32_e32 v3, v3
+; GCN-NEXT:    v_mac_f32_e32 v2, 0xcf800000, v3
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v2, v2
-; GCN-NEXT:    v_mul_lo_u32 v8, v6, v5
-; GCN-NEXT:    v_mul_lo_u32 v9, v7, v2
-; GCN-NEXT:    v_mul_hi_u32 v10, v6, v2
-; GCN-NEXT:    v_mul_lo_u32 v11, v6, v2
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v10, v8
-; GCN-NEXT:    v_mul_hi_u32 v10, v2, v11
-; GCN-NEXT:    v_mul_hi_u32 v12, v5, v11
-; GCN-NEXT:    v_mul_lo_u32 v11, v5, v11
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v8, v9
-; GCN-NEXT:    v_mul_hi_u32 v9, v2, v8
-; GCN-NEXT:    v_mul_lo_u32 v13, v2, v8
-; GCN-NEXT:    v_mul_hi_u32 v14, v5, v8
-; GCN-NEXT:    v_mul_lo_u32 v8, v5, v8
-; GCN-NEXT:    v_add_i32_e32 v10, vcc, v10, v13
-; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v4, v9, vcc
-; GCN-NEXT:    v_add_i32_e32 v10, vcc, v11, v10
-; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v9, v12, vcc
-; GCN-NEXT:    v_addc_u32_e32 v10, vcc, v14, v3, vcc
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v9, v8
-; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v4, v10, vcc
-; GCN-NEXT:    v_add_i32_e64 v2, s[4:5], v2, v8
-; GCN-NEXT:    v_addc_u32_e64 v8, vcc, v5, v9, s[4:5]
-; GCN-NEXT:    v_mul_hi_u32 v10, v6, v2
-; GCN-NEXT:    v_mul_lo_u32 v7, v7, v2
-; GCN-NEXT:    v_mul_lo_u32 v11, v6, v2
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v9
-; GCN-NEXT:    v_mul_lo_u32 v6, v6, v8
-; GCN-NEXT:    v_mul_hi_u32 v9, v8, v11
-; GCN-NEXT:    v_mul_lo_u32 v12, v8, v11
-; GCN-NEXT:    v_mul_hi_u32 v11, v2, v11
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, v10, v6
+; GCN-NEXT:    v_cvt_u32_f32_e32 v3, v3
+; GCN-NEXT:    v_mul_hi_u32 v6, v4, v2
+; GCN-NEXT:    v_mul_lo_u32 v7, v4, v3
+; GCN-NEXT:    v_mul_lo_u32 v8, v5, v2
 ; GCN-NEXT:    v_add_i32_e32 v6, vcc, v6, v7
-; GCN-NEXT:    v_mul_hi_u32 v7, v8, v6
-; GCN-NEXT:    v_mul_hi_u32 v10, v2, v6
-; GCN-NEXT:    v_mul_lo_u32 v13, v2, v6
-; GCN-NEXT:    v_mul_lo_u32 v6, v8, v6
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v11, v13
-; GCN-NEXT:    v_addc_u32_e32 v10, vcc, v4, v10, vcc
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v12, v8
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v10, v9, vcc
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v7, v3, vcc
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, v8, v6
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v4, v7, vcc
-; GCN-NEXT:    v_addc_u32_e64 v5, vcc, v5, v7, s[4:5]
-; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v6
-; GCN-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; GCN-NEXT:    v_mul_hi_u32 v6, 24, v2
-; GCN-NEXT:    v_mul_hi_u32 v2, 0, v2
-; GCN-NEXT:    v_mul_hi_u32 v7, 24, v5
-; GCN-NEXT:    v_mul_lo_u32 v8, v5, 24
-; GCN-NEXT:    v_mul_hi_u32 v5, 0, v5
+; GCN-NEXT:    v_mul_lo_u32 v7, v4, v2
 ; GCN-NEXT:    v_add_i32_e32 v6, vcc, v6, v8
-; GCN-NEXT:    v_addc_u32_e32 v4, vcc, v4, v7, vcc
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, 0, v6
-; GCN-NEXT:    v_addc_u32_e32 v2, vcc, v4, v2, vcc
-; GCN-NEXT:    v_addc_u32_e32 v3, vcc, v5, v3, vcc
+; GCN-NEXT:    v_mul_lo_u32 v9, v2, v6
+; GCN-NEXT:    v_mul_hi_u32 v8, v2, v6
+; GCN-NEXT:    v_mul_hi_u32 v10, v2, v7
+; GCN-NEXT:    v_mul_hi_u32 v13, v3, v6
+; GCN-NEXT:    v_mul_lo_u32 v6, v3, v6
+; GCN-NEXT:    v_add_i32_e32 v9, vcc, v10, v9
+; GCN-NEXT:    v_mul_lo_u32 v10, v3, v7
+; GCN-NEXT:    v_mul_hi_u32 v7, v3, v7
+; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v12, v8, vcc
+; GCN-NEXT:    v_add_i32_e32 v9, vcc, v10, v9
+; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v8, v7, vcc
+; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v13, v11, vcc
+; GCN-NEXT:    v_add_i32_e32 v6, vcc, v7, v6
+; GCN-NEXT:    v_add_i32_e64 v2, s[4:5], v2, v6
+; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v12, v8, vcc
+; GCN-NEXT:    v_addc_u32_e64 v6, vcc, v3, v7, s[4:5]
+; GCN-NEXT:    v_mul_lo_u32 v8, v4, v6
+; GCN-NEXT:    v_mul_hi_u32 v9, v4, v2
+; GCN-NEXT:    v_mul_lo_u32 v5, v5, v2
+; GCN-NEXT:    v_mul_lo_u32 v4, v4, v2
+; GCN-NEXT:    v_add_i32_e32 v8, vcc, v9, v8
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v8, v5
+; GCN-NEXT:    v_mul_lo_u32 v10, v2, v5
+; GCN-NEXT:    v_mul_hi_u32 v13, v2, v4
+; GCN-NEXT:    v_mul_hi_u32 v14, v2, v5
+; GCN-NEXT:    v_mul_hi_u32 v9, v6, v4
+; GCN-NEXT:    v_mul_lo_u32 v4, v6, v4
+; GCN-NEXT:    v_add_i32_e32 v10, vcc, v13, v10
+; GCN-NEXT:    v_mul_hi_u32 v8, v6, v5
+; GCN-NEXT:    v_addc_u32_e32 v13, vcc, v12, v14, vcc
+; GCN-NEXT:    v_mul_lo_u32 v5, v6, v5
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v4, v10
+; GCN-NEXT:    v_addc_u32_e32 v4, vcc, v13, v9, vcc
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v8, v11, vcc
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v4, v5
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v12, v6, vcc
+; GCN-NEXT:    v_add_i32_e32 v3, vcc, v3, v7
+; GCN-NEXT:    v_addc_u32_e64 v3, vcc, v3, v5, s[4:5]
+; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v4
+; GCN-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; GCN-NEXT:    v_mul_hi_u32 v5, 24, v2
+; GCN-NEXT:    v_mul_lo_u32 v4, v3, 24
+; GCN-NEXT:    v_mul_hi_u32 v6, 24, v3
+; GCN-NEXT:    v_mul_hi_u32 v2, 0, v2
+; GCN-NEXT:    v_mul_hi_u32 v3, 0, v3
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v5, v4
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v12, v6, vcc
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, 0, v4
+; GCN-NEXT:    v_addc_u32_e32 v2, vcc, v5, v2, vcc
+; GCN-NEXT:    v_addc_u32_e32 v3, vcc, v3, v11, vcc
+; GCN-NEXT:    v_mul_lo_u32 v3, v0, v3
 ; GCN-NEXT:    v_mul_hi_u32 v4, v0, v2
 ; GCN-NEXT:    v_mul_lo_u32 v5, v1, v2
 ; GCN-NEXT:    v_mul_lo_u32 v2, v0, v2
-; GCN-NEXT:    v_mul_lo_u32 v3, v0, v3
 ; GCN-NEXT:    v_add_i32_e32 v3, vcc, v4, v3
 ; GCN-NEXT:    v_add_i32_e32 v3, vcc, v3, v5
 ; GCN-NEXT:    v_sub_i32_e32 v4, vcc, 0, v3
 ; GCN-NEXT:    v_sub_i32_e32 v2, vcc, 24, v2
 ; GCN-NEXT:    v_subb_u32_e64 v4, s[4:5], v4, v1, vcc
+; GCN-NEXT:    v_sub_i32_e64 v5, s[4:5], v2, v0
+; GCN-NEXT:    v_subb_u32_e64 v6, s[6:7], v4, v1, s[4:5]
+; GCN-NEXT:    v_subbrev_u32_e64 v4, s[4:5], 0, v4, s[4:5]
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[4:5], v4, v1
+; GCN-NEXT:    v_cndmask_b32_e64 v7, 0, -1, s[4:5]
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[4:5], v5, v0
 ; GCN-NEXT:    v_subb_u32_e32 v3, vcc, 0, v3, vcc
+; GCN-NEXT:    v_cndmask_b32_e64 v8, 0, -1, s[4:5]
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[4:5], v4, v1
+; GCN-NEXT:    v_cndmask_b32_e64 v7, v7, v8, s[4:5]
+; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v3, v1
+; GCN-NEXT:    v_cmp_ne_u32_e64 s[6:7], 0, v7
+; GCN-NEXT:    v_cndmask_b32_e64 v7, 0, -1, vcc
 ; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v2, v0
-; GCN-NEXT:    v_cndmask_b32_e64 v5, 0, -1, vcc
-; GCN-NEXT:    v_sub_i32_e32 v6, vcc, v2, v0
-; GCN-NEXT:    v_subb_u32_e64 v7, s[4:5], v4, v1, vcc
-; GCN-NEXT:    v_subbrev_u32_e32 v4, vcc, 0, v4, vcc
-; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v6, v0
-; GCN-NEXT:    v_cndmask_b32_e64 v8, 0, -1, vcc
-; GCN-NEXT:    v_sub_i32_e32 v0, vcc, v6, v0
-; GCN-NEXT:    v_cmp_ge_u32_e64 s[4:5], v3, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v9, 0, -1, s[4:5]
-; GCN-NEXT:    v_cmp_ge_u32_e64 s[4:5], v4, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v10, 0, -1, s[4:5]
-; GCN-NEXT:    v_cmp_eq_u32_e64 s[4:5], v3, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v5, v9, v5, s[4:5]
-; GCN-NEXT:    v_subbrev_u32_e32 v7, vcc, 0, v7, vcc
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v1
-; GCN-NEXT:    v_cndmask_b32_e32 v1, v10, v8, vcc
-; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v1
-; GCN-NEXT:    v_cndmask_b32_e32 v0, v6, v0, vcc
-; GCN-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, v5
-; GCN-NEXT:    v_cndmask_b32_e64 v0, v2, v0, s[4:5]
-; GCN-NEXT:    v_cndmask_b32_e32 v1, v4, v7, vcc
-; GCN-NEXT:    v_cndmask_b32_e64 v1, v3, v1, s[4:5]
+; GCN-NEXT:    v_sub_i32_e64 v8, s[4:5], v5, v0
+; GCN-NEXT:    v_cndmask_b32_e64 v0, 0, -1, vcc
+; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v1
+; GCN-NEXT:    v_cndmask_b32_e32 v0, v7, v0, vcc
+; GCN-NEXT:    v_subbrev_u32_e64 v1, s[4:5], 0, v6, s[4:5]
+; GCN-NEXT:    v_cndmask_b32_e64 v5, v5, v8, s[6:7]
+; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
+; GCN-NEXT:    v_cndmask_b32_e64 v1, v4, v1, s[6:7]
+; GCN-NEXT:    v_cndmask_b32_e32 v0, v2, v5, vcc
+; GCN-NEXT:    v_cndmask_b32_e32 v1, v3, v1, vcc
 ; GCN-NEXT:    s_setpc_b64 s[30:31]
   %result = srem i64 24, %x
   ret i64 %result
 }
 
 define i64 @v_test_srem_pow2_k_num_i64(i64 %x) {
 ; GCN-LABEL: v_test_srem_pow2_k_num_i64:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GCN-NEXT:    v_ashrrev_i32_e32 v2, 31, v1
-; GCN-NEXT:    v_mov_b32_e32 v3, 0
-; GCN-NEXT:    v_mov_b32_e32 v4, 0
-; GCN-NEXT:    s_mov_b32 s6, 0x8000
 ; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v2
 ; GCN-NEXT:    v_addc_u32_e32 v1, vcc, v1, v2, vcc
 ; GCN-NEXT:    v_xor_b32_e32 v1, v1, v2
 ; GCN-NEXT:    v_xor_b32_e32 v0, v0, v2
 ; GCN-NEXT:    v_cvt_f32_u32_e32 v2, v0
-; GCN-NEXT:    v_cvt_f32_u32_e32 v5, v1
-; GCN-NEXT:    v_sub_i32_e32 v6, vcc, 0, v0
-; GCN-NEXT:    v_subb_u32_e32 v7, vcc, 0, v1, vcc
-; GCN-NEXT:    v_mac_f32_e32 v2, 0x4f800000, v5
+; GCN-NEXT:    v_cvt_f32_u32_e32 v3, v1
+; GCN-NEXT:    v_sub_i32_e32 v4, vcc, 0, v0
+; GCN-NEXT:    v_subb_u32_e32 v5, vcc, 0, v1, vcc
+; GCN-NEXT:    v_mov_b32_e32 v12, 0
+; GCN-NEXT:    v_mac_f32_e32 v2, 0x4f800000, v3
 ; GCN-NEXT:    v_rcp_f32_e32 v2, v2
+; GCN-NEXT:    v_mov_b32_e32 v11, 0
 ; GCN-NEXT:    v_mul_f32_e32 v2, 0x5f7ffffc, v2
-; GCN-NEXT:    v_mul_f32_e32 v5, 0x2f800000, v2
-; GCN-NEXT:    v_trunc_f32_e32 v5, v5
-; GCN-NEXT:    v_mac_f32_e32 v2, 0xcf800000, v5
-; GCN-NEXT:    v_cvt_u32_f32_e32 v5, v5
+; GCN-NEXT:    v_mul_f32_e32 v3, 0x2f800000, v2
+; GCN-NEXT:    v_trunc_f32_e32 v3, v3
+; GCN-NEXT:    v_mac_f32_e32 v2, 0xcf800000, v3
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v2, v2
-; GCN-NEXT:    v_mul_lo_u32 v8, v6, v5
-; GCN-NEXT:    v_mul_lo_u32 v9, v7, v2
-; GCN-NEXT:    v_mul_hi_u32 v10, v6, v2
-; GCN-NEXT:    v_mul_lo_u32 v11, v6, v2
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v10, v8
-; GCN-NEXT:    v_mul_hi_u32 v10, v2, v11
-; GCN-NEXT:    v_mul_hi_u32 v12, v5, v11
-; GCN-NEXT:    v_mul_lo_u32 v11, v5, v11
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v8, v9
-; GCN-NEXT:    v_mul_hi_u32 v9, v2, v8
-; GCN-NEXT:    v_mul_lo_u32 v13, v2, v8
-; GCN-NEXT:    v_mul_hi_u32 v14, v5, v8
-; GCN-NEXT:    v_mul_lo_u32 v8, v5, v8
-; GCN-NEXT:    v_add_i32_e32 v10, vcc, v10, v13
-; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v4, v9, vcc
-; GCN-NEXT:    v_add_i32_e32 v10, vcc, v11, v10
-; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v9, v12, vcc
-; GCN-NEXT:    v_addc_u32_e32 v10, vcc, v14, v3, vcc
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v9, v8
-; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v4, v10, vcc
-; GCN-NEXT:    v_add_i32_e64 v2, s[4:5], v2, v8
-; GCN-NEXT:    v_addc_u32_e64 v8, vcc, v5, v9, s[4:5]
-; GCN-NEXT:    v_mul_hi_u32 v10, v6, v2
-; GCN-NEXT:    v_mul_lo_u32 v7, v7, v2
-; GCN-NEXT:    v_mul_lo_u32 v11, v6, v2
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v9
-; GCN-NEXT:    v_mul_lo_u32 v6, v6, v8
-; GCN-NEXT:    v_mul_hi_u32 v9, v8, v11
-; GCN-NEXT:    v_mul_lo_u32 v12, v8, v11
-; GCN-NEXT:    v_mul_hi_u32 v11, v2, v11
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, v10, v6
+; GCN-NEXT:    v_cvt_u32_f32_e32 v3, v3
+; GCN-NEXT:    v_mul_hi_u32 v6, v4, v2
+; GCN-NEXT:    v_mul_lo_u32 v7, v4, v3
+; GCN-NEXT:    v_mul_lo_u32 v8, v5, v2
 ; GCN-NEXT:    v_add_i32_e32 v6, vcc, v6, v7
-; GCN-NEXT:    v_mul_hi_u32 v7, v8, v6
-; GCN-NEXT:    v_mul_hi_u32 v10, v2, v6
-; GCN-NEXT:    v_mul_lo_u32 v13, v2, v6
-; GCN-NEXT:    v_mul_lo_u32 v6, v8, v6
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v11, v13
-; GCN-NEXT:    v_addc_u32_e32 v10, vcc, v4, v10, vcc
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v12, v8
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v10, v9, vcc
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v7, v3, vcc
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, v8, v6
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v4, v7, vcc
-; GCN-NEXT:    v_addc_u32_e64 v5, vcc, v5, v7, s[4:5]
-; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v6
-; GCN-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; GCN-NEXT:    v_mul_hi_u32 v6, s6, v2
-; GCN-NEXT:    v_mul_hi_u32 v2, 0, v2
-; GCN-NEXT:    v_mul_hi_u32 v7, s6, v5
-; GCN-NEXT:    v_lshlrev_b32_e32 v8, 15, v5
-; GCN-NEXT:    v_mul_hi_u32 v5, 0, v5
+; GCN-NEXT:    v_mul_lo_u32 v7, v4, v2
 ; GCN-NEXT:    v_add_i32_e32 v6, vcc, v6, v8
-; GCN-NEXT:    v_addc_u32_e32 v4, vcc, v4, v7, vcc
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, 0, v6
-; GCN-NEXT:    v_addc_u32_e32 v2, vcc, v4, v2, vcc
-; GCN-NEXT:    v_addc_u32_e32 v3, vcc, v5, v3, vcc
+; GCN-NEXT:    v_mul_lo_u32 v9, v2, v6
+; GCN-NEXT:    v_mul_hi_u32 v8, v2, v6
+; GCN-NEXT:    v_mul_hi_u32 v10, v2, v7
+; GCN-NEXT:    v_mul_hi_u32 v13, v3, v6
+; GCN-NEXT:    v_mul_lo_u32 v6, v3, v6
+; GCN-NEXT:    v_add_i32_e32 v9, vcc, v10, v9
+; GCN-NEXT:    v_mul_lo_u32 v10, v3, v7
+; GCN-NEXT:    v_mul_hi_u32 v7, v3, v7
+; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v12, v8, vcc
+; GCN-NEXT:    v_add_i32_e32 v9, vcc, v10, v9
+; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v8, v7, vcc
+; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v13, v11, vcc
+; GCN-NEXT:    v_add_i32_e32 v6, vcc, v7, v6
+; GCN-NEXT:    v_add_i32_e64 v2, s[4:5], v2, v6
+; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v12, v8, vcc
+; GCN-NEXT:    v_addc_u32_e64 v6, vcc, v3, v7, s[4:5]
+; GCN-NEXT:    v_mul_lo_u32 v8, v4, v6
+; GCN-NEXT:    v_mul_hi_u32 v9, v4, v2
+; GCN-NEXT:    v_mul_lo_u32 v5, v5, v2
+; GCN-NEXT:    v_mul_lo_u32 v4, v4, v2
+; GCN-NEXT:    v_add_i32_e32 v8, vcc, v9, v8
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v8, v5
+; GCN-NEXT:    v_mul_lo_u32 v10, v2, v5
+; GCN-NEXT:    v_mul_hi_u32 v13, v2, v4
+; GCN-NEXT:    v_mul_hi_u32 v14, v2, v5
+; GCN-NEXT:    v_mul_hi_u32 v9, v6, v4
+; GCN-NEXT:    v_mul_lo_u32 v4, v6, v4
+; GCN-NEXT:    v_add_i32_e32 v10, vcc, v13, v10
+; GCN-NEXT:    v_mul_hi_u32 v8, v6, v5
+; GCN-NEXT:    v_addc_u32_e32 v13, vcc, v12, v14, vcc
+; GCN-NEXT:    v_mul_lo_u32 v5, v6, v5
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v4, v10
+; GCN-NEXT:    v_addc_u32_e32 v4, vcc, v13, v9, vcc
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v8, v11, vcc
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v4, v5
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v12, v6, vcc
+; GCN-NEXT:    v_add_i32_e32 v3, vcc, v3, v7
+; GCN-NEXT:    v_addc_u32_e64 v3, vcc, v3, v5, s[4:5]
+; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v4
+; GCN-NEXT:    s_mov_b32 s4, 0x8000
+; GCN-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; GCN-NEXT:    v_mul_hi_u32 v4, s4, v2
+; GCN-NEXT:    v_mul_hi_u32 v5, s4, v3
+; GCN-NEXT:    v_lshlrev_b32_e32 v6, 15, v3
+; GCN-NEXT:    v_mul_hi_u32 v2, 0, v2
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v4, v6
+; GCN-NEXT:    v_mul_hi_u32 v3, 0, v3
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v12, v5, vcc
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, 0, v4
+; GCN-NEXT:    v_addc_u32_e32 v2, vcc, v5, v2, vcc
+; GCN-NEXT:    v_addc_u32_e32 v3, vcc, v3, v11, vcc
+; GCN-NEXT:    v_mul_lo_u32 v3, v0, v3
 ; GCN-NEXT:    v_mul_hi_u32 v4, v0, v2
 ; GCN-NEXT:    v_mul_lo_u32 v5, v1, v2
 ; GCN-NEXT:    v_mul_lo_u32 v2, v0, v2
-; GCN-NEXT:    v_mul_lo_u32 v3, v0, v3
 ; GCN-NEXT:    v_add_i32_e32 v3, vcc, v4, v3
 ; GCN-NEXT:    v_add_i32_e32 v3, vcc, v3, v5
 ; GCN-NEXT:    v_sub_i32_e32 v4, vcc, 0, v3
-; GCN-NEXT:    v_sub_i32_e32 v2, vcc, s6, v2
+; GCN-NEXT:    v_sub_i32_e32 v2, vcc, s4, v2
 ; GCN-NEXT:    v_subb_u32_e64 v4, s[4:5], v4, v1, vcc
+; GCN-NEXT:    v_sub_i32_e64 v5, s[4:5], v2, v0
+; GCN-NEXT:    v_subb_u32_e64 v6, s[6:7], v4, v1, s[4:5]
+; GCN-NEXT:    v_subbrev_u32_e64 v4, s[4:5], 0, v4, s[4:5]
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[4:5], v4, v1
+; GCN-NEXT:    v_cndmask_b32_e64 v7, 0, -1, s[4:5]
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[4:5], v5, v0
 ; GCN-NEXT:    v_subb_u32_e32 v3, vcc, 0, v3, vcc
+; GCN-NEXT:    v_cndmask_b32_e64 v8, 0, -1, s[4:5]
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[4:5], v4, v1
+; GCN-NEXT:    v_cndmask_b32_e64 v7, v7, v8, s[4:5]
+; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v3, v1
+; GCN-NEXT:    v_cmp_ne_u32_e64 s[6:7], 0, v7
+; GCN-NEXT:    v_cndmask_b32_e64 v7, 0, -1, vcc
 ; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v2, v0
-; GCN-NEXT:    v_cndmask_b32_e64 v5, 0, -1, vcc
-; GCN-NEXT:    v_sub_i32_e32 v6, vcc, v2, v0
-; GCN-NEXT:    v_subb_u32_e64 v7, s[4:5], v4, v1, vcc
-; GCN-NEXT:    v_subbrev_u32_e32 v4, vcc, 0, v4, vcc
-; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v6, v0
-; GCN-NEXT:    v_cndmask_b32_e64 v8, 0, -1, vcc
-; GCN-NEXT:    v_sub_i32_e32 v0, vcc, v6, v0
-; GCN-NEXT:    v_cmp_ge_u32_e64 s[4:5], v3, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v9, 0, -1, s[4:5]
-; GCN-NEXT:    v_cmp_ge_u32_e64 s[4:5], v4, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v10, 0, -1, s[4:5]
-; GCN-NEXT:    v_cmp_eq_u32_e64 s[4:5], v3, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v5, v9, v5, s[4:5]
-; GCN-NEXT:    v_subbrev_u32_e32 v7, vcc, 0, v7, vcc
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v1
-; GCN-NEXT:    v_cndmask_b32_e32 v1, v10, v8, vcc
-; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v1
-; GCN-NEXT:    v_cndmask_b32_e32 v0, v6, v0, vcc
-; GCN-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, v5
-; GCN-NEXT:    v_cndmask_b32_e64 v0, v2, v0, s[4:5]
-; GCN-NEXT:    v_cndmask_b32_e32 v1, v4, v7, vcc
-; GCN-NEXT:    v_cndmask_b32_e64 v1, v3, v1, s[4:5]
+; GCN-NEXT:    v_sub_i32_e64 v8, s[4:5], v5, v0
+; GCN-NEXT:    v_cndmask_b32_e64 v0, 0, -1, vcc
+; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v1
+; GCN-NEXT:    v_cndmask_b32_e32 v0, v7, v0, vcc
+; GCN-NEXT:    v_subbrev_u32_e64 v1, s[4:5], 0, v6, s[4:5]
+; GCN-NEXT:    v_cndmask_b32_e64 v5, v5, v8, s[6:7]
+; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
+; GCN-NEXT:    v_cndmask_b32_e64 v1, v4, v1, s[6:7]
+; GCN-NEXT:    v_cndmask_b32_e32 v0, v2, v5, vcc
+; GCN-NEXT:    v_cndmask_b32_e32 v1, v3, v1, vcc
 ; GCN-NEXT:    s_setpc_b64 s[30:31]
   %result = srem i64 32768, %x
   ret i64 %result
 }
 
 define i64 @v_test_srem_pow2_k_den_i64(i64 %x) {
 ; GCN-LABEL: v_test_srem_pow2_k_den_i64:
 ; GCN:       ; %bb.0:
; GCN-NEXT:    s_setpc_b64 s[30:31]
   %result = srem i64 %x, 32768
   ret i64 %result
 }
 
 define amdgpu_kernel void @s_test_srem24_k_num_i64(i64 addrspace(1)* %out, i64 %x) {
 ; GCN-LABEL: s_test_srem24_k_num_i64:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
-; GCN-NEXT:    s_mov_b32 s7, 0xf000
-; GCN-NEXT:    s_mov_b32 s6, -1
+; GCN-NEXT:    s_mov_b32 s6, 0x41c00000
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_mov_b32 s2, 0x41c00000
-; GCN-NEXT:    s_mov_b32 s4, s0
-; GCN-NEXT:    s_mov_b32 s5, s1
-; GCN-NEXT:    s_ashr_i64 s[0:1], s[2:3], 40
-; GCN-NEXT:    v_cvt_f32_i32_e32 v0, s0
-; GCN-NEXT:    s_ashr_i32 s1, s0, 30
+; GCN-NEXT:    s_ashr_i64 s[4:5], s[2:3], 40
+; GCN-NEXT:    v_cvt_f32_i32_e32 v0, s4
+; GCN-NEXT:    s_ashr_i32 s5, s4, 30
+; GCN-NEXT:    s_or_b32 s5, s5, 1
+; GCN-NEXT:    v_mov_b32_e32 v2, s5
 ; GCN-NEXT:    v_rcp_iflag_f32_e32 v1, v0
-; GCN-NEXT:    s_or_b32 s1, s1, 1
-; GCN-NEXT:    v_mul_f32_e32 v1, s2, v1
-; GCN-NEXT:    v_mov_b32_e32 v2, s1
+; GCN-NEXT:    s_mov_b32 s3, 0xf000
+; GCN-NEXT:    s_mov_b32 s2, -1
+; GCN-NEXT:    v_mul_f32_e32 v1, s6, v1
 ; GCN-NEXT:    v_trunc_f32_e32 v1, v1
-; GCN-NEXT:    v_mad_f32 v3, -v1, v0, s2
+; GCN-NEXT:    v_mad_f32 v3, -v1, v0, s6
+; GCN-NEXT:    v_cvt_i32_f32_e32 v1, v1
 ; GCN-NEXT:    v_cmp_ge_f32_e64 vcc, |v3|, |v0|
 ; GCN-NEXT:    v_cndmask_b32_e32 v0, 0, v2, vcc
-; GCN-NEXT:    v_cvt_i32_f32_e32 v1, v1
 ; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v1
-; GCN-NEXT:    v_mul_lo_u32 v0, v0, s0
+; GCN-NEXT:    v_mul_lo_u32 v0, v0, s4
 ; GCN-NEXT:    v_sub_i32_e32 v0, vcc, 24, v0
 ; GCN-NEXT:    v_bfe_i32 v0, v0, 0, 24
 ; GCN-NEXT:    v_ashrrev_i32_e32 v1, 31, v0
-; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
+; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 ; GCN-NEXT:    s_endpgm
   %x.shr = ashr i64 %x, 40
   %result = srem i64 24, %x.shr
   store i64 %result, i64 addrspace(1)* %out
   ret void
 }
 
 define amdgpu_kernel void @s_test_srem24_k_den_i64(i64 addrspace(1)* %out, i64 %x) {
 ; GCN-LABEL: s_test_srem24_k_den_i64:
 ; GCN:       ; %bb.0:
-; GCN-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
-; GCN-NEXT:    s_mov_b32 s7, 0xf000
-; GCN-NEXT:    s_mov_b32 s6, -1
+; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN-NEXT:    s_mov_b32 s8, 0x46b6fe00
+; GCN-NEXT:    s_mov_b32 s3, 0xf000
+; GCN-NEXT:    s_mov_b32 s2, -1
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_mov_b32 s2, 0x46b6fe00
-; GCN-NEXT:    s_mov_b32 s4, s0
-; GCN-NEXT:    s_mov_b32 s5, s1
-; GCN-NEXT:    s_ashr_i64 s[0:1], s[2:3], 40
-; GCN-NEXT:    v_cvt_f32_i32_e32 v0, s0
-; GCN-NEXT:    s_ashr_i32 s1, s0, 30
-; GCN-NEXT:    v_mul_f32_e32 v1, 0x38331158, v0
-; GCN-NEXT:    s_or_b32 s1, s1, 1
-; GCN-NEXT:    v_trunc_f32_e32 v1, v1
-; GCN-NEXT:    v_mov_b32_e32 v2, s1
-; GCN-NEXT:    v_mad_f32 v0, -v1, s2, v0
-; GCN-NEXT:    v_cvt_i32_f32_e32 v1, v1
-; GCN-NEXT:    v_cmp_ge_f32_e64 vcc, |v0|, s2
-; GCN-NEXT:    v_cndmask_b32_e32 v0, 0, v2, vcc
-; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v1
-; GCN-NEXT:    s_movk_i32 s1, 0x5b7f
-; GCN-NEXT:    v_mul_lo_u32 v0, v0, s1
-; GCN-NEXT:    v_sub_i32_e32 v0, vcc, s0, v0
+; GCN-NEXT:    s_ashr_i64 s[6:7], s[6:7], 40
+; GCN-NEXT:    v_cvt_f32_i32_e32 v0, s6
+; GCN-NEXT:    s_ashr_i32 s0, s6, 30
+; GCN-NEXT:    s_or_b32 s0, s0, 1
+; GCN-NEXT:    v_mov_b32_e32 v1, s0
+; GCN-NEXT:    v_mul_f32_e32 v2, 0x38331158, v0
+; GCN-NEXT:    v_trunc_f32_e32 v2, v2
+; GCN-NEXT:    v_mad_f32 v0, -v2, s8, v0
+; GCN-NEXT:    v_cvt_i32_f32_e32 v2, v2
+; GCN-NEXT:    v_cmp_ge_f32_e64 vcc, |v0|, s8
+; GCN-NEXT:    v_cndmask_b32_e32 v0, 0, v1, vcc
+; GCN-NEXT:    s_movk_i32 s0, 0x5b7f
+; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v2
+; GCN-NEXT:    v_mul_lo_u32 v0, v0, s0
+; GCN-NEXT:    s_mov_b32 s0, s4
+; GCN-NEXT:    s_mov_b32 s1, s5
+; GCN-NEXT:    v_sub_i32_e32 v0, vcc, s6, v0
 ; GCN-NEXT:    v_bfe_i32 v0, v0, 0, 24
 ; GCN-NEXT:    v_ashrrev_i32_e32 v1, 31, v0
-; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
+; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 ; GCN-NEXT:    s_endpgm
   %x.shr = ashr i64 %x, 40
   %result = srem i64 %x.shr, 23423
   store i64 %result, i64 addrspace(1)* %out
   ret void
 }
 
 define i64 @v_test_srem24_k_num_i64(i64 %x) {
 ; GCN-LABEL: v_test_srem24_k_num_i64:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GCN-NEXT:    v_ashr_i64 v[0:1], v[0:1], 40
 ; GCN-NEXT:    s_mov_b32 s4, 0x41c00000
 ; GCN-NEXT:    v_cvt_f32_i32_e32 v1, v0
-; GCN-NEXT:    v_ashrrev_i32_e32 v2, 30, v0
-; GCN-NEXT:    v_rcp_iflag_f32_e32 v3, v1
-; GCN-NEXT:    v_or_b32_e32 v2, 1, v2
-; GCN-NEXT:    v_mul_f32_e32 v3, s4, v3
-; GCN-NEXT:    v_trunc_f32_e32 v3, v3
-; GCN-NEXT:    v_mad_f32 v4, -v3, v1, s4
+; GCN-NEXT:    v_ashrrev_i32_e32 v3, 30, v0
+; GCN-NEXT:    v_or_b32_e32 v3, 1, v3
+; GCN-NEXT:    v_rcp_iflag_f32_e32 v2, v1
+; GCN-NEXT:    v_mul_f32_e32 v2, s4, v2
+; GCN-NEXT:    v_trunc_f32_e32 v2, v2
+; GCN-NEXT:    v_mad_f32 v4, -v2, v1, s4
+; GCN-NEXT:    v_cvt_i32_f32_e32 v2, v2
 ; GCN-NEXT:    v_cmp_ge_f32_e64 vcc, |v4|, |v1|
-; GCN-NEXT:    v_cndmask_b32_e32 v1, 0, v2, vcc
-; GCN-NEXT:    v_cvt_i32_f32_e32 v2, v3
+; GCN-NEXT:    v_cndmask_b32_e32 v1, 0, v3, vcc
 ; GCN-NEXT:    v_add_i32_e32 v1, vcc, v2, v1
 ; GCN-NEXT:    v_mul_lo_u32 v0, v1, v0
 ; GCN-NEXT:    v_sub_i32_e32 v0, vcc, 24, v0
 ; GCN-NEXT:    v_bfe_i32 v0, v0, 0, 24
 ; GCN-NEXT:    v_ashrrev_i32_e32 v1, 31, v0
 ; GCN-NEXT:    s_setpc_b64 s[30:31]
   %x.shr = ashr i64 %x, 40
   %result = srem i64 24, %x.shr
   ret i64 %result
 }
 
 define i64 @v_test_srem24_pow2_k_num_i64(i64 %x) {
 ; GCN-LABEL: v_test_srem24_pow2_k_num_i64:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GCN-NEXT:    v_ashr_i64 v[0:1], v[0:1], 40
 ; GCN-NEXT:    s_mov_b32 s4, 0x47000000
 ; GCN-NEXT:    v_cvt_f32_i32_e32 v1, v0
-; GCN-NEXT:    v_ashrrev_i32_e32 v2, 30, v0
-; GCN-NEXT:    v_rcp_iflag_f32_e32 v3, v1
-; GCN-NEXT:    v_or_b32_e32 v2, 1, v2
-; GCN-NEXT:    v_mul_f32_e32 v3, s4, v3
-; GCN-NEXT:    v_trunc_f32_e32 v3, v3
-; GCN-NEXT:    v_mad_f32 v4, -v3, v1, s4
+; GCN-NEXT:    v_ashrrev_i32_e32 v3, 30, v0
+; GCN-NEXT:    v_or_b32_e32 v3, 1, v3
+; GCN-NEXT:    v_rcp_iflag_f32_e32 v2, v1
+; GCN-NEXT:    v_mul_f32_e32 v2, s4, v2
+; GCN-NEXT:    v_trunc_f32_e32 v2, v2
+; GCN-NEXT:    v_mad_f32 v4, -v2, v1, s4
+; GCN-NEXT:    v_cvt_i32_f32_e32 v2, v2
 ; GCN-NEXT:    v_cmp_ge_f32_e64 vcc, |v4|, |v1|
-; GCN-NEXT:    v_cndmask_b32_e32 v1, 0, v2, vcc
-; GCN-NEXT:    v_cvt_i32_f32_e32 v2, v3
+; GCN-NEXT:    v_cndmask_b32_e32 v1, 0, v3, vcc
 ; GCN-NEXT:    v_add_i32_e32 v1, vcc, v2, v1
 ; GCN-NEXT:    v_mul_lo_u32 v0, v1, v0
 ; GCN-NEXT:    v_sub_i32_e32 v0, vcc, 0x8000, v0
 ; GCN-NEXT:    v_bfe_i32 v0, v0, 0, 24
 ; GCN-NEXT:    v_ashrrev_i32_e32 v1, 31, v0
 ; GCN-NEXT:    s_setpc_b64 s[30:31]
   %x.shr = ashr i64 %x, 40
   %result = srem i64 32768, %x.shr

llvm/test/CodeGen/AMDGPU/udiv64.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
+; RUN: llc -march=amdgcn -mcpu=gfx600 -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
 
 define amdgpu_kernel void @s_test_udiv_i64(i64 addrspace(1)* %out, i64 %x, i64 %y) {
 ; GCN-LABEL: s_test_udiv_i64:
 ; GCN:       ; %bb.0:
-; GCN-NEXT:    s_load_dwordx4 s[8:11], s[0:1], 0x9
 ; GCN-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xd
+; GCN-NEXT:    v_mov_b32_e32 v2, 0
+; GCN-NEXT:    s_load_dwordx4 s[8:11], s[0:1], 0x9
 ; GCN-NEXT:    s_mov_b32 s7, 0xf000
 ; GCN-NEXT:    s_mov_b32 s6, -1
-; GCN-NEXT:    v_mov_b32_e32 v0, 0
-; GCN-NEXT:    v_mov_b32_e32 v1, 0
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_mov_b32 s4, s8
-; GCN-NEXT:    s_mov_b32 s5, s9
-; GCN-NEXT:    v_cvt_f32_u32_e32 v2, s2
-; GCN-NEXT:    v_cvt_f32_u32_e32 v3, s3
-; GCN-NEXT:    s_sub_u32 s8, 0, s2
-; GCN-NEXT:    v_mov_b32_e32 v4, s3
-; GCN-NEXT:    v_mov_b32_e32 v5, s11
-; GCN-NEXT:    v_mac_f32_e32 v2, 0x4f800000, v3
-; GCN-NEXT:    s_subb_u32 s9, 0, s3
-; GCN-NEXT:    v_rcp_f32_e32 v2, v2
-; GCN-NEXT:    v_mul_f32_e32 v2, 0x5f7ffffc, v2
-; GCN-NEXT:    v_mul_f32_e32 v3, 0x2f800000, v2
+; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s2
+; GCN-NEXT:    v_cvt_f32_u32_e32 v1, s3
+; GCN-NEXT:    s_sub_u32 s4, 0, s2
+; GCN-NEXT:    s_subb_u32 s5, 0, s3
+; GCN-NEXT:    v_mac_f32_e32 v0, 0x4f800000, v1
+; GCN-NEXT:    v_rcp_f32_e32 v0, v0
+; GCN-NEXT:    v_mov_b32_e32 v1, 0
+; GCN-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v0
+; GCN-NEXT:    v_mul_f32_e32 v3, 0x2f800000, v0
 ; GCN-NEXT:    v_trunc_f32_e32 v3, v3
-; GCN-NEXT:    v_mac_f32_e32 v2, 0xcf800000, v3
+; GCN-NEXT:    v_mac_f32_e32 v0, 0xcf800000, v3
+; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v0
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v3, v3
-; GCN-NEXT:    v_cvt_u32_f32_e32 v2, v2
-; GCN-NEXT:    v_mul_lo_u32 v6, s8, v3
-; GCN-NEXT:    v_mul_lo_u32 v7, s9, v2
-; GCN-NEXT:    v_mul_hi_u32 v8, s8, v2
-; GCN-NEXT:    v_mul_lo_u32 v9, s8, v2
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, v8, v6
-; GCN-NEXT:    v_mul_hi_u32 v8, v2, v9
-; GCN-NEXT:    v_mul_hi_u32 v10, v3, v9
-; GCN-NEXT:    v_mul_lo_u32 v9, v3, v9
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, v6, v7
-; GCN-NEXT:    v_mul_hi_u32 v7, v2, v6
-; GCN-NEXT:    v_mul_lo_u32 v11, v2, v6
-; GCN-NEXT:    v_mul_hi_u32 v12, v3, v6
-; GCN-NEXT:    v_mul_lo_u32 v6, v3, v6
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v8, v11
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v1, v7, vcc
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v9, v8
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v7, v10, vcc
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v12, v0, vcc
+; GCN-NEXT:    v_mul_hi_u32 v5, s4, v0
+; GCN-NEXT:    v_mul_lo_u32 v4, s4, v3
+; GCN-NEXT:    v_mul_lo_u32 v7, s5, v0
+; GCN-NEXT:    v_mul_lo_u32 v6, s4, v0
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v5, v4
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v4, v7
+; GCN-NEXT:    v_mul_hi_u32 v5, v0, v6
+; GCN-NEXT:    v_mul_lo_u32 v7, v0, v4
+; GCN-NEXT:    v_mul_hi_u32 v9, v0, v4
+; GCN-NEXT:    v_mul_lo_u32 v8, v3, v6
+; GCN-NEXT:    v_mul_hi_u32 v6, v3, v6
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v7
+; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v2, v9, vcc
+; GCN-NEXT:    v_mul_hi_u32 v9, v3, v4
+; GCN-NEXT:    v_mul_lo_u32 v4, v3, v4
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v8, v5
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v7, v6, vcc
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v9, v1, vcc
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v5, v4
+; GCN-NEXT:    v_add_i32_e64 v0, s[0:1], v0, v4
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v2, v6, vcc
+; GCN-NEXT:    v_addc_u32_e64 v4, vcc, v3, v5, s[0:1]
+; GCN-NEXT:    v_mul_lo_u32 v6, s4, v4
+; GCN-NEXT:    v_mul_hi_u32 v7, s4, v0
+; GCN-NEXT:    v_mul_lo_u32 v8, s5, v0
+; GCN-NEXT:    s_mov_b32 s5, s9
 ; GCN-NEXT:    v_add_i32_e32 v6, vcc, v7, v6
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v1, v8, vcc
-; GCN-NEXT:    v_add_i32_e64 v2, s[0:1], v2, v6
-; GCN-NEXT:    v_addc_u32_e64 v6, vcc, v3, v7, s[0:1]
-; GCN-NEXT:    v_mul_hi_u32 v8, s8, v2
-; GCN-NEXT:    v_mul_lo_u32 v9, s9, v2
-; GCN-NEXT:    v_mul_lo_u32 v10, s8, v2
-; GCN-NEXT:    v_mul_lo_u32 v11, s8, v6
-; GCN-NEXT:    v_mul_hi_u32 v12, v6, v10
-; GCN-NEXT:    v_mul_lo_u32 v13, v6, v10
-; GCN-NEXT:    v_mul_hi_u32 v10, v2, v10
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v8, v11
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v9, v8
-; GCN-NEXT:    v_mul_hi_u32 v9, v6, v8
-; GCN-NEXT:    v_mul_hi_u32 v11, v2, v8
-; GCN-NEXT:    v_mul_lo_u32 v14, v2, v8
-; GCN-NEXT:    v_mul_lo_u32 v6, v6, v8
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v10, v14
-; GCN-NEXT:    v_addc_u32_e32 v10, vcc, v1, v11, vcc
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v13, v8
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v10, v12, vcc
-; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v9, v0, vcc
-; GCN-NEXT:    v_add_i32_e32 v3, vcc, v3, v7
+; GCN-NEXT:    v_mul_lo_u32 v7, s4, v0
 ; GCN-NEXT:    v_add_i32_e32 v6, vcc, v8, v6
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v1, v9, vcc
-; GCN-NEXT:    v_addc_u32_e64 v3, vcc, v3, v7, s[0:1]
-; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v6
+; GCN-NEXT:    v_mul_lo_u32 v10, v0, v6
+; GCN-NEXT:    v_mul_hi_u32 v12, v0, v6
+; GCN-NEXT:    v_mul_hi_u32 v11, v0, v7
+; GCN-NEXT:    v_mul_hi_u32 v9, v4, v7
+; GCN-NEXT:    v_mul_lo_u32 v7, v4, v7
+; GCN-NEXT:    v_mul_hi_u32 v8, v4, v6
+; GCN-NEXT:    v_add_i32_e32 v10, vcc, v11, v10
+; GCN-NEXT:    v_addc_u32_e32 v11, vcc, v2, v12, vcc
+; GCN-NEXT:    v_mul_lo_u32 v4, v4, v6
+; GCN-NEXT:    v_add_i32_e32 v7, vcc, v7, v10
+; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v11, v9, vcc
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v8, v1, vcc
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v7, v4
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v2, v6, vcc
+; GCN-NEXT:    v_add_i32_e32 v3, vcc, v3, v5
+; GCN-NEXT:    v_addc_u32_e64 v3, vcc, v3, v6, s[0:1]
+; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v4
 ; GCN-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; GCN-NEXT:    v_mul_hi_u32 v6, s10, v2
-; GCN-NEXT:    v_mul_hi_u32 v7, s11, v2
-; GCN-NEXT:    v_mul_lo_u32 v2, s11, v2
-; GCN-NEXT:    v_mul_hi_u32 v8, s10, v3
-; GCN-NEXT:    v_mul_lo_u32 v9, s10, v3
-; GCN-NEXT:    v_mul_hi_u32 v10, s11, v3
+; GCN-NEXT:    v_mul_lo_u32 v4, s10, v3
+; GCN-NEXT:    v_mul_hi_u32 v5, s10, v0
+; GCN-NEXT:    v_mul_hi_u32 v6, s10, v3
+; GCN-NEXT:    v_mul_hi_u32 v7, s11, v3
 ; GCN-NEXT:    v_mul_lo_u32 v3, s11, v3
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, v6, v9
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v1, v8, vcc
-; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v6
-; GCN-NEXT:    v_addc_u32_e32 v2, vcc, v8, v7, vcc
-; GCN-NEXT:    v_addc_u32_e32 v0, vcc, v10, v0, vcc
-; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v3
-; GCN-NEXT:    v_addc_u32_e32 v0, vcc, v1, v0, vcc
-; GCN-NEXT:    v_mul_hi_u32 v1, s2, v2
-; GCN-NEXT:    v_mul_lo_u32 v3, s3, v2
-; GCN-NEXT:    v_mul_lo_u32 v6, s2, v2
-; GCN-NEXT:    v_mul_lo_u32 v7, s2, v0
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, 2, v2
-; GCN-NEXT:    v_addc_u32_e32 v9, vcc, 0, v0, vcc
-; GCN-NEXT:    v_add_i32_e32 v10, vcc, 1, v2
-; GCN-NEXT:    v_addc_u32_e32 v11, vcc, 0, v0, vcc
-; GCN-NEXT:    v_add_i32_e32 v1, vcc, v1, v7
-; GCN-NEXT:    v_add_i32_e32 v1, vcc, v1, v3
-; GCN-NEXT:    v_sub_i32_e32 v3, vcc, s11, v1
-; GCN-NEXT:    v_sub_i32_e32 v6, vcc, s10, v6
-; GCN-NEXT:    v_subb_u32_e64 v3, s[0:1], v3, v4, vcc
-; GCN-NEXT:    v_subrev_i32_e64 v4, s[0:1], s2, v6
-; GCN-NEXT:    v_subb_u32_e32 v1, vcc, v5, v1, vcc
-; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s2, v6
-; GCN-NEXT:    v_cndmask_b32_e64 v5, 0, -1, vcc
-; GCN-NEXT:    v_subbrev_u32_e64 v3, vcc, 0, v3, s[0:1]
-; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s2, v4
-; GCN-NEXT:    v_cndmask_b32_e64 v4, 0, -1, vcc
-; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s3, v1
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v5, v4
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v2, v6, vcc
+; GCN-NEXT:    v_mul_lo_u32 v6, s11, v0
+; GCN-NEXT:    v_mul_hi_u32 v0, s11, v0
+; GCN-NEXT:    s_mov_b32 s4, s8
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v6, v4
+; GCN-NEXT:    v_addc_u32_e32 v0, vcc, v5, v0, vcc
+; GCN-NEXT:    v_addc_u32_e32 v1, vcc, v7, v1, vcc
+; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v3
+; GCN-NEXT:    v_addc_u32_e32 v1, vcc, v2, v1, vcc
+; GCN-NEXT:    v_mul_lo_u32 v2, s2, v1
+; GCN-NEXT:    v_mul_hi_u32 v3, s2, v0
+; GCN-NEXT:    v_mul_lo_u32 v4, s3, v0
+; GCN-NEXT:    v_mov_b32_e32 v5, s3
+; GCN-NEXT:    v_add_i32_e32 v2, vcc, v3, v2
+; GCN-NEXT:    v_mul_lo_u32 v3, s2, v0
+; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v4
+; GCN-NEXT:    v_sub_i32_e32 v4, vcc, s11, v2
+; GCN-NEXT:    v_sub_i32_e32 v3, vcc, s10, v3
+; GCN-NEXT:    v_subb_u32_e64 v4, s[0:1], v4, v5, vcc
+; GCN-NEXT:    v_subrev_i32_e64 v5, s[0:1], s2, v3
+; GCN-NEXT:    v_subbrev_u32_e64 v4, s[0:1], 0, v4, s[0:1]
+; GCN-NEXT:    v_cmp_le_u32_e64 s[0:1], s3, v4
+; GCN-NEXT:    v_cndmask_b32_e64 v6, 0, -1, s[0:1]
+; GCN-NEXT:    v_cmp_le_u32_e64 s[0:1], s2, v5
+; GCN-NEXT:    v_cndmask_b32_e64 v5, 0, -1, s[0:1]
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], s3, v4
+; GCN-NEXT:    v_cndmask_b32_e64 v4, v6, v5, s[0:1]
+; GCN-NEXT:    v_add_i32_e64 v5, s[0:1], 2, v0
+; GCN-NEXT:    v_addc_u32_e64 v6, s[0:1], 0, v1, s[0:1]
+; GCN-NEXT:    v_add_i32_e64 v7, s[0:1], 1, v0
+; GCN-NEXT:    v_addc_u32_e64 v8, s[0:1], 0, v1, s[0:1]
+; GCN-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v4
+; GCN-NEXT:    v_cndmask_b32_e64 v4, v8, v6, s[0:1]
+; GCN-NEXT:    v_mov_b32_e32 v6, s11
+; GCN-NEXT:    v_subb_u32_e32 v2, vcc, v6, v2, vcc
+; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s3, v2
 ; GCN-NEXT:    v_cndmask_b32_e64 v6, 0, -1, vcc
-; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s3, v3
-; GCN-NEXT:    v_cndmask_b32_e64 v7, 0, -1, vcc
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, s3, v1
-; GCN-NEXT:    v_cndmask_b32_e32 v1, v6, v5, vcc
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, s3, v3
-; GCN-NEXT:    v_cndmask_b32_e32 v3, v7, v4, vcc
-; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v3
-; GCN-NEXT:    v_cndmask_b32_e32 v3, v11, v9, vcc
-; GCN-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v1, v0, v3, s[0:1]
-; GCN-NEXT:    v_cndmask_b32_e32 v0, v10, v8, vcc
-; GCN-NEXT:    v_cndmask_b32_e64 v0, v2, v0, s[0:1]
+; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s2, v3
+; GCN-NEXT:    v_cndmask_b32_e64 v3, 0, -1, vcc
+; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, s3, v2
+; GCN-NEXT:    v_cndmask_b32_e32 v2, v6, v3, vcc
+; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v2, v7, v5, s[0:1]
+; GCN-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; GCN-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
 ; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
 ; GCN-NEXT:    s_endpgm
   %result = udiv i64 %x, %y
   store i64 %result, i64 addrspace(1)* %out
   ret void
 }
 
 define i64 @v_test_udiv_i64(i64 %x, i64 %y) {
 ; GCN-LABEL: v_test_udiv_i64:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GCN-NEXT:    v_cvt_f32_u32_e32 v4, v2
 ; GCN-NEXT:    v_cvt_f32_u32_e32 v5, v3
 ; GCN-NEXT:    v_sub_i32_e32 v6, vcc, 0, v2
 ; GCN-NEXT:    v_subb_u32_e32 v7, vcc, 0, v3, vcc
-; GCN-NEXT:    v_mov_b32_e32 v8, 0
-; GCN-NEXT:    v_mov_b32_e32 v9, 0
+; GCN-NEXT:    v_mov_b32_e32 v14, 0
 ; GCN-NEXT:    v_mac_f32_e32 v4, 0x4f800000, v5
 ; GCN-NEXT:    v_rcp_f32_e32 v4, v4
+; GCN-NEXT:    v_mov_b32_e32 v13, 0
 ; GCN-NEXT:    v_mul_f32_e32 v4, 0x5f7ffffc, v4
 ; GCN-NEXT:    v_mul_f32_e32 v5, 0x2f800000, v4
 ; GCN-NEXT:    v_trunc_f32_e32 v5, v5
 ; GCN-NEXT:    v_mac_f32_e32 v4, 0xcf800000, v5
-; GCN-NEXT:    v_cvt_u32_f32_e32 v5, v5
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v4, v4
-; GCN-NEXT:    v_mul_lo_u32 v10, v6, v5
-; GCN-NEXT:    v_mul_lo_u32 v11, v7, v4
-; GCN-NEXT:    v_mul_hi_u32 v12, v6, v4
-; GCN-NEXT:    v_mul_lo_u32 v13, v6, v4
-; GCN-NEXT:    v_add_i32_e32 v10, vcc, v12, v10
-; GCN-NEXT:    v_mul_hi_u32 v12, v4, v13
-; GCN-NEXT:    v_mul_hi_u32 v14, v5, v13
-; GCN-NEXT:    v_mul_lo_u32 v13, v5, v13
-; GCN-NEXT:    v_add_i32_e32 v10, vcc, v10, v11
-; GCN-NEXT:    v_mul_hi_u32 v11, v4, v10
-; GCN-NEXT:    v_mul_lo_u32 v15, v4, v10
-; GCN-NEXT:    v_mul_hi_u32 v16, v5, v10
-; GCN-NEXT:    v_mul_lo_u32 v10, v5, v10
-; GCN-NEXT:    v_add_i32_e32 v12, vcc, v12, v15
-; GCN-NEXT:    v_addc_u32_e32 v11, vcc, v9, v11, vcc
-; GCN-NEXT:    v_add_i32_e32 v12, vcc, v13, v12
-; GCN-NEXT:    v_addc_u32_e32 v11, vcc, v11, v14, vcc
-; GCN-NEXT:    v_addc_u32_e32 v12, vcc, v16, v8, vcc
-; GCN-NEXT:    v_add_i32_e32 v10, vcc, v11, v10
-; GCN-NEXT:    v_addc_u32_e32 v11, vcc, v9, v12, vcc
-; GCN-NEXT:    v_add_i32_e64 v4, s[4:5], v4, v10
-; GCN-NEXT:    v_addc_u32_e64 v10, vcc, v5, v11, s[4:5]
-; GCN-NEXT:    v_mul_hi_u32 v12, v6, v4
+; GCN-NEXT:    v_cvt_u32_f32_e32 v5, v5
+; GCN-NEXT:    v_mul_hi_u32 v8, v6, v4
+; GCN-NEXT:    v_mul_lo_u32 v9, v6, v5
+; GCN-NEXT:    v_mul_lo_u32 v10, v7, v4
+; GCN-NEXT:    v_add_i32_e32 v8, vcc, v8, v9
+; GCN-NEXT:    v_mul_lo_u32 v9, v6, v4
+; GCN-NEXT:    v_add_i32_e32 v8, vcc, v8, v10
+; GCN-NEXT:    v_mul_lo_u32 v11, v4, v8
+; GCN-NEXT:    v_mul_hi_u32 v10, v4, v8
+; GCN-NEXT:    v_mul_hi_u32 v12, v4, v9
+; GCN-NEXT:    v_mul_hi_u32 v15, v5, v8
+; GCN-NEXT:    v_mul_lo_u32 v8, v5, v8
+; GCN-NEXT:    v_add_i32_e32 v11, vcc, v12, v11
+; GCN-NEXT:    v_mul_lo_u32 v12, v5, v9
+; GCN-NEXT:    v_mul_hi_u32 v9, v5, v9
+; GCN-NEXT:    v_addc_u32_e32 v10, vcc, v14, v10, vcc
+; GCN-NEXT:    v_add_i32_e32 v11, vcc, v12, v11
+; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v10, v9, vcc
+; GCN-NEXT:    v_addc_u32_e32 v10, vcc, v15, v13, vcc
+; GCN-NEXT:    v_add_i32_e32 v8, vcc, v9, v8
+; GCN-NEXT:    v_add_i32_e64 v4, s[4:5], v4, v8
+; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v14, v10, vcc
+; GCN-NEXT:    v_addc_u32_e64 v8, vcc, v5, v9, s[4:5]
+; GCN-NEXT:    v_mul_lo_u32 v10, v6, v8
+; GCN-NEXT:    v_mul_hi_u32 v11, v6, v4
 ; GCN-NEXT:    v_mul_lo_u32 v7, v7, v4
-; GCN-NEXT:    v_mul_lo_u32 v13, v6, v4
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v11
-; GCN-NEXT:    v_mul_lo_u32 v6, v6, v10
-; GCN-NEXT:    v_mul_hi_u32 v11, v10, v13
-; GCN-NEXT:    v_mul_lo_u32 v14, v10, v13
-; GCN-NEXT:    v_mul_hi_u32 v13, v4, v13
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, v12, v6
+; GCN-NEXT:    v_mul_lo_u32 v6, v6, v4
+; GCN-NEXT:    v_add_i32_e32 v10, vcc, v11, v10
+; GCN-NEXT:    v_add_i32_e32 v7, vcc, v10, v7
+; GCN-NEXT:    v_mul_lo_u32 v12, v4, v7
+; GCN-NEXT:    v_mul_hi_u32 v15, v4, v6
+; GCN-NEXT:    v_mul_hi_u32 v16, v4, v7
+; GCN-NEXT:    v_mul_hi_u32 v11, v8, v6
+; GCN-NEXT:    v_mul_lo_u32 v6, v8, v6
+; GCN-NEXT:    v_add_i32_e32 v12, vcc, v15, v12
+; GCN-NEXT:    v_mul_hi_u32 v10, v8, v7
+; GCN-NEXT:    v_addc_u32_e32 v15, vcc, v14, v16, vcc
+; GCN-NEXT:    v_mul_lo_u32 v7, v8, v7
+; GCN-NEXT:    v_add_i32_e32 v6, vcc, v6, v12
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v15, v11, vcc
+; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v10, v13, vcc
 ; GCN-NEXT:    v_add_i32_e32 v6, vcc, v6, v7
-; GCN-NEXT:    v_mul_hi_u32 v7, v10, v6
-; GCN-NEXT:    v_mul_hi_u32 v12, v4, v6
-; GCN-NEXT:    v_mul_lo_u32 v15, v4, v6
-; GCN-NEXT:    v_mul_lo_u32 v6, v10, v6
-; GCN-NEXT:    v_add_i32_e32 v10, vcc, v13, v15
-; GCN-NEXT:    v_addc_u32_e32 v12, vcc, v9, v12, vcc
-; GCN-NEXT:    v_add_i32_e32 v10, vcc, v14, v10
-; GCN-NEXT:    v_addc_u32_e32 v10, vcc, v12, v11, vcc
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v7, v8, vcc
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, v10, v6
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v9, v7, vcc
+; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v14, v8, vcc
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v9
 ; GCN-NEXT:    v_addc_u32_e64 v5, vcc, v5, v7, s[4:5]
 ; GCN-NEXT:    v_add_i32_e32 v4, vcc, v4, v6
 ; GCN-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; GCN-NEXT:    v_mul_hi_u32 v6, v0, v4
-; GCN-NEXT:    v_mul_hi_u32 v7, v1, v4
-; GCN-NEXT:    v_mul_lo_u32 v4, v1, v4
-; GCN-NEXT:    v_mul_hi_u32 v10, v0, v5
-; GCN-NEXT:    v_mul_lo_u32 v11, v0, v5
-; GCN-NEXT:    v_mul_hi_u32 v12, v1, v5
+; GCN-NEXT:    v_mul_lo_u32 v6, v0, v5
+; GCN-NEXT:    v_mul_hi_u32 v7, v0, v4
+; GCN-NEXT:    v_mul_hi_u32 v8, v0, v5
+; GCN-NEXT:    v_mul_hi_u32 v9, v1, v5
 ; GCN-NEXT:    v_mul_lo_u32 v5, v1, v5
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, v6, v11
-; GCN-NEXT:    v_addc_u32_e32 v10, vcc, v9, v10, vcc
-; GCN-NEXT:    v_add_i32_e32 v4, vcc, v4, v6
-; GCN-NEXT:    v_addc_u32_e32 v4, vcc, v10, v7, vcc
-; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v12, v8, vcc
+; GCN-NEXT:    v_add_i32_e32 v6, vcc, v7, v6
+; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v14, v8, vcc
+; GCN-NEXT:    v_mul_lo_u32 v8, v1, v4
+; GCN-NEXT:    v_mul_hi_u32 v4, v1, v4
+; GCN-NEXT:    v_add_i32_e32 v6, vcc, v8, v6
+; GCN-NEXT:    v_addc_u32_e32 v4, vcc, v7, v4, vcc
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v9, v13, vcc
 ; GCN-NEXT:    v_add_i32_e32 v4, vcc, v4, v5
-; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v9, v6, vcc
-; GCN-NEXT:    v_mul_hi_u32 v6, v2, v4
-; GCN-NEXT:    v_mul_lo_u32 v7, v3, v4
-; GCN-NEXT:    v_mul_lo_u32 v8, v2, v4
-; GCN-NEXT:    v_mul_lo_u32 v9, v2, v5
-; GCN-NEXT:    v_add_i32_e32 v10, vcc, 2, v4
-; GCN-NEXT:    v_addc_u32_e32 v11, vcc, 0, v5, vcc
-; GCN-NEXT:    v_add_i32_e32 v12, vcc, 1, v4
-; GCN-NEXT:    v_addc_u32_e32 v13, vcc, 0, v5, vcc
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, v6, v9
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, v6, v7
-; GCN-NEXT:    v_sub_i32_e32 v7, vcc, v1, v6
-; GCN-NEXT:    v_sub_i32_e32 v0, vcc, v0, v8
-; GCN-NEXT:    v_subb_u32_e64 v7, s[4:5], v7, v3, vcc
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v14, v6, vcc
+; GCN-NEXT:    v_mul_lo_u32 v6, v2, v5
+; GCN-NEXT:    v_mul_hi_u32 v7, v2, v4
+; GCN-NEXT:    v_mul_lo_u32 v8, v3, v4
+; GCN-NEXT:    v_add_i32_e32 v6, vcc, v7, v6
+; GCN-NEXT:    v_mul_lo_u32 v7, v2, v4
+; GCN-NEXT:    v_add_i32_e32 v6, vcc, v6, v8
+; GCN-NEXT:    v_sub_i32_e32 v8, vcc, v1, v6
+; GCN-NEXT:    v_sub_i32_e32 v0, vcc, v0, v7
+; GCN-NEXT:    v_subb_u32_e64 v7, s[4:5], v8, v3, vcc
 ; GCN-NEXT:    v_sub_i32_e64 v8, s[4:5], v0, v2
+; GCN-NEXT:    v_subbrev_u32_e64 v7, s[4:5], 0, v7, s[4:5]
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[4:5], v7, v3
+; GCN-NEXT:    v_cndmask_b32_e64 v9, 0, -1, s[4:5]
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[4:5], v8, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v8, 0, -1, s[4:5]
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[4:5], v7, v3
 ; GCN-NEXT:    v_subb_u32_e32 v1, vcc, v1, v6, vcc
+; GCN-NEXT:    v_cndmask_b32_e64 v7, v9, v8, s[4:5]
+; GCN-NEXT:    v_add_i32_e64 v8, s[4:5], 2, v4
+; GCN-NEXT:    v_addc_u32_e64 v9, s[4:5], 0, v5, s[4:5]
+; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v1, v3
+; GCN-NEXT:    v_add_i32_e64 v10, s[4:5], 1, v4
+; GCN-NEXT:    v_cndmask_b32_e64 v6, 0, -1, vcc
 ; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v0, v2
+; GCN-NEXT:    v_addc_u32_e64 v11, s[4:5], 0, v5, s[4:5]
 ; GCN-NEXT:    v_cndmask_b32_e64 v0, 0, -1, vcc
-; GCN-NEXT:    v_subbrev_u32_e64 v6, vcc, 0, v7, s[4:5]
-; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v8, v2
-; GCN-NEXT:    v_cndmask_b32_e64 v2, 0, -1, vcc
-; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v1, v3
-; GCN-NEXT:    v_cndmask_b32_e64 v7, 0, -1, vcc
-; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v6, v3
-; GCN-NEXT:    v_cndmask_b32_e64 v8, 0, -1, vcc
 ; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, v1, v3
-; GCN-NEXT:    v_cndmask_b32_e32 v0, v7, v0, vcc
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, v6, v3
-; GCN-NEXT:    v_cndmask_b32_e32 v1, v8, v2, vcc
-; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v1
-; GCN-NEXT:    v_cndmask_b32_e32 v1, v12, v10, vcc
-; GCN-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, v0
-; GCN-NEXT:    v_cndmask_b32_e64 v0, v4, v1, s[4:5]
-; GCN-NEXT:    v_cndmask_b32_e32 v1, v13, v11, vcc
-; GCN-NEXT:    v_cndmask_b32_e64 v1, v5, v1, s[4:5]
+; GCN-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, v7
+; GCN-NEXT:    v_cndmask_b32_e32 v0, v6, v0, vcc
+; GCN-NEXT:    v_cndmask_b32_e64 v7, v10, v8, s[4:5]
+; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
+; GCN-NEXT:    v_cndmask_b32_e64 v1, v11, v9, s[4:5]
+; GCN-NEXT:    v_cndmask_b32_e32 v0, v4, v7, vcc
+; GCN-NEXT:    v_cndmask_b32_e32 v1, v5, v1, vcc
 ; GCN-NEXT:    s_setpc_b64 s[30:31]
   %result = udiv i64 %x, %y
   ret i64 %result
 }
 
 define amdgpu_kernel void @s_test_udiv24_64(i64 addrspace(1)* %out, i64 %x, i64 %y) {
 ; GCN-LABEL: s_test_udiv24_64:
 ; GCN:       ; %bb.0:
-; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GCN-NEXT:    s_load_dword s2, s[0:1], 0xe
+; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GCN-NEXT:    s_mov_b32 s11, 0xf000
 ; GCN-NEXT:    s_mov_b32 s10, -1
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN-NEXT:    s_lshr_b32 s3, s2, 8
+; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s3
+; GCN-NEXT:    s_lshr_b32 s2, s7, 8
 ; GCN-NEXT:    s_mov_b32 s8, s4
 ; GCN-NEXT:    s_mov_b32 s9, s5
-; GCN-NEXT:    s_lshr_b32 s3, s7, 8
-; GCN-NEXT:    s_lshr_b32 s2, s2, 8
-; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s2
 ; GCN-NEXT:    v_rcp_iflag_f32_e32 v0, v0
 ; GCN-NEXT:    v_mul_f32_e32 v0, 0x4f800000, v0
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v0
-; GCN-NEXT:    v_mul_hi_u32 v1, v0, s2
-; GCN-NEXT:    v_mul_lo_u32 v2, v0, s2
-; GCN-NEXT:    v_sub_i32_e32 v3, vcc, 0, v2
-; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v1, v2, v3, s[0:1]
+; GCN-NEXT:    v_mul_lo_u32 v1, v0, s3
+; GCN-NEXT:    v_mul_hi_u32 v2, v0, s3
+; GCN-NEXT:    v_sub_i32_e32 v3, vcc, 0, v1
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v1, v1, v3, s[0:1]
 ; GCN-NEXT:    v_mul_hi_u32 v1, v1, v0
 ; GCN-NEXT:    v_add_i32_e32 v2, vcc, v1, v0
 ; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, v1, v0
 ; GCN-NEXT:    v_cndmask_b32_e64 v0, v0, v2, s[0:1]
-; GCN-NEXT:    v_mul_hi_u32 v0, v0, s3
-; GCN-NEXT:    v_mul_lo_u32 v1, v0, s2
+; GCN-NEXT:    v_mul_hi_u32 v0, v0, s2
+; GCN-NEXT:    v_mul_lo_u32 v1, v0, s3
 ; GCN-NEXT:    v_add_i32_e32 v2, vcc, 1, v0
 ; GCN-NEXT:    v_add_i32_e32 v3, vcc, -1, v0
-; GCN-NEXT:    v_cmp_ge_u32_e64 s[0:1], s3, v1
-; GCN-NEXT:    v_sub_i32_e32 v1, vcc, s3, v1
-; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s2, v1
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[0:1], s2, v1
+; GCN-NEXT:    v_sub_i32_e32 v1, vcc, s2, v1
+; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s3, v1
 ; GCN-NEXT:    s_and_b64 vcc, vcc, s[0:1]
 ; GCN-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
 ; GCN-NEXT:    v_cndmask_b32_e64 v0, v3, v0, s[0:1]
 ; GCN-NEXT:    v_mov_b32_e32 v1, 0
 ; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[8:11], 0
 ; GCN-NEXT:    s_endpgm
   %1 = lshr i64 %x, 40
   %2 = lshr i64 %y, 40
   %result = udiv i64 %1, %2
   store i64 %result, i64 addrspace(1)* %out
   ret void
 }
 
 define i64 @v_test_udiv24_i64(i64 %x, i64 %y) {
 ; GCN-LABEL: v_test_udiv24_i64:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT:    v_lshrrev_b32_e32 v0, 8, v1
-; GCN-NEXT:    v_lshrrev_b32_e32 v1, 8, v3
-; GCN-NEXT:    v_cvt_f32_u32_e32 v2, v1
+; GCN-NEXT:    v_lshrrev_b32_e32 v0, 8, v3
+; GCN-NEXT:    v_cvt_f32_u32_e32 v2, v0
+; GCN-NEXT:    v_lshrrev_b32_e32 v1, 8, v1
 ; GCN-NEXT:    v_rcp_iflag_f32_e32 v2, v2
 ; GCN-NEXT:    v_mul_f32_e32 v2, 0x4f800000, v2
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v2, v2
-; GCN-NEXT:    v_mul_hi_u32 v3, v2, v1
-; GCN-NEXT:    v_mul_lo_u32 v4, v2, v1
-; GCN-NEXT:    v_sub_i32_e32 v5, vcc, 0, v4
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v3
-; GCN-NEXT:    v_cndmask_b32_e32 v3, v4, v5, vcc
+; GCN-NEXT:    v_mul_lo_u32 v3, v2, v0
+; GCN-NEXT:    v_mul_hi_u32 v4, v2, v0
+; GCN-NEXT:    v_sub_i32_e32 v5, vcc, 0, v3
+; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v4
+; GCN-NEXT:    v_cndmask_b32_e32 v3, v3, v5, vcc
 ; GCN-NEXT:    v_mul_hi_u32 v3, v3, v2
 ; GCN-NEXT:    v_add_i32_e64 v4, s[4:5], v2, v3
 ; GCN-NEXT:    v_sub_i32_e64 v2, s[4:5], v2, v3
 ; GCN-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
-; GCN-NEXT:    v_mul_hi_u32 v2, v2, v0
-; GCN-NEXT:    v_mul_lo_u32 v3, v2, v1
+; GCN-NEXT:    v_mul_hi_u32 v2, v2, v1
+; GCN-NEXT:    v_mul_lo_u32 v3, v2, v0
 ; GCN-NEXT:    v_add_i32_e32 v4, vcc, 1, v2
 ; GCN-NEXT:    v_add_i32_e32 v5, vcc, -1, v2
-; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v0, v3
-; GCN-NEXT:    v_sub_i32_e64 v0, s[4:5], v0, v3
-; GCN-NEXT:    v_cmp_ge_u32_e64 s[4:5], v0, v1
+; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v1, v3
+; GCN-NEXT:    v_sub_i32_e64 v1, s[4:5], v1, v3
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[4:5], v1, v0
 ; GCN-NEXT:    s_and_b64 s[4:5], s[4:5], vcc
 ; GCN-NEXT:    v_cndmask_b32_e64 v0, v2, v4, s[4:5]
 ; GCN-NEXT:    v_cndmask_b32_e32 v0, v5, v0, vcc
 ; GCN-NEXT:    v_mov_b32_e32 v1, 0
 ; GCN-NEXT:    s_setpc_b64 s[30:31]
   %1 = lshr i64 %x, 40
   %2 = lshr i64 %y, 40
   %result = udiv i64 %1, %2
   ret i64 %result
 }
 
 define amdgpu_kernel void @s_test_udiv32_i64(i64 addrspace(1)* %out, i64 %x, i64 %y) {
 ; GCN-LABEL: s_test_udiv32_i64:
 ; GCN:       ; %bb.0:
+; GCN-NEXT:    s_load_dword s12, s[0:1], 0xe
 ; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; GCN-NEXT:    s_load_dword s2, s[0:1], 0xe
 ; GCN-NEXT:    s_mov_b32 s11, 0xf000
 ; GCN-NEXT:    s_mov_b32 s10, -1
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s12
 ; GCN-NEXT:    s_mov_b32 s8, s4
 ; GCN-NEXT:    s_mov_b32 s9, s5
-; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s2
 ; GCN-NEXT:    v_rcp_iflag_f32_e32 v0, v0
 ; GCN-NEXT:    v_mul_f32_e32 v0, 0x4f800000, v0
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v0
-; GCN-NEXT:    v_mul_hi_u32 v1, v0, s2
-; GCN-NEXT:    v_mul_lo_u32 v2, v0, s2
-; GCN-NEXT:    v_sub_i32_e32 v3, vcc, 0, v2
-; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v1, v2, v3, s[0:1]
+; GCN-NEXT:    v_mul_lo_u32 v1, v0, s12
+; GCN-NEXT:    v_mul_hi_u32 v2, v0, s12
+; GCN-NEXT:    v_sub_i32_e32 v3, vcc, 0, v1
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[2:3], 0, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v1, v1, v3, s[2:3]
 ; GCN-NEXT:    v_mul_hi_u32 v1, v1, v0
 ; GCN-NEXT:    v_add_i32_e32 v2, vcc, v1, v0
 ; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, v1, v0
-; GCN-NEXT:    v_cndmask_b32_e64 v0, v0, v2, s[0:1]
+; GCN-NEXT:    v_cndmask_b32_e64 v0, v0, v2, s[2:3]
 ; GCN-NEXT:    v_mul_hi_u32 v0, v0, s7
-; GCN-NEXT:    v_mul_lo_u32 v1, v0, s2
+; GCN-NEXT:    v_mul_lo_u32 v1, v0, s12
 ; GCN-NEXT:    v_add_i32_e32 v2, vcc, 1, v0
 ; GCN-NEXT:    v_add_i32_e32 v3, vcc, -1, v0
 ; GCN-NEXT:    v_cmp_ge_u32_e64 s[0:1], s7, v1
 ; GCN-NEXT:    v_sub_i32_e32 v1, vcc, s7, v1
-; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s2, v1
+; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s12, v1
 ; GCN-NEXT:    s_and_b64 vcc, vcc, s[0:1]
 ; GCN-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
 ; GCN-NEXT:    v_cndmask_b32_e64 v0, v3, v0, s[0:1]
 ; GCN-NEXT:    v_mov_b32_e32 v1, 0
 ; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[8:11], 0
 ; GCN-NEXT:    s_endpgm
   %1 = lshr i64 %x, 32
   %2 = lshr i64 %y, 32
   %result = udiv i64 %1, %2
   store i64 %result, i64 addrspace(1)* %out
   ret void
 }
 
 define amdgpu_kernel void @s_test_udiv31_i64(i64 addrspace(1)* %out, i64 %x, i64 %y) {
 ; GCN-LABEL: s_test_udiv31_i64:
 ; GCN:       ; %bb.0:
-; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GCN-NEXT:    s_load_dword s2, s[0:1], 0xe
+; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GCN-NEXT:    s_mov_b32 s11, 0xf000
 ; GCN-NEXT:    s_mov_b32 s10, -1
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN-NEXT:    s_lshr_b32 s3, s2, 1
+; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s3
+; GCN-NEXT:    s_lshr_b32 s2, s7, 1
 ; GCN-NEXT:    s_mov_b32 s8, s4
 ; GCN-NEXT:    s_mov_b32 s9, s5
-; GCN-NEXT:    s_lshr_b32 s3, s7, 1
-; GCN-NEXT:    s_lshr_b32 s2, s2, 1
-; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s2
 ; GCN-NEXT:    v_rcp_iflag_f32_e32 v0, v0
 ; GCN-NEXT:    v_mul_f32_e32 v0, 0x4f800000, v0
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v0
-; GCN-NEXT:    v_mul_hi_u32 v1, v0, s2
-; GCN-NEXT:    v_mul_lo_u32 v2, v0, s2
-; GCN-NEXT:    v_sub_i32_e32 v3, vcc, 0, v2
-; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v1, v2, v3, s[0:1]
+; GCN-NEXT:    v_mul_lo_u32 v1, v0, s3
+; GCN-NEXT:    v_mul_hi_u32 v2, v0, s3
+; GCN-NEXT:    v_sub_i32_e32 v3, vcc, 0, v1
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v1, v1, v3, s[0:1]
 ; GCN-NEXT:    v_mul_hi_u32 v1, v1, v0
 ; GCN-NEXT:    v_add_i32_e32 v2, vcc, v1, v0
 ; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, v1, v0
 ; GCN-NEXT:    v_cndmask_b32_e64 v0, v0, v2, s[0:1]
-; GCN-NEXT:    v_mul_hi_u32 v0, v0, s3
-; GCN-NEXT:    v_mul_lo_u32 v1, v0, s2
+; GCN-NEXT:    v_mul_hi_u32 v0, v0, s2
+; GCN-NEXT:    v_mul_lo_u32 v1, v0, s3
 ; GCN-NEXT:    v_add_i32_e32 v2, vcc, 1, v0
 ; GCN-NEXT:    v_add_i32_e32 v3, vcc, -1, v0
-; GCN-NEXT:    v_cmp_ge_u32_e64 s[0:1], s3, v1
-; GCN-NEXT:    v_sub_i32_e32 v1, vcc, s3, v1
-; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s2, v1
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[0:1], s2, v1
+; GCN-NEXT:    v_sub_i32_e32 v1, vcc, s2, v1
+; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s3, v1
 ; GCN-NEXT:    s_and_b64 vcc, vcc, s[0:1]
 ; GCN-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
 ; GCN-NEXT:    v_cndmask_b32_e64 v0, v3, v0, s[0:1]
 ; GCN-NEXT:    v_mov_b32_e32 v1, 0
 ; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[8:11], 0
 ; GCN-NEXT:    s_endpgm
   %1 = lshr i64 %x, 33
   %2 = lshr i64 %y, 33
   %result = udiv i64 %1, %2
   store i64 %result, i64 addrspace(1)* %out
   ret void
 }
 
 define amdgpu_kernel void @s_test_udiv23_i64(i64 addrspace(1)* %out, i64 %x, i64 %y) {
 ; GCN-LABEL: s_test_udiv23_i64:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; GCN-NEXT:    s_load_dword s8, s[0:1], 0xe
+; GCN-NEXT:    s_load_dword s0, s[0:1], 0xe
 ; GCN-NEXT:    s_mov_b32 s3, 0xf000
 ; GCN-NEXT:    s_mov_b32 s2, -1
-; GCN-NEXT:    v_mov_b32_e32 v1, 0
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_mov_b32 s0, s4
 ; GCN-NEXT:    s_mov_b32 s1, s5
-; GCN-NEXT:    s_lshr_b32 s4, s7, 9
-; GCN-NEXT:    s_lshr_b32 s5, s8, 9
-; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s5
-; GCN-NEXT:    v_cvt_f32_u32_e32 v2, s4
-; GCN-NEXT:    v_rcp_iflag_f32_e32 v3, v0
-; GCN-NEXT:    v_mul_f32_e32 v3, v2, v3
-; GCN-NEXT:    v_trunc_f32_e32 v3, v3
-; GCN-NEXT:    v_mad_f32 v2, -v3, v0, v2
-; GCN-NEXT:    v_cvt_u32_f32_e32 v3, v3
-; GCN-NEXT:    v_cmp_ge_f32_e64 vcc, |v2|, |v0|
-; GCN-NEXT:    v_addc_u32_e32 v0, vcc, 0, v3, vcc
+; GCN-NEXT:    s_lshr_b32 s0, s0, 9
+; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s0
+; GCN-NEXT:    s_lshr_b32 s0, s7, 9
+; GCN-NEXT:    v_cvt_f32_u32_e32 v1, s0
+; GCN-NEXT:    s_mov_b32 s0, s4
+; GCN-NEXT:    v_rcp_iflag_f32_e32 v2, v0
+; GCN-NEXT:    v_mul_f32_e32 v2, v1, v2
+; GCN-NEXT:    v_trunc_f32_e32 v2, v2
+; GCN-NEXT:    v_mad_f32 v3, -v2, v0, v1
+; GCN-NEXT:    v_cvt_u32_f32_e32 v2, v2
+; GCN-NEXT:    v_cmp_ge_f32_e64 vcc, |v3|, |v0|
+; GCN-NEXT:    v_mov_b32_e32 v1, 0
+; GCN-NEXT:    v_addc_u32_e32 v0, vcc, 0, v2, vcc
 ; GCN-NEXT:    v_and_b32_e32 v0, 0x7fffff, v0
 ; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 ; GCN-NEXT:    s_endpgm
   %1 = lshr i64 %x, 41
   %2 = lshr i64 %y, 41
   %result = udiv i64 %1, %2
   store i64 %result, i64 addrspace(1)* %out
   ret void
 }
 
 define amdgpu_kernel void @s_test_udiv24_i48(i48 addrspace(1)* %out, i48 %x, i48 %y) {
 ; GCN-LABEL: s_test_udiv24_i48:
 ; GCN:       ; %bb.0:
+; GCN-NEXT:    s_load_dword s2, s[0:1], 0xd
+; GCN-NEXT:    s_load_dword s3, s[0:1], 0xe
+; GCN-NEXT:    s_mov_b32 s5, 0xff000000
+; GCN-NEXT:    s_mov_b32 s4, 0xffff
+; GCN-NEXT:    s_load_dword s6, s[0:1], 0xb
+; GCN-NEXT:    s_load_dword s7, s[0:1], 0xc
+; GCN-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN-NEXT:    s_and_b32 s2, s2, s5
+; GCN-NEXT:    s_and_b32 s3, s3, s4
+; GCN-NEXT:    v_mov_b32_e32 v0, s2
+; GCN-NEXT:    v_alignbit_b32 v0, s3, v0, 24
+; GCN-NEXT:    v_cvt_f32_u32_e32 v1, v0
+; GCN-NEXT:    v_cvt_f32_ubyte3_e32 v2, s3
+; GCN-NEXT:    s_and_b32 s7, s7, s4
+; GCN-NEXT:    s_and_b32 s6, s6, s5
+; GCN-NEXT:    v_mac_f32_e32 v1, 0x4f800000, v2
+; GCN-NEXT:    v_rcp_f32_e32 v1, v1
+; GCN-NEXT:    s_lshr_b64 s[2:3], s[2:3], 24
+; GCN-NEXT:    s_sub_u32 s8, 0, s2
+; GCN-NEXT:    s_subb_u32 s9, 0, s3
+; GCN-NEXT:    v_mul_f32_e32 v1, 0x5f7ffffc, v1
+; GCN-NEXT:    v_mul_f32_e32 v2, 0x2f800000, v1
+; GCN-NEXT:    v_trunc_f32_e32 v2, v2
+; GCN-NEXT:    v_mac_f32_e32 v1, 0xcf800000, v2
+; GCN-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GCN-NEXT:    v_cvt_u32_f32_e32 v2, v2
+; GCN-NEXT:    v_mov_b32_e32 v9, 0
+; GCN-NEXT:    v_mov_b32_e32 v8, 0
+; GCN-NEXT:    v_mul_hi_u32 v4, s8, v1
+; GCN-NEXT:    v_mul_lo_u32 v3, s8, v2
+; GCN-NEXT:    v_mul_lo_u32 v5, s9, v1
 ; GCN-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
-; GCN-NEXT:    s_load_dword s2, s[0:1], 0xb
-; GCN-NEXT:    s_load_dword s3, s[0:1], 0xc
-; GCN-NEXT:    s_load_dword s8, s[0:1], 0xd
-; GCN-NEXT:    s_load_dword s9, s[0:1], 0xe
+; GCN-NEXT:    v_add_i32_e32 v3, vcc, v4, v3
+; GCN-NEXT:    v_mul_lo_u32 v4, s8, v1
+; GCN-NEXT:    v_add_i32_e32 v3, vcc, v3, v5
+; GCN-NEXT:    v_mul_lo_u32 v6, v1, v3
+; GCN-NEXT:    v_mul_hi_u32 v5, v1, v3
+; GCN-NEXT:    v_mul_hi_u32 v7, v1, v4
+; GCN-NEXT:    v_mul_hi_u32 v10, v2, v3
+; GCN-NEXT:    v_mul_lo_u32 v3, v2, v3
+; GCN-NEXT:    v_add_i32_e32 v6, vcc, v7, v6
+; GCN-NEXT:    v_mul_lo_u32 v7, v2, v4
+; GCN-NEXT:    v_mul_hi_u32 v4, v2, v4
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v9, v5, vcc
+; GCN-NEXT:    v_add_i32_e32 v6, vcc, v7, v6
+; GCN-NEXT:    v_addc_u32_e32 v4, vcc, v5, v4, vcc
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v10, v8, vcc
+; GCN-NEXT:    v_add_i32_e32 v3, vcc, v4, v3
+; GCN-NEXT:    v_add_i32_e64 v1, s[2:3], v1, v3
+; GCN-NEXT:    v_addc_u32_e32 v4, vcc, v9, v5, vcc
+; GCN-NEXT:    v_addc_u32_e64 v3, vcc, v2, v4, s[2:3]
+; GCN-NEXT:    v_mul_lo_u32 v5, s8, v3
+; GCN-NEXT:    v_mul_hi_u32 v6, s8, v1
+; GCN-NEXT:    v_mul_lo_u32 v7, s9, v1
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v6, v5
+; GCN-NEXT:    v_mul_lo_u32 v6, s8, v1
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v7, v5
+; GCN-NEXT:    v_mul_lo_u32 v11, v1, v5
+; GCN-NEXT:    v_mul_hi_u32 v13, v1, v5
+; GCN-NEXT:    v_mul_hi_u32 v12, v1, v6
+; GCN-NEXT:    v_mul_hi_u32 v10, v3, v6
+; GCN-NEXT:    v_mul_lo_u32 v6, v3, v6
+; GCN-NEXT:    v_mul_hi_u32 v7, v3, v5
+; GCN-NEXT:    v_add_i32_e32 v11, vcc, v12, v11
+; GCN-NEXT:    v_addc_u32_e32 v12, vcc, v9, v13, vcc
+; GCN-NEXT:    v_mul_lo_u32 v3, v3, v5
+; GCN-NEXT:    v_add_i32_e32 v6, vcc, v6, v11
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v12, v10, vcc
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v7, v8, vcc
+; GCN-NEXT:    v_add_i32_e32 v3, vcc, v6, v3
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v9, v5, vcc
+; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v4
+; GCN-NEXT:    v_addc_u32_e64 v2, vcc, v2, v5, s[2:3]
+; GCN-NEXT:    v_add_i32_e32 v1, vcc, v1, v3
+; GCN-NEXT:    v_mov_b32_e32 v3, s6
+; GCN-NEXT:    v_alignbit_b32 v3, s7, v3, 24
+; GCN-NEXT:    v_addc_u32_e32 v2, vcc, 0, v2, vcc
+; GCN-NEXT:    v_mul_hi_u32 v5, v3, v1
+; GCN-NEXT:    v_mul_lo_u32 v4, v3, v2
+; GCN-NEXT:    v_mul_hi_u32 v6, v3, v2
+; GCN-NEXT:    v_mul_hi_u32 v1, 0, v1
+; GCN-NEXT:    v_mul_hi_u32 v2, 0, v2
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v5, v4
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v9, v6, vcc
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, 0, v4
+; GCN-NEXT:    v_addc_u32_e32 v1, vcc, v5, v1, vcc
+; GCN-NEXT:    v_addc_u32_e32 v2, vcc, v2, v8, vcc
+; GCN-NEXT:    v_add_i32_e32 v1, vcc, 0, v1
+; GCN-NEXT:    v_addc_u32_e32 v2, vcc, v9, v2, vcc
+; GCN-NEXT:    v_mul_lo_u32 v4, v0, v2
+; GCN-NEXT:    v_mul_hi_u32 v5, v0, v1
+; GCN-NEXT:    v_mul_lo_u32 v6, v0, v1
 ; GCN-NEXT:    s_mov_b32 s7, 0xf000
 ; GCN-NEXT:    s_mov_b32 s6, -1
-; GCN-NEXT:    s_mov_b32 s0, 0xffff
-; GCN-NEXT:    s_mov_b32 s10, 0xff000000
-; GCN-NEXT:    v_mov_b32_e32 v0, 0
-; GCN-NEXT:    v_mov_b32_e32 v1, 0
-; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_and_b32 s11, s3, s0
-; GCN-NEXT:    s_and_b32 s12, s2, s10
-; GCN-NEXT:    s_and_b32 s1, s9, s0
-; GCN-NEXT:    s_and_b32 s0, s8, s10
-; GCN-NEXT:    s_lshr_b64 s[2:3], s[0:1], 24
-; GCN-NEXT:    v_mov_b32_e32 v2, s0
-; GCN-NEXT:    v_alignbit_b32 v2, s1, v2, 24
-; GCN-NEXT:    v_cvt_f32_ubyte3_e32 v3, s1
-; GCN-NEXT:    v_mov_b32_e32 v4, s12
-; GCN-NEXT:    v_alignbit_b32 v4, s11, v4, 24
-; GCN-NEXT:    v_cvt_f32_u32_e32 v5, v2
-; GCN-NEXT:    s_sub_u32 s2, 0, s2
-; GCN-NEXT:    v_mac_f32_e32 v5, 0x4f800000, v3
-; GCN-NEXT:    s_subb_u32 s3, 0, s3
-; GCN-NEXT:    v_rcp_f32_e32 v3, v5
-; GCN-NEXT:    v_mul_f32_e32 v3, 0x5f7ffffc, v3
-; GCN-NEXT:    v_mul_f32_e32 v5, 0x2f800000, v3
-; GCN-NEXT:    v_trunc_f32_e32 v5, v5
-; GCN-NEXT:    v_mac_f32_e32 v3, 0xcf800000, v5
-; GCN-NEXT:    v_cvt_u32_f32_e32 v5, v5
-; GCN-NEXT:    v_cvt_u32_f32_e32 v3, v3
-; GCN-NEXT:    v_mul_lo_u32 v6, s2, v5
-; GCN-NEXT:    v_mul_lo_u32 v7, s3, v3
-; GCN-NEXT:    v_mul_hi_u32 v8, s2, v3
-; GCN-NEXT:    v_mul_lo_u32 v9, s2, v3
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, v8, v6
-; GCN-NEXT:    v_mul_hi_u32 v8, v3, v9
-; GCN-NEXT:    v_mul_hi_u32 v10, v5, v9
-; GCN-NEXT:    v_mul_lo_u32 v9, v5, v9
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, v6, v7
-; GCN-NEXT:    v_mul_hi_u32 v7, v3, v6
-; GCN-NEXT:    v_mul_lo_u32 v11, v3, v6
-; GCN-NEXT:    v_mul_hi_u32 v12, v5, v6
-; GCN-NEXT:    v_mul_lo_u32 v6, v5, v6
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v8, v11
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v1, v7, vcc
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v9, v8
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v7, v10, vcc
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v12, v0, vcc
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, v7, v6
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v1, v8, vcc
-; GCN-NEXT:    v_add_i32_e64 v3, s[0:1], v3, v6
-; GCN-NEXT:    v_addc_u32_e64 v6, vcc, v5, v7, s[0:1]
-; GCN-NEXT:    v_mul_hi_u32 v8, s2, v3
-; GCN-NEXT:    v_mul_lo_u32 v9, s3, v3
-; GCN-NEXT:    v_mul_lo_u32 v10, s2, v3
-; GCN-NEXT:    v_mul_lo_u32 v11, s2, v6
-; GCN-NEXT:    v_mul_hi_u32 v12, v6, v10
-; GCN-NEXT:    v_mul_lo_u32 v13, v6, v10
-; GCN-NEXT:    v_mul_hi_u32 v10, v3, v10
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v8, v11
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v9, v8
-; GCN-NEXT:    v_mul_hi_u32 v9, v6, v8
-; GCN-NEXT:    v_mul_hi_u32 v11, v3, v8
-; GCN-NEXT:    v_mul_lo_u32 v14, v3, v8
-; GCN-NEXT:    v_mul_lo_u32 v6, v6, v8
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v10, v14
-; GCN-NEXT:    v_addc_u32_e32 v10, vcc, v1, v11, vcc
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v13, v8
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v10, v12, vcc
-; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v9, v0, vcc
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v7
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, v8, v6
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v1, v9, vcc
-; GCN-NEXT:    v_addc_u32_e64 v5, vcc, v5, v7, s[0:1]
-; GCN-NEXT:    v_add_i32_e32 v3, vcc, v3, v6
-; GCN-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; GCN-NEXT:    v_mul_hi_u32 v6, v4, v3
-; GCN-NEXT:    v_mul_hi_u32 v3, 0, v3
-; GCN-NEXT:    v_mul_hi_u32 v7, v4, v5
-; GCN-NEXT:    v_mul_lo_u32 v8, v4, v5
-; GCN-NEXT:    v_mul_hi_u32 v5, 0, v5
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, v6, v8
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v1, v7, vcc
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, 0, v6
-; GCN-NEXT:    v_addc_u32_e32 v3, vcc, v7, v3, vcc
-; GCN-NEXT:    v_addc_u32_e32 v0, vcc, v5, v0, vcc
-; GCN-NEXT:    v_add_i32_e32 v3, vcc, 0, v3
-; GCN-NEXT:    v_addc_u32_e32 v0, vcc, v1, v0, vcc
-; GCN-NEXT:    v_mul_hi_u32 v1, v2, v3
-; GCN-NEXT:    v_mul_lo_u32 v5, v2, v3
-; GCN-NEXT:    v_mul_lo_u32 v6, v2, v0
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, 2, v3
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, 0, v0, vcc
-; GCN-NEXT:    v_add_i32_e32 v9, vcc, 1, v3
-; GCN-NEXT:    v_addc_u32_e32 v10, vcc, 0, v0, vcc
-; GCN-NEXT:    v_add_i32_e32 v1, vcc, v1, v6
-; GCN-NEXT:    v_sub_i32_e32 v4, vcc, v4, v5
-; GCN-NEXT:    v_subb_u32_e32 v1, vcc, 0, v1, vcc
-; GCN-NEXT:    v_sub_i32_e32 v5, vcc, v4, v2
-; GCN-NEXT:    v_cmp_ge_u32_e64 s[0:1], v4, v2
-; GCN-NEXT:    v_cndmask_b32_e64 v4, 0, -1, s[0:1]
-; GCN-NEXT:    v_subbrev_u32_e32 v6, vcc, 0, v1, vcc
-; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v5, v2
-; GCN-NEXT:    v_cndmask_b32_e64 v2, 0, -1, vcc
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v1
-; GCN-NEXT:    v_cndmask_b32_e32 v1, -1, v4, vcc
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v5, v4
+; GCN-NEXT:    v_sub_i32_e32 v3, vcc, v3, v6
+; GCN-NEXT:    v_subb_u32_e32 v4, vcc, 0, v4, vcc
+; GCN-NEXT:    v_sub_i32_e32 v5, vcc, v3, v0
+; GCN-NEXT:    v_subbrev_u32_e32 v6, vcc, 0, v4, vcc
+; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v5, v0
+; GCN-NEXT:    v_cndmask_b32_e64 v5, 0, -1, vcc
 ; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v6
-; GCN-NEXT:    v_cndmask_b32_e32 v2, -1, v2, vcc
-; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
-; GCN-NEXT:    v_cndmask_b32_e32 v2, v9, v7, vcc
-; GCN-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v1, v3, v2, s[0:1]
-; GCN-NEXT:    v_cndmask_b32_e32 v2, v10, v8, vcc
-; GCN-NEXT:    v_cndmask_b32_e64 v0, v0, v2, s[0:1]
-; GCN-NEXT:    buffer_store_short v0, off, s[4:7], 0 offset:4
-; GCN-NEXT:    buffer_store_dword v1, off, s[4:7], 0
+; GCN-NEXT:    v_cndmask_b32_e32 v5, -1, v5, vcc
+; GCN-NEXT:    v_add_i32_e32 v6, vcc, 2, v1
+; GCN-NEXT:    v_addc_u32_e32 v7, vcc, 0, v2, vcc
+; GCN-NEXT:    v_add_i32_e32 v8, vcc, 1, v1
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[0:1], v3, v0
+; GCN-NEXT:    v_addc_u32_e32 v9, vcc, 0, v2, vcc
+; GCN-NEXT:    v_cndmask_b32_e64 v0, 0, -1, s[0:1]
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v4
+; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v5
+; GCN-NEXT:    v_cndmask_b32_e64 v0, -1, v0, s[0:1]
+; GCN-NEXT:    v_cndmask_b32_e32 v5, v8, v6, vcc
+; GCN-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v0
+; GCN-NEXT:    v_cndmask_b32_e64 v0, v1, v5, s[0:1]
+; GCN-NEXT:    v_cndmask_b32_e32 v1, v9, v7, vcc
+; GCN-NEXT:    v_cndmask_b32_e64 v1, v2, v1, s[0:1]
+; GCN-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN-NEXT:    buffer_store_short v1, off, s[4:7], 0 offset:4
+; GCN-NEXT:    buffer_store_dword v0, off, s[4:7], 0
 ; GCN-NEXT:    s_endpgm
   %1 = lshr i48 %x, 24
   %2 = lshr i48 %y, 24
   %result = udiv i48 %1, %2
   store i48 %result, i48 addrspace(1)* %out
   ret void
 }
 
 define amdgpu_kernel void @s_test_udiv_k_num_i64(i64 addrspace(1)* %out, i64 %x) {
 ; GCN-LABEL: s_test_udiv_k_num_i64:
 ; GCN:       ; %bb.0:
-; GCN-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
-; GCN-NEXT:    s_mov_b32 s7, 0xf000
-; GCN-NEXT:    s_mov_b32 s6, -1
-; GCN-NEXT:    v_mov_b32_e32 v0, 0
-; GCN-NEXT:    v_mov_b32_e32 v1, 0
+; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN-NEXT:    v_mov_b32_e32 v2, 0
+; GCN-NEXT:    s_mov_b32 s11, 0xf000
+; GCN-NEXT:    s_mov_b32 s10, -1
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_mov_b32 s4, s0
-; GCN-NEXT:    s_mov_b32 s5, s1
-; GCN-NEXT:    v_cvt_f32_u32_e32 v2, s2
-; GCN-NEXT:    v_cvt_f32_u32_e32 v3, s3
-; GCN-NEXT:    s_sub_u32 s8, 0, s2
-; GCN-NEXT:    v_mov_b32_e32 v4, s3
-; GCN-NEXT:    v_mac_f32_e32 v2, 0x4f800000, v3
-; GCN-NEXT:    s_subb_u32 s9, 0, s3
-; GCN-NEXT:    v_rcp_f32_e32 v2, v2
-; GCN-NEXT:    v_mul_f32_e32 v2, 0x5f7ffffc, v2
-; GCN-NEXT:    v_mul_f32_e32 v3, 0x2f800000, v2
+; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s6
+; GCN-NEXT:    v_cvt_f32_u32_e32 v1, s7
+; GCN-NEXT:    s_sub_u32 s2, 0, s6
+; GCN-NEXT:    s_subb_u32 s3, 0, s7
+; GCN-NEXT:    s_mov_b32 s8, s4
+; GCN-NEXT:    v_mac_f32_e32 v0, 0x4f800000, v1
+; GCN-NEXT:    v_rcp_f32_e32 v0, v0
+; GCN-NEXT:    v_mov_b32_e32 v1, 0
+; GCN-NEXT:    s_mov_b32 s9, s5
+; GCN-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v0
+; GCN-NEXT:    v_mul_f32_e32 v3, 0x2f800000, v0
 ; GCN-NEXT:    v_trunc_f32_e32 v3, v3
-; GCN-NEXT:    v_mac_f32_e32 v2, 0xcf800000, v3
+; GCN-NEXT:    v_mac_f32_e32 v0, 0xcf800000, v3
+; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v0
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v3, v3
-; GCN-NEXT:    v_cvt_u32_f32_e32 v2, v2
-; GCN-NEXT:    v_mul_lo_u32 v5, s8, v3
-; GCN-NEXT:    v_mul_lo_u32 v6, s9, v2
-; GCN-NEXT:    v_mul_hi_u32 v7, s8, v2
-; GCN-NEXT:    v_mul_lo_u32 v8, s8, v2
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, v7, v5
-; GCN-NEXT:    v_mul_hi_u32 v7, v2, v8
-; GCN-NEXT:    v_mul_hi_u32 v9, v3, v8
-; GCN-NEXT:    v_mul_lo_u32 v8, v3, v8
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v6
-; GCN-NEXT:    v_mul_hi_u32 v6, v2, v5
-; GCN-NEXT:    v_mul_lo_u32 v10, v2, v5
-; GCN-NEXT:    v_mul_hi_u32 v11, v3, v5
-; GCN-NEXT:    v_mul_lo_u32 v5, v3, v5
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v7, v10
-; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v1, v6, vcc
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v8, v7
-; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v6, v9, vcc
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v11, v0, vcc
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, v6, v5
-; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v1, v7, vcc
-; GCN-NEXT:    v_add_i32_e64 v2, s[0:1], v2, v5
-; GCN-NEXT:    v_addc_u32_e64 v5, vcc, v3, v6, s[0:1]
-; GCN-NEXT:    v_mul_hi_u32 v7, s8, v2
-; GCN-NEXT:    v_mul_lo_u32 v8, s9, v2
-; GCN-NEXT:    v_mul_lo_u32 v9, s8, v2
-; GCN-NEXT:    v_mul_lo_u32 v10, s8, v5
-; GCN-NEXT:    v_mul_hi_u32 v11, v5, v9
-; GCN-NEXT:    v_mul_lo_u32 v12, v5, v9
-; GCN-NEXT:    v_mul_hi_u32 v9, v2, v9
+; GCN-NEXT:    v_mul_hi_u32 v5, s2, v0
+; GCN-NEXT:    v_mul_lo_u32 v4, s2, v3
+; GCN-NEXT:    v_mul_lo_u32 v7, s3, v0
+; GCN-NEXT:    v_mul_lo_u32 v6, s2, v0
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v5, v4
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v4, v7
+; GCN-NEXT:    v_mul_hi_u32 v5, v0, v6
+; GCN-NEXT:    v_mul_lo_u32 v7, v0, v4
+; GCN-NEXT:    v_mul_hi_u32 v9, v0, v4
+; GCN-NEXT:    v_mul_lo_u32 v8, v3, v6
+; GCN-NEXT:    v_mul_hi_u32 v6, v3, v6
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v7
+; GCN-NEXT:    v_mul_hi_u32 v10, v3, v4
+; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v2, v9, vcc
+; GCN-NEXT:    v_mul_lo_u32 v4, v3, v4
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v8, v5
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v7, v6, vcc
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v10, v1, vcc
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v5, v4
+; GCN-NEXT:    v_add_i32_e64 v0, s[0:1], v0, v4
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v2, v6, vcc
+; GCN-NEXT:    v_addc_u32_e64 v4, vcc, v3, v5, s[0:1]
+; GCN-NEXT:    v_mul_lo_u32 v6, s2, v4
+; GCN-NEXT:    v_mul_hi_u32 v7, s2, v0
+; GCN-NEXT:    v_mul_lo_u32 v8, s3, v0
+; GCN-NEXT:    v_add_i32_e32 v6, vcc, v7, v6
+; GCN-NEXT:    v_mul_lo_u32 v7, s2, v0
+; GCN-NEXT:    v_add_i32_e32 v6, vcc, v8, v6
+; GCN-NEXT:    v_mul_lo_u32 v10, v0, v6
+; GCN-NEXT:    v_mul_hi_u32 v12, v0, v6
+; GCN-NEXT:    v_mul_hi_u32 v11, v0, v7
+; GCN-NEXT:    v_mul_hi_u32 v9, v4, v7
+; GCN-NEXT:    v_mul_lo_u32 v7, v4, v7
+; GCN-NEXT:    v_mul_hi_u32 v8, v4, v6
+; GCN-NEXT:    v_add_i32_e32 v10, vcc, v11, v10
+; GCN-NEXT:    v_addc_u32_e32 v11, vcc, v2, v12, vcc
+; GCN-NEXT:    v_mul_lo_u32 v4, v4, v6
 ; GCN-NEXT:    v_add_i32_e32 v7, vcc, v7, v10
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v8, v7
-; GCN-NEXT:    v_mul_hi_u32 v8, v5, v7
-; GCN-NEXT:    v_mul_hi_u32 v10, v2, v7
-; GCN-NEXT:    v_mul_lo_u32 v13, v2, v7
-; GCN-NEXT:    v_mul_lo_u32 v5, v5, v7
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v9, v13
-; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v1, v10, vcc
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v12, v7
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v9, v11, vcc
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v8, v0, vcc
-; GCN-NEXT:    v_add_i32_e32 v3, vcc, v3, v6
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, v7, v5
-; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v1, v8, vcc
+; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v11, v9, vcc
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v8, v1, vcc
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v7, v4
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v2, v6, vcc
+; GCN-NEXT:    v_add_i32_e32 v3, vcc, v3, v5
 ; GCN-NEXT:    v_addc_u32_e64 v3, vcc, v3, v6, s[0:1]
-; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v5
+; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v4
 ; GCN-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; GCN-NEXT:    v_mul_hi_u32 v2, v2, 24
-; GCN-NEXT:    v_mul_hi_u32 v5, v3, 24
-; GCN-NEXT:    v_mul_lo_u32 v3, v3, 24
-; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v3
-; GCN-NEXT:    v_addc_u32_e32 v2, vcc, v1, v5, vcc
-; GCN-NEXT:    v_addc_u32_e32 v0, vcc, 0, v0, vcc
-; GCN-NEXT:    v_mul_hi_u32 v1, s2, v2
-; GCN-NEXT:    v_mul_lo_u32 v3, s3, v2
-; GCN-NEXT:    v_mul_lo_u32 v5, s2, v2
-; GCN-NEXT:    v_mul_lo_u32 v6, s2, v0
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, 2, v2
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, 0, v0, vcc
-; GCN-NEXT:    v_add_i32_e32 v9, vcc, 1, v2
-; GCN-NEXT:    v_addc_u32_e32 v10, vcc, 0, v0, vcc
-; GCN-NEXT:    v_add_i32_e32 v1, vcc, v1, v6
-; GCN-NEXT:    v_add_i32_e32 v1, vcc, v1, v3
-; GCN-NEXT:    v_sub_i32_e32 v3, vcc, 0, v1
-; GCN-NEXT:    v_sub_i32_e32 v5, vcc, 24, v5
-; GCN-NEXT:    v_subb_u32_e64 v3, s[0:1], v3, v4, vcc
-; GCN-NEXT:    v_subrev_i32_e64 v4, s[0:1], s2, v5
-; GCN-NEXT:    v_subb_u32_e32 v1, vcc, 0, v1, vcc
-; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s2, v5
-; GCN-NEXT:    v_cndmask_b32_e64 v5, 0, -1, vcc
-; GCN-NEXT:    v_subbrev_u32_e64 v3, vcc, 0, v3, s[0:1]
-; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s2, v4
-; GCN-NEXT:    v_cndmask_b32_e64 v4, 0, -1, vcc
-; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s3, v1
+; GCN-NEXT:    v_mul_lo_u32 v4, v3, 24
+; GCN-NEXT:    v_mul_hi_u32 v0, v0, 24
+; GCN-NEXT:    v_mul_hi_u32 v3, v3, 24
+; GCN-NEXT:    v_mov_b32_e32 v5, s7
+; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v4
+; GCN-NEXT:    v_addc_u32_e32 v0, vcc, v2, v3, vcc
+; GCN-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GCN-NEXT:    v_mul_lo_u32 v2, s6, v1
+; GCN-NEXT:    v_mul_hi_u32 v3, s6, v0
+; GCN-NEXT:    v_mul_lo_u32 v4, s7, v0
+; GCN-NEXT:    v_add_i32_e32 v2, vcc, v3, v2
+; GCN-NEXT:    v_mul_lo_u32 v3, s6, v0
+; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v4
+; GCN-NEXT:    v_sub_i32_e32 v4, vcc, 0, v2
+; GCN-NEXT:    v_sub_i32_e32 v3, vcc, 24, v3
+; GCN-NEXT:    v_subb_u32_e64 v4, s[0:1], v4, v5, vcc
+; GCN-NEXT:    v_subrev_i32_e64 v5, s[0:1], s6, v3
+; GCN-NEXT:    v_subbrev_u32_e64 v4, s[0:1], 0, v4, s[0:1]
+; GCN-NEXT:    v_cmp_le_u32_e64 s[0:1], s7, v4
+; GCN-NEXT:    v_cndmask_b32_e64 v6, 0, -1, s[0:1]
+; GCN-NEXT:    v_cmp_le_u32_e64 s[0:1], s6, v5
+; GCN-NEXT:    v_cndmask_b32_e64 v5, 0, -1, s[0:1]
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], s7, v4
+; GCN-NEXT:    v_cndmask_b32_e64 v4, v6, v5, s[0:1]
+; GCN-NEXT:    v_add_i32_e64 v5, s[0:1], 2, v0
+; GCN-NEXT:    v_addc_u32_e64 v6, s[0:1], 0, v1, s[0:1]
+; GCN-NEXT:    v_add_i32_e64 v7, s[0:1], 1, v0
+; GCN-NEXT:    v_subb_u32_e32 v2, vcc, 0, v2, vcc
+; GCN-NEXT:    v_addc_u32_e64 v8, s[0:1], 0, v1, s[0:1]
+; GCN-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v4
+; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s7, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v4, v8, v6, s[0:1]
 ; GCN-NEXT:    v_cndmask_b32_e64 v6, 0, -1, vcc
-; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s3, v3
-; GCN-NEXT:    v_cndmask_b32_e64 v11, 0, -1, vcc
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, s3, v1
-; GCN-NEXT:    v_cndmask_b32_e32 v1, v6, v5, vcc
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, s3, v3
-; GCN-NEXT:    v_cndmask_b32_e32 v3, v11, v4, vcc
-; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v3
-; GCN-NEXT:    v_cndmask_b32_e32 v3, v10, v8, vcc
-; GCN-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v1, v0, v3, s[0:1]
-; GCN-NEXT:    v_cndmask_b32_e32 v0, v9, v7, vcc
-; GCN-NEXT:    v_cndmask_b32_e64 v0, v2, v0, s[0:1]
-; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
+; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s6, v3
+; GCN-NEXT:    v_cndmask_b32_e64 v3, 0, -1, vcc
+; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, s7, v2
+; GCN-NEXT:    v_cndmask_b32_e32 v2, v6, v3, vcc
+; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v2, v7, v5, s[0:1]
+; GCN-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; GCN-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
+; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[8:11], 0
 ; GCN-NEXT:    s_endpgm
   %result = udiv i64 24, %x
   store i64 %result, i64 addrspace(1)* %out
   ret void
 }
 
 ; define i64 @v_test_udiv_k_num_i64(i64 %x) {
 ;   %result = udiv i64 24, %x
 ;   ret i64 %result
 ; }
 
 define i64 @v_test_udiv_pow2_k_num_i64(i64 %x) {
 ; GCN-LABEL: v_test_udiv_pow2_k_num_i64:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GCN-NEXT:    v_cvt_f32_u32_e32 v2, v0
 ; GCN-NEXT:    v_cvt_f32_u32_e32 v3, v1
 ; GCN-NEXT:    v_sub_i32_e32 v4, vcc, 0, v0
 ; GCN-NEXT:    v_subb_u32_e32 v5, vcc, 0, v1, vcc
-; GCN-NEXT:    v_mov_b32_e32 v6, 0
-; GCN-NEXT:    v_mov_b32_e32 v7, 0
-; GCN-NEXT:    s_mov_b32 s6, 0x8000
+; GCN-NEXT:    v_mov_b32_e32 v12, 0
 ; GCN-NEXT:    v_mac_f32_e32 v2, 0x4f800000, v3
 ; GCN-NEXT:    v_rcp_f32_e32 v2, v2
+; GCN-NEXT:    v_mov_b32_e32 v11, 0
 ; GCN-NEXT:    v_mul_f32_e32 v2, 0x5f7ffffc, v2
 ; GCN-NEXT:    v_mul_f32_e32 v3, 0x2f800000, v2
 ; GCN-NEXT:    v_trunc_f32_e32 v3, v3
 ; GCN-NEXT:    v_mac_f32_e32 v2, 0xcf800000, v3
-; GCN-NEXT:    v_cvt_u32_f32_e32 v3, v3
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v2, v2
-; GCN-NEXT:    v_mul_lo_u32 v8, v4, v3
-; GCN-NEXT:    v_mul_lo_u32 v9, v5, v2
-; GCN-NEXT:    v_mul_hi_u32 v10, v4, v2
-; GCN-NEXT:    v_mul_lo_u32 v11, v4, v2
+; GCN-NEXT:    v_cvt_u32_f32_e32 v3, v3
+; GCN-NEXT:    v_mul_hi_u32 v7, v4, v2
+; GCN-NEXT:    v_mul_lo_u32 v6, v4, v3
+; GCN-NEXT:    v_mul_lo_u32 v8, v5, v2
+; GCN-NEXT:    v_mul_lo_u32 v9, v4, v2
+; GCN-NEXT:    v_add_i32_e32 v6, vcc, v7, v6
+; GCN-NEXT:    v_add_i32_e32 v6, vcc, v6, v8
+; GCN-NEXT:    v_mul_lo_u32 v8, v2, v6
+; GCN-NEXT:    v_mul_hi_u32 v10, v2, v9
+; GCN-NEXT:    v_mul_hi_u32 v7, v2, v6
+; GCN-NEXT:    v_mul_hi_u32 v13, v3, v6
+; GCN-NEXT:    v_mul_lo_u32 v6, v3, v6
 ; GCN-NEXT:    v_add_i32_e32 v8, vcc, v10, v8
-; GCN-NEXT:    v_mul_hi_u32 v10, v2, v11
-; GCN-NEXT:    v_mul_hi_u32 v12, v3, v11
-; GCN-NEXT:    v_mul_lo_u32 v11, v3, v11
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v8, v9
-; GCN-NEXT:    v_mul_hi_u32 v9, v2, v8
-; GCN-NEXT:    v_mul_lo_u32 v13, v2, v8
-; GCN-NEXT:    v_mul_hi_u32 v14, v3, v8
-; GCN-NEXT:    v_mul_lo_u32 v8, v3, v8
-; GCN-NEXT:    v_add_i32_e32 v10, vcc, v10, v13
-; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v7, v9, vcc
-; GCN-NEXT:    v_add_i32_e32 v10, vcc, v11, v10
-; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v9, v12, vcc
-; GCN-NEXT:    v_addc_u32_e32 v10, vcc, v14, v6, vcc
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v9, v8
-; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v7, v10, vcc
-; GCN-NEXT:    v_add_i32_e64 v2, s[4:5], v2, v8
-; GCN-NEXT:    v_addc_u32_e64 v8, vcc, v3, v9, s[4:5]
-; GCN-NEXT:    v_mul_hi_u32 v10, v4, v2
+; GCN-NEXT:    v_mul_lo_u32 v10, v3, v9
+; GCN-NEXT:    v_mul_hi_u32 v9, v3, v9
+; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v12, v7, vcc
+; GCN-NEXT:    v_add_i32_e32 v8, vcc, v10, v8
+; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v7, v9, vcc
+; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v13, v11, vcc
+; GCN-NEXT:    v_add_i32_e32 v6, vcc, v7, v6
+; GCN-NEXT:    v_add_i32_e64 v2, s[4:5], v2, v6
+; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v12, v8, vcc
+; GCN-NEXT:    v_addc_u32_e64 v6, vcc, v3, v7, s[4:5]
+; GCN-NEXT:    v_mul_lo_u32 v8, v4, v6
+; GCN-NEXT:    v_mul_hi_u32 v9, v4, v2
 ; GCN-NEXT:    v_mul_lo_u32 v5, v5, v2
-; GCN-NEXT:    v_mul_lo_u32 v11, v4, v2
-; GCN-NEXT:    v_add_i32_e32 v3, vcc, v3, v9
-; GCN-NEXT:    v_mul_lo_u32 v4, v4, v8
-; GCN-NEXT:    v_mul_hi_u32 v9, v8, v11
-; GCN-NEXT:    v_mul_lo_u32 v12, v8, v11
-; GCN-NEXT:    v_mul_hi_u32 v11, v2, v11
-; GCN-NEXT:    v_add_i32_e32 v4, vcc, v10, v4
+; GCN-NEXT:    v_mul_lo_u32 v4, v4, v2
+; GCN-NEXT:    v_add_i32_e32 v8, vcc, v9, v8
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v8, v5
+; GCN-NEXT:    v_mul_lo_u32 v10, v2, v5
+; GCN-NEXT:    v_mul_hi_u32 v13, v2, v4
+; GCN-NEXT:    v_mul_hi_u32 v14, v2, v5
+; GCN-NEXT:    v_mul_hi_u32 v9, v6, v4
+; GCN-NEXT:    v_mul_lo_u32 v4, v6, v4
+; GCN-NEXT:    v_add_i32_e32 v10, vcc, v13, v10
+; GCN-NEXT:    v_mul_hi_u32 v8, v6, v5
+; GCN-NEXT:    v_addc_u32_e32 v13, vcc, v12, v14, vcc
+; GCN-NEXT:    v_mul_lo_u32 v5, v6, v5
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v4, v10
+; GCN-NEXT:    v_addc_u32_e32 v4, vcc, v13, v9, vcc
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v8, v11, vcc
 ; GCN-NEXT:    v_add_i32_e32 v4, vcc, v4, v5
-; GCN-NEXT:    v_mul_hi_u32 v5, v8, v4
-; GCN-NEXT:    v_mul_hi_u32 v10, v2, v4
-; GCN-NEXT:    v_mul_lo_u32 v13, v2, v4
-; GCN-NEXT:    v_mul_lo_u32 v4, v8, v4
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v11, v13
-; GCN-NEXT:    v_addc_u32_e32 v10, vcc, v7, v10, vcc
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v12, v8
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v10, v9, vcc
-; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v5, v6, vcc
-; GCN-NEXT:    v_add_i32_e32 v4, vcc, v8, v4
-; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v7, v5, vcc
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v12, v6, vcc
+; GCN-NEXT:    v_add_i32_e32 v3, vcc, v3, v7
 ; GCN-NEXT:    v_addc_u32_e64 v3, vcc, v3, v5, s[4:5]
 ; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v4
 ; GCN-NEXT:    v_addc_u32_e32 v2, vcc, 0, v3, vcc
 ; GCN-NEXT:    v_lshrrev_b32_e32 v2, 17, v2
 ; GCN-NEXT:    v_mul_lo_u32 v3, v1, v2
 ; GCN-NEXT:    v_mul_hi_u32 v4, v0, v2
-; GCN-NEXT:    v_mul_lo_u32 v5, v0, v2
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, 2, v2
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, 0, v7, vcc
-; GCN-NEXT:    v_add_i32_e32 v9, vcc, 1, v2
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
+; GCN-NEXT:    s_mov_b32 s4, 0x8000
 ; GCN-NEXT:    v_add_i32_e32 v3, vcc, v4, v3
-; GCN-NEXT:    v_sub_i32_e32 v4, vcc, s6, v5
-; GCN-NEXT:    v_sub_i32_e64 v5, s[4:5], 0, v3
-; GCN-NEXT:    v_sub_i32_e64 v10, s[4:5], v4, v0
-; GCN-NEXT:    v_cmp_ge_u32_e64 s[6:7], v4, v0
-; GCN-NEXT:    v_cndmask_b32_e64 v4, 0, -1, s[6:7]
-; GCN-NEXT:    v_subb_u32_e64 v5, s[6:7], v5, v1, vcc
-; GCN-NEXT:    v_cmp_ge_u32_e64 s[6:7], v10, v0
-; GCN-NEXT:    v_cndmask_b32_e64 v0, 0, -1, s[6:7]
+; GCN-NEXT:    v_mul_lo_u32 v4, v0, v2
+; GCN-NEXT:    v_sub_i32_e32 v5, vcc, 0, v3
+; GCN-NEXT:    v_sub_i32_e32 v4, vcc, s4, v4
+; GCN-NEXT:    v_subb_u32_e64 v5, s[4:5], v5, v1, vcc
+; GCN-NEXT:    v_sub_i32_e64 v6, s[4:5], v4, v0
+; GCN-NEXT:    v_subbrev_u32_e64 v5, s[4:5], 0, v5, s[4:5]
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[4:5], v5, v1
+; GCN-NEXT:    v_cndmask_b32_e64 v7, 0, -1, s[4:5]
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[4:5], v6, v0
+; GCN-NEXT:    v_cndmask_b32_e64 v6, 0, -1, s[4:5]
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[4:5], v5, v1
+; GCN-NEXT:    v_cndmask_b32_e64 v5, v7, v6, s[4:5]
+; GCN-NEXT:    v_add_i32_e64 v6, s[4:5], 2, v2
+; GCN-NEXT:    v_addc_u32_e64 v7, s[4:5], 0, v12, s[4:5]
+; GCN-NEXT:    v_add_i32_e64 v8, s[4:5], 1, v2
 ; GCN-NEXT:    v_subb_u32_e32 v3, vcc, 0, v3, vcc
-; GCN-NEXT:    v_subbrev_u32_e64 v5, vcc, 0, v5, s[4:5]
+; GCN-NEXT:    v_addc_u32_e64 v9, s[4:5], 0, v12, s[4:5]
+; GCN-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, v5
 ; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v3, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v10, 0, -1, vcc
-; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v5, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v11, 0, -1, vcc
+; GCN-NEXT:    v_cndmask_b32_e64 v5, v8, v6, s[4:5]
+; GCN-NEXT:    v_cndmask_b32_e64 v6, 0, -1, vcc
+; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v4, v0
+; GCN-NEXT:    v_cndmask_b32_e64 v0, 0, -1, vcc
 ; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v1
-; GCN-NEXT:    v_cndmask_b32_e32 v3, v10, v4, vcc
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, v5, v1
-; GCN-NEXT:    v_cndmask_b32_e32 v0, v11, v0, vcc
+; GCN-NEXT:    v_cndmask_b32_e32 v0, v6, v0, vcc
 ; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
-; GCN-NEXT:    v_cndmask_b32_e32 v0, v9, v6, vcc
-; GCN-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, v3
-; GCN-NEXT:    v_cndmask_b32_e64 v0, v2, v0, s[4:5]
-; GCN-NEXT:    v_cndmask_b32_e32 v1, v7, v8, vcc
-; GCN-NEXT:    v_cndmask_b32_e64 v1, 0, v1, s[4:5]
+; GCN-NEXT:    v_cndmask_b32_e64 v1, v9, v7, s[4:5]
+; GCN-NEXT:    v_cndmask_b32_e32 v0, v2, v5, vcc
+; GCN-NEXT:    v_cndmask_b32_e32 v1, 0, v1, vcc
 ; GCN-NEXT:    s_setpc_b64 s[30:31]
   %result = udiv i64 32768, %x
   ret i64 %result
 }
 
 define i64 @v_test_udiv_pow2_k_den_i64(i64 %x) {
 ; GCN-LABEL: v_test_udiv_pow2_k_den_i64:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GCN-NEXT:    v_alignbit_b32 v0, v1, v0, 15
 ; GCN-NEXT:    v_lshrrev_b32_e32 v1, 15, v1
 ; GCN-NEXT:    s_setpc_b64 s[30:31]
   %result = udiv i64 %x, 32768
   ret i64 %result
 }
 
 define amdgpu_kernel void @s_test_udiv_k_den_i64(i64 addrspace(1)* %out, i64 %x) {
 ; GCN-LABEL: s_test_udiv_k_den_i64:
 ; GCN:       ; %bb.0:
-; GCN-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
-; GCN-NEXT:    s_mov_b32 s7, 0xf000
-; GCN-NEXT:    s_mov_b32 s6, -1
 ; GCN-NEXT:    v_mov_b32_e32 v0, 0x4f800000
-; GCN-NEXT:    s_movk_i32 s8, 0xffe8
-; GCN-NEXT:    v_mov_b32_e32 v1, 0
-; GCN-NEXT:    v_mov_b32_e32 v2, 0
 ; GCN-NEXT:    v_madak_f32 v0, 0, v0, 0x41c00000
 ; GCN-NEXT:    v_rcp_f32_e32 v0, v0
+; GCN-NEXT:    s_movk_i32 s2, 0xffe8
+; GCN-NEXT:    v_mov_b32_e32 v8, 0
+; GCN-NEXT:    v_mov_b32_e32 v7, 0
 ; GCN-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v0
-; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_mov_b32 s4, s0
-; GCN-NEXT:    s_mov_b32 s5, s1
-; GCN-NEXT:    v_mul_f32_e32 v3, 0x2f800000, v0
-; GCN-NEXT:    v_mov_b32_e32 v4, s3
-; GCN-NEXT:    v_trunc_f32_e32 v3, v3
-; GCN-NEXT:    v_mac_f32_e32 v0, 0xcf800000, v3
-; GCN-NEXT:    v_cvt_u32_f32_e32 v3, v3
+; GCN-NEXT:    v_mul_f32_e32 v1, 0x2f800000, v0
+; GCN-NEXT:    v_trunc_f32_e32 v1, v1
+; GCN-NEXT:    v_mac_f32_e32 v0, 0xcf800000, v1
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v0
-; GCN-NEXT:    v_mul_lo_u32 v5, v3, s8
-; GCN-NEXT:    v_mul_hi_u32 v6, v0, s8
-; GCN-NEXT:    v_mul_lo_u32 v7, v0, s8
-; GCN-NEXT:    v_subrev_i32_e32 v6, vcc, v0, v6
-; GCN-NEXT:    v_mul_hi_u32 v8, v0, v7
-; GCN-NEXT:    v_mul_hi_u32 v9, v3, v7
-; GCN-NEXT:    v_mul_lo_u32 v7, v3, v7
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v6
-; GCN-NEXT:    v_mul_hi_u32 v6, v0, v5
-; GCN-NEXT:    v_mul_lo_u32 v10, v0, v5
-; GCN-NEXT:    v_mul_hi_u32 v11, v3, v5
-; GCN-NEXT:    v_mul_lo_u32 v5, v3, v5
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v8, v10
-; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v2, v6, vcc
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v7, v8
-; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v6, v9, vcc
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v11, v1, vcc
+; GCN-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GCN-NEXT:    s_load_dwordx4 s[8:11], s[0:1], 0x9
+; GCN-NEXT:    s_mov_b32 s7, 0xf000
+; GCN-NEXT:    v_mul_hi_u32 v2, v0, s2
+; GCN-NEXT:    v_mul_lo_u32 v3, v1, s2
+; GCN-NEXT:    v_mul_lo_u32 v4, v0, s2
+; GCN-NEXT:    s_mov_b32 s6, -1
+; GCN-NEXT:    v_subrev_i32_e32 v2, vcc, v0, v2
+; GCN-NEXT:    v_add_i32_e32 v2, vcc, v3, v2
+; GCN-NEXT:    v_mul_lo_u32 v5, v0, v2
+; GCN-NEXT:    v_mul_hi_u32 v6, v0, v4
+; GCN-NEXT:    v_mul_hi_u32 v3, v0, v2
+; GCN-NEXT:    v_mul_hi_u32 v9, v1, v2
+; GCN-NEXT:    v_mul_lo_u32 v2, v1, v2
 ; GCN-NEXT:    v_add_i32_e32 v5, vcc, v6, v5
-; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v2, v7, vcc
-; GCN-NEXT:    v_add_i32_e64 v0, s[0:1], v0, v5
-; GCN-NEXT:    v_addc_u32_e64 v5, vcc, v3, v6, s[0:1]
-; GCN-NEXT:    v_mul_hi_u32 v7, v0, s8
-; GCN-NEXT:    v_mul_lo_u32 v8, v0, s8
-; GCN-NEXT:    v_mul_lo_u32 v9, v5, s8
-; GCN-NEXT:    v_subrev_i32_e32 v7, vcc, v0, v7
-; GCN-NEXT:    v_mul_hi_u32 v10, v0, v8
-; GCN-NEXT:    v_mul_hi_u32 v11, v5, v8
-; GCN-NEXT:    v_mul_lo_u32 v8, v5, v8
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v7, v9
-; GCN-NEXT:    v_mul_hi_u32 v9, v0, v7
-; GCN-NEXT:    v_mul_lo_u32 v12, v0, v7
-; GCN-NEXT:    v_mul_hi_u32 v13, v5, v7
-; GCN-NEXT:    v_mul_lo_u32 v5, v5, v7
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v10, v12
-; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v2, v9, vcc
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v8, v7
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v9, v11, vcc
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v13, v1, vcc
-; GCN-NEXT:    v_add_i32_e32 v3, vcc, v3, v6
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, v7, v5
-; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v2, v8, vcc
-; GCN-NEXT:    v_addc_u32_e64 v3, vcc, v3, v6, s[0:1]
-; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v5
-; GCN-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; GCN-NEXT:    v_mul_hi_u32 v5, s2, v0
-; GCN-NEXT:    v_mul_hi_u32 v6, s3, v0
-; GCN-NEXT:    v_mul_lo_u32 v0, s3, v0
-; GCN-NEXT:    v_mul_hi_u32 v7, s2, v3
-; GCN-NEXT:    v_mul_lo_u32 v8, s2, v3
-; GCN-NEXT:    v_mul_hi_u32 v9, s3, v3
-; GCN-NEXT:    v_mul_lo_u32 v3, s3, v3
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v8
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v2, v7, vcc
-; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v5
-; GCN-NEXT:    v_addc_u32_e32 v0, vcc, v7, v6, vcc
-; GCN-NEXT:    v_addc_u32_e32 v1, vcc, v9, v1, vcc
-; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v3
-; GCN-NEXT:    v_addc_u32_e32 v1, vcc, v2, v1, vcc
-; GCN-NEXT:    v_mul_hi_u32 v2, v0, 24
-; GCN-NEXT:    v_mul_lo_u32 v3, v0, 24
-; GCN-NEXT:    v_mul_lo_u32 v5, v1, 24
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, 2, v0
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, 0, v1, vcc
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, 1, v0
-; GCN-NEXT:    v_addc_u32_e32 v9, vcc, 0, v1, vcc
-; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v5
-; GCN-NEXT:    v_sub_i32_e32 v3, vcc, s2, v3
-; GCN-NEXT:    v_subb_u32_e32 v2, vcc, v4, v2, vcc
-; GCN-NEXT:    v_subrev_i32_e32 v4, vcc, 24, v3
-; GCN-NEXT:    v_cmp_lt_u32_e64 s[0:1], 23, v3
-; GCN-NEXT:    v_cndmask_b32_e64 v3, 0, -1, s[0:1]
+; GCN-NEXT:    v_mul_lo_u32 v6, v1, v4
+; GCN-NEXT:    v_mul_hi_u32 v4, v1, v4
+; GCN-NEXT:    v_addc_u32_e32 v3, vcc, v8, v3, vcc
+; GCN-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN-NEXT:    s_mov_b32 s4, s8
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v6, v5
+; GCN-NEXT:    v_addc_u32_e32 v3, vcc, v3, v4, vcc
+; GCN-NEXT:    v_addc_u32_e32 v4, vcc, v9, v7, vcc
+; GCN-NEXT:    v_add_i32_e32 v2, vcc, v3, v2
+; GCN-NEXT:    v_add_i32_e64 v0, s[0:1], v0, v2
+; GCN-NEXT:    v_addc_u32_e32 v3, vcc, v8, v4, vcc
+; GCN-NEXT:    v_mul_hi_u32 v4, v0, s2
+; GCN-NEXT:    v_addc_u32_e64 v2, vcc, v1, v3, s[0:1]
+; GCN-NEXT:    v_mul_lo_u32 v5, v2, s2
+; GCN-NEXT:    v_mul_lo_u32 v6, v0, s2
+; GCN-NEXT:    v_subrev_i32_e32 v4, vcc, v0, v4
+; GCN-NEXT:    s_mov_b32 s5, s9
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v4, v5
+; GCN-NEXT:    v_mul_lo_u32 v5, v0, v4
+; GCN-NEXT:    v_mul_hi_u32 v9, v0, v6
+; GCN-NEXT:    v_mul_hi_u32 v10, v0, v4
+; GCN-NEXT:    v_mul_hi_u32 v11, v2, v4
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v9, v5
+; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v8, v10, vcc
+; GCN-NEXT:    v_mul_lo_u32 v10, v2, v6
+; GCN-NEXT:    v_mul_hi_u32 v6, v2, v6
+; GCN-NEXT:    v_mul_lo_u32 v2, v2, v4
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v10, v5
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v9, v6, vcc
+; GCN-NEXT:    v_addc_u32_e32 v4, vcc, v11, v7, vcc
+; GCN-NEXT:    v_add_i32_e32 v2, vcc, v5, v2
+; GCN-NEXT:    v_addc_u32_e32 v4, vcc, v8, v4, vcc
+; GCN-NEXT:    v_add_i32_e32 v1, vcc, v1, v3
+; GCN-NEXT:    v_addc_u32_e64 v1, vcc, v1, v4, s[0:1]
+; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v2
+; GCN-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GCN-NEXT:    v_mul_lo_u32 v2, s10, v1
+; GCN-NEXT:    v_mul_hi_u32 v3, s10, v0
+; GCN-NEXT:    v_mul_hi_u32 v4, s10, v1
+; GCN-NEXT:    v_mul_hi_u32 v5, s11, v1
+; GCN-NEXT:    v_mul_lo_u32 v1, s11, v1
+; GCN-NEXT:    v_add_i32_e32 v2, vcc, v3, v2
+; GCN-NEXT:    v_addc_u32_e32 v3, vcc, v8, v4, vcc
+; GCN-NEXT:    v_mul_lo_u32 v4, s11, v0
+; GCN-NEXT:    v_mul_hi_u32 v0, s11, v0
+; GCN-NEXT:    v_add_i32_e32 v2, vcc, v4, v2
+; GCN-NEXT:    v_addc_u32_e32 v0, vcc, v3, v0, vcc
+; GCN-NEXT:    v_addc_u32_e32 v2, vcc, v5, v7, vcc
+; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v1
+; GCN-NEXT:    v_addc_u32_e32 v1, vcc, v8, v2, vcc
+; GCN-NEXT:    v_mul_lo_u32 v2, v1, 24
+; GCN-NEXT:    v_mul_hi_u32 v3, v0, 24
+; GCN-NEXT:    v_mul_lo_u32 v4, v0, 24
+; GCN-NEXT:    v_add_i32_e32 v2, vcc, v3, v2
+; GCN-NEXT:    v_sub_i32_e32 v4, vcc, s10, v4
+; GCN-NEXT:    v_mov_b32_e32 v3, s11
+; GCN-NEXT:    v_subb_u32_e32 v2, vcc, v3, v2, vcc
+; GCN-NEXT:    v_subrev_i32_e32 v3, vcc, 24, v4
 ; GCN-NEXT:    v_subbrev_u32_e32 v5, vcc, 0, v2, vcc
-; GCN-NEXT:    v_cmp_lt_u32_e32 vcc, 23, v4
-; GCN-NEXT:    v_cndmask_b32_e64 v4, 0, -1, vcc
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v2
-; GCN-NEXT:    v_cndmask_b32_e32 v2, -1, v3, vcc
+; GCN-NEXT:    v_cmp_lt_u32_e32 vcc, 23, v3
+; GCN-NEXT:    v_cndmask_b32_e64 v3, 0, -1, vcc
 ; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v5
-; GCN-NEXT:    v_cndmask_b32_e32 v3, -1, v4, vcc
+; GCN-NEXT:    v_cndmask_b32_e32 v3, -1, v3, vcc
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, 2, v0
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, 0, v1, vcc
+; GCN-NEXT:    v_add_i32_e32 v7, vcc, 1, v0
+; GCN-NEXT:    v_cmp_lt_u32_e64 s[0:1], 23, v4
+; GCN-NEXT:    v_addc_u32_e32 v8, vcc, 0, v1, vcc
+; GCN-NEXT:    v_cndmask_b32_e64 v4, 0, -1, s[0:1]
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v2
 ; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v3
-; GCN-NEXT:    v_cndmask_b32_e32 v3, v9, v7, vcc
+; GCN-NEXT:    v_cndmask_b32_e64 v2, -1, v4, s[0:1]
 ; GCN-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v2
+; GCN-NEXT:    v_cndmask_b32_e32 v3, v8, v6, vcc
+; GCN-NEXT:    v_cndmask_b32_e32 v2, v7, v5, vcc
 ; GCN-NEXT:    v_cndmask_b32_e64 v1, v1, v3, s[0:1]
-; GCN-NEXT:    v_cndmask_b32_e32 v2, v8, v6, vcc
 ; GCN-NEXT:    v_cndmask_b32_e64 v0, v0, v2, s[0:1]
 ; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
 ; GCN-NEXT:    s_endpgm
   %result = udiv i64 %x, 24
   store i64 %result, i64 addrspace(1)* %out
   ret void
 }
 
 define i64 @v_test_udiv_k_den_i64(i64 %x) {
 ; GCN-LABEL: v_test_udiv_k_den_i64:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GCN-NEXT:    v_mov_b32_e32 v2, 0x4f800000
-; GCN-NEXT:    s_movk_i32 s6, 0xffe8
-; GCN-NEXT:    v_mov_b32_e32 v3, 0
-; GCN-NEXT:    v_mov_b32_e32 v4, 0
 ; GCN-NEXT:    v_madak_f32 v2, 0, v2, 0x41c00000
 ; GCN-NEXT:    v_rcp_f32_e32 v2, v2
+; GCN-NEXT:    s_movk_i32 s6, 0xffe8
+; GCN-NEXT:    v_mov_b32_e32 v10, 0
+; GCN-NEXT:    v_mov_b32_e32 v9, 0
 ; GCN-NEXT:    v_mul_f32_e32 v2, 0x5f7ffffc, v2
-; GCN-NEXT:    v_mul_f32_e32 v5, 0x2f800000, v2
-; GCN-NEXT:    v_trunc_f32_e32 v5, v5
-; GCN-NEXT:    v_mac_f32_e32 v2, 0xcf800000, v5
-; GCN-NEXT:    v_cvt_u32_f32_e32 v5, v5
+; GCN-NEXT:    v_mul_f32_e32 v3, 0x2f800000, v2
+; GCN-NEXT:    v_trunc_f32_e32 v3, v3
+; GCN-NEXT:    v_mac_f32_e32 v2, 0xcf800000, v3
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v2, v2
-; GCN-NEXT:    v_mul_lo_u32 v6, v5, s6
-; GCN-NEXT:    v_mul_hi_u32 v7, v2, s6
+; GCN-NEXT:    v_cvt_u32_f32_e32 v3, v3
+; GCN-NEXT:    v_mul_hi_u32 v4, v2, s6
+; GCN-NEXT:    v_mul_lo_u32 v5, v3, s6
+; GCN-NEXT:    v_mul_lo_u32 v6, v2, s6
+; GCN-NEXT:    v_subrev_i32_e32 v4, vcc, v2, v4
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v5, v4
+; GCN-NEXT:    v_mul_lo_u32 v7, v2, v4
+; GCN-NEXT:    v_mul_hi_u32 v8, v2, v6
+; GCN-NEXT:    v_mul_hi_u32 v5, v2, v4
+; GCN-NEXT:    v_mul_hi_u32 v11, v3, v4
+; GCN-NEXT:    v_mul_lo_u32 v4, v3, v4
+; GCN-NEXT:    v_add_i32_e32 v7, vcc, v8, v7
+; GCN-NEXT:    v_mul_lo_u32 v8, v3, v6
+; GCN-NEXT:    v_mul_hi_u32 v6, v3, v6
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v10, v5, vcc
+; GCN-NEXT:    v_add_i32_e32 v7, vcc, v8, v7
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v5, v6, vcc
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v11, v9, vcc
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v5, v4
+; GCN-NEXT:    v_add_i32_e64 v2, s[4:5], v2, v4
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v10, v6, vcc
+; GCN-NEXT:    v_mul_hi_u32 v6, v2, s6
+; GCN-NEXT:    v_addc_u32_e64 v4, vcc, v3, v5, s[4:5]
+; GCN-NEXT:    v_mul_lo_u32 v7, v4, s6
 ; GCN-NEXT:    v_mul_lo_u32 v8, v2, s6
-; GCN-NEXT:    v_subrev_i32_e32 v7, vcc, v2, v7
-; GCN-NEXT:    v_mul_hi_u32 v9, v2, v8
-; GCN-NEXT:    v_mul_hi_u32 v10, v5, v8
-; GCN-NEXT:    v_mul_lo_u32 v8, v5, v8
+; GCN-NEXT:    v_subrev_i32_e32 v6, vcc, v2, v6
 ; GCN-NEXT:    v_add_i32_e32 v6, vcc, v6, v7
-; GCN-NEXT:    v_mul_hi_u32 v7, v2, v6
-; GCN-NEXT:    v_mul_lo_u32 v11, v2, v6
-; GCN-NEXT:    v_mul_hi_u32 v12, v5, v6
-; GCN-NEXT:    v_mul_lo_u32 v6, v5, v6
-; GCN-NEXT:    v_add_i32_e32 v9, vcc, v9, v11
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v4, v7, vcc
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v8, v9
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v7, v10, vcc
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v12, v3, vcc
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, v7, v6
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v4, v8, vcc
-; GCN-NEXT:    v_add_i32_e64 v2, s[4:5], v2, v6
-; GCN-NEXT:    v_addc_u32_e64 v6, vcc, v5, v7, s[4:5]
-; GCN-NEXT:    v_mul_hi_u32 v8, v2, s6
-; GCN-NEXT:    v_mul_lo_u32 v9, v2, s6
-; GCN-NEXT:    v_mul_lo_u32 v10, v6, s6
-; GCN-NEXT:    v_subrev_i32_e32 v8, vcc, v2, v8
-; GCN-NEXT:    v_mul_hi_u32 v11, v2, v9
-; GCN-NEXT:    v_mul_hi_u32 v12, v6, v9
-; GCN-NEXT:    v_mul_lo_u32 v9, v6, v9
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v8, v10
-; GCN-NEXT:    v_mul_hi_u32 v10, v2, v8
-; GCN-NEXT:    v_mul_lo_u32 v13, v2, v8
-; GCN-NEXT:    v_mul_hi_u32 v14, v6, v8
-; GCN-NEXT:    v_mul_lo_u32 v6, v6, v8
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v11, v13
-; GCN-NEXT:    v_addc_u32_e32 v10, vcc, v4, v10, vcc
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v9, v8
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v10, v12, vcc
-; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v14, v3, vcc
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v7
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, v8, v6
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v4, v9, vcc
-; GCN-NEXT:    v_addc_u32_e64 v5, vcc, v5, v7, s[4:5]
-; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v6
-; GCN-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; GCN-NEXT:    v_mul_hi_u32 v6, v0, v2
-; GCN-NEXT:    v_mul_hi_u32 v7, v1, v2
-; GCN-NEXT:    v_mul_lo_u32 v2, v1, v2
-; GCN-NEXT:    v_mul_hi_u32 v8, v0, v5
-; GCN-NEXT:    v_mul_lo_u32 v9, v0, v5
-; GCN-NEXT:    v_mul_hi_u32 v10, v1, v5
-; GCN-NEXT:    v_mul_lo_u32 v5, v1, v5
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, v6, v9
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v4, v8, vcc
-; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v6
-; GCN-NEXT:    v_addc_u32_e32 v2, vcc, v8, v7, vcc
-; GCN-NEXT:    v_addc_u32_e32 v3, vcc, v10, v3, vcc
-; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v5
-; GCN-NEXT:    v_addc_u32_e32 v3, vcc, v4, v3, vcc
-; GCN-NEXT:    v_mul_hi_u32 v4, v2, 24
-; GCN-NEXT:    v_mul_lo_u32 v5, v2, 24
-; GCN-NEXT:    v_mul_lo_u32 v6, v3, 24
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, 2, v2
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, 0, v3, vcc
-; GCN-NEXT:    v_add_i32_e32 v9, vcc, 1, v2
-; GCN-NEXT:    v_addc_u32_e32 v10, vcc, 0, v3, vcc
-; GCN-NEXT:    v_add_i32_e32 v4, vcc, v4, v6
-; GCN-NEXT:    v_sub_i32_e32 v0, vcc, v0, v5
+; GCN-NEXT:    v_mul_lo_u32 v7, v2, v6
+; GCN-NEXT:    v_mul_hi_u32 v11, v2, v8
+; GCN-NEXT:    v_mul_hi_u32 v12, v2, v6
+; GCN-NEXT:    v_mul_hi_u32 v13, v4, v6
+; GCN-NEXT:    v_add_i32_e32 v7, vcc, v11, v7
+; GCN-NEXT:    v_addc_u32_e32 v11, vcc, v10, v12, vcc
+; GCN-NEXT:    v_mul_lo_u32 v12, v4, v8
+; GCN-NEXT:    v_mul_hi_u32 v8, v4, v8
+; GCN-NEXT:    v_mul_lo_u32 v4, v4, v6
+; GCN-NEXT:    v_add_i32_e32 v7, vcc, v12, v7
+; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v11, v8, vcc
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v13, v9, vcc
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v7, v4
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v10, v6, vcc
+; GCN-NEXT:    v_add_i32_e32 v3, vcc, v3, v5
+; GCN-NEXT:    v_addc_u32_e64 v3, vcc, v3, v6, s[4:5]
+; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v4
+; GCN-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; GCN-NEXT:    v_mul_lo_u32 v4, v0, v3
+; GCN-NEXT:    v_mul_hi_u32 v5, v0, v2
+; GCN-NEXT:    v_mul_hi_u32 v6, v0, v3
+; GCN-NEXT:    v_mul_hi_u32 v7, v1, v3
+; GCN-NEXT:    v_mul_lo_u32 v3, v1, v3
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v5, v4
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v10, v6, vcc
+; GCN-NEXT:    v_mul_lo_u32 v6, v1, v2
+; GCN-NEXT:    v_mul_hi_u32 v2, v1, v2
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v6, v4
+; GCN-NEXT:    v_addc_u32_e32 v2, vcc, v5, v2, vcc
+; GCN-NEXT:    v_addc_u32_e32 v4, vcc, v7, v9, vcc
+; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v3
+; GCN-NEXT:    v_addc_u32_e32 v3, vcc, v10, v4, vcc
+; GCN-NEXT:    v_mul_lo_u32 v4, v3, 24
+; GCN-NEXT:    v_mul_hi_u32 v5, v2, 24
+; GCN-NEXT:    v_mul_lo_u32 v6, v2, 24
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v5, v4
+; GCN-NEXT:    v_sub_i32_e32 v0, vcc, v0, v6
 ; GCN-NEXT:    v_subb_u32_e32 v1, vcc, v1, v4, vcc
 ; GCN-NEXT:    v_subrev_i32_e32 v4, vcc, 24, v0
-; GCN-NEXT:    v_cmp_lt_u32_e64 s[4:5], 23, v0
-; GCN-NEXT:    v_cndmask_b32_e64 v0, 0, -1, s[4:5]
 ; GCN-NEXT:    v_subbrev_u32_e32 v5, vcc, 0, v1, vcc
 ; GCN-NEXT:    v_cmp_lt_u32_e32 vcc, 23, v4
 ; GCN-NEXT:    v_cndmask_b32_e64 v4, 0, -1, vcc
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v1
-; GCN-NEXT:    v_cndmask_b32_e32 v0, -1, v0, vcc
 ; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v5
-; GCN-NEXT:    v_cndmask_b32_e32 v1, -1, v4, vcc
-; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v1
-; GCN-NEXT:    v_cndmask_b32_e32 v1, v9, v7, vcc
+; GCN-NEXT:    v_cndmask_b32_e32 v4, -1, v4, vcc
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, 2, v2
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, 0, v3, vcc
+; GCN-NEXT:    v_add_i32_e32 v7, vcc, 1, v2
+; GCN-NEXT:    v_cmp_lt_u32_e64 s[4:5], 23, v0
+; GCN-NEXT:    v_addc_u32_e32 v8, vcc, 0, v3, vcc
+; GCN-NEXT:    v_cndmask_b32_e64 v0, 0, -1, s[4:5]
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[4:5], 0, v1
+; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v4
+; GCN-NEXT:    v_cndmask_b32_e64 v0, -1, v0, s[4:5]
+; GCN-NEXT:    v_cndmask_b32_e32 v4, v7, v5, vcc
 ; GCN-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, v0
-; GCN-NEXT:    v_cndmask_b32_e64 v0, v2, v1, s[4:5]
-; GCN-NEXT:    v_cndmask_b32_e32 v1, v10, v8, vcc
+; GCN-NEXT:    v_cndmask_b32_e32 v1, v8, v6, vcc
+; GCN-NEXT:    v_cndmask_b32_e64 v0, v2, v4, s[4:5]
 ; GCN-NEXT:    v_cndmask_b32_e64 v1, v3, v1, s[4:5]
 ; GCN-NEXT:    s_setpc_b64 s[30:31]
   %result = udiv i64 %x, 24
   ret i64 %result
 }
 
 define amdgpu_kernel void @s_test_udiv24_k_num_i64(i64 addrspace(1)* %out, i64 %x) {
 ; GCN-LABEL: s_test_udiv24_k_num_i64:
 ; GCN:       ; %bb.0:
-; GCN-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
-; GCN-NEXT:    s_mov_b32 s7, 0xf000
-; GCN-NEXT:    s_mov_b32 s6, -1
+; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_mov_b32 s4, s0
-; GCN-NEXT:    s_mov_b32 s5, s1
-; GCN-NEXT:    s_lshr_b32 s2, s3, 8
+; GCN-NEXT:    s_mov_b32 s6, -1
+; GCN-NEXT:    s_lshr_b32 s2, s7, 8
 ; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s2
+; GCN-NEXT:    s_mov_b32 s7, 0xf000
 ; GCN-NEXT:    v_rcp_iflag_f32_e32 v0, v0
 ; GCN-NEXT:    v_mul_f32_e32 v0, 0x4f800000, v0
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v0
-; GCN-NEXT:    v_mul_hi_u32 v1, v0, s2
-; GCN-NEXT:    v_mul_lo_u32 v2, v0, s2
-; GCN-NEXT:    v_sub_i32_e32 v3, vcc, 0, v2
-; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v1, v2, v3, s[0:1]
+; GCN-NEXT:    v_mul_lo_u32 v1, v0, s2
+; GCN-NEXT:    v_mul_hi_u32 v2, v0, s2
+; GCN-NEXT:    v_sub_i32_e32 v3, vcc, 0, v1
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v1, v1, v3, s[0:1]
 ; GCN-NEXT:    v_mul_hi_u32 v1, v1, v0
 ; GCN-NEXT:    v_add_i32_e32 v2, vcc, v1, v0
 ; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, v1, v0
 ; GCN-NEXT:    v_cndmask_b32_e64 v0, v0, v2, s[0:1]
 ; GCN-NEXT:    v_mul_hi_u32 v0, v0, 24
 ; GCN-NEXT:    v_mul_lo_u32 v1, v0, s2
 ; GCN-NEXT:    v_add_i32_e32 v2, vcc, 1, v0
 ; GCN-NEXT:    v_add_i32_e32 v3, vcc, -1, v0
; GCN-NEXT:    s_endpgm
   %result = udiv i64 24, %x.shr
   store i64 %result, i64 addrspace(1)* %out
   ret void
 }
 
 define amdgpu_kernel void @s_test_udiv24_k_den_i64(i64 addrspace(1)* %out, i64 %x) {
 ; GCN-LABEL: s_test_udiv24_k_den_i64:
 ; GCN:       ; %bb.0:
+; GCN-NEXT:    s_movk_i32 s8, 0x5b7f
+; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s8
 ; GCN-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; GCN-NEXT:    s_movk_i32 s9, 0x5b7e
 ; GCN-NEXT:    s_mov_b32 s7, 0xf000
+; GCN-NEXT:    v_rcp_iflag_f32_e32 v0, v0
 ; GCN-NEXT:    s_mov_b32 s6, -1
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_movk_i32 s2, 0x5b7f
-; GCN-NEXT:    s_movk_i32 s8, 0x5b7e
-; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s2
-; GCN-NEXT:    v_rcp_iflag_f32_e32 v0, v0
-; GCN-NEXT:    v_mul_f32_e32 v0, 0x4f800000, v0
 ; GCN-NEXT:    s_mov_b32 s4, s0
 ; GCN-NEXT:    s_mov_b32 s5, s1
-; GCN-NEXT:    s_lshr_b32 s3, s3, 8
+; GCN-NEXT:    v_mul_f32_e32 v0, 0x4f800000, v0
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v0
-; GCN-NEXT:    v_mul_hi_u32 v1, v0, s2
-; GCN-NEXT:    v_mul_lo_u32 v2, v0, s2
-; GCN-NEXT:    v_sub_i32_e32 v3, vcc, 0, v2
-; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v1, v2, v3, s[0:1]
+; GCN-NEXT:    s_lshr_b32 s2, s3, 8
+; GCN-NEXT:    v_mul_lo_u32 v1, v0, s8
+; GCN-NEXT:    v_mul_hi_u32 v2, v0, s8
+; GCN-NEXT:    v_sub_i32_e32 v3, vcc, 0, v1
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v1, v1, v3, s[0:1]
 ; GCN-NEXT:    v_mul_hi_u32 v1, v1, v0
 ; GCN-NEXT:    v_add_i32_e32 v2, vcc, v1, v0
 ; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, v1, v0
 ; GCN-NEXT:    v_cndmask_b32_e64 v0, v0, v2, s[0:1]
-; GCN-NEXT:    v_mul_hi_u32 v0, v0, s3
-; GCN-NEXT:    v_mul_lo_u32 v1, v0, s2
+; GCN-NEXT:    v_mul_hi_u32 v0, v0, s2
+; GCN-NEXT:    v_mul_lo_u32 v1, v0, s8
 ; GCN-NEXT:    v_add_i32_e32 v2, vcc, 1, v0
 ; GCN-NEXT:    v_add_i32_e32 v3, vcc, -1, v0
-; GCN-NEXT:    v_cmp_ge_u32_e64 s[0:1], s3, v1
-; GCN-NEXT:    v_sub_i32_e32 v1, vcc, s3, v1
-; GCN-NEXT:    v_cmp_lt_u32_e32 vcc, s8, v1
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[0:1], s2, v1
+; GCN-NEXT:    v_sub_i32_e32 v1, vcc, s2, v1
+; GCN-NEXT:    v_cmp_lt_u32_e32 vcc, s9, v1
 ; GCN-NEXT:    s_and_b64 vcc, vcc, s[0:1]
 ; GCN-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
 ; GCN-NEXT:    v_cndmask_b32_e64 v0, v3, v0, s[0:1]
 ; GCN-NEXT:    v_mov_b32_e32 v1, 0
 ; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
 ; GCN-NEXT:    s_endpgm
   %x.shr = lshr i64 %x, 40
   %result = udiv i64 %x.shr, 23423
   store i64 %result, i64 addrspace(1)* %out
   ret void
 }
 
 define i64 @v_test_udiv24_k_num_i64(i64 %x) {
 ; GCN-LABEL: v_test_udiv24_k_num_i64:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GCN-NEXT:    v_lshrrev_b32_e32 v0, 8, v1
 ; GCN-NEXT:    v_cvt_f32_u32_e32 v1, v0
 ; GCN-NEXT:    v_rcp_iflag_f32_e32 v1, v1
 ; GCN-NEXT:    v_mul_f32_e32 v1, 0x4f800000, v1
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v1, v1
-; GCN-NEXT:    v_mul_hi_u32 v2, v1, v0
-; GCN-NEXT:    v_mul_lo_u32 v3, v1, v0
-; GCN-NEXT:    v_sub_i32_e32 v4, vcc, 0, v3
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v2
-; GCN-NEXT:    v_cndmask_b32_e32 v2, v3, v4, vcc
+; GCN-NEXT:    v_mul_lo_u32 v2, v1, v0
+; GCN-NEXT:    v_mul_hi_u32 v3, v1, v0
+; GCN-NEXT:    v_sub_i32_e32 v4, vcc, 0, v2
+; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v3
+; GCN-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
 ; GCN-NEXT:    v_mul_hi_u32 v2, v2, v1
 ; GCN-NEXT:    v_add_i32_e64 v3, s[4:5], v1, v2
 ; GCN-NEXT:    v_sub_i32_e64 v1, s[4:5], v1, v2
 ; GCN-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
 ; GCN-NEXT:    v_mul_hi_u32 v1, v1, 24
 ; GCN-NEXT:    v_mul_lo_u32 v2, v1, v0
 ; GCN-NEXT:    v_add_i32_e32 v3, vcc, 1, v1
 ; GCN-NEXT:    v_add_i32_e32 v4, vcc, -1, v1
; GCN-NEXT:    s_setpc_b64 s[30:31]
   ret i64 %result
 }
 
 define i64 @v_test_udiv24_pow2_k_num_i64(i64 %x) {
 ; GCN-LABEL: v_test_udiv24_pow2_k_num_i64:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GCN-NEXT:    v_lshrrev_b32_e32 v0, 8, v1
-; GCN-NEXT:    s_mov_b32 s6, 0x8001
 ; GCN-NEXT:    v_cvt_f32_u32_e32 v1, v0
+; GCN-NEXT:    s_mov_b32 s6, 0x8001
 ; GCN-NEXT:    v_rcp_iflag_f32_e32 v1, v1
 ; GCN-NEXT:    v_mul_f32_e32 v1, 0x4f800000, v1
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v1, v1
-; GCN-NEXT:    v_mul_hi_u32 v2, v1, v0
-; GCN-NEXT:    v_mul_lo_u32 v3, v1, v0
-; GCN-NEXT:    v_sub_i32_e32 v4, vcc, 0, v3
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v2
-; GCN-NEXT:    v_cndmask_b32_e32 v2, v3, v4, vcc
+; GCN-NEXT:    v_mul_lo_u32 v2, v1, v0
+; GCN-NEXT:    v_mul_hi_u32 v3, v1, v0
+; GCN-NEXT:    v_sub_i32_e32 v4, vcc, 0, v2
+; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v3
+; GCN-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
 ; GCN-NEXT:    v_mul_hi_u32 v2, v2, v1
 ; GCN-NEXT:    v_add_i32_e64 v3, s[4:5], v1, v2
 ; GCN-NEXT:    v_sub_i32_e64 v1, s[4:5], v1, v2
 ; GCN-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
 ; GCN-NEXT:    v_lshrrev_b32_e32 v1, 17, v1
 ; GCN-NEXT:    v_mul_u32_u24_e32 v2, v1, v0
+; GCN-NEXT:    v_cmp_gt_u32_e64 s[4:5], s6, v2
 ; GCN-NEXT:    v_add_i32_e32 v3, vcc, 1, v1
 ; GCN-NEXT:    v_add_i32_e32 v4, vcc, -1, v1
-; GCN-NEXT:    v_cmp_gt_u32_e64 s[4:5], s6, v2
 ; GCN-NEXT:    v_sub_i32_e32 v2, vcc, 0x8000, v2
 ; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v2, v0
 ; GCN-NEXT:    s_and_b64 vcc, vcc, s[4:5]
 ; GCN-NEXT:    v_cndmask_b32_e32 v0, v1, v3, vcc
 ; GCN-NEXT:    v_cndmask_b32_e64 v0, v4, v0, s[4:5]
 ; GCN-NEXT:    v_mov_b32_e32 v1, 0
 ; GCN-NEXT:    s_setpc_b64 s[30:31]
   %x.shr = lshr i64 %x, 40

llvm/test/CodeGen/AMDGPU/urem64.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
+; RUN: llc -march=amdgcn -mcpu=gfx600 -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
 
 define amdgpu_kernel void @s_test_urem_i64(i64 addrspace(1)* %out, i64 %x, i64 %y) {
 ; GCN-LABEL: s_test_urem_i64:
 ; GCN:       ; %bb.0:
+; GCN-NEXT:    s_load_dwordx2 s[12:13], s[0:1], 0xd
+; GCN-NEXT:    v_mov_b32_e32 v2, 0
 ; GCN-NEXT:    s_load_dwordx4 s[8:11], s[0:1], 0x9
-; GCN-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xd
 ; GCN-NEXT:    s_mov_b32 s7, 0xf000
 ; GCN-NEXT:    s_mov_b32 s6, -1
-; GCN-NEXT:    v_mov_b32_e32 v0, 0
-; GCN-NEXT:    v_mov_b32_e32 v1, 0
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s12
+; GCN-NEXT:    v_cvt_f32_u32_e32 v1, s13
+; GCN-NEXT:    s_sub_u32 s2, 0, s12
+; GCN-NEXT:    s_subb_u32 s3, 0, s13
 ; GCN-NEXT:    s_mov_b32 s4, s8
+; GCN-NEXT:    v_mac_f32_e32 v0, 0x4f800000, v1
+; GCN-NEXT:    v_rcp_f32_e32 v0, v0
+; GCN-NEXT:    v_mov_b32_e32 v1, 0
 ; GCN-NEXT:    s_mov_b32 s5, s9
-; GCN-NEXT:    v_cvt_f32_u32_e32 v2, s2
-; GCN-NEXT:    v_cvt_f32_u32_e32 v3, s3
-; GCN-NEXT:    s_sub_u32 s8, 0, s2
-; GCN-NEXT:    v_mov_b32_e32 v4, s3
-; GCN-NEXT:    v_mov_b32_e32 v5, s11
-; GCN-NEXT:    v_mac_f32_e32 v2, 0x4f800000, v3
-; GCN-NEXT:    s_subb_u32 s9, 0, s3
-; GCN-NEXT:    v_rcp_f32_e32 v2, v2
-; GCN-NEXT:    v_mul_f32_e32 v2, 0x5f7ffffc, v2
-; GCN-NEXT:    v_mul_f32_e32 v3, 0x2f800000, v2
+; GCN-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v0
+; GCN-NEXT:    v_mul_f32_e32 v3, 0x2f800000, v0
 ; GCN-NEXT:    v_trunc_f32_e32 v3, v3
-; GCN-NEXT:    v_mac_f32_e32 v2, 0xcf800000, v3
+; GCN-NEXT:    v_mac_f32_e32 v0, 0xcf800000, v3
+; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v0
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v3, v3
-; GCN-NEXT:    v_cvt_u32_f32_e32 v2, v2
-; GCN-NEXT:    v_mul_lo_u32 v6, s8, v3
-; GCN-NEXT:    v_mul_lo_u32 v7, s9, v2
-; GCN-NEXT:    v_mul_hi_u32 v8, s8, v2
-; GCN-NEXT:    v_mul_lo_u32 v9, s8, v2
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, v8, v6
-; GCN-NEXT:    v_mul_hi_u32 v8, v2, v9
-; GCN-NEXT:    v_mul_hi_u32 v10, v3, v9
-; GCN-NEXT:    v_mul_lo_u32 v9, v3, v9
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, v6, v7
-; GCN-NEXT:    v_mul_hi_u32 v7, v2, v6
-; GCN-NEXT:    v_mul_lo_u32 v11, v2, v6
-; GCN-NEXT:    v_mul_hi_u32 v12, v3, v6
-; GCN-NEXT:    v_mul_lo_u32 v6, v3, v6
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v8, v11
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v1, v7, vcc
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v9, v8
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v7, v10, vcc
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v12, v0, vcc
+; GCN-NEXT:    v_mul_hi_u32 v5, s2, v0
+; GCN-NEXT:    v_mul_lo_u32 v4, s2, v3
+; GCN-NEXT:    v_mul_lo_u32 v7, s3, v0
+; GCN-NEXT:    v_mul_lo_u32 v6, s2, v0
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v5, v4
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v4, v7
+; GCN-NEXT:    v_mul_hi_u32 v5, v0, v6
+; GCN-NEXT:    v_mul_lo_u32 v7, v0, v4
+; GCN-NEXT:    v_mul_hi_u32 v9, v0, v4
+; GCN-NEXT:    v_mul_lo_u32 v8, v3, v6
+; GCN-NEXT:    v_mul_hi_u32 v6, v3, v6
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v7
+; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v2, v9, vcc
+; GCN-NEXT:    v_mul_hi_u32 v9, v3, v4
+; GCN-NEXT:    v_mul_lo_u32 v4, v3, v4
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v8, v5
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v7, v6, vcc
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v9, v1, vcc
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v5, v4
+; GCN-NEXT:    v_add_i32_e64 v0, s[0:1], v0, v4
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v2, v6, vcc
+; GCN-NEXT:    v_addc_u32_e64 v4, vcc, v3, v5, s[0:1]
+; GCN-NEXT:    v_mul_lo_u32 v6, s2, v4
+; GCN-NEXT:    v_mul_hi_u32 v7, s2, v0
+; GCN-NEXT:    v_mul_lo_u32 v8, s3, v0
 ; GCN-NEXT:    v_add_i32_e32 v6, vcc, v7, v6
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v1, v8, vcc
-; GCN-NEXT:    v_add_i32_e64 v2, s[0:1], v2, v6
-; GCN-NEXT:    v_addc_u32_e64 v6, vcc, v3, v7, s[0:1]
-; GCN-NEXT:    v_mul_hi_u32 v8, s8, v2
-; GCN-NEXT:    v_mul_lo_u32 v9, s9, v2
-; GCN-NEXT:    v_mul_lo_u32 v10, s8, v2
-; GCN-NEXT:    v_mul_lo_u32 v11, s8, v6
-; GCN-NEXT:    v_mul_hi_u32 v12, v6, v10
-; GCN-NEXT:    v_mul_lo_u32 v13, v6, v10
-; GCN-NEXT:    v_mul_hi_u32 v10, v2, v10
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v8, v11
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v9, v8
-; GCN-NEXT:    v_mul_hi_u32 v9, v6, v8
-; GCN-NEXT:    v_mul_hi_u32 v11, v2, v8
-; GCN-NEXT:    v_mul_lo_u32 v14, v2, v8
-; GCN-NEXT:    v_mul_lo_u32 v6, v6, v8
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v10, v14
-; GCN-NEXT:    v_addc_u32_e32 v10, vcc, v1, v11, vcc
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v13, v8
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v10, v12, vcc
-; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v9, v0, vcc
-; GCN-NEXT:    v_add_i32_e32 v3, vcc, v3, v7
+; GCN-NEXT:    v_mul_lo_u32 v7, s2, v0
 ; GCN-NEXT:    v_add_i32_e32 v6, vcc, v8, v6
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v1, v9, vcc
-; GCN-NEXT:    v_addc_u32_e64 v3, vcc, v3, v7, s[0:1]
-; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v6
+; GCN-NEXT:    v_mul_lo_u32 v10, v0, v6
+; GCN-NEXT:    v_mul_hi_u32 v12, v0, v6
+; GCN-NEXT:    v_mul_hi_u32 v11, v0, v7
+; GCN-NEXT:    v_mul_hi_u32 v9, v4, v7
+; GCN-NEXT:    v_mul_lo_u32 v7, v4, v7
+; GCN-NEXT:    v_mul_hi_u32 v8, v4, v6
+; GCN-NEXT:    v_add_i32_e32 v10, vcc, v11, v10
+; GCN-NEXT:    v_addc_u32_e32 v11, vcc, v2, v12, vcc
+; GCN-NEXT:    v_mul_lo_u32 v4, v4, v6
+; GCN-NEXT:    v_add_i32_e32 v7, vcc, v7, v10
+; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v11, v9, vcc
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v8, v1, vcc
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v7, v4
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v2, v6, vcc
+; GCN-NEXT:    v_add_i32_e32 v3, vcc, v3, v5
+; GCN-NEXT:    v_addc_u32_e64 v3, vcc, v3, v6, s[0:1]
+; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v4
 ; GCN-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; GCN-NEXT:    v_mul_hi_u32 v6, s10, v2
-; GCN-NEXT:    v_mul_hi_u32 v7, s11, v2
-; GCN-NEXT:    v_mul_lo_u32 v2, s11, v2
-; GCN-NEXT:    v_mul_hi_u32 v8, s10, v3
-; GCN-NEXT:    v_mul_lo_u32 v9, s10, v3
-; GCN-NEXT:    v_mul_hi_u32 v10, s11, v3
+; GCN-NEXT:    v_mul_lo_u32 v4, s10, v3
+; GCN-NEXT:    v_mul_hi_u32 v5, s10, v0
+; GCN-NEXT:    v_mul_hi_u32 v6, s10, v3
+; GCN-NEXT:    v_mul_hi_u32 v7, s11, v3
 ; GCN-NEXT:    v_mul_lo_u32 v3, s11, v3
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, v6, v9
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v1, v8, vcc
-; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v6
-; GCN-NEXT:    v_addc_u32_e32 v2, vcc, v8, v7, vcc
-; GCN-NEXT:    v_addc_u32_e32 v0, vcc, v10, v0, vcc
-; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v3
-; GCN-NEXT:    v_addc_u32_e32 v0, vcc, v1, v0, vcc
-; GCN-NEXT:    v_mul_hi_u32 v1, s2, v2
-; GCN-NEXT:    v_mul_lo_u32 v3, s3, v2
-; GCN-NEXT:    v_mul_lo_u32 v2, s2, v2
-; GCN-NEXT:    v_mul_lo_u32 v0, s2, v0
-; GCN-NEXT:    v_add_i32_e32 v0, vcc, v1, v0
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v5, v4
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v2, v6, vcc
+; GCN-NEXT:    v_mul_lo_u32 v6, s11, v0
+; GCN-NEXT:    v_mul_hi_u32 v0, s11, v0
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v6, v4
+; GCN-NEXT:    v_addc_u32_e32 v0, vcc, v5, v0, vcc
+; GCN-NEXT:    v_addc_u32_e32 v1, vcc, v7, v1, vcc
 ; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v3
-; GCN-NEXT:    v_sub_i32_e32 v1, vcc, s11, v0
-; GCN-NEXT:    v_sub_i32_e32 v2, vcc, s10, v2
-; GCN-NEXT:    v_subb_u32_e64 v1, s[0:1], v1, v4, vcc
-; GCN-NEXT:    v_subb_u32_e32 v0, vcc, v5, v0, vcc
-; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s2, v2
+; GCN-NEXT:    v_addc_u32_e32 v1, vcc, v2, v1, vcc
+; GCN-NEXT:    v_mul_lo_u32 v1, s12, v1
+; GCN-NEXT:    v_mul_hi_u32 v2, s12, v0
+; GCN-NEXT:    v_mul_lo_u32 v3, s13, v0
+; GCN-NEXT:    v_mul_lo_u32 v0, s12, v0
+; GCN-NEXT:    v_add_i32_e32 v1, vcc, v2, v1
+; GCN-NEXT:    v_add_i32_e32 v1, vcc, v1, v3
+; GCN-NEXT:    v_sub_i32_e32 v2, vcc, s11, v1
+; GCN-NEXT:    v_mov_b32_e32 v3, s13
+; GCN-NEXT:    v_sub_i32_e32 v0, vcc, s10, v0
+; GCN-NEXT:    v_subb_u32_e64 v2, s[0:1], v2, v3, vcc
+; GCN-NEXT:    v_subrev_i32_e64 v4, s[0:1], s12, v0
+; GCN-NEXT:    v_subb_u32_e64 v3, s[2:3], v2, v3, s[0:1]
+; GCN-NEXT:    v_subbrev_u32_e64 v2, s[0:1], 0, v2, s[0:1]
+; GCN-NEXT:    v_cmp_le_u32_e64 s[0:1], s13, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v5, 0, -1, s[0:1]
+; GCN-NEXT:    v_cmp_le_u32_e64 s[0:1], s12, v4
+; GCN-NEXT:    v_cndmask_b32_e64 v6, 0, -1, s[0:1]
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], s13, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v5, v5, v6, s[0:1]
+; GCN-NEXT:    v_subrev_i32_e64 v6, s[0:1], s12, v4
+; GCN-NEXT:    v_subbrev_u32_e64 v3, s[0:1], 0, v3, s[0:1]
+; GCN-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v5
+; GCN-NEXT:    v_cndmask_b32_e64 v2, v2, v3, s[0:1]
+; GCN-NEXT:    v_mov_b32_e32 v3, s11
+; GCN-NEXT:    v_subb_u32_e32 v1, vcc, v3, v1, vcc
+; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s13, v1
 ; GCN-NEXT:    v_cndmask_b32_e64 v3, 0, -1, vcc
-; GCN-NEXT:    v_subrev_i32_e32 v5, vcc, s2, v2
-; GCN-NEXT:    v_subb_u32_e64 v4, s[0:1], v1, v4, vcc
-; GCN-NEXT:    v_subbrev_u32_e32 v1, vcc, 0, v1, vcc
-; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s2, v5
-; GCN-NEXT:    v_cndmask_b32_e64 v6, 0, -1, vcc
-; GCN-NEXT:    v_subrev_i32_e32 v7, vcc, s2, v5
-; GCN-NEXT:    v_cmp_le_u32_e64 s[0:1], s3, v0
-; GCN-NEXT:    v_cndmask_b32_e64 v8, 0, -1, s[0:1]
-; GCN-NEXT:    v_cmp_le_u32_e64 s[0:1], s3, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v9, 0, -1, s[0:1]
-; GCN-NEXT:    v_subbrev_u32_e32 v4, vcc, 0, v4, vcc
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, s3, v0
-; GCN-NEXT:    v_cndmask_b32_e32 v3, v8, v3, vcc
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, s3, v1
-; GCN-NEXT:    v_cndmask_b32_e32 v6, v9, v6, vcc
-; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v6
-; GCN-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
-; GCN-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v3
-; GCN-NEXT:    v_cndmask_b32_e64 v1, v0, v1, s[0:1]
-; GCN-NEXT:    v_cndmask_b32_e32 v0, v5, v7, vcc
-; GCN-NEXT:    v_cndmask_b32_e64 v0, v2, v0, s[0:1]
+; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s12, v0
+; GCN-NEXT:    v_cndmask_b32_e64 v5, 0, -1, vcc
+; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, s13, v1
+; GCN-NEXT:    v_cndmask_b32_e32 v3, v3, v5, vcc
+; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v3
+; GCN-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc
+; GCN-NEXT:    v_cndmask_b32_e64 v2, v4, v6, s[0:1]
+; GCN-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
 ; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
 ; GCN-NEXT:    s_endpgm
   %result = urem i64 %x, %y
   store i64 %result, i64 addrspace(1)* %out
   ret void
 }
 
 define i64 @v_test_urem_i64(i64 %x, i64 %y) {
 ; GCN-LABEL: v_test_urem_i64:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GCN-NEXT:    v_cvt_f32_u32_e32 v4, v2
 ; GCN-NEXT:    v_cvt_f32_u32_e32 v5, v3
 ; GCN-NEXT:    v_sub_i32_e32 v6, vcc, 0, v2
 ; GCN-NEXT:    v_subb_u32_e32 v7, vcc, 0, v3, vcc
-; GCN-NEXT:    v_mov_b32_e32 v8, 0
-; GCN-NEXT:    v_mov_b32_e32 v9, 0
+; GCN-NEXT:    v_mov_b32_e32 v14, 0
 ; GCN-NEXT:    v_mac_f32_e32 v4, 0x4f800000, v5
 ; GCN-NEXT:    v_rcp_f32_e32 v4, v4
+; GCN-NEXT:    v_mov_b32_e32 v13, 0
 ; GCN-NEXT:    v_mul_f32_e32 v4, 0x5f7ffffc, v4
 ; GCN-NEXT:    v_mul_f32_e32 v5, 0x2f800000, v4
 ; GCN-NEXT:    v_trunc_f32_e32 v5, v5
 ; GCN-NEXT:    v_mac_f32_e32 v4, 0xcf800000, v5
-; GCN-NEXT:    v_cvt_u32_f32_e32 v5, v5
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v4, v4
-; GCN-NEXT:    v_mul_lo_u32 v10, v6, v5
-; GCN-NEXT:    v_mul_lo_u32 v11, v7, v4
-; GCN-NEXT:    v_mul_hi_u32 v12, v6, v4
-; GCN-NEXT:    v_mul_lo_u32 v13, v6, v4
-; GCN-NEXT:    v_add_i32_e32 v10, vcc, v12, v10
-; GCN-NEXT:    v_mul_hi_u32 v12, v4, v13
-; GCN-NEXT:    v_mul_hi_u32 v14, v5, v13
-; GCN-NEXT:    v_mul_lo_u32 v13, v5, v13
-; GCN-NEXT:    v_add_i32_e32 v10, vcc, v10, v11
-; GCN-NEXT:    v_mul_hi_u32 v11, v4, v10
-; GCN-NEXT:    v_mul_lo_u32 v15, v4, v10
-; GCN-NEXT:    v_mul_hi_u32 v16, v5, v10
-; GCN-NEXT:    v_mul_lo_u32 v10, v5, v10
-; GCN-NEXT:    v_add_i32_e32 v12, vcc, v12, v15
-; GCN-NEXT:    v_addc_u32_e32 v11, vcc, v9, v11, vcc
-; GCN-NEXT:    v_add_i32_e32 v12, vcc, v13, v12
-; GCN-NEXT:    v_addc_u32_e32 v11, vcc, v11, v14, vcc
-; GCN-NEXT:    v_addc_u32_e32 v12, vcc, v16, v8, vcc
-; GCN-NEXT:    v_add_i32_e32 v10, vcc, v11, v10
-; GCN-NEXT:    v_addc_u32_e32 v11, vcc, v9, v12, vcc
-; GCN-NEXT:    v_add_i32_e64 v4, s[4:5], v4, v10
-; GCN-NEXT:    v_addc_u32_e64 v10, vcc, v5, v11, s[4:5]
-; GCN-NEXT:    v_mul_hi_u32 v12, v6, v4
+; GCN-NEXT:    v_cvt_u32_f32_e32 v5, v5
+; GCN-NEXT:    v_mul_hi_u32 v8, v6, v4
+; GCN-NEXT:    v_mul_lo_u32 v9, v6, v5
+; GCN-NEXT:    v_mul_lo_u32 v10, v7, v4
+; GCN-NEXT:    v_add_i32_e32 v8, vcc, v8, v9
+; GCN-NEXT:    v_mul_lo_u32 v9, v6, v4
+; GCN-NEXT:    v_add_i32_e32 v8, vcc, v8, v10
+; GCN-NEXT:    v_mul_lo_u32 v11, v4, v8
+; GCN-NEXT:    v_mul_hi_u32 v10, v4, v8
+; GCN-NEXT:    v_mul_hi_u32 v12, v4, v9
+; GCN-NEXT:    v_mul_hi_u32 v15, v5, v8
+; GCN-NEXT:    v_mul_lo_u32 v8, v5, v8
+; GCN-NEXT:    v_add_i32_e32 v11, vcc, v12, v11
+; GCN-NEXT:    v_mul_lo_u32 v12, v5, v9
+; GCN-NEXT:    v_mul_hi_u32 v9, v5, v9
+; GCN-NEXT:    v_addc_u32_e32 v10, vcc, v14, v10, vcc
+; GCN-NEXT:    v_add_i32_e32 v11, vcc, v12, v11
+; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v10, v9, vcc
+; GCN-NEXT:    v_addc_u32_e32 v10, vcc, v15, v13, vcc
+; GCN-NEXT:    v_add_i32_e32 v8, vcc, v9, v8
+; GCN-NEXT:    v_add_i32_e64 v4, s[4:5], v4, v8
+; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v14, v10, vcc
+; GCN-NEXT:    v_addc_u32_e64 v8, vcc, v5, v9, s[4:5]
+; GCN-NEXT:    v_mul_lo_u32 v10, v6, v8
+; GCN-NEXT:    v_mul_hi_u32 v11, v6, v4
 ; GCN-NEXT:    v_mul_lo_u32 v7, v7, v4
-; GCN-NEXT:    v_mul_lo_u32 v13, v6, v4
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v11
-; GCN-NEXT:    v_mul_lo_u32 v6, v6, v10
-; GCN-NEXT:    v_mul_hi_u32 v11, v10, v13
-; GCN-NEXT:    v_mul_lo_u32 v14, v10, v13
-; GCN-NEXT:    v_mul_hi_u32 v13, v4, v13
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, v12, v6
+; GCN-NEXT:    v_mul_lo_u32 v6, v6, v4
+; GCN-NEXT:    v_add_i32_e32 v10, vcc, v11, v10
+; GCN-NEXT:    v_add_i32_e32 v7, vcc, v10, v7
+; GCN-NEXT:    v_mul_lo_u32 v12, v4, v7
+; GCN-NEXT:    v_mul_hi_u32 v15, v4, v6
+; GCN-NEXT:    v_mul_hi_u32 v16, v4, v7
+; GCN-NEXT:    v_mul_hi_u32 v11, v8, v6
+; GCN-NEXT:    v_mul_lo_u32 v6, v8, v6
+; GCN-NEXT:    v_add_i32_e32 v12, vcc, v15, v12
+; GCN-NEXT:    v_mul_hi_u32 v10, v8, v7
+; GCN-NEXT:    v_addc_u32_e32 v15, vcc, v14, v16, vcc
+; GCN-NEXT:    v_mul_lo_u32 v7, v8, v7
+; GCN-NEXT:    v_add_i32_e32 v6, vcc, v6, v12
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v15, v11, vcc
+; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v10, v13, vcc
 ; GCN-NEXT:    v_add_i32_e32 v6, vcc, v6, v7
-; GCN-NEXT:    v_mul_hi_u32 v7, v10, v6
-; GCN-NEXT:    v_mul_hi_u32 v12, v4, v6
-; GCN-NEXT:    v_mul_lo_u32 v15, v4, v6
-; GCN-NEXT:    v_mul_lo_u32 v6, v10, v6
-; GCN-NEXT:    v_add_i32_e32 v10, vcc, v13, v15
-; GCN-NEXT:    v_addc_u32_e32 v12, vcc, v9, v12, vcc
-; GCN-NEXT:    v_add_i32_e32 v10, vcc, v14, v10
-; GCN-NEXT:    v_addc_u32_e32 v10, vcc, v12, v11, vcc
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v7, v8, vcc
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, v10, v6
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v9, v7, vcc
+; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v14, v8, vcc
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v9
 ; GCN-NEXT:    v_addc_u32_e64 v5, vcc, v5, v7, s[4:5]
 ; GCN-NEXT:    v_add_i32_e32 v4, vcc, v4, v6
 ; GCN-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
-; GCN-NEXT:    v_mul_hi_u32 v6, v0, v4
-; GCN-NEXT:    v_mul_hi_u32 v7, v1, v4
-; GCN-NEXT:    v_mul_lo_u32 v4, v1, v4
-; GCN-NEXT:    v_mul_hi_u32 v10, v0, v5
-; GCN-NEXT:    v_mul_lo_u32 v11, v0, v5
-; GCN-NEXT:    v_mul_hi_u32 v12, v1, v5
+; GCN-NEXT:    v_mul_lo_u32 v6, v0, v5
+; GCN-NEXT:    v_mul_hi_u32 v7, v0, v4
+; GCN-NEXT:    v_mul_hi_u32 v8, v0, v5
+; GCN-NEXT:    v_mul_hi_u32 v9, v1, v5
 ; GCN-NEXT:    v_mul_lo_u32 v5, v1, v5
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, v6, v11
-; GCN-NEXT:    v_addc_u32_e32 v10, vcc, v9, v10, vcc
-; GCN-NEXT:    v_add_i32_e32 v4, vcc, v4, v6
-; GCN-NEXT:    v_addc_u32_e32 v4, vcc, v10, v7, vcc
-; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v12, v8, vcc
+; GCN-NEXT:    v_add_i32_e32 v6, vcc, v7, v6
+; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v14, v8, vcc
+; GCN-NEXT:    v_mul_lo_u32 v8, v1, v4
+; GCN-NEXT:    v_mul_hi_u32 v4, v1, v4
+; GCN-NEXT:    v_add_i32_e32 v6, vcc, v8, v6
+; GCN-NEXT:    v_addc_u32_e32 v4, vcc, v7, v4, vcc
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v9, v13, vcc
 ; GCN-NEXT:    v_add_i32_e32 v4, vcc, v4, v5
-; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v9, v6, vcc
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v14, v6, vcc
+; GCN-NEXT:    v_mul_lo_u32 v5, v2, v5
 ; GCN-NEXT:    v_mul_hi_u32 v6, v2, v4
 ; GCN-NEXT:    v_mul_lo_u32 v7, v3, v4
 ; GCN-NEXT:    v_mul_lo_u32 v4, v2, v4
-; GCN-NEXT:    v_mul_lo_u32 v5, v2, v5
 ; GCN-NEXT:    v_add_i32_e32 v5, vcc, v6, v5
 ; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v7
 ; GCN-NEXT:    v_sub_i32_e32 v6, vcc, v1, v5
 ; GCN-NEXT:    v_sub_i32_e32 v0, vcc, v0, v4
 ; GCN-NEXT:    v_subb_u32_e64 v4, s[4:5], v6, v3, vcc
+; GCN-NEXT:    v_sub_i32_e64 v6, s[4:5], v0, v2
+; GCN-NEXT:    v_subb_u32_e64 v7, s[6:7], v4, v3, s[4:5]
+; GCN-NEXT:    v_subbrev_u32_e64 v4, s[4:5], 0, v4, s[4:5]
 ; GCN-NEXT:    v_subb_u32_e32 v1, vcc, v1, v5, vcc
-; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v0, v2
-; GCN-NEXT:    v_cndmask_b32_e64 v5, 0, -1, vcc
-; GCN-NEXT:    v_sub_i32_e32 v6, vcc, v0, v2
-; GCN-NEXT:    v_subb_u32_e64 v7, s[4:5], v4, v3, vcc
-; GCN-NEXT:    v_subbrev_u32_e32 v4, vcc, 0, v4, vcc
-; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v6, v2
-; GCN-NEXT:    v_cndmask_b32_e64 v8, 0, -1, vcc
-; GCN-NEXT:    v_sub_i32_e32 v2, vcc, v6, v2
-; GCN-NEXT:    v_cmp_ge_u32_e64 s[4:5], v1, v3
-; GCN-NEXT:    v_cndmask_b32_e64 v9, 0, -1, s[4:5]
 ; GCN-NEXT:    v_cmp_ge_u32_e64 s[4:5], v4, v3
-; GCN-NEXT:    v_cndmask_b32_e64 v10, 0, -1, s[4:5]
-; GCN-NEXT:    v_cmp_eq_u32_e64 s[4:5], v1, v3
-; GCN-NEXT:    v_cndmask_b32_e64 v5, v9, v5, s[4:5]
-; GCN-NEXT:    v_subbrev_u32_e32 v7, vcc, 0, v7, vcc
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
-; GCN-NEXT:    v_cndmask_b32_e32 v3, v10, v8, vcc
-; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v3
-; GCN-NEXT:    v_cndmask_b32_e32 v2, v6, v2, vcc
-; GCN-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, v5
-; GCN-NEXT:    v_cndmask_b32_e64 v0, v0, v2, s[4:5]
-; GCN-NEXT:    v_cndmask_b32_e32 v2, v4, v7, vcc
-; GCN-NEXT:    v_cndmask_b32_e64 v1, v1, v2, s[4:5]
+; GCN-NEXT:    v_cndmask_b32_e64 v8, 0, -1, s[4:5]
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[4:5], v6, v2
+; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v1, v3
+; GCN-NEXT:    v_cndmask_b32_e64 v9, 0, -1, s[4:5]
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[4:5], v4, v3
+; GCN-NEXT:    v_cndmask_b32_e64 v5, 0, -1, vcc
+; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v0, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v8, v8, v9, s[4:5]
+; GCN-NEXT:    v_sub_i32_e64 v9, s[4:5], v6, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v2, 0, -1, vcc
+; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, v1, v3
+; GCN-NEXT:    v_cndmask_b32_e32 v2, v5, v2, vcc
+; GCN-NEXT:    v_cmp_ne_u32_e64 s[6:7], 0, v8
+; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
+; GCN-NEXT:    v_subbrev_u32_e64 v2, s[4:5], 0, v7, s[4:5]
+; GCN-NEXT:    v_cndmask_b32_e64 v6, v6, v9, s[6:7]
+; GCN-NEXT:    v_cndmask_b32_e64 v2, v4, v2, s[6:7]
+; GCN-NEXT:    v_cndmask_b32_e32 v0, v0, v6, vcc
+; GCN-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc
 ; GCN-NEXT:    s_setpc_b64 s[30:31]
   %result = urem i64 %x, %y
   ret i64 %result
 }
 
 define amdgpu_kernel void @s_test_urem31_i64(i64 addrspace(1)* %out, i64 %x, i64 %y) {
 ; GCN-LABEL: s_test_urem31_i64:
 ; GCN:       ; %bb.0:
-; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GCN-NEXT:    s_load_dword s2, s[0:1], 0xe
-; GCN-NEXT:    s_mov_b32 s11, 0xf000
-; GCN-NEXT:    s_mov_b32 s10, -1
+; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_mov_b32 s8, s4
-; GCN-NEXT:    s_mov_b32 s9, s5
-; GCN-NEXT:    s_lshr_b32 s3, s7, 1
-; GCN-NEXT:    s_lshr_b32 s4, s2, 1
-; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s4
+; GCN-NEXT:    s_mov_b32 s6, -1
+; GCN-NEXT:    s_lshr_b32 s8, s2, 1
+; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s8
+; GCN-NEXT:    s_lshr_b32 s0, s7, 1
+; GCN-NEXT:    s_mov_b32 s7, 0xf000
 ; GCN-NEXT:    v_rcp_iflag_f32_e32 v0, v0
 ; GCN-NEXT:    v_mul_f32_e32 v0, 0x4f800000, v0
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v0
-; GCN-NEXT:    v_mul_hi_u32 v1, v0, s4
-; GCN-NEXT:    v_mul_lo_u32 v2, v0, s4
-; GCN-NEXT:    v_sub_i32_e32 v3, vcc, 0, v2
-; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v1, v2, v3, s[0:1]
+; GCN-NEXT:    v_mul_lo_u32 v1, v0, s8
+; GCN-NEXT:    v_mul_hi_u32 v2, v0, s8
+; GCN-NEXT:    v_sub_i32_e32 v3, vcc, 0, v1
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[2:3], 0, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v1, v1, v3, s[2:3]
 ; GCN-NEXT:    v_mul_hi_u32 v1, v1, v0
 ; GCN-NEXT:    v_add_i32_e32 v2, vcc, v1, v0
 ; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, v1, v0
-; GCN-NEXT:    v_cndmask_b32_e64 v0, v0, v2, s[0:1]
-; GCN-NEXT:    v_mul_hi_u32 v0, v0, s3
-; GCN-NEXT:    v_mul_lo_u32 v0, v0, s4
-; GCN-NEXT:    v_sub_i32_e32 v1, vcc, s3, v0
-; GCN-NEXT:    v_add_i32_e32 v2, vcc, s4, v1
-; GCN-NEXT:    v_cmp_ge_u32_e64 s[0:1], s3, v0
-; GCN-NEXT:    v_cmp_le_u32_e64 s[2:3], s4, v1
-; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, s4, v1
+; GCN-NEXT:    v_cndmask_b32_e64 v0, v0, v2, s[2:3]
+; GCN-NEXT:    v_mul_hi_u32 v0, v0, s0
+; GCN-NEXT:    v_mul_lo_u32 v0, v0, s8
+; GCN-NEXT:    v_sub_i32_e32 v1, vcc, s0, v0
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[0:1], s0, v0
+; GCN-NEXT:    v_add_i32_e32 v2, vcc, s8, v1
+; GCN-NEXT:    v_cmp_le_u32_e64 s[2:3], s8, v1
+; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, s8, v1
 ; GCN-NEXT:    s_and_b64 vcc, s[2:3], s[0:1]
 ; GCN-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc
 ; GCN-NEXT:    v_cndmask_b32_e64 v0, v2, v0, s[0:1]
 ; GCN-NEXT:    v_mov_b32_e32 v1, 0
-; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[8:11], 0
+; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
 ; GCN-NEXT:    s_endpgm
   %1 = lshr i64 %x, 33
   %2 = lshr i64 %y, 33
   %result = urem i64 %1, %2
   store i64 %result, i64 addrspace(1)* %out
   ret void
 }
 
 define amdgpu_kernel void @s_test_urem31_v2i64(<2 x i64> addrspace(1)* %out, <2 x i64> %x, <2 x i64> %y) {
 ; GCN-LABEL: s_test_urem31_v2i64:
 ; GCN:       ; %bb.0:
-; GCN-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x9
-; GCN-NEXT:    s_load_dwordx4 s[12:15], s[0:1], 0xd
 ; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x11
-; GCN-NEXT:    s_mov_b32 s11, 0xf000
-; GCN-NEXT:    s_mov_b32 s10, -1
-; GCN-NEXT:    v_mov_b32_e32 v1, 0
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_lshr_b32 s4, s13, 1
-; GCN-NEXT:    s_lshr_b32 s6, s15, 1
-; GCN-NEXT:    s_lshr_b32 s12, s5, 1
-; GCN-NEXT:    s_lshr_b32 s5, s7, 1
-; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s5
-; GCN-NEXT:    v_cvt_f32_u32_e32 v2, s12
+; GCN-NEXT:    s_mov_b32 s6, -1
+; GCN-NEXT:    s_lshr_b32 s9, s7, 1
+; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s9
+; GCN-NEXT:    s_lshr_b32 s10, s5, 1
+; GCN-NEXT:    v_cvt_f32_u32_e32 v1, s10
+; GCN-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GCN-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0xd
 ; GCN-NEXT:    v_rcp_iflag_f32_e32 v0, v0
-; GCN-NEXT:    v_rcp_iflag_f32_e32 v2, v2
+; GCN-NEXT:    s_mov_b32 s7, 0xf000
+; GCN-NEXT:    v_rcp_iflag_f32_e32 v1, v1
 ; GCN-NEXT:    v_mul_f32_e32 v0, 0x4f800000, v0
-; GCN-NEXT:    v_mul_f32_e32 v2, 0x4f800000, v2
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v0
-; GCN-NEXT:    v_cvt_u32_f32_e32 v2, v2
-; GCN-NEXT:    v_mul_hi_u32 v3, v0, s5
-; GCN-NEXT:    v_mul_lo_u32 v4, v0, s5
-; GCN-NEXT:    v_mul_hi_u32 v5, v2, s12
-; GCN-NEXT:    v_mul_lo_u32 v6, v2, s12
-; GCN-NEXT:    v_sub_i32_e32 v7, vcc, 0, v4
-; GCN-NEXT:    v_sub_i32_e32 v8, vcc, 0, v6
+; GCN-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN-NEXT:    s_lshr_b32 s8, s1, 1
+; GCN-NEXT:    v_mul_f32_e32 v1, 0x4f800000, v1
+; GCN-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GCN-NEXT:    v_mul_lo_u32 v2, v0, s9
+; GCN-NEXT:    v_mul_hi_u32 v3, v0, s9
+; GCN-NEXT:    s_lshr_b32 s2, s3, 1
+; GCN-NEXT:    v_sub_i32_e32 v4, vcc, 0, v2
 ; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v3
-; GCN-NEXT:    v_cndmask_b32_e64 v3, v4, v7, s[0:1]
-; GCN-NEXT:    v_cmp_eq_u32_e64 s[2:3], 0, v5
-; GCN-NEXT:    v_cndmask_b32_e64 v4, v6, v8, s[2:3]
-; GCN-NEXT:    v_mul_hi_u32 v3, v3, v0
-; GCN-NEXT:    v_mul_hi_u32 v4, v4, v2
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, v3, v0
-; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, v3, v0
-; GCN-NEXT:    v_add_i32_e32 v3, vcc, v4, v2
-; GCN-NEXT:    v_subrev_i32_e32 v2, vcc, v4, v2
-; GCN-NEXT:    v_cndmask_b32_e64 v0, v0, v5, s[0:1]
-; GCN-NEXT:    v_cndmask_b32_e64 v2, v2, v3, s[2:3]
-; GCN-NEXT:    v_mul_hi_u32 v0, v0, s6
-; GCN-NEXT:    v_mul_hi_u32 v2, v2, s4
-; GCN-NEXT:    v_mul_lo_u32 v0, v0, s5
-; GCN-NEXT:    v_mul_lo_u32 v2, v2, s12
-; GCN-NEXT:    v_sub_i32_e32 v3, vcc, s6, v0
-; GCN-NEXT:    v_sub_i32_e32 v4, vcc, s4, v2
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, s5, v3
-; GCN-NEXT:    v_cmp_ge_u32_e64 s[0:1], s6, v0
-; GCN-NEXT:    v_cmp_le_u32_e64 s[2:3], s5, v3
-; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, s5, v3
-; GCN-NEXT:    v_add_i32_e32 v6, vcc, s12, v4
-; GCN-NEXT:    v_cmp_ge_u32_e64 s[4:5], s4, v2
-; GCN-NEXT:    v_cmp_le_u32_e64 s[6:7], s12, v4
-; GCN-NEXT:    v_subrev_i32_e32 v2, vcc, s12, v4
+; GCN-NEXT:    v_cndmask_b32_e64 v2, v2, v4, s[0:1]
+; GCN-NEXT:    v_mul_hi_u32 v2, v2, v0
+; GCN-NEXT:    v_mul_lo_u32 v3, v1, s10
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v2, v0
+; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, v2, v0
+; GCN-NEXT:    v_mul_hi_u32 v2, v1, s10
+; GCN-NEXT:    v_cndmask_b32_e64 v0, v0, v4, s[0:1]
+; GCN-NEXT:    v_mul_hi_u32 v0, v0, s2
+; GCN-NEXT:    v_sub_i32_e32 v4, vcc, 0, v3
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v2, v3, v4, s[0:1]
+; GCN-NEXT:    v_mul_lo_u32 v0, v0, s9
+; GCN-NEXT:    v_mul_hi_u32 v2, v2, v1
+; GCN-NEXT:    v_sub_i32_e32 v3, vcc, s2, v0
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[2:3], s2, v0
+; GCN-NEXT:    v_add_i32_e32 v0, vcc, v2, v1
+; GCN-NEXT:    v_subrev_i32_e32 v1, vcc, v2, v1
+; GCN-NEXT:    v_cndmask_b32_e64 v0, v1, v0, s[0:1]
+; GCN-NEXT:    v_mul_hi_u32 v0, v0, s8
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, s9, v3
+; GCN-NEXT:    v_cmp_le_u32_e64 s[0:1], s9, v3
+; GCN-NEXT:    v_subrev_i32_e32 v1, vcc, s9, v3
+; GCN-NEXT:    v_mul_lo_u32 v0, v0, s10
+; GCN-NEXT:    s_and_b64 vcc, s[0:1], s[2:3]
+; GCN-NEXT:    v_cndmask_b32_e32 v1, v3, v1, vcc
+; GCN-NEXT:    v_cndmask_b32_e64 v2, v4, v1, s[2:3]
+; GCN-NEXT:    v_sub_i32_e32 v1, vcc, s8, v0
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[0:1], s8, v0
+; GCN-NEXT:    v_add_i32_e32 v3, vcc, s10, v1
+; GCN-NEXT:    v_cmp_le_u32_e64 s[2:3], s10, v1
+; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, s10, v1
 ; GCN-NEXT:    s_and_b64 vcc, s[2:3], s[0:1]
-; GCN-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
-; GCN-NEXT:    s_and_b64 vcc, s[6:7], s[4:5]
-; GCN-NEXT:    v_cndmask_b32_e32 v3, v4, v2, vcc
-; GCN-NEXT:    v_cndmask_b32_e64 v2, v5, v0, s[0:1]
-; GCN-NEXT:    v_cndmask_b32_e64 v0, v6, v3, s[4:5]
+; GCN-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc
+; GCN-NEXT:    v_mov_b32_e32 v1, 0
+; GCN-NEXT:    v_cndmask_b32_e64 v0, v3, v0, s[0:1]
 ; GCN-NEXT:    v_mov_b32_e32 v3, v1
-; GCN-NEXT:    buffer_store_dwordx4 v[0:3], off, s[8:11], 0
+; GCN-NEXT:    buffer_store_dwordx4 v[0:3], off, s[4:7], 0
 ; GCN-NEXT:    s_endpgm
   %1 = lshr <2 x i64> %x, <i64 33, i64 33>
   %2 = lshr <2 x i64> %y, <i64 33, i64 33>
   %result = urem <2 x i64> %1, %2
   store <2 x i64> %result, <2 x i64> addrspace(1)* %out
   ret void
 }
 
 define amdgpu_kernel void @s_test_urem24_i64(i64 addrspace(1)* %out, i64 %x, i64 %y) {
 ; GCN-LABEL: s_test_urem24_i64:
 ; GCN:       ; %bb.0:
-; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GCN-NEXT:    s_load_dword s2, s[0:1], 0xe
-; GCN-NEXT:    s_mov_b32 s11, 0xf000
-; GCN-NEXT:    s_mov_b32 s10, -1
+; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_mov_b32 s8, s4
-; GCN-NEXT:    s_mov_b32 s9, s5
-; GCN-NEXT:    s_lshr_b32 s3, s7, 8
-; GCN-NEXT:    s_lshr_b32 s4, s2, 8
-; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s4
+; GCN-NEXT:    s_mov_b32 s6, -1
+; GCN-NEXT:    s_lshr_b32 s8, s2, 8
+; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s8
+; GCN-NEXT:    s_lshr_b32 s0, s7, 8
+; GCN-NEXT:    s_mov_b32 s7, 0xf000
 ; GCN-NEXT:    v_rcp_iflag_f32_e32 v0, v0
 ; GCN-NEXT:    v_mul_f32_e32 v0, 0x4f800000, v0
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v0
-; GCN-NEXT:    v_mul_hi_u32 v1, v0, s4
-; GCN-NEXT:    v_mul_lo_u32 v2, v0, s4
-; GCN-NEXT:    v_sub_i32_e32 v3, vcc, 0, v2
-; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v1, v2, v3, s[0:1]
+; GCN-NEXT:    v_mul_lo_u32 v1, v0, s8
+; GCN-NEXT:    v_mul_hi_u32 v2, v0, s8
+; GCN-NEXT:    v_sub_i32_e32 v3, vcc, 0, v1
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[2:3], 0, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v1, v1, v3, s[2:3]
 ; GCN-NEXT:    v_mul_hi_u32 v1, v1, v0
 ; GCN-NEXT:    v_add_i32_e32 v2, vcc, v1, v0
 ; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, v1, v0
-; GCN-NEXT:    v_cndmask_b32_e64 v0, v0, v2, s[0:1]
-; GCN-NEXT:    v_mul_hi_u32 v0, v0, s3
-; GCN-NEXT:    v_mul_lo_u32 v0, v0, s4
-; GCN-NEXT:    v_sub_i32_e32 v1, vcc, s3, v0
-; GCN-NEXT:    v_add_i32_e32 v2, vcc, s4, v1
-; GCN-NEXT:    v_cmp_ge_u32_e64 s[0:1], s3, v0
-; GCN-NEXT:    v_cmp_le_u32_e64 s[2:3], s4, v1
-; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, s4, v1
+; GCN-NEXT:    v_cndmask_b32_e64 v0, v0, v2, s[2:3]
+; GCN-NEXT:    v_mul_hi_u32 v0, v0, s0
+; GCN-NEXT:    v_mul_lo_u32 v0, v0, s8
+; GCN-NEXT:    v_sub_i32_e32 v1, vcc, s0, v0
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[0:1], s0, v0
+; GCN-NEXT:    v_add_i32_e32 v2, vcc, s8, v1
+; GCN-NEXT:    v_cmp_le_u32_e64 s[2:3], s8, v1
+; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, s8, v1
 ; GCN-NEXT:    s_and_b64 vcc, s[2:3], s[0:1]
 ; GCN-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc
 ; GCN-NEXT:    v_cndmask_b32_e64 v0, v2, v0, s[0:1]
 ; GCN-NEXT:    v_mov_b32_e32 v1, 0
-; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[8:11], 0
+; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
 ; GCN-NEXT:    s_endpgm
   %1 = lshr i64 %x, 40
   %2 = lshr i64 %y, 40
   %result = urem i64 %1, %2
   store i64 %result, i64 addrspace(1)* %out
   ret void
 }
 
 define amdgpu_kernel void @s_test_urem23_64_v2i64(<2 x i64> addrspace(1)* %out, <2 x i64> %x, <2 x i64> %y) {
 ; GCN-LABEL: s_test_urem23_64_v2i64:
 ; GCN:       ; %bb.0:
+; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x11
+; GCN-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN-NEXT:    s_mov_b32 s6, -1
+; GCN-NEXT:    s_lshr_b32 s10, s5, 1
+; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s10
 ; GCN-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
-; GCN-NEXT:    s_load_dwordx4 s[8:11], s[0:1], 0xd
-; GCN-NEXT:    s_load_dwordx4 s[12:15], s[0:1], 0x11
+; GCN-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0xd
+; GCN-NEXT:    s_lshr_b32 s9, s7, 9
 ; GCN-NEXT:    s_mov_b32 s7, 0xf000
-; GCN-NEXT:    s_mov_b32 s6, -1
-; GCN-NEXT:    v_mov_b32_e32 v1, 0
-; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_lshr_b32 s2, s11, 9
-; GCN-NEXT:    s_lshr_b32 s3, s9, 1
-; GCN-NEXT:    s_lshr_b32 s8, s15, 9
-; GCN-NEXT:    s_lshr_b32 s9, s13, 1
-; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s9
-; GCN-NEXT:    v_cvt_f32_u32_e32 v2, s2
-; GCN-NEXT:    v_cvt_f32_u32_e32 v3, s8
 ; GCN-NEXT:    v_rcp_iflag_f32_e32 v0, v0
-; GCN-NEXT:    v_rcp_iflag_f32_e32 v4, v3
+; GCN-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN-NEXT:    s_lshr_b32 s2, s1, 1
+; GCN-NEXT:    s_lshr_b32 s8, s3, 9
 ; GCN-NEXT:    v_mul_f32_e32 v0, 0x4f800000, v0
-; GCN-NEXT:    v_mul_f32_e32 v4, v2, v4
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v0
-; GCN-NEXT:    v_trunc_f32_e32 v4, v4
-; GCN-NEXT:    v_mul_hi_u32 v5, v0, s9
-; GCN-NEXT:    v_mul_lo_u32 v6, v0, s9
-; GCN-NEXT:    v_mad_f32 v2, -v4, v3, v2
-; GCN-NEXT:    v_cvt_u32_f32_e32 v4, v4
-; GCN-NEXT:    v_sub_i32_e32 v7, vcc, 0, v6
-; GCN-NEXT:    v_cmp_ge_f32_e64 vcc, |v2|, |v3|
-; GCN-NEXT:    v_addc_u32_e32 v2, vcc, 0, v4, vcc
-; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v5
-; GCN-NEXT:    v_cndmask_b32_e64 v3, v6, v7, s[0:1]
-; GCN-NEXT:    v_mul_lo_u32 v2, v2, s8
-; GCN-NEXT:    v_mul_hi_u32 v3, v3, v0
-; GCN-NEXT:    v_sub_i32_e32 v2, vcc, s2, v2
-; GCN-NEXT:    v_add_i32_e32 v4, vcc, v3, v0
-; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, v3, v0
+; GCN-NEXT:    v_mul_lo_u32 v1, v0, s10
+; GCN-NEXT:    v_mul_hi_u32 v2, v0, s10
+; GCN-NEXT:    v_sub_i32_e32 v3, vcc, 0, v1
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v1, v1, v3, s[0:1]
+; GCN-NEXT:    v_mul_hi_u32 v1, v1, v0
+; GCN-NEXT:    v_cvt_f32_u32_e32 v3, s9
+; GCN-NEXT:    v_cvt_f32_u32_e32 v2, s8
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v1, v0
+; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, v1, v0
 ; GCN-NEXT:    v_cndmask_b32_e64 v0, v0, v4, s[0:1]
-; GCN-NEXT:    v_mul_hi_u32 v0, v0, s3
-; GCN-NEXT:    v_mul_lo_u32 v0, v0, s9
-; GCN-NEXT:    v_sub_i32_e32 v3, vcc, s3, v0
-; GCN-NEXT:    v_add_i32_e32 v4, vcc, s9, v3
-; GCN-NEXT:    v_cmp_ge_u32_e64 s[0:1], s3, v0
-; GCN-NEXT:    v_cmp_le_u32_e64 s[2:3], s9, v3
-; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, s9, v3
+; GCN-NEXT:    v_mul_hi_u32 v0, v0, s2
+; GCN-NEXT:    v_rcp_iflag_f32_e32 v1, v3
+; GCN-NEXT:    v_mul_lo_u32 v0, v0, s10
+; GCN-NEXT:    v_mul_f32_e32 v1, v2, v1
+; GCN-NEXT:    v_trunc_f32_e32 v1, v1
+; GCN-NEXT:    v_sub_i32_e32 v4, vcc, s2, v0
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[0:1], s2, v0
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, s10, v4
+; GCN-NEXT:    v_cmp_le_u32_e64 s[2:3], s10, v4
+; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, s10, v4
 ; GCN-NEXT:    s_and_b64 vcc, s[2:3], s[0:1]
-; GCN-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
-; GCN-NEXT:    v_cndmask_b32_e64 v0, v4, v0, s[0:1]
-; GCN-NEXT:    v_and_b32_e32 v2, 0x7fffff, v2
+; GCN-NEXT:    v_cndmask_b32_e32 v0, v4, v0, vcc
+; GCN-NEXT:    v_cvt_u32_f32_e32 v4, v1
+; GCN-NEXT:    v_mad_f32 v1, -v1, v3, v2
+; GCN-NEXT:    v_cmp_ge_f32_e64 vcc, |v1|, |v3|
+; GCN-NEXT:    v_cndmask_b32_e64 v0, v5, v0, s[0:1]
+; GCN-NEXT:    v_addc_u32_e32 v1, vcc, 0, v4, vcc
+; GCN-NEXT:    v_mul_lo_u32 v2, v1, s9
+; GCN-NEXT:    v_mov_b32_e32 v1, 0
 ; GCN-NEXT:    v_mov_b32_e32 v3, v1
+; GCN-NEXT:    v_sub_i32_e32 v2, vcc, s8, v2
+; GCN-NEXT:    v_and_b32_e32 v2, 0x7fffff, v2
 ; GCN-NEXT:    buffer_store_dwordx4 v[0:3], off, s[4:7], 0
 ; GCN-NEXT:    s_endpgm
   %1 = lshr <2 x i64> %x, <i64 33, i64 41>
   %2 = lshr <2 x i64> %y, <i64 33, i64 41>
   %result = urem <2 x i64> %1, %2
   store <2 x i64> %result, <2 x i64> addrspace(1)* %out
   ret void
 }
 
 define amdgpu_kernel void @s_test_urem_k_num_i64(i64 addrspace(1)* %out, i64 %x) {
 ; GCN-LABEL: s_test_urem_k_num_i64:
 ; GCN:       ; %bb.0:
-; GCN-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
-; GCN-NEXT:    s_mov_b32 s7, 0xf000
-; GCN-NEXT:    s_mov_b32 s6, -1
-; GCN-NEXT:    v_mov_b32_e32 v0, 0
-; GCN-NEXT:    v_mov_b32_e32 v1, 0
+; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN-NEXT:    v_mov_b32_e32 v2, 0
+; GCN-NEXT:    s_mov_b32 s11, 0xf000
+; GCN-NEXT:    s_mov_b32 s10, -1
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_mov_b32 s4, s0
-; GCN-NEXT:    s_mov_b32 s5, s1
-; GCN-NEXT:    v_cvt_f32_u32_e32 v2, s2
-; GCN-NEXT:    v_cvt_f32_u32_e32 v3, s3
-; GCN-NEXT:    s_sub_u32 s8, 0, s2
-; GCN-NEXT:    v_mov_b32_e32 v4, s3
-; GCN-NEXT:    v_mac_f32_e32 v2, 0x4f800000, v3
-; GCN-NEXT:    s_subb_u32 s9, 0, s3
-; GCN-NEXT:    v_rcp_f32_e32 v2, v2
-; GCN-NEXT:    v_mul_f32_e32 v2, 0x5f7ffffc, v2
-; GCN-NEXT:    v_mul_f32_e32 v3, 0x2f800000, v2
+; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s6
+; GCN-NEXT:    v_cvt_f32_u32_e32 v1, s7
+; GCN-NEXT:    s_sub_u32 s2, 0, s6
+; GCN-NEXT:    s_subb_u32 s3, 0, s7
+; GCN-NEXT:    s_mov_b32 s8, s4
+; GCN-NEXT:    v_mac_f32_e32 v0, 0x4f800000, v1
+; GCN-NEXT:    v_rcp_f32_e32 v0, v0
+; GCN-NEXT:    v_mov_b32_e32 v1, 0
+; GCN-NEXT:    s_mov_b32 s9, s5
+; GCN-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v0
+; GCN-NEXT:    v_mul_f32_e32 v3, 0x2f800000, v0
 ; GCN-NEXT:    v_trunc_f32_e32 v3, v3
-; GCN-NEXT:    v_mac_f32_e32 v2, 0xcf800000, v3
+; GCN-NEXT:    v_mac_f32_e32 v0, 0xcf800000, v3
+; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v0
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v3, v3
-; GCN-NEXT:    v_cvt_u32_f32_e32 v2, v2
-; GCN-NEXT:    v_mul_lo_u32 v5, s8, v3
-; GCN-NEXT:    v_mul_lo_u32 v6, s9, v2
-; GCN-NEXT:    v_mul_hi_u32 v7, s8, v2
-; GCN-NEXT:    v_mul_lo_u32 v8, s8, v2
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, v7, v5
-; GCN-NEXT:    v_mul_hi_u32 v7, v2, v8
-; GCN-NEXT:    v_mul_hi_u32 v9, v3, v8
-; GCN-NEXT:    v_mul_lo_u32 v8, v3, v8
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v6
-; GCN-NEXT:    v_mul_hi_u32 v6, v2, v5
-; GCN-NEXT:    v_mul_lo_u32 v10, v2, v5
-; GCN-NEXT:    v_mul_hi_u32 v11, v3, v5
-; GCN-NEXT:    v_mul_lo_u32 v5, v3, v5
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v7, v10
-; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v1, v6, vcc
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v8, v7
-; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v6, v9, vcc
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v11, v0, vcc
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, v6, v5
-; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v1, v7, vcc
-; GCN-NEXT:    v_add_i32_e64 v2, s[0:1], v2, v5
-; GCN-NEXT:    v_addc_u32_e64 v5, vcc, v3, v6, s[0:1]
-; GCN-NEXT:    v_mul_hi_u32 v7, s8, v2
-; GCN-NEXT:    v_mul_lo_u32 v8, s9, v2
-; GCN-NEXT:    v_mul_lo_u32 v9, s8, v2
-; GCN-NEXT:    v_mul_lo_u32 v10, s8, v5
-; GCN-NEXT:    v_mul_hi_u32 v11, v5, v9
-; GCN-NEXT:    v_mul_lo_u32 v12, v5, v9
-; GCN-NEXT:    v_mul_hi_u32 v9, v2, v9
+; GCN-NEXT:    v_mul_hi_u32 v5, s2, v0
+; GCN-NEXT:    v_mul_lo_u32 v4, s2, v3
+; GCN-NEXT:    v_mul_lo_u32 v7, s3, v0
+; GCN-NEXT:    v_mul_lo_u32 v6, s2, v0
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v5, v4
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v4, v7
+; GCN-NEXT:    v_mul_hi_u32 v5, v0, v6
+; GCN-NEXT:    v_mul_lo_u32 v7, v0, v4
+; GCN-NEXT:    v_mul_hi_u32 v9, v0, v4
+; GCN-NEXT:    v_mul_lo_u32 v8, v3, v6
+; GCN-NEXT:    v_mul_hi_u32 v6, v3, v6
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v7
+; GCN-NEXT:    v_mul_hi_u32 v10, v3, v4
+; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v2, v9, vcc
+; GCN-NEXT:    v_mul_lo_u32 v4, v3, v4
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v8, v5
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v7, v6, vcc
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v10, v1, vcc
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v5, v4
+; GCN-NEXT:    v_add_i32_e64 v0, s[0:1], v0, v4
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v2, v6, vcc
+; GCN-NEXT:    v_addc_u32_e64 v4, vcc, v3, v5, s[0:1]
+; GCN-NEXT:    v_mul_lo_u32 v6, s2, v4
+; GCN-NEXT:    v_mul_hi_u32 v7, s2, v0
+; GCN-NEXT:    v_mul_lo_u32 v8, s3, v0
+; GCN-NEXT:    v_add_i32_e32 v6, vcc, v7, v6
+; GCN-NEXT:    v_mul_lo_u32 v7, s2, v0
+; GCN-NEXT:    v_add_i32_e32 v6, vcc, v8, v6
+; GCN-NEXT:    v_mul_lo_u32 v10, v0, v6
+; GCN-NEXT:    v_mul_hi_u32 v12, v0, v6
+; GCN-NEXT:    v_mul_hi_u32 v11, v0, v7
+; GCN-NEXT:    v_mul_hi_u32 v9, v4, v7
+; GCN-NEXT:    v_mul_lo_u32 v7, v4, v7
+; GCN-NEXT:    v_mul_hi_u32 v8, v4, v6
+; GCN-NEXT:    v_add_i32_e32 v10, vcc, v11, v10
+; GCN-NEXT:    v_addc_u32_e32 v11, vcc, v2, v12, vcc
+; GCN-NEXT:    v_mul_lo_u32 v4, v4, v6
 ; GCN-NEXT:    v_add_i32_e32 v7, vcc, v7, v10
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v8, v7
-; GCN-NEXT:    v_mul_hi_u32 v8, v5, v7
-; GCN-NEXT:    v_mul_hi_u32 v10, v2, v7
-; GCN-NEXT:    v_mul_lo_u32 v13, v2, v7
-; GCN-NEXT:    v_mul_lo_u32 v5, v5, v7
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v9, v13
-; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v1, v10, vcc
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v12, v7
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v9, v11, vcc
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v8, v0, vcc
-; GCN-NEXT:    v_add_i32_e32 v3, vcc, v3, v6
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, v7, v5
-; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v1, v8, vcc
+; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v11, v9, vcc
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v8, v1, vcc
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v7, v4
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v2, v6, vcc
+; GCN-NEXT:    v_add_i32_e32 v3, vcc, v3, v5
 ; GCN-NEXT:    v_addc_u32_e64 v3, vcc, v3, v6, s[0:1]
-; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v5
+; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v4
 ; GCN-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; GCN-NEXT:    v_mul_hi_u32 v2, v2, 24
-; GCN-NEXT:    v_mul_hi_u32 v5, v3, 24
-; GCN-NEXT:    v_mul_lo_u32 v3, v3, 24
-; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v3
-; GCN-NEXT:    v_addc_u32_e32 v1, vcc, v1, v5, vcc
-; GCN-NEXT:    v_addc_u32_e32 v0, vcc, 0, v0, vcc
-; GCN-NEXT:    v_mul_hi_u32 v2, s2, v1
-; GCN-NEXT:    v_mul_lo_u32 v3, s3, v1
-; GCN-NEXT:    v_mul_lo_u32 v1, s2, v1
-; GCN-NEXT:    v_mul_lo_u32 v0, s2, v0
-; GCN-NEXT:    v_add_i32_e32 v0, vcc, v2, v0
-; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v3
-; GCN-NEXT:    v_sub_i32_e32 v2, vcc, 0, v0
-; GCN-NEXT:    v_sub_i32_e32 v3, vcc, 24, v1
-; GCN-NEXT:    v_subb_u32_e64 v1, s[0:1], v2, v4, vcc
-; GCN-NEXT:    v_subb_u32_e32 v0, vcc, 0, v0, vcc
-; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s2, v3
-; GCN-NEXT:    v_cndmask_b32_e64 v2, 0, -1, vcc
-; GCN-NEXT:    v_subrev_i32_e32 v5, vcc, s2, v3
-; GCN-NEXT:    v_subb_u32_e64 v4, s[0:1], v1, v4, vcc
-; GCN-NEXT:    v_subbrev_u32_e32 v1, vcc, 0, v1, vcc
-; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s2, v5
-; GCN-NEXT:    v_cndmask_b32_e64 v6, 0, -1, vcc
-; GCN-NEXT:    v_subrev_i32_e32 v7, vcc, s2, v5
-; GCN-NEXT:    v_cmp_le_u32_e64 s[0:1], s3, v0
-; GCN-NEXT:    v_cndmask_b32_e64 v8, 0, -1, s[0:1]
-; GCN-NEXT:    v_cmp_le_u32_e64 s[0:1], s3, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v9, 0, -1, s[0:1]
-; GCN-NEXT:    v_subbrev_u32_e32 v4, vcc, 0, v4, vcc
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, s3, v0
-; GCN-NEXT:    v_cndmask_b32_e32 v2, v8, v2, vcc
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, s3, v1
-; GCN-NEXT:    v_cndmask_b32_e32 v6, v9, v6, vcc
-; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v6
-; GCN-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
-; GCN-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v2
-; GCN-NEXT:    v_cndmask_b32_e64 v1, v0, v1, s[0:1]
-; GCN-NEXT:    v_cndmask_b32_e32 v0, v5, v7, vcc
-; GCN-NEXT:    v_cndmask_b32_e64 v0, v3, v0, s[0:1]
-; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
+; GCN-NEXT:    v_mul_lo_u32 v4, v3, 24
+; GCN-NEXT:    v_mul_hi_u32 v0, v0, 24
+; GCN-NEXT:    v_mul_hi_u32 v3, v3, 24
+; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v4
+; GCN-NEXT:    v_addc_u32_e32 v0, vcc, v2, v3, vcc
+; GCN-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GCN-NEXT:    v_mul_lo_u32 v1, s6, v1
+; GCN-NEXT:    v_mul_hi_u32 v2, s6, v0
+; GCN-NEXT:    v_mul_lo_u32 v3, s7, v0
+; GCN-NEXT:    v_mul_lo_u32 v0, s6, v0
+; GCN-NEXT:    v_add_i32_e32 v1, vcc, v2, v1
+; GCN-NEXT:    v_add_i32_e32 v1, vcc, v1, v3
+; GCN-NEXT:    v_sub_i32_e32 v2, vcc, 0, v1
+; GCN-NEXT:    v_mov_b32_e32 v3, s7
+; GCN-NEXT:    v_sub_i32_e32 v0, vcc, 24, v0
+; GCN-NEXT:    v_subb_u32_e64 v2, s[0:1], v2, v3, vcc
+; GCN-NEXT:    v_subrev_i32_e64 v4, s[0:1], s6, v0
+; GCN-NEXT:    v_subb_u32_e64 v3, s[2:3], v2, v3, s[0:1]
+; GCN-NEXT:    v_subbrev_u32_e64 v2, s[0:1], 0, v2, s[0:1]
+; GCN-NEXT:    v_cmp_le_u32_e64 s[0:1], s7, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v5, 0, -1, s[0:1]
+; GCN-NEXT:    v_cmp_le_u32_e64 s[0:1], s6, v4
+; GCN-NEXT:    v_cndmask_b32_e64 v6, 0, -1, s[0:1]
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], s7, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v5, v5, v6, s[0:1]
+; GCN-NEXT:    v_subrev_i32_e64 v6, s[0:1], s6, v4
+; GCN-NEXT:    v_subb_u32_e32 v1, vcc, 0, v1, vcc
+; GCN-NEXT:    v_subbrev_u32_e64 v3, s[0:1], 0, v3, s[0:1]
+; GCN-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v5
+; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s7, v1
+; GCN-NEXT:    v_cndmask_b32_e64 v2, v2, v3, s[0:1]
+; GCN-NEXT:    v_cndmask_b32_e64 v3, 0, -1, vcc
+; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s6, v0
+; GCN-NEXT:    v_cndmask_b32_e64 v5, 0, -1, vcc
+; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, s7, v1
+; GCN-NEXT:    v_cndmask_b32_e32 v3, v3, v5, vcc
+; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v3
+; GCN-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc
+; GCN-NEXT:    v_cndmask_b32_e64 v2, v4, v6, s[0:1]
+; GCN-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
+; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[8:11], 0
 ; GCN-NEXT:    s_endpgm
   %result = urem i64 24, %x
   store i64 %result, i64 addrspace(1)* %out
   ret void
 }
 
 define amdgpu_kernel void @s_test_urem_k_den_i64(i64 addrspace(1)* %out, i64 %x) {
 ; GCN-LABEL: s_test_urem_k_den_i64:
 ; GCN:       ; %bb.0:
-; GCN-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
-; GCN-NEXT:    s_mov_b32 s7, 0xf000
-; GCN-NEXT:    s_mov_b32 s6, -1
 ; GCN-NEXT:    v_mov_b32_e32 v0, 0x4f800000
-; GCN-NEXT:    s_movk_i32 s8, 0xffe8
-; GCN-NEXT:    v_mov_b32_e32 v1, 0
-; GCN-NEXT:    v_mov_b32_e32 v2, 0
 ; GCN-NEXT:    v_madak_f32 v0, 0, v0, 0x41c00000
 ; GCN-NEXT:    v_rcp_f32_e32 v0, v0
+; GCN-NEXT:    s_movk_i32 s2, 0xffe8
+; GCN-NEXT:    v_mov_b32_e32 v8, 0
+; GCN-NEXT:    v_mov_b32_e32 v7, 0
 ; GCN-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v0
-; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_mov_b32 s4, s0
-; GCN-NEXT:    s_mov_b32 s5, s1
-; GCN-NEXT:    v_mul_f32_e32 v3, 0x2f800000, v0
-; GCN-NEXT:    v_mov_b32_e32 v4, s3
-; GCN-NEXT:    v_trunc_f32_e32 v3, v3
-; GCN-NEXT:    v_mac_f32_e32 v0, 0xcf800000, v3
-; GCN-NEXT:    v_cvt_u32_f32_e32 v3, v3
+; GCN-NEXT:    v_mul_f32_e32 v1, 0x2f800000, v0
+; GCN-NEXT:    v_trunc_f32_e32 v1, v1
+; GCN-NEXT:    v_mac_f32_e32 v0, 0xcf800000, v1
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v0
-; GCN-NEXT:    v_mul_lo_u32 v5, v3, s8
-; GCN-NEXT:    v_mul_hi_u32 v6, v0, s8
-; GCN-NEXT:    v_mul_lo_u32 v7, v0, s8
-; GCN-NEXT:    v_subrev_i32_e32 v6, vcc, v0, v6
-; GCN-NEXT:    v_mul_hi_u32 v8, v0, v7
-; GCN-NEXT:    v_mul_hi_u32 v9, v3, v7
-; GCN-NEXT:    v_mul_lo_u32 v7, v3, v7
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v6
-; GCN-NEXT:    v_mul_hi_u32 v6, v0, v5
-; GCN-NEXT:    v_mul_lo_u32 v10, v0, v5
-; GCN-NEXT:    v_mul_hi_u32 v11, v3, v5
-; GCN-NEXT:    v_mul_lo_u32 v5, v3, v5
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v8, v10
-; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v2, v6, vcc
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v7, v8
-; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v6, v9, vcc
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v11, v1, vcc
+; GCN-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN-NEXT:    s_mov_b32 s11, 0xf000
+; GCN-NEXT:    v_mul_hi_u32 v2, v0, s2
+; GCN-NEXT:    v_mul_lo_u32 v3, v1, s2
+; GCN-NEXT:    v_mul_lo_u32 v4, v0, s2
+; GCN-NEXT:    s_mov_b32 s10, -1
+; GCN-NEXT:    v_subrev_i32_e32 v2, vcc, v0, v2
+; GCN-NEXT:    v_add_i32_e32 v2, vcc, v3, v2
+; GCN-NEXT:    v_mul_lo_u32 v5, v0, v2
+; GCN-NEXT:    v_mul_hi_u32 v6, v0, v4
+; GCN-NEXT:    v_mul_hi_u32 v3, v0, v2
+; GCN-NEXT:    v_mul_hi_u32 v9, v1, v2
+; GCN-NEXT:    v_mul_lo_u32 v2, v1, v2
 ; GCN-NEXT:    v_add_i32_e32 v5, vcc, v6, v5
-; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v2, v7, vcc
-; GCN-NEXT:    v_add_i32_e64 v0, s[0:1], v0, v5
-; GCN-NEXT:    v_addc_u32_e64 v5, vcc, v3, v6, s[0:1]
-; GCN-NEXT:    v_mul_hi_u32 v7, v0, s8
-; GCN-NEXT:    v_mul_lo_u32 v8, v0, s8
-; GCN-NEXT:    v_mul_lo_u32 v9, v5, s8
-; GCN-NEXT:    v_subrev_i32_e32 v7, vcc, v0, v7
-; GCN-NEXT:    v_mul_hi_u32 v10, v0, v8
-; GCN-NEXT:    v_mul_hi_u32 v11, v5, v8
-; GCN-NEXT:    v_mul_lo_u32 v8, v5, v8
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v7, v9
-; GCN-NEXT:    v_mul_hi_u32 v9, v0, v7
-; GCN-NEXT:    v_mul_lo_u32 v12, v0, v7
-; GCN-NEXT:    v_mul_hi_u32 v13, v5, v7
-; GCN-NEXT:    v_mul_lo_u32 v5, v5, v7
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v10, v12
-; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v2, v9, vcc
-; GCN-NEXT:    v_add_i32_e32 v7, vcc, v8, v7
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v9, v11, vcc
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v13, v1, vcc
-; GCN-NEXT:    v_add_i32_e32 v3, vcc, v3, v6
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, v7, v5
-; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v2, v8, vcc
-; GCN-NEXT:    v_addc_u32_e64 v3, vcc, v3, v6, s[0:1]
-; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v5
-; GCN-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
-; GCN-NEXT:    v_mul_hi_u32 v5, s2, v0
-; GCN-NEXT:    v_mul_hi_u32 v6, s3, v0
-; GCN-NEXT:    v_mul_lo_u32 v0, s3, v0
-; GCN-NEXT:    v_mul_hi_u32 v7, s2, v3
-; GCN-NEXT:    v_mul_lo_u32 v8, s2, v3
-; GCN-NEXT:    v_mul_hi_u32 v9, s3, v3
-; GCN-NEXT:    v_mul_lo_u32 v3, s3, v3
-; GCN-NEXT:    v_add_i32_e32 v5, vcc, v5, v8
-; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v2, v7, vcc
-; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v5
-; GCN-NEXT:    v_addc_u32_e32 v0, vcc, v7, v6, vcc
-; GCN-NEXT:    v_addc_u32_e32 v1, vcc, v9, v1, vcc
-; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v3
-; GCN-NEXT:    v_addc_u32_e32 v1, vcc, v2, v1, vcc
+; GCN-NEXT:    v_mul_lo_u32 v6, v1, v4
+; GCN-NEXT:    v_mul_hi_u32 v4, v1, v4
+; GCN-NEXT:    v_addc_u32_e32 v3, vcc, v8, v3, vcc
+; GCN-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN-NEXT:    s_mov_b32 s8, s4
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v6, v5
+; GCN-NEXT:    v_addc_u32_e32 v3, vcc, v3, v4, vcc
+; GCN-NEXT:    v_addc_u32_e32 v4, vcc, v9, v7, vcc
+; GCN-NEXT:    v_add_i32_e32 v2, vcc, v3, v2
+; GCN-NEXT:    v_add_i32_e64 v0, s[0:1], v0, v2
+; GCN-NEXT:    v_addc_u32_e32 v3, vcc, v8, v4, vcc
+; GCN-NEXT:    v_mul_hi_u32 v4, v0, s2
+; GCN-NEXT:    v_addc_u32_e64 v2, vcc, v1, v3, s[0:1]
+; GCN-NEXT:    v_mul_lo_u32 v5, v2, s2
+; GCN-NEXT:    v_mul_lo_u32 v6, v0, s2
+; GCN-NEXT:    v_subrev_i32_e32 v4, vcc, v0, v4
+; GCN-NEXT:    s_mov_b32 s9, s5
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v4, v5
+; GCN-NEXT:    v_mul_lo_u32 v5, v0, v4
+; GCN-NEXT:    v_mul_hi_u32 v9, v0, v6
+; GCN-NEXT:    v_mul_hi_u32 v10, v0, v4
+; GCN-NEXT:    v_mul_hi_u32 v11, v2, v4
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v9, v5
+; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v8, v10, vcc
+; GCN-NEXT:    v_mul_lo_u32 v10, v2, v6
+; GCN-NEXT:    v_mul_hi_u32 v6, v2, v6
+; GCN-NEXT:    v_mul_lo_u32 v2, v2, v4
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v10, v5
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v9, v6, vcc
+; GCN-NEXT:    v_addc_u32_e32 v4, vcc, v11, v7, vcc
+; GCN-NEXT:    v_add_i32_e32 v2, vcc, v5, v2
+; GCN-NEXT:    v_addc_u32_e32 v4, vcc, v8, v4, vcc
+; GCN-NEXT:    v_add_i32_e32 v1, vcc, v1, v3
+; GCN-NEXT:    v_addc_u32_e64 v1, vcc, v1, v4, s[0:1]
+; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v2
+; GCN-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GCN-NEXT:    v_mul_lo_u32 v2, s6, v1
+; GCN-NEXT:    v_mul_hi_u32 v3, s6, v0
+; GCN-NEXT:    v_mul_hi_u32 v4, s6, v1
+; GCN-NEXT:    v_mul_hi_u32 v5, s7, v1
+; GCN-NEXT:    v_mul_lo_u32 v1, s7, v1
+; GCN-NEXT:    v_add_i32_e32 v2, vcc, v3, v2
+; GCN-NEXT:    v_addc_u32_e32 v3, vcc, v8, v4, vcc
+; GCN-NEXT:    v_mul_lo_u32 v4, s7, v0
+; GCN-NEXT:    v_mul_hi_u32 v0, s7, v0
+; GCN-NEXT:    v_add_i32_e32 v2, vcc, v4, v2
+; GCN-NEXT:    v_addc_u32_e32 v0, vcc, v3, v0, vcc
+; GCN-NEXT:    v_addc_u32_e32 v2, vcc, v5, v7, vcc
+; GCN-NEXT:    v_add_i32_e32 v0, vcc, v0, v1
+; GCN-NEXT:    v_addc_u32_e32 v1, vcc, v8, v2, vcc
 ; GCN-NEXT:    v_mul_hi_u32 v2, v0, 24
-; GCN-NEXT:    v_mul_lo_u32 v0, v0, 24
 ; GCN-NEXT:    v_mul_lo_u32 v1, v1, 24
+; GCN-NEXT:    v_mul_lo_u32 v0, v0, 24
 ; GCN-NEXT:    v_add_i32_e32 v1, vcc, v2, v1
-; GCN-NEXT:    v_sub_i32_e32 v0, vcc, s2, v0
-; GCN-NEXT:    v_subb_u32_e32 v1, vcc, v4, v1, vcc
+; GCN-NEXT:    v_sub_i32_e32 v0, vcc, s6, v0
+; GCN-NEXT:    v_mov_b32_e32 v2, s7
+; GCN-NEXT:    v_subb_u32_e32 v1, vcc, v2, v1, vcc
 ; GCN-NEXT:    v_subrev_i32_e32 v2, vcc, 24, v0
-; GCN-NEXT:    v_cmp_lt_u32_e64 s[0:1], 23, v0
-; GCN-NEXT:    v_cndmask_b32_e64 v3, 0, -1, s[0:1]
-; GCN-NEXT:    v_subbrev_u32_e32 v4, vcc, 0, v1, vcc
-; GCN-NEXT:    v_subrev_i32_e32 v5, vcc, 24, v2
-; GCN-NEXT:    v_cmp_lt_u32_e64 s[0:1], 23, v2
-; GCN-NEXT:    v_cndmask_b32_e64 v6, 0, -1, s[0:1]
-; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v3, -1, v3, s[0:1]
-; GCN-NEXT:    v_subbrev_u32_e32 v7, vcc, 0, v4, vcc
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v4
+; GCN-NEXT:    v_subbrev_u32_e32 v3, vcc, 0, v1, vcc
+; GCN-NEXT:    v_subrev_i32_e32 v4, vcc, 24, v2
+; GCN-NEXT:    v_subbrev_u32_e32 v5, vcc, 0, v3, vcc
+; GCN-NEXT:    v_cmp_lt_u32_e32 vcc, 23, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v6, 0, -1, vcc
+; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v3
 ; GCN-NEXT:    v_cndmask_b32_e32 v6, -1, v6, vcc
 ; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v6
-; GCN-NEXT:    v_cndmask_b32_e32 v4, v4, v7, vcc
-; GCN-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v3
-; GCN-NEXT:    v_cndmask_b32_e64 v1, v1, v4, s[0:1]
-; GCN-NEXT:    v_cndmask_b32_e32 v2, v2, v5, vcc
+; GCN-NEXT:    v_cmp_lt_u32_e64 s[0:1], 23, v0
+; GCN-NEXT:    v_cndmask_b32_e32 v3, v3, v5, vcc
+; GCN-NEXT:    v_cndmask_b32_e64 v5, 0, -1, s[0:1]
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v1
+; GCN-NEXT:    v_cndmask_b32_e64 v5, -1, v5, s[0:1]
+; GCN-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v5
+; GCN-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
+; GCN-NEXT:    v_cndmask_b32_e64 v1, v1, v3, s[0:1]
 ; GCN-NEXT:    v_cndmask_b32_e64 v0, v0, v2, s[0:1]
-; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
+; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[8:11], 0
 ; GCN-NEXT:    s_endpgm
   %result = urem i64 %x, 24
   store i64 %result, i64 addrspace(1)* %out
   ret void
 }
 
 ; FIXME: Constant bus violation
 ; define i64 @v_test_urem_k_num_i64(i64 %x) {
 ;   %result = urem i64 24, %x
 ;   ret i64 %result
 ; }
 
 define i64 @v_test_urem_pow2_k_num_i64(i64 %x) {
 ; GCN-LABEL: v_test_urem_pow2_k_num_i64:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GCN-NEXT:    v_cvt_f32_u32_e32 v2, v0
 ; GCN-NEXT:    v_cvt_f32_u32_e32 v3, v1
 ; GCN-NEXT:    v_sub_i32_e32 v4, vcc, 0, v0
 ; GCN-NEXT:    v_subb_u32_e32 v5, vcc, 0, v1, vcc
-; GCN-NEXT:    v_mov_b32_e32 v6, 0
-; GCN-NEXT:    v_mov_b32_e32 v7, 0
-; GCN-NEXT:    s_mov_b32 s6, 0x8000
+; GCN-NEXT:    v_mov_b32_e32 v12, 0
 ; GCN-NEXT:    v_mac_f32_e32 v2, 0x4f800000, v3
 ; GCN-NEXT:    v_rcp_f32_e32 v2, v2
+; GCN-NEXT:    v_mov_b32_e32 v11, 0
 ; GCN-NEXT:    v_mul_f32_e32 v2, 0x5f7ffffc, v2
 ; GCN-NEXT:    v_mul_f32_e32 v3, 0x2f800000, v2
 ; GCN-NEXT:    v_trunc_f32_e32 v3, v3
 ; GCN-NEXT:    v_mac_f32_e32 v2, 0xcf800000, v3
-; GCN-NEXT:    v_cvt_u32_f32_e32 v3, v3
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v2, v2
-; GCN-NEXT:    v_mul_lo_u32 v8, v4, v3
-; GCN-NEXT:    v_mul_lo_u32 v9, v5, v2
-; GCN-NEXT:    v_mul_hi_u32 v10, v4, v2
-; GCN-NEXT:    v_mul_lo_u32 v11, v4, v2
+; GCN-NEXT:    v_cvt_u32_f32_e32 v3, v3
+; GCN-NEXT:    v_mul_hi_u32 v7, v4, v2
+; GCN-NEXT:    v_mul_lo_u32 v6, v4, v3
+; GCN-NEXT:    v_mul_lo_u32 v8, v5, v2
+; GCN-NEXT:    v_mul_lo_u32 v9, v4, v2
+; GCN-NEXT:    v_add_i32_e32 v6, vcc, v7, v6
+; GCN-NEXT:    v_add_i32_e32 v6, vcc, v6, v8
+; GCN-NEXT:    v_mul_lo_u32 v8, v2, v6
+; GCN-NEXT:    v_mul_hi_u32 v10, v2, v9
+; GCN-NEXT:    v_mul_hi_u32 v7, v2, v6
+; GCN-NEXT:    v_mul_hi_u32 v13, v3, v6
+; GCN-NEXT:    v_mul_lo_u32 v6, v3, v6
 ; GCN-NEXT:    v_add_i32_e32 v8, vcc, v10, v8
-; GCN-NEXT:    v_mul_hi_u32 v10, v2, v11
-; GCN-NEXT:    v_mul_hi_u32 v12, v3, v11
-; GCN-NEXT:    v_mul_lo_u32 v11, v3, v11
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v8, v9
-; GCN-NEXT:    v_mul_hi_u32 v9, v2, v8
-; GCN-NEXT:    v_mul_lo_u32 v13, v2, v8
-; GCN-NEXT:    v_mul_hi_u32 v14, v3, v8
-; GCN-NEXT:    v_mul_lo_u32 v8, v3, v8
-; GCN-NEXT:    v_add_i32_e32 v10, vcc, v10, v13
-; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v7, v9, vcc
-; GCN-NEXT:    v_add_i32_e32 v10, vcc, v11, v10
-; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v9, v12, vcc
-; GCN-NEXT:    v_addc_u32_e32 v10, vcc, v14, v6, vcc
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v9, v8
-; GCN-NEXT:    v_addc_u32_e32 v9, vcc, v7, v10, vcc
-; GCN-NEXT:    v_add_i32_e64 v2, s[4:5], v2, v8
-; GCN-NEXT:    v_addc_u32_e64 v8, vcc, v3, v9, s[4:5]
-; GCN-NEXT:    v_mul_hi_u32 v10, v4, v2
+; GCN-NEXT:    v_mul_lo_u32 v10, v3, v9
+; GCN-NEXT:    v_mul_hi_u32 v9, v3, v9
+; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v12, v7, vcc
+; GCN-NEXT:    v_add_i32_e32 v8, vcc, v10, v8
+; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v7, v9, vcc
+; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v13, v11, vcc
+; GCN-NEXT:    v_add_i32_e32 v6, vcc, v7, v6
+; GCN-NEXT:    v_add_i32_e64 v2, s[4:5], v2, v6
+; GCN-NEXT:    v_addc_u32_e32 v7, vcc, v12, v8, vcc
+; GCN-NEXT:    v_addc_u32_e64 v6, vcc, v3, v7, s[4:5]
+; GCN-NEXT:    v_mul_lo_u32 v8, v4, v6
+; GCN-NEXT:    v_mul_hi_u32 v9, v4, v2
 ; GCN-NEXT:    v_mul_lo_u32 v5, v5, v2
-; GCN-NEXT:    v_mul_lo_u32 v11, v4, v2
-; GCN-NEXT:    v_add_i32_e32 v3, vcc, v3, v9
-; GCN-NEXT:    v_mul_lo_u32 v4, v4, v8
-; GCN-NEXT:    v_mul_hi_u32 v9, v8, v11
-; GCN-NEXT:    v_mul_lo_u32 v12, v8, v11
-; GCN-NEXT:    v_mul_hi_u32 v11, v2, v11
-; GCN-NEXT:    v_add_i32_e32 v4, vcc, v10, v4
+; GCN-NEXT:    v_mul_lo_u32 v4, v4, v2
+; GCN-NEXT:    v_add_i32_e32 v8, vcc, v9, v8
+; GCN-NEXT:    v_add_i32_e32 v5, vcc, v8, v5
+; GCN-NEXT:    v_mul_lo_u32 v10, v2, v5
+; GCN-NEXT:    v_mul_hi_u32 v13, v2, v4
+; GCN-NEXT:    v_mul_hi_u32 v14, v2, v5
+; GCN-NEXT:    v_mul_hi_u32 v9, v6, v4
+; GCN-NEXT:    v_mul_lo_u32 v4, v6, v4
+; GCN-NEXT:    v_add_i32_e32 v10, vcc, v13, v10
+; GCN-NEXT:    v_mul_hi_u32 v8, v6, v5
+; GCN-NEXT:    v_addc_u32_e32 v13, vcc, v12, v14, vcc
+; GCN-NEXT:    v_mul_lo_u32 v5, v6, v5
+; GCN-NEXT:    v_add_i32_e32 v4, vcc, v4, v10
+; GCN-NEXT:    v_addc_u32_e32 v4, vcc, v13, v9, vcc
+; GCN-NEXT:    v_addc_u32_e32 v6, vcc, v8, v11, vcc
 ; GCN-NEXT:    v_add_i32_e32 v4, vcc, v4, v5
-; GCN-NEXT:    v_mul_hi_u32 v5, v8, v4
-; GCN-NEXT:    v_mul_hi_u32 v10, v2, v4
-; GCN-NEXT:    v_mul_lo_u32 v13, v2, v4
-; GCN-NEXT:    v_mul_lo_u32 v4, v8, v4
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v11, v13
-; GCN-NEXT:    v_addc_u32_e32 v10, vcc, v7, v10, vcc
-; GCN-NEXT:    v_add_i32_e32 v8, vcc, v12, v8
-; GCN-NEXT:    v_addc_u32_e32 v8, vcc, v10, v9, vcc
-; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v5, v6, vcc
-; GCN-NEXT:    v_add_i32_e32 v4, vcc, v8, v4
-; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v7, v5, vcc
+; GCN-NEXT:    v_addc_u32_e32 v5, vcc, v12, v6, vcc
+; GCN-NEXT:    v_add_i32_e32 v3, vcc, v3, v7
 ; GCN-NEXT:    v_addc_u32_e64 v3, vcc, v3, v5, s[4:5]
 ; GCN-NEXT:    v_add_i32_e32 v2, vcc, v2, v4
 ; GCN-NEXT:    v_addc_u32_e32 v2, vcc, 0, v3, vcc
 ; GCN-NEXT:    v_lshrrev_b32_e32 v2, 17, v2
 ; GCN-NEXT:    v_mul_lo_u32 v3, v1, v2
 ; GCN-NEXT:    v_mul_hi_u32 v4, v0, v2
 ; GCN-NEXT:    v_mul_lo_u32 v2, v0, v2
+; GCN-NEXT:    s_mov_b32 s4, 0x8000
 ; GCN-NEXT:    v_add_i32_e32 v3, vcc, v4, v3
-; GCN-NEXT:    v_sub_i32_e32 v2, vcc, s6, v2
-; GCN-NEXT:    v_sub_i32_e64 v4, s[4:5], 0, v3
-; GCN-NEXT:    v_cmp_ge_u32_e64 s[4:5], v2, v0
-; GCN-NEXT:    v_cndmask_b32_e64 v5, 0, -1, s[4:5]
+; GCN-NEXT:    v_sub_i32_e32 v4, vcc, 0, v3
+; GCN-NEXT:    v_sub_i32_e32 v2, vcc, s4, v2
 ; GCN-NEXT:    v_subb_u32_e64 v4, s[4:5], v4, v1, vcc
-; GCN-NEXT:    v_subb_u32_e32 v3, vcc, 0, v3, vcc
-; GCN-NEXT:    v_sub_i32_e32 v6, vcc, v2, v0
-; GCN-NEXT:    v_subb_u32_e64 v7, s[4:5], v4, v1, vcc
-; GCN-NEXT:    v_subbrev_u32_e32 v4, vcc, 0, v4, vcc
-; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v6, v0
-; GCN-NEXT:    v_cndmask_b32_e64 v8, 0, -1, vcc
-; GCN-NEXT:    v_sub_i32_e32 v0, vcc, v6, v0
-; GCN-NEXT:    v_cmp_ge_u32_e64 s[4:5], v3, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v9, 0, -1, s[4:5]
+; GCN-NEXT:    v_sub_i32_e64 v5, s[4:5], v2, v0
+; GCN-NEXT:    v_subb_u32_e64 v6, s[6:7], v4, v1, s[4:5]
+; GCN-NEXT:    v_subbrev_u32_e64 v4, s[4:5], 0, v4, s[4:5]
 ; GCN-NEXT:    v_cmp_ge_u32_e64 s[4:5], v4, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v10, 0, -1, s[4:5]
-; GCN-NEXT:    v_cmp_eq_u32_e64 s[4:5], v3, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v5, v9, v5, s[4:5]
-; GCN-NEXT:    v_subbrev_u32_e32 v7, vcc, 0, v7, vcc
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v1
-; GCN-NEXT:    v_cndmask_b32_e32 v1, v10, v8, vcc
-; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v1
-; GCN-NEXT:    v_cndmask_b32_e32 v0, v6, v0, vcc
-; GCN-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, v5
-; GCN-NEXT:    v_cndmask_b32_e64 v0, v2, v0, s[4:5]
-; GCN-NEXT:    v_cndmask_b32_e32 v1, v4, v7, vcc
-; GCN-NEXT:    v_cndmask_b32_e64 v1, v3, v1, s[4:5]
+; GCN-NEXT:    v_cndmask_b32_e64 v7, 0, -1, s[4:5]
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[4:5], v5, v0
+; GCN-NEXT:    v_subb_u32_e32 v3, vcc, 0, v3, vcc
+; GCN-NEXT:    v_cndmask_b32_e64 v8, 0, -1, s[4:5]
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[4:5], v4, v1
+; GCN-NEXT:    v_cndmask_b32_e64 v7, v7, v8, s[4:5]
+; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v3, v1
+; GCN-NEXT:    v_cmp_ne_u32_e64 s[6:7], 0, v7
+; GCN-NEXT:    v_cndmask_b32_e64 v7, 0, -1, vcc
+; GCN-NEXT:    v_cmp_ge_u32_e32 vcc, v2, v0
+; GCN-NEXT:    v_sub_i32_e64 v8, s[4:5], v5, v0
+; GCN-NEXT:    v_cndmask_b32_e64 v0, 0, -1, vcc
+; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v1
+; GCN-NEXT:    v_cndmask_b32_e32 v0, v7, v0, vcc
+; GCN-NEXT:    v_subbrev_u32_e64 v1, s[4:5], 0, v6, s[4:5]
+; GCN-NEXT:    v_cndmask_b32_e64 v5, v5, v8, s[6:7]
+; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
+; GCN-NEXT:    v_cndmask_b32_e64 v1, v4, v1, s[6:7]
+; GCN-NEXT:    v_cndmask_b32_e32 v0, v2, v5, vcc
+; GCN-NEXT:    v_cndmask_b32_e32 v1, v3, v1, vcc
 ; GCN-NEXT:    s_setpc_b64 s[30:31]
   %result = urem i64 32768, %x
   ret i64 %result
 }
 
 define i64 @v_test_urem_pow2_k_den_i64(i64 %x) {
 ; GCN-LABEL: v_test_urem_pow2_k_den_i64:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GCN-NEXT:    v_and_b32_e32 v0, 0x7fff, v0
 ; GCN-NEXT:    v_mov_b32_e32 v1, 0
 ; GCN-NEXT:    s_setpc_b64 s[30:31]
   %result = urem i64 %x, 32768
   ret i64 %result
 }
 
 define amdgpu_kernel void @s_test_urem24_k_num_i64(i64 addrspace(1)* %out, i64 %x) {
 ; GCN-LABEL: s_test_urem24_k_num_i64:
 ; GCN:       ; %bb.0:
-; GCN-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
-; GCN-NEXT:    s_mov_b32 s7, 0xf000
-; GCN-NEXT:    s_mov_b32 s6, -1
+; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_mov_b32 s4, s0
-; GCN-NEXT:    s_mov_b32 s5, s1
-; GCN-NEXT:    s_lshr_b32 s8, s3, 8
+; GCN-NEXT:    s_mov_b32 s6, -1
+; GCN-NEXT:    s_lshr_b32 s8, s7, 8
 ; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s8
+; GCN-NEXT:    s_mov_b32 s7, 0xf000
 ; GCN-NEXT:    v_rcp_iflag_f32_e32 v0, v0
 ; GCN-NEXT:    v_mul_f32_e32 v0, 0x4f800000, v0
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v0
-; GCN-NEXT:    v_mul_hi_u32 v1, v0, s8
-; GCN-NEXT:    v_mul_lo_u32 v2, v0, s8
-; GCN-NEXT:    v_sub_i32_e32 v3, vcc, 0, v2
-; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v1, v2, v3, s[0:1]
+; GCN-NEXT:    v_mul_lo_u32 v1, v0, s8
+; GCN-NEXT:    v_mul_hi_u32 v2, v0, s8
+; GCN-NEXT:    v_sub_i32_e32 v3, vcc, 0, v1
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v1, v1, v3, s[0:1]
 ; GCN-NEXT:    v_mul_hi_u32 v1, v1, v0
 ; GCN-NEXT:    v_add_i32_e32 v2, vcc, v1, v0
 ; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, v1, v0
 ; GCN-NEXT:    v_cndmask_b32_e64 v0, v0, v2, s[0:1]
 ; GCN-NEXT:    v_mul_hi_u32 v0, v0, 24
 ; GCN-NEXT:    v_mul_lo_u32 v0, v0, s8
 ; GCN-NEXT:    v_sub_i32_e32 v1, vcc, 24, v0
-; GCN-NEXT:    v_add_i32_e32 v2, vcc, s8, v1
 ; GCN-NEXT:    v_cmp_gt_u32_e64 s[0:1], 25, v0
+; GCN-NEXT:    v_add_i32_e32 v2, vcc, s8, v1
 ; GCN-NEXT:    v_cmp_le_u32_e64 s[2:3], s8, v1
 ; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, s8, v1
 ; GCN-NEXT:    s_and_b64 vcc, s[2:3], s[0:1]
 ; GCN-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc
 ; GCN-NEXT:    v_cndmask_b32_e64 v0, v2, v0, s[0:1]
 ; GCN-NEXT:    v_mov_b32_e32 v1, 0
 ; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
 ; GCN-NEXT:    s_endpgm
   %x.shr = lshr i64 %x, 40
   %result = urem i64 24, %x.shr
   store i64 %result, i64 addrspace(1)* %out
   ret void
 }
 
 define amdgpu_kernel void @s_test_urem24_k_den_i64(i64 addrspace(1)* %out, i64 %x) {
 ; GCN-LABEL: s_test_urem24_k_den_i64:
 ; GCN:       ; %bb.0:
+; GCN-NEXT:    s_movk_i32 s8, 0x5b7f
+; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s8
 ; GCN-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; GCN-NEXT:    s_movk_i32 s9, 0x5b7e
 ; GCN-NEXT:    s_mov_b32 s7, 0xf000
+; GCN-NEXT:    v_rcp_iflag_f32_e32 v0, v0
 ; GCN-NEXT:    s_mov_b32 s6, -1
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_movk_i32 s2, 0x5b7f
-; GCN-NEXT:    s_movk_i32 s8, 0x5b7e
-; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s2
-; GCN-NEXT:    v_rcp_iflag_f32_e32 v0, v0
-; GCN-NEXT:    v_mul_f32_e32 v0, 0x4f800000, v0
 ; GCN-NEXT:    s_mov_b32 s4, s0
 ; GCN-NEXT:    s_mov_b32 s5, s1
-; GCN-NEXT:    s_lshr_b32 s3, s3, 8
+; GCN-NEXT:    v_mul_f32_e32 v0, 0x4f800000, v0
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v0
-; GCN-NEXT:    v_mul_hi_u32 v1, v0, s2
-; GCN-NEXT:    v_mul_lo_u32 v2, v0, s2
-; GCN-NEXT:    v_sub_i32_e32 v3, vcc, 0, v2
-; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v1
-; GCN-NEXT:    v_cndmask_b32_e64 v1, v2, v3, s[0:1]
+; GCN-NEXT:    s_lshr_b32 s2, s3, 8
+; GCN-NEXT:    v_mul_lo_u32 v1, v0, s8
+; GCN-NEXT:    v_mul_hi_u32 v2, v0, s8
+; GCN-NEXT:    v_sub_i32_e32 v3, vcc, 0, v1
+; GCN-NEXT:    v_cmp_eq_u32_e64 s[0:1], 0, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v1, v1, v3, s[0:1]
 ; GCN-NEXT:    v_mul_hi_u32 v1, v1, v0
 ; GCN-NEXT:    v_add_i32_e32 v2, vcc, v1, v0
 ; GCN-NEXT:    v_subrev_i32_e32 v0, vcc, v1, v0
 ; GCN-NEXT:    v_cndmask_b32_e64 v0, v0, v2, s[0:1]
-; GCN-NEXT:    v_mul_hi_u32 v0, v0, s3
-; GCN-NEXT:    v_mul_lo_u32 v0, v0, s2
-; GCN-NEXT:    v_sub_i32_e32 v1, vcc, s3, v0
-; GCN-NEXT:    v_add_i32_e32 v2, vcc, s2, v1
-; GCN-NEXT:    v_cmp_ge_u32_e64 s[0:1], s3, v0
-; GCN-NEXT:    v_cmp_lt_u32_e64 s[2:3], s8, v1
+; GCN-NEXT:    v_mul_hi_u32 v0, v0, s2
+; GCN-NEXT:    v_mul_lo_u32 v0, v0, s8
+; GCN-NEXT:    v_sub_i32_e32 v1, vcc, s2, v0
+; GCN-NEXT:    v_cmp_ge_u32_e64 s[0:1], s2, v0
+; GCN-NEXT:    v_add_i32_e32 v2, vcc, s8, v1
+; GCN-NEXT:    v_cmp_lt_u32_e64 s[2:3], s9, v1
 ; GCN-NEXT:    v_add_i32_e32 v0, vcc, 0xffffa481, v1
 ; GCN-NEXT:    s_and_b64 vcc, s[2:3], s[0:1]
 ; GCN-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc
 ; GCN-NEXT:    v_cndmask_b32_e64 v0, v2, v0, s[0:1]
 ; GCN-NEXT:    v_mov_b32_e32 v1, 0
 ; GCN-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
 ; GCN-NEXT:    s_endpgm
   %x.shr = lshr i64 %x, 40
   %result = urem i64 %x.shr, 23423
   store i64 %result, i64 addrspace(1)* %out
   ret void
 }
 
 define i64 @v_test_urem24_k_num_i64(i64 %x) {
 ; GCN-LABEL: v_test_urem24_k_num_i64:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GCN-NEXT:    v_lshrrev_b32_e32 v0, 8, v1
 ; GCN-NEXT:    v_cvt_f32_u32_e32 v1, v0
 ; GCN-NEXT:    v_rcp_iflag_f32_e32 v1, v1
 ; GCN-NEXT:    v_mul_f32_e32 v1, 0x4f800000, v1
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v1, v1
-; GCN-NEXT:    v_mul_hi_u32 v2, v1, v0
-; GCN-NEXT:    v_mul_lo_u32 v3, v1, v0
-; GCN-NEXT:    v_sub_i32_e32 v4, vcc, 0, v3
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v2
-; GCN-NEXT:    v_cndmask_b32_e32 v2, v3, v4, vcc
+; GCN-NEXT:    v_mul_lo_u32 v2, v1, v0
+; GCN-NEXT:    v_mul_hi_u32 v3, v1, v0
+; GCN-NEXT:    v_sub_i32_e32 v4, vcc, 0, v2
+; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v3
+; GCN-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
 ; GCN-NEXT:    v_mul_hi_u32 v2, v2, v1
 ; GCN-NEXT:    v_add_i32_e64 v3, s[4:5], v1, v2
 ; GCN-NEXT:    v_sub_i32_e64 v1, s[4:5], v1, v2
 ; GCN-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
 ; GCN-NEXT:    v_mul_hi_u32 v1, v1, 24
 ; GCN-NEXT:    v_mul_lo_u32 v1, v1, v0
 ; GCN-NEXT:    v_sub_i32_e32 v2, vcc, 24, v1
 ; GCN-NEXT:    v_add_i32_e32 v3, vcc, v2, v0
; GCN-NEXT:    s_setpc_b64 s[30:31]
   ret i64 %result
 }
 
 define i64 @v_test_urem24_pow2_k_num_i64(i64 %x) {
 ; GCN-LABEL: v_test_urem24_pow2_k_num_i64:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GCN-NEXT:    v_lshrrev_b32_e32 v0, 8, v1
-; GCN-NEXT:    s_mov_b32 s6, 0x8001
 ; GCN-NEXT:    v_cvt_f32_u32_e32 v1, v0
+; GCN-NEXT:    s_mov_b32 s6, 0x8001
 ; GCN-NEXT:    v_rcp_iflag_f32_e32 v1, v1
 ; GCN-NEXT:    v_mul_f32_e32 v1, 0x4f800000, v1
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v1, v1
-; GCN-NEXT:    v_mul_hi_u32 v2, v1, v0
-; GCN-NEXT:    v_mul_lo_u32 v3, v1, v0
-; GCN-NEXT:    v_sub_i32_e32 v4, vcc, 0, v3
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v2
-; GCN-NEXT:    v_cndmask_b32_e32 v2, v3, v4, vcc
+; GCN-NEXT:    v_mul_lo_u32 v2, v1, v0
+; GCN-NEXT:    v_mul_hi_u32 v3, v1, v0
+; GCN-NEXT:    v_sub_i32_e32 v4, vcc, 0, v2
+; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v3
+; GCN-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
 ; GCN-NEXT:    v_mul_hi_u32 v2, v2, v1
 ; GCN-NEXT:    v_add_i32_e64 v3, s[4:5], v1, v2
 ; GCN-NEXT:    v_sub_i32_e64 v1, s[4:5], v1, v2
 ; GCN-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
 ; GCN-NEXT:    v_lshrrev_b32_e32 v1, 17, v1
 ; GCN-NEXT:    v_mul_u32_u24_e32 v1, v1, v0
 ; GCN-NEXT:    v_sub_i32_e32 v2, vcc, 0x8000, v1
 ; GCN-NEXT:    v_add_i32_e32 v3, vcc, v2, v0

llvm/test/CodeGen/AMDGPU/v_mac.ll

-; RUN:  llc -amdgpu-scalarize-global-loads=false  -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=GCN %s
+; RUN:  llc -amdgpu-scalarize-global-loads=false  -march=amdgcn -mcpu=tahiti -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=GCN %s
 ; RUN:  llc -amdgpu-scalarize-global-loads=false  -march=amdgcn -mcpu=tonga -mattr=-fp64-fp16-denormals,-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefix=VI -check-prefix=VI-FLUSH -check-prefix=GCN %s
 ; RUN:  llc -amdgpu-scalarize-global-loads=false  -march=amdgcn -mcpu=tonga -mattr=+fp64-fp16-denormals,-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefix=VI -check-prefix=VI-DENORM -check-prefix=GCN %s
 
 ; GCN-LABEL: {{^}}mac_vvv:
 ; GCN: buffer_load_dword [[A:v[0-9]+]], off, s[{{[0-9]+:[0-9]+}}], 0{{$}}
 ; GCN: buffer_load_dword [[B:v[0-9]+]], off, s[{{[0-9]+:[0-9]+}}], 0 offset:4
 ; GCN: buffer_load_dword [[C:v[0-9]+]], off, s[{{[0-9]+:[0-9]+}}], 0 offset:8
 ; GCN: v_mac_f32_e32 [[C]], [[A]], [[B]]