[CodeGen] MachineVerifier to check early-clobber constraint (#151421)

Currently MachineVerifier is missing verifying early-clobber operand
constraint.
The only other machine operand constraint -  TiedTo is already verified.
This commit is contained in:
Abhay Kanhere 2025-11-04 18:39:31 -08:00 committed by GitHub
parent d6fdfe0a27
commit d998f92a00
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
14 changed files with 5145 additions and 5073 deletions

View File

@ -2584,6 +2584,14 @@ MachineVerifier::visitMachineOperand(const MachineOperand *MO, unsigned MONum) {
report("Extra explicit operand on non-variadic instruction", MO, MONum);
}
// Verify earlyClobber def operand
if (MCID.getOperandConstraint(MONum, MCOI::EARLY_CLOBBER) != -1) {
if (!MO->isReg())
report("Early clobber must be a register", MI);
if (!MO->isEarlyClobber())
report("Missing earlyClobber flag", MI);
}
switch (MO->getType()) {
case MachineOperand::MO_Register: {
// Verify debug flag on debug instructions. Check this first because reg0

View File

@ -116,8 +116,14 @@ bool AMDGPUInstructionSelector::constrainCopyLikeIntrin(MachineInstr &MI,
if (!DstRC || DstRC != SrcRC)
return false;
return RBI.constrainGenericRegister(Dst.getReg(), *DstRC, *MRI) &&
RBI.constrainGenericRegister(Src.getReg(), *SrcRC, *MRI);
if (!RBI.constrainGenericRegister(Dst.getReg(), *DstRC, *MRI) ||
!RBI.constrainGenericRegister(Src.getReg(), *SrcRC, *MRI))
return false;
const MCInstrDesc &MCID = MI.getDesc();
if (MCID.getOperandConstraint(0, MCOI::EARLY_CLOBBER) != -1) {
MI.getOperand(0).setIsEarlyClobber(true);
}
return true;
}
bool AMDGPUInstructionSelector::selectCOPY(MachineInstr &I) const {
@ -602,6 +608,7 @@ bool AMDGPUInstructionSelector::selectG_AMDGPU_MAD_64_32(
I.setDesc(TII.get(Opc));
I.addOperand(*MF, MachineOperand::CreateImm(0));
I.addImplicitDefUseOperands(*MF);
I.getOperand(0).setIsEarlyClobber(true);
return constrainSelectedInstRegOperands(I, TII, TRI, RBI);
}
@ -3787,6 +3794,10 @@ bool AMDGPUInstructionSelector::selectSMFMACIntrin(MachineInstr &MI) const {
MI.removeOperand(1); // Intrinsic ID
MI.addOperand(VDst_In); // Readd VDst_In to the end
MI.addImplicitDefUseOperands(*MI.getParent()->getParent());
const MCInstrDesc &MCID = MI.getDesc();
if (MCID.getOperandConstraint(0, MCOI::EARLY_CLOBBER) != -1) {
MI.getOperand(0).setIsEarlyClobber(true);
}
return true;
}

View File

@ -27,11 +27,11 @@ define amdgpu_kernel void @set_inactive_imm_poison(ptr addrspace(1) %out) {
; GCN: ; %bb.0:
; GCN-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN-NEXT: v_mov_b32_e32 v0, 1
; GCN-NEXT: v_mov_b32_e32 v0, v0
; GCN-NEXT: v_mov_b32_e32 v1, v0
; GCN-NEXT: s_mov_b32 s2, -1
; GCN-NEXT: s_mov_b32 s3, 0xf000
; GCN-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0
; GCN-NEXT: buffer_store_dword v1, off, s[0:3], 0
; GCN-NEXT: s_endpgm
%tmp.0 = call i32 @llvm.amdgcn.set.inactive.i32(i32 1, i32 poison) #0
%tmp = call i32 @llvm.amdgcn.strict.wwm.i32(i32 %tmp.0)
@ -68,12 +68,12 @@ define amdgpu_kernel void @set_inactive_imm_poison_64(ptr addrspace(1) %out) {
; GCN-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN-NEXT: v_mov_b32_e32 v0, 1
; GCN-NEXT: v_mov_b32_e32 v1, 0
; GCN-NEXT: v_mov_b32_e32 v0, v0
; GCN-NEXT: v_mov_b32_e32 v1, v1
; GCN-NEXT: v_mov_b32_e32 v2, v0
; GCN-NEXT: v_mov_b32_e32 v3, v1
; GCN-NEXT: s_mov_b32 s2, -1
; GCN-NEXT: s_mov_b32 s3, 0xf000
; GCN-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0
; GCN-NEXT: buffer_store_dwordx2 v[2:3], off, s[0:3], 0
; GCN-NEXT: s_endpgm
%tmp.0 = call i64 @llvm.amdgcn.set.inactive.i64(i64 1, i64 poison) #0
%tmp = call i64 @llvm.amdgcn.strict.wwm.i64(i64 %tmp.0)

View File

@ -165,10 +165,10 @@ define amdgpu_kernel void @v_mul_i64_zext_src0_src1(ptr addrspace(1) %out, ptr a
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 2, v0
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: s_clause 0x1
; GFX10-NEXT: global_load_dword v1, v0, s[2:3]
; GFX10-NEXT: global_load_dword v2, v0, s[6:7]
; GFX10-NEXT: global_load_dword v2, v0, s[2:3]
; GFX10-NEXT: global_load_dword v3, v0, s[6:7]
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: v_mad_u64_u32 v[0:1], s2, v1, v2, 0
; GFX10-NEXT: v_mad_u64_u32 v[0:1], s2, v2, v3, 0
; GFX10-NEXT: v_mov_b32_e32 v2, 0
; GFX10-NEXT: global_store_dwordx2 v2, v[0:1], s[0:1]
; GFX10-NEXT: s_endpgm
@ -179,15 +179,15 @@ define amdgpu_kernel void @v_mul_i64_zext_src0_src1(ptr addrspace(1) %out, ptr a
; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT: s_load_b64 s[4:5], s[4:5], 0x34
; GFX11-NEXT: v_and_b32_e32 v0, 0x3ff, v0
; GFX11-NEXT: v_mov_b32_e32 v2, 0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 2, v0
; GFX11-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-NEXT: s_clause 0x1
; GFX11-NEXT: global_load_b32 v1, v0, s[2:3]
; GFX11-NEXT: global_load_b32 v0, v0, s[4:5]
; GFX11-NEXT: global_load_b32 v2, v0, s[2:3]
; GFX11-NEXT: global_load_b32 v3, v0, s[4:5]
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: v_mad_u64_u32 v[0:1], null, v1, v0, 0
; GFX11-NEXT: v_mad_u64_u32 v[0:1], null, v2, v3, 0
; GFX11-NEXT: v_mov_b32_e32 v2, 0
; GFX11-NEXT: global_store_b64 v2, v[0:1], s[0:1]
; GFX11-NEXT: s_endpgm
%tid = call i32 @llvm.amdgcn.workitem.id.x()

File diff suppressed because it is too large Load Diff

File diff suppressed because it is too large Load Diff

File diff suppressed because it is too large Load Diff

File diff suppressed because it is too large Load Diff

File diff suppressed because it is too large Load Diff

View File

@ -2407,51 +2407,52 @@ define <2 x i128> @v_srem_v2i128_vv(<2 x i128> %lhs, <2 x i128> %rhs) {
; GISEL-NEXT: s_or_b64 exec, exec, s[12:13]
; GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v30, v31, 0
; GISEL-NEXT: v_mad_u64_u32 v[14:15], s[4:5], v30, v18, 0
; GISEL-NEXT: v_mul_lo_u32 v24, v30, v19
; GISEL-NEXT: v_mul_lo_u32 v25, v29, v18
; GISEL-NEXT: v_mul_lo_u32 v27, v30, v19
; GISEL-NEXT: v_mul_lo_u32 v36, v29, v18
; GISEL-NEXT: v_mad_u64_u32 v[18:19], s[4:5], v35, v20, 0
; GISEL-NEXT: v_mad_u64_u32 v[22:23], s[4:5], v35, v2, 0
; GISEL-NEXT: v_mul_lo_u32 v26, v35, v3
; GISEL-NEXT: v_mul_lo_u32 v27, v34, v2
; GISEL-NEXT: v_mul_lo_u32 v37, v35, v3
; GISEL-NEXT: v_mul_lo_u32 v38, v34, v2
; GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v29, v32, v[14:15]
; GISEL-NEXT: v_mad_u64_u32 v[14:15], s[4:5], v34, v21, v[22:23]
; GISEL-NEXT: v_mov_b32_e32 v22, v19
; GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v10, v31, v[2:3]
; GISEL-NEXT: v_mad_u64_u32 v[14:15], s[4:5], v4, v20, v[14:15]
; GISEL-NEXT: v_mad_u64_u32 v[1:2], vcc, v30, v32, v[1:2]
; GISEL-NEXT: v_mov_b32_e32 v23, v14
; GISEL-NEXT: v_mad_u64_u32 v[22:23], s[4:5], v35, v21, v[22:23]
; GISEL-NEXT: v_mad_u64_u32 v[1:2], s[6:7], v29, v31, v[1:2]
; GISEL-NEXT: v_addc_u32_e64 v3, s[6:7], v3, v24, s[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[22:23], s[6:7], v34, v20, v[22:23]
; GISEL-NEXT: v_addc_u32_e64 v14, s[6:7], v15, v26, s[6:7]
; GISEL-NEXT: v_addc_u32_e32 v3, vcc, v3, v25, vcc
; GISEL-NEXT: v_mad_u64_u32 v[23:24], s[4:5], v10, v31, v[2:3]
; GISEL-NEXT: v_mad_u64_u32 v[25:26], s[4:5], v4, v20, v[14:15]
; GISEL-NEXT: v_mov_b32_e32 v2, v23
; GISEL-NEXT: v_mad_u64_u32 v[14:15], vcc, v30, v32, v[1:2]
; GISEL-NEXT: v_mov_b32_e32 v23, v25
; GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v35, v21, v[22:23]
; GISEL-NEXT: v_mad_u64_u32 v[22:23], s[6:7], v29, v31, v[14:15]
; GISEL-NEXT: v_addc_u32_e64 v3, s[6:7], v24, v27, s[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[14:15], s[6:7], v34, v20, v[1:2]
; GISEL-NEXT: v_addc_u32_e64 v2, s[6:7], v26, v37, s[6:7]
; GISEL-NEXT: v_addc_u32_e32 v1, vcc, v3, v36, vcc
; GISEL-NEXT: v_sub_i32_e32 v0, vcc, v16, v0
; GISEL-NEXT: v_subb_u32_e32 v1, vcc, v17, v1, vcc
; GISEL-NEXT: v_xor_b32_e32 v15, v0, v28
; GISEL-NEXT: v_addc_u32_e64 v0, s[4:5], v14, v27, s[4:5]
; GISEL-NEXT: v_sub_i32_e64 v12, s[4:5], v12, v18
; GISEL-NEXT: v_subb_u32_e64 v14, s[4:5], v13, v22, s[4:5]
; GISEL-NEXT: v_xor_b32_e32 v16, v12, v33
; GISEL-NEXT: v_mad_u64_u32 v[12:13], s[6:7], v10, v32, v[3:4]
; GISEL-NEXT: v_xor_b32_e32 v1, v1, v28
; GISEL-NEXT: v_mad_u64_u32 v[3:4], s[6:7], v4, v21, v[0:1]
; GISEL-NEXT: v_xor_b32_e32 v14, v14, v33
; GISEL-NEXT: v_mad_u64_u32 v[10:11], s[6:7], v11, v31, v[12:13]
; GISEL-NEXT: v_sub_i32_e64 v0, s[6:7], v15, v28
; GISEL-NEXT: v_subb_u32_e32 v16, vcc, v17, v22, vcc
; GISEL-NEXT: v_xor_b32_e32 v19, v0, v28
; GISEL-NEXT: v_addc_u32_e64 v0, s[4:5], v2, v38, s[4:5]
; GISEL-NEXT: v_sub_i32_e64 v2, s[4:5], v12, v18
; GISEL-NEXT: v_subb_u32_e64 v14, s[4:5], v13, v14, s[4:5]
; GISEL-NEXT: v_xor_b32_e32 v18, v2, v33
; GISEL-NEXT: v_mad_u64_u32 v[2:3], s[6:7], v10, v32, v[1:2]
; GISEL-NEXT: v_xor_b32_e32 v1, v16, v28
; GISEL-NEXT: v_mad_u64_u32 v[12:13], s[6:7], v4, v21, v[0:1]
; GISEL-NEXT: v_xor_b32_e32 v10, v14, v33
; GISEL-NEXT: v_mad_u64_u32 v[16:17], s[6:7], v11, v31, v[2:3]
; GISEL-NEXT: v_sub_i32_e64 v0, s[6:7], v19, v28
; GISEL-NEXT: v_subb_u32_e64 v1, s[6:7], v1, v28, s[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[3:4], s[8:9], v5, v20, v[3:4]
; GISEL-NEXT: v_sub_i32_e64 v4, s[8:9], v16, v33
; GISEL-NEXT: v_subb_u32_e64 v5, s[8:9], v14, v33, s[8:9]
; GISEL-NEXT: v_subb_u32_e32 v2, vcc, v8, v2, vcc
; GISEL-NEXT: v_subb_u32_e32 v8, vcc, v9, v10, vcc
; GISEL-NEXT: v_xor_b32_e32 v2, v2, v28
; GISEL-NEXT: v_subb_u32_e64 v6, vcc, v6, v23, s[4:5]
; GISEL-NEXT: v_subb_u32_e32 v3, vcc, v7, v3, vcc
; GISEL-NEXT: v_mad_u64_u32 v[2:3], s[8:9], v5, v20, v[12:13]
; GISEL-NEXT: v_sub_i32_e64 v4, s[8:9], v18, v33
; GISEL-NEXT: v_subb_u32_e64 v5, s[8:9], v10, v33, s[8:9]
; GISEL-NEXT: v_subb_u32_e32 v3, vcc, v8, v23, vcc
; GISEL-NEXT: v_subb_u32_e32 v8, vcc, v9, v16, vcc
; GISEL-NEXT: v_xor_b32_e32 v3, v3, v28
; GISEL-NEXT: v_subb_u32_e64 v6, vcc, v6, v15, s[4:5]
; GISEL-NEXT: v_subb_u32_e32 v2, vcc, v7, v2, vcc
; GISEL-NEXT: v_xor_b32_e32 v6, v6, v33
; GISEL-NEXT: v_xor_b32_e32 v7, v8, v28
; GISEL-NEXT: v_xor_b32_e32 v8, v3, v33
; GISEL-NEXT: v_subb_u32_e64 v2, vcc, v2, v28, s[6:7]
; GISEL-NEXT: v_xor_b32_e32 v8, v2, v33
; GISEL-NEXT: v_subb_u32_e64 v2, vcc, v3, v28, s[6:7]
; GISEL-NEXT: v_subb_u32_e32 v3, vcc, v7, v28, vcc
; GISEL-NEXT: v_subb_u32_e64 v6, vcc, v6, v33, s[8:9]
; GISEL-NEXT: v_subb_u32_e32 v7, vcc, v8, v33, vcc
@ -3216,36 +3217,38 @@ define <2 x i128> @v_urem_v2i128_vv(<2 x i128> %lhs, <2 x i128> %rhs) {
; GISEL-NEXT: s_or_b64 exec, exec, s[12:13]
; GISEL-NEXT: v_mad_u64_u32 v[16:17], s[4:5], v8, v32, 0
; GISEL-NEXT: v_mad_u64_u32 v[22:23], s[4:5], v8, v20, 0
; GISEL-NEXT: v_mul_lo_u32 v28, v8, v21
; GISEL-NEXT: v_mul_lo_u32 v29, v9, v20
; GISEL-NEXT: v_mul_lo_u32 v34, v8, v21
; GISEL-NEXT: v_mul_lo_u32 v35, v9, v20
; GISEL-NEXT: v_mad_u64_u32 v[20:21], s[4:5], v12, v24, 0
; GISEL-NEXT: v_mad_u64_u32 v[26:27], s[4:5], v12, v18, 0
; GISEL-NEXT: v_mul_lo_u32 v30, v12, v19
; GISEL-NEXT: v_mul_lo_u32 v31, v13, v18
; GISEL-NEXT: v_mul_lo_u32 v36, v12, v19
; GISEL-NEXT: v_mul_lo_u32 v37, v13, v18
; GISEL-NEXT: v_mad_u64_u32 v[18:19], s[4:5], v9, v33, v[22:23]
; GISEL-NEXT: v_mad_u64_u32 v[22:23], s[4:5], v13, v25, v[26:27]
; GISEL-NEXT: v_mad_u64_u32 v[18:19], s[4:5], v10, v32, v[18:19]
; GISEL-NEXT: v_mad_u64_u32 v[22:23], s[4:5], v14, v24, v[22:23]
; GISEL-NEXT: v_mad_u64_u32 v[17:18], vcc, v8, v33, v[17:18]
; GISEL-NEXT: v_mad_u64_u32 v[21:22], s[4:5], v12, v25, v[21:22]
; GISEL-NEXT: v_mad_u64_u32 v[8:9], s[6:7], v9, v32, v[17:18]
; GISEL-NEXT: v_addc_u32_e64 v17, s[6:7], v19, v28, s[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[12:13], s[6:7], v13, v24, v[21:22]
; GISEL-NEXT: v_addc_u32_e64 v18, s[6:7], v23, v30, s[6:7]
; GISEL-NEXT: v_addc_u32_e32 v17, vcc, v17, v29, vcc
; GISEL-NEXT: v_mad_u64_u32 v[26:27], s[4:5], v10, v32, v[18:19]
; GISEL-NEXT: v_mad_u64_u32 v[28:29], s[4:5], v14, v24, v[22:23]
; GISEL-NEXT: v_mov_b32_e32 v18, v26
; GISEL-NEXT: v_mad_u64_u32 v[30:31], vcc, v8, v33, v[17:18]
; GISEL-NEXT: v_mov_b32_e32 v22, v28
; GISEL-NEXT: v_mad_u64_u32 v[17:18], s[4:5], v12, v25, v[21:22]
; GISEL-NEXT: v_mad_u64_u32 v[21:22], s[6:7], v9, v32, v[30:31]
; GISEL-NEXT: v_addc_u32_e64 v12, s[6:7], v27, v34, s[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[8:9], s[6:7], v13, v24, v[17:18]
; GISEL-NEXT: v_addc_u32_e64 v13, s[6:7], v29, v36, s[6:7]
; GISEL-NEXT: v_addc_u32_e32 v12, vcc, v12, v35, vcc
; GISEL-NEXT: v_sub_i32_e32 v0, vcc, v0, v16
; GISEL-NEXT: v_subb_u32_e32 v1, vcc, v1, v8, vcc
; GISEL-NEXT: v_addc_u32_e64 v8, s[4:5], v18, v31, s[4:5]
; GISEL-NEXT: v_subb_u32_e32 v1, vcc, v1, v21, vcc
; GISEL-NEXT: v_addc_u32_e64 v13, s[4:5], v13, v37, s[4:5]
; GISEL-NEXT: v_sub_i32_e64 v4, s[4:5], v4, v20
; GISEL-NEXT: v_subb_u32_e64 v5, s[4:5], v5, v12, s[4:5]
; GISEL-NEXT: v_mad_u64_u32 v[16:17], s[6:7], v10, v33, v[17:18]
; GISEL-NEXT: v_mad_u64_u32 v[18:19], s[6:7], v14, v25, v[8:9]
; GISEL-NEXT: v_mad_u64_u32 v[10:11], s[6:7], v11, v32, v[16:17]
; GISEL-NEXT: v_mad_u64_u32 v[11:12], s[6:7], v15, v24, v[18:19]
; GISEL-NEXT: v_subb_u32_e32 v2, vcc, v2, v9, vcc
; GISEL-NEXT: v_subb_u32_e32 v3, vcc, v3, v10, vcc
; GISEL-NEXT: v_subb_u32_e64 v6, vcc, v6, v13, s[4:5]
; GISEL-NEXT: v_subb_u32_e32 v7, vcc, v7, v11, vcc
; GISEL-NEXT: v_subb_u32_e64 v5, s[4:5], v5, v8, s[4:5]
; GISEL-NEXT: v_mad_u64_u32 v[16:17], s[6:7], v10, v33, v[12:13]
; GISEL-NEXT: v_mad_u64_u32 v[18:19], s[6:7], v14, v25, v[13:14]
; GISEL-NEXT: v_mad_u64_u32 v[12:13], s[6:7], v11, v32, v[16:17]
; GISEL-NEXT: v_mad_u64_u32 v[10:11], s[6:7], v15, v24, v[18:19]
; GISEL-NEXT: v_subb_u32_e32 v2, vcc, v2, v22, vcc
; GISEL-NEXT: v_subb_u32_e32 v3, vcc, v3, v12, vcc
; GISEL-NEXT: v_subb_u32_e64 v6, vcc, v6, v9, s[4:5]
; GISEL-NEXT: v_subb_u32_e32 v7, vcc, v7, v10, vcc
; GISEL-NEXT: s_setpc_b64 s[30:31]
%shl = urem <2 x i128> %lhs, %rhs
ret <2 x i128> %shl

View File

@ -237,31 +237,31 @@ define i128 @fptosi_f64_to_i128(double %x) {
; GISEL-NEXT: v_add_u32_e32 v7, 0xfffffbcd, v6
; GISEL-NEXT: v_lshlrev_b64 v[0:1], v7, v[4:5]
; GISEL-NEXT: v_cmp_gt_u32_e32 vcc, 64, v7
; GISEL-NEXT: v_cndmask_b32_e32 v10, 0, v0, vcc
; GISEL-NEXT: v_cndmask_b32_e32 v11, 0, v1, vcc
; GISEL-NEXT: v_mad_u64_u32 v[0:1], s[6:7], v10, v9, 0
; GISEL-NEXT: v_cndmask_b32_e32 v12, 0, v0, vcc
; GISEL-NEXT: v_cndmask_b32_e32 v13, 0, v1, vcc
; GISEL-NEXT: v_mad_u64_u32 v[0:1], s[6:7], v12, v9, 0
; GISEL-NEXT: v_add_u32_e32 v6, 0xfffffb8d, v6
; GISEL-NEXT: v_sub_u32_e32 v2, 64, v7
; GISEL-NEXT: v_lshrrev_b64 v[2:3], v2, v[4:5]
; GISEL-NEXT: v_lshlrev_b64 v[4:5], v6, v[4:5]
; GISEL-NEXT: v_cmp_eq_u32_e64 s[6:7], 0, v7
; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v11, v9, v[0:1]
; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v13, v9, v[0:1]
; GISEL-NEXT: v_cndmask_b32_e32 v2, v4, v2, vcc
; GISEL-NEXT: v_cndmask_b32_e64 v12, v2, 0, s[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v12, v8, v[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[0:1], s[8:9], v10, v8, 0
; GISEL-NEXT: v_mov_b32_e32 v2, v6
; GISEL-NEXT: v_mul_lo_u32 v6, v10, v9
; GISEL-NEXT: v_mad_u64_u32 v[1:2], s[8:9], v10, v9, v[1:2]
; GISEL-NEXT: v_mul_lo_u32 v4, v11, v9
; GISEL-NEXT: v_cndmask_b32_e64 v14, v2, 0, s[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[10:11], s[8:9], v14, v8, v[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[0:1], s[8:9], v12, v8, 0
; GISEL-NEXT: v_mov_b32_e32 v2, v10
; GISEL-NEXT: v_mul_lo_u32 v10, v12, v9
; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v12, v9, v[1:2]
; GISEL-NEXT: v_mul_lo_u32 v4, v13, v9
; GISEL-NEXT: v_cndmask_b32_e32 v3, v5, v3, vcc
; GISEL-NEXT: v_mad_u64_u32 v[1:2], s[10:11], v11, v8, v[1:2]
; GISEL-NEXT: v_addc_co_u32_e64 v6, s[10:11], v7, v6, s[10:11]
; GISEL-NEXT: v_mad_u64_u32 v[1:2], s[10:11], v13, v8, v[6:7]
; GISEL-NEXT: v_addc_co_u32_e64 v6, s[10:11], v11, v10, s[10:11]
; GISEL-NEXT: v_addc_co_u32_e64 v4, s[8:9], v6, v4, s[8:9]
; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v12, v9, v[4:5]
; GISEL-NEXT: v_cndmask_b32_e64 v3, v3, 0, s[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v14, v9, v[4:5]
; GISEL-NEXT: v_cndmask_b32_e64 v5, v3, 0, s[6:7]
; GISEL-NEXT: ; implicit-def: $vgpr9
; GISEL-NEXT: v_mad_u64_u32 v[3:4], s[6:7], v3, v8, v[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[3:4], s[6:7], v5, v8, v[6:7]
; GISEL-NEXT: ; implicit-def: $vgpr6
; GISEL-NEXT: ; implicit-def: $vgpr4_vgpr5
; GISEL-NEXT: ; implicit-def: $vgpr8
@ -275,17 +275,18 @@ define i128 @fptosi_f64_to_i128(double %x) {
; GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
; GISEL-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc
; GISEL-NEXT: v_cmp_eq_u32_e32 vcc, 0, v2
; GISEL-NEXT: v_cndmask_b32_e32 v4, v0, v4, vcc
; GISEL-NEXT: v_mad_u64_u32 v[2:3], s[6:7], v4, v9, 0
; GISEL-NEXT: v_cndmask_b32_e32 v5, v1, v5, vcc
; GISEL-NEXT: v_mad_u64_u32 v[0:1], s[6:7], v4, v8, 0
; GISEL-NEXT: v_mad_u64_u32 v[2:3], s[6:7], v5, v9, v[2:3]
; GISEL-NEXT: v_mul_lo_u32 v6, v5, v9
; GISEL-NEXT: v_mad_u64_u32 v[1:2], vcc, v4, v9, v[1:2]
; GISEL-NEXT: v_mul_lo_u32 v4, v4, v9
; GISEL-NEXT: v_mad_u64_u32 v[1:2], s[6:7], v5, v8, v[1:2]
; GISEL-NEXT: v_addc_co_u32_e64 v3, s[6:7], v3, v4, s[6:7]
; GISEL-NEXT: v_addc_co_u32_e32 v3, vcc, v3, v6, vcc
; GISEL-NEXT: v_cndmask_b32_e32 v6, v0, v4, vcc
; GISEL-NEXT: v_mad_u64_u32 v[2:3], s[6:7], v6, v9, 0
; GISEL-NEXT: v_cndmask_b32_e32 v7, v1, v5, vcc
; GISEL-NEXT: v_mad_u64_u32 v[0:1], s[6:7], v6, v8, 0
; GISEL-NEXT: v_mad_u64_u32 v[4:5], s[6:7], v7, v9, v[2:3]
; GISEL-NEXT: v_mul_lo_u32 v10, v7, v9
; GISEL-NEXT: v_mov_b32_e32 v2, v4
; GISEL-NEXT: v_mad_u64_u32 v[3:4], vcc, v6, v9, v[1:2]
; GISEL-NEXT: v_mul_lo_u32 v6, v6, v9
; GISEL-NEXT: v_mad_u64_u32 v[1:2], s[6:7], v7, v8, v[3:4]
; GISEL-NEXT: v_addc_co_u32_e64 v3, s[6:7], v5, v6, s[6:7]
; GISEL-NEXT: v_addc_co_u32_e32 v3, vcc, v3, v10, vcc
; GISEL-NEXT: .LBB0_6: ; %Flow1
; GISEL-NEXT: s_or_b64 exec, exec, s[8:9]
; GISEL-NEXT: .LBB0_7: ; %Flow2
@ -604,31 +605,31 @@ define i128 @fptoui_f64_to_i128(double %x) {
; GISEL-NEXT: v_add_u32_e32 v7, 0xfffffbcd, v6
; GISEL-NEXT: v_lshlrev_b64 v[0:1], v7, v[4:5]
; GISEL-NEXT: v_cmp_gt_u32_e32 vcc, 64, v7
; GISEL-NEXT: v_cndmask_b32_e32 v10, 0, v0, vcc
; GISEL-NEXT: v_cndmask_b32_e32 v11, 0, v1, vcc
; GISEL-NEXT: v_mad_u64_u32 v[0:1], s[6:7], v10, v9, 0
; GISEL-NEXT: v_cndmask_b32_e32 v12, 0, v0, vcc
; GISEL-NEXT: v_cndmask_b32_e32 v13, 0, v1, vcc
; GISEL-NEXT: v_mad_u64_u32 v[0:1], s[6:7], v12, v9, 0
; GISEL-NEXT: v_add_u32_e32 v6, 0xfffffb8d, v6
; GISEL-NEXT: v_sub_u32_e32 v2, 64, v7
; GISEL-NEXT: v_lshrrev_b64 v[2:3], v2, v[4:5]
; GISEL-NEXT: v_lshlrev_b64 v[4:5], v6, v[4:5]
; GISEL-NEXT: v_cmp_eq_u32_e64 s[6:7], 0, v7
; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v11, v9, v[0:1]
; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v13, v9, v[0:1]
; GISEL-NEXT: v_cndmask_b32_e32 v2, v4, v2, vcc
; GISEL-NEXT: v_cndmask_b32_e64 v12, v2, 0, s[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v12, v8, v[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[0:1], s[8:9], v10, v8, 0
; GISEL-NEXT: v_mov_b32_e32 v2, v6
; GISEL-NEXT: v_mul_lo_u32 v6, v10, v9
; GISEL-NEXT: v_mad_u64_u32 v[1:2], s[8:9], v10, v9, v[1:2]
; GISEL-NEXT: v_mul_lo_u32 v4, v11, v9
; GISEL-NEXT: v_cndmask_b32_e64 v14, v2, 0, s[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[10:11], s[8:9], v14, v8, v[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[0:1], s[8:9], v12, v8, 0
; GISEL-NEXT: v_mov_b32_e32 v2, v10
; GISEL-NEXT: v_mul_lo_u32 v10, v12, v9
; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v12, v9, v[1:2]
; GISEL-NEXT: v_mul_lo_u32 v4, v13, v9
; GISEL-NEXT: v_cndmask_b32_e32 v3, v5, v3, vcc
; GISEL-NEXT: v_mad_u64_u32 v[1:2], s[10:11], v11, v8, v[1:2]
; GISEL-NEXT: v_addc_co_u32_e64 v6, s[10:11], v7, v6, s[10:11]
; GISEL-NEXT: v_mad_u64_u32 v[1:2], s[10:11], v13, v8, v[6:7]
; GISEL-NEXT: v_addc_co_u32_e64 v6, s[10:11], v11, v10, s[10:11]
; GISEL-NEXT: v_addc_co_u32_e64 v4, s[8:9], v6, v4, s[8:9]
; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v12, v9, v[4:5]
; GISEL-NEXT: v_cndmask_b32_e64 v3, v3, 0, s[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v14, v9, v[4:5]
; GISEL-NEXT: v_cndmask_b32_e64 v5, v3, 0, s[6:7]
; GISEL-NEXT: ; implicit-def: $vgpr9
; GISEL-NEXT: v_mad_u64_u32 v[3:4], s[6:7], v3, v8, v[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[3:4], s[6:7], v5, v8, v[6:7]
; GISEL-NEXT: ; implicit-def: $vgpr6
; GISEL-NEXT: ; implicit-def: $vgpr4_vgpr5
; GISEL-NEXT: ; implicit-def: $vgpr8
@ -642,17 +643,18 @@ define i128 @fptoui_f64_to_i128(double %x) {
; GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
; GISEL-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc
; GISEL-NEXT: v_cmp_eq_u32_e32 vcc, 0, v2
; GISEL-NEXT: v_cndmask_b32_e32 v4, v0, v4, vcc
; GISEL-NEXT: v_mad_u64_u32 v[2:3], s[6:7], v4, v9, 0
; GISEL-NEXT: v_cndmask_b32_e32 v5, v1, v5, vcc
; GISEL-NEXT: v_mad_u64_u32 v[0:1], s[6:7], v4, v8, 0
; GISEL-NEXT: v_mad_u64_u32 v[2:3], s[6:7], v5, v9, v[2:3]
; GISEL-NEXT: v_mul_lo_u32 v6, v5, v9
; GISEL-NEXT: v_mad_u64_u32 v[1:2], vcc, v4, v9, v[1:2]
; GISEL-NEXT: v_mul_lo_u32 v4, v4, v9
; GISEL-NEXT: v_mad_u64_u32 v[1:2], s[6:7], v5, v8, v[1:2]
; GISEL-NEXT: v_addc_co_u32_e64 v3, s[6:7], v3, v4, s[6:7]
; GISEL-NEXT: v_addc_co_u32_e32 v3, vcc, v3, v6, vcc
; GISEL-NEXT: v_cndmask_b32_e32 v6, v0, v4, vcc
; GISEL-NEXT: v_mad_u64_u32 v[2:3], s[6:7], v6, v9, 0
; GISEL-NEXT: v_cndmask_b32_e32 v7, v1, v5, vcc
; GISEL-NEXT: v_mad_u64_u32 v[0:1], s[6:7], v6, v8, 0
; GISEL-NEXT: v_mad_u64_u32 v[4:5], s[6:7], v7, v9, v[2:3]
; GISEL-NEXT: v_mul_lo_u32 v10, v7, v9
; GISEL-NEXT: v_mov_b32_e32 v2, v4
; GISEL-NEXT: v_mad_u64_u32 v[3:4], vcc, v6, v9, v[1:2]
; GISEL-NEXT: v_mul_lo_u32 v6, v6, v9
; GISEL-NEXT: v_mad_u64_u32 v[1:2], s[6:7], v7, v8, v[3:4]
; GISEL-NEXT: v_addc_co_u32_e64 v3, s[6:7], v5, v6, s[6:7]
; GISEL-NEXT: v_addc_co_u32_e32 v3, vcc, v3, v10, vcc
; GISEL-NEXT: .LBB1_6: ; %Flow1
; GISEL-NEXT: s_or_b64 exec, exec, s[8:9]
; GISEL-NEXT: .LBB1_7: ; %Flow2
@ -962,31 +964,31 @@ define i128 @fptosi_f32_to_i128(float %x) {
; GISEL-NEXT: v_add_u32_e32 v7, 0xffffff6a, v6
; GISEL-NEXT: v_lshlrev_b64 v[0:1], v7, v[4:5]
; GISEL-NEXT: v_cmp_gt_u32_e32 vcc, 64, v7
; GISEL-NEXT: v_cndmask_b32_e32 v10, 0, v0, vcc
; GISEL-NEXT: v_cndmask_b32_e32 v11, 0, v1, vcc
; GISEL-NEXT: v_mad_u64_u32 v[0:1], s[6:7], v10, v8, 0
; GISEL-NEXT: v_cndmask_b32_e32 v12, 0, v0, vcc
; GISEL-NEXT: v_cndmask_b32_e32 v13, 0, v1, vcc
; GISEL-NEXT: v_mad_u64_u32 v[0:1], s[6:7], v12, v8, 0
; GISEL-NEXT: v_add_u32_e32 v6, 0xffffff2a, v6
; GISEL-NEXT: v_sub_u32_e32 v2, 64, v7
; GISEL-NEXT: v_lshrrev_b64 v[2:3], v2, v[4:5]
; GISEL-NEXT: v_lshlrev_b64 v[4:5], v6, v[4:5]
; GISEL-NEXT: v_cmp_eq_u32_e64 s[6:7], 0, v7
; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v11, v8, v[0:1]
; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v13, v8, v[0:1]
; GISEL-NEXT: v_cndmask_b32_e32 v2, v4, v2, vcc
; GISEL-NEXT: v_cndmask_b32_e64 v12, v2, 0, s[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v12, v9, v[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[0:1], s[8:9], v10, v9, 0
; GISEL-NEXT: v_mov_b32_e32 v2, v6
; GISEL-NEXT: v_mul_lo_u32 v6, v10, v8
; GISEL-NEXT: v_mad_u64_u32 v[1:2], s[8:9], v10, v8, v[1:2]
; GISEL-NEXT: v_mul_lo_u32 v4, v11, v8
; GISEL-NEXT: v_cndmask_b32_e64 v14, v2, 0, s[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[10:11], s[8:9], v14, v9, v[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[0:1], s[8:9], v12, v9, 0
; GISEL-NEXT: v_mov_b32_e32 v2, v10
; GISEL-NEXT: v_mul_lo_u32 v10, v12, v8
; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v12, v8, v[1:2]
; GISEL-NEXT: v_mul_lo_u32 v4, v13, v8
; GISEL-NEXT: v_cndmask_b32_e32 v3, v5, v3, vcc
; GISEL-NEXT: v_mad_u64_u32 v[1:2], s[10:11], v11, v9, v[1:2]
; GISEL-NEXT: v_addc_co_u32_e64 v6, s[10:11], v7, v6, s[10:11]
; GISEL-NEXT: v_mad_u64_u32 v[1:2], s[10:11], v13, v9, v[6:7]
; GISEL-NEXT: v_addc_co_u32_e64 v6, s[10:11], v11, v10, s[10:11]
; GISEL-NEXT: v_addc_co_u32_e64 v4, s[8:9], v6, v4, s[8:9]
; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v12, v8, v[4:5]
; GISEL-NEXT: v_cndmask_b32_e64 v3, v3, 0, s[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v14, v8, v[4:5]
; GISEL-NEXT: v_cndmask_b32_e64 v5, v3, 0, s[6:7]
; GISEL-NEXT: ; implicit-def: $vgpr8
; GISEL-NEXT: v_mad_u64_u32 v[3:4], s[6:7], v3, v9, v[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[3:4], s[6:7], v5, v9, v[6:7]
; GISEL-NEXT: ; implicit-def: $vgpr6
; GISEL-NEXT: ; implicit-def: $vgpr4_vgpr5
; GISEL-NEXT: ; implicit-def: $vgpr9
@ -999,12 +1001,14 @@ define i128 @fptosi_f32_to_i128(float %x) {
; GISEL-NEXT: v_cmp_gt_u32_e32 vcc, 64, v2
; GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
; GISEL-NEXT: v_cmp_eq_u32_e32 vcc, 0, v2
; GISEL-NEXT: v_cndmask_b32_e32 v4, v0, v4, vcc
; GISEL-NEXT: v_mad_u64_u32 v[0:1], s[8:9], v4, v9, 0
; GISEL-NEXT: v_mad_u64_u32 v[2:3], s[8:9], v4, v8, 0
; GISEL-NEXT: v_mul_lo_u32 v5, v4, v8
; GISEL-NEXT: v_mad_u64_u32 v[1:2], vcc, v4, v8, v[1:2]
; GISEL-NEXT: v_addc_co_u32_e32 v3, vcc, v3, v5, vcc
; GISEL-NEXT: v_cndmask_b32_e32 v6, v0, v4, vcc
; GISEL-NEXT: v_mad_u64_u32 v[0:1], s[8:9], v6, v9, 0
; GISEL-NEXT: v_mad_u64_u32 v[2:3], s[8:9], v6, v8, 0
; GISEL-NEXT: v_mov_b32_e32 v4, v1
; GISEL-NEXT: v_mul_lo_u32 v7, v6, v8
; GISEL-NEXT: v_mov_b32_e32 v5, v2
; GISEL-NEXT: v_mad_u64_u32 v[1:2], vcc, v6, v8, v[4:5]
; GISEL-NEXT: v_addc_co_u32_e32 v3, vcc, v3, v7, vcc
; GISEL-NEXT: .LBB2_6: ; %Flow1
; GISEL-NEXT: s_or_b64 exec, exec, s[6:7]
; GISEL-NEXT: .LBB2_7: ; %Flow2
@ -1314,31 +1318,31 @@ define i128 @fptoui_f32_to_i128(float %x) {
; GISEL-NEXT: v_add_u32_e32 v7, 0xffffff6a, v6
; GISEL-NEXT: v_lshlrev_b64 v[0:1], v7, v[4:5]
; GISEL-NEXT: v_cmp_gt_u32_e32 vcc, 64, v7
; GISEL-NEXT: v_cndmask_b32_e32 v10, 0, v0, vcc
; GISEL-NEXT: v_cndmask_b32_e32 v11, 0, v1, vcc
; GISEL-NEXT: v_mad_u64_u32 v[0:1], s[6:7], v10, v8, 0
; GISEL-NEXT: v_cndmask_b32_e32 v12, 0, v0, vcc
; GISEL-NEXT: v_cndmask_b32_e32 v13, 0, v1, vcc
; GISEL-NEXT: v_mad_u64_u32 v[0:1], s[6:7], v12, v8, 0
; GISEL-NEXT: v_add_u32_e32 v6, 0xffffff2a, v6
; GISEL-NEXT: v_sub_u32_e32 v2, 64, v7
; GISEL-NEXT: v_lshrrev_b64 v[2:3], v2, v[4:5]
; GISEL-NEXT: v_lshlrev_b64 v[4:5], v6, v[4:5]
; GISEL-NEXT: v_cmp_eq_u32_e64 s[6:7], 0, v7
; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v11, v8, v[0:1]
; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v13, v8, v[0:1]
; GISEL-NEXT: v_cndmask_b32_e32 v2, v4, v2, vcc
; GISEL-NEXT: v_cndmask_b32_e64 v12, v2, 0, s[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v12, v9, v[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[0:1], s[8:9], v10, v9, 0
; GISEL-NEXT: v_mov_b32_e32 v2, v6
; GISEL-NEXT: v_mul_lo_u32 v6, v10, v8
; GISEL-NEXT: v_mad_u64_u32 v[1:2], s[8:9], v10, v8, v[1:2]
; GISEL-NEXT: v_mul_lo_u32 v4, v11, v8
; GISEL-NEXT: v_cndmask_b32_e64 v14, v2, 0, s[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[10:11], s[8:9], v14, v9, v[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[0:1], s[8:9], v12, v9, 0
; GISEL-NEXT: v_mov_b32_e32 v2, v10
; GISEL-NEXT: v_mul_lo_u32 v10, v12, v8
; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v12, v8, v[1:2]
; GISEL-NEXT: v_mul_lo_u32 v4, v13, v8
; GISEL-NEXT: v_cndmask_b32_e32 v3, v5, v3, vcc
; GISEL-NEXT: v_mad_u64_u32 v[1:2], s[10:11], v11, v9, v[1:2]
; GISEL-NEXT: v_addc_co_u32_e64 v6, s[10:11], v7, v6, s[10:11]
; GISEL-NEXT: v_mad_u64_u32 v[1:2], s[10:11], v13, v9, v[6:7]
; GISEL-NEXT: v_addc_co_u32_e64 v6, s[10:11], v11, v10, s[10:11]
; GISEL-NEXT: v_addc_co_u32_e64 v4, s[8:9], v6, v4, s[8:9]
; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v12, v8, v[4:5]
; GISEL-NEXT: v_cndmask_b32_e64 v3, v3, 0, s[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v14, v8, v[4:5]
; GISEL-NEXT: v_cndmask_b32_e64 v5, v3, 0, s[6:7]
; GISEL-NEXT: ; implicit-def: $vgpr8
; GISEL-NEXT: v_mad_u64_u32 v[3:4], s[6:7], v3, v9, v[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[3:4], s[6:7], v5, v9, v[6:7]
; GISEL-NEXT: ; implicit-def: $vgpr6
; GISEL-NEXT: ; implicit-def: $vgpr4_vgpr5
; GISEL-NEXT: ; implicit-def: $vgpr9
@ -1351,12 +1355,14 @@ define i128 @fptoui_f32_to_i128(float %x) {
; GISEL-NEXT: v_cmp_gt_u32_e32 vcc, 64, v2
; GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
; GISEL-NEXT: v_cmp_eq_u32_e32 vcc, 0, v2
; GISEL-NEXT: v_cndmask_b32_e32 v4, v0, v4, vcc
; GISEL-NEXT: v_mad_u64_u32 v[0:1], s[8:9], v4, v9, 0
; GISEL-NEXT: v_mad_u64_u32 v[2:3], s[8:9], v4, v8, 0
; GISEL-NEXT: v_mul_lo_u32 v5, v4, v8
; GISEL-NEXT: v_mad_u64_u32 v[1:2], vcc, v4, v8, v[1:2]
; GISEL-NEXT: v_addc_co_u32_e32 v3, vcc, v3, v5, vcc
; GISEL-NEXT: v_cndmask_b32_e32 v6, v0, v4, vcc
; GISEL-NEXT: v_mad_u64_u32 v[0:1], s[8:9], v6, v9, 0
; GISEL-NEXT: v_mad_u64_u32 v[2:3], s[8:9], v6, v8, 0
; GISEL-NEXT: v_mov_b32_e32 v4, v1
; GISEL-NEXT: v_mul_lo_u32 v7, v6, v8
; GISEL-NEXT: v_mov_b32_e32 v5, v2
; GISEL-NEXT: v_mad_u64_u32 v[1:2], vcc, v6, v8, v[4:5]
; GISEL-NEXT: v_addc_co_u32_e32 v3, vcc, v3, v7, vcc
; GISEL-NEXT: .LBB3_6: ; %Flow1
; GISEL-NEXT: s_or_b64 exec, exec, s[6:7]
; GISEL-NEXT: .LBB3_7: ; %Flow2
@ -1702,31 +1708,31 @@ define i128 @fptosi_bf16_to_i128(bfloat %x) {
; GISEL-NEXT: v_add_u32_e32 v10, 0xffffff7a, v5
; GISEL-NEXT: v_lshlrev_b64 v[0:1], v10, v[6:7]
; GISEL-NEXT: v_cmp_gt_u32_e32 vcc, 64, v10
; GISEL-NEXT: v_cndmask_b32_e32 v11, 0, v0, vcc
; GISEL-NEXT: v_cndmask_b32_e32 v12, 0, v1, vcc
; GISEL-NEXT: v_mad_u64_u32 v[0:1], s[6:7], v11, v9, 0
; GISEL-NEXT: v_cndmask_b32_e32 v12, 0, v0, vcc
; GISEL-NEXT: v_cndmask_b32_e32 v13, 0, v1, vcc
; GISEL-NEXT: v_mad_u64_u32 v[0:1], s[6:7], v12, v9, 0
; GISEL-NEXT: v_add_u32_e32 v4, 0xffffff3a, v5
; GISEL-NEXT: v_sub_u32_e32 v2, 64, v10
; GISEL-NEXT: v_lshrrev_b64 v[2:3], v2, v[6:7]
; GISEL-NEXT: v_lshlrev_b64 v[4:5], v4, v[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v12, v9, v[0:1]
; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v13, v9, v[0:1]
; GISEL-NEXT: v_cndmask_b32_e32 v2, v4, v2, vcc
; GISEL-NEXT: v_cmp_eq_u32_e64 s[6:7], 0, v10
; GISEL-NEXT: v_cndmask_b32_e64 v10, v2, 0, s[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v10, v8, v[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[0:1], s[8:9], v11, v8, 0
; GISEL-NEXT: v_mov_b32_e32 v2, v6
; GISEL-NEXT: v_mul_lo_u32 v6, v11, v9
; GISEL-NEXT: v_mad_u64_u32 v[1:2], s[8:9], v11, v9, v[1:2]
; GISEL-NEXT: v_mul_lo_u32 v4, v12, v9
; GISEL-NEXT: v_cndmask_b32_e64 v14, v2, 0, s[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[10:11], s[8:9], v14, v8, v[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[0:1], s[8:9], v12, v8, 0
; GISEL-NEXT: v_mov_b32_e32 v2, v10
; GISEL-NEXT: v_mul_lo_u32 v10, v12, v9
; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v12, v9, v[1:2]
; GISEL-NEXT: v_mul_lo_u32 v4, v13, v9
; GISEL-NEXT: v_cndmask_b32_e32 v3, v5, v3, vcc
; GISEL-NEXT: v_mad_u64_u32 v[1:2], s[10:11], v12, v8, v[1:2]
; GISEL-NEXT: v_addc_co_u32_e64 v6, s[10:11], v7, v6, s[10:11]
; GISEL-NEXT: v_mad_u64_u32 v[1:2], s[10:11], v13, v8, v[6:7]
; GISEL-NEXT: v_addc_co_u32_e64 v6, s[10:11], v11, v10, s[10:11]
; GISEL-NEXT: v_addc_co_u32_e64 v4, s[8:9], v6, v4, s[8:9]
; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v10, v9, v[4:5]
; GISEL-NEXT: v_cndmask_b32_e64 v3, v3, 0, s[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v14, v9, v[4:5]
; GISEL-NEXT: v_cndmask_b32_e64 v5, v3, 0, s[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[3:4], s[6:7], v5, v8, v[6:7]
; GISEL-NEXT: ; implicit-def: $vgpr5
; GISEL-NEXT: v_mad_u64_u32 v[3:4], s[6:7], v3, v8, v[6:7]
; GISEL-NEXT: ; implicit-def: $vgpr6_vgpr7
; GISEL-NEXT: ; implicit-def: $vgpr8
; GISEL-NEXT: .LBB6_4: ; %Flow
@ -2050,31 +2056,31 @@ define i128 @fptoui_bf16_to_i128(bfloat %x) {
; GISEL-NEXT: v_add_u32_e32 v10, 0xffffff7a, v5
; GISEL-NEXT: v_lshlrev_b64 v[0:1], v10, v[6:7]
; GISEL-NEXT: v_cmp_gt_u32_e32 vcc, 64, v10
; GISEL-NEXT: v_cndmask_b32_e32 v11, 0, v0, vcc
; GISEL-NEXT: v_cndmask_b32_e32 v12, 0, v1, vcc
; GISEL-NEXT: v_mad_u64_u32 v[0:1], s[6:7], v11, v9, 0
; GISEL-NEXT: v_cndmask_b32_e32 v12, 0, v0, vcc
; GISEL-NEXT: v_cndmask_b32_e32 v13, 0, v1, vcc
; GISEL-NEXT: v_mad_u64_u32 v[0:1], s[6:7], v12, v9, 0
; GISEL-NEXT: v_add_u32_e32 v4, 0xffffff3a, v5
; GISEL-NEXT: v_sub_u32_e32 v2, 64, v10
; GISEL-NEXT: v_lshrrev_b64 v[2:3], v2, v[6:7]
; GISEL-NEXT: v_lshlrev_b64 v[4:5], v4, v[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v12, v9, v[0:1]
; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v13, v9, v[0:1]
; GISEL-NEXT: v_cndmask_b32_e32 v2, v4, v2, vcc
; GISEL-NEXT: v_cmp_eq_u32_e64 s[6:7], 0, v10
; GISEL-NEXT: v_cndmask_b32_e64 v10, v2, 0, s[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v10, v8, v[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[0:1], s[8:9], v11, v8, 0
; GISEL-NEXT: v_mov_b32_e32 v2, v6
; GISEL-NEXT: v_mul_lo_u32 v6, v11, v9
; GISEL-NEXT: v_mad_u64_u32 v[1:2], s[8:9], v11, v9, v[1:2]
; GISEL-NEXT: v_mul_lo_u32 v4, v12, v9
; GISEL-NEXT: v_cndmask_b32_e64 v14, v2, 0, s[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[10:11], s[8:9], v14, v8, v[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[0:1], s[8:9], v12, v8, 0
; GISEL-NEXT: v_mov_b32_e32 v2, v10
; GISEL-NEXT: v_mul_lo_u32 v10, v12, v9
; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v12, v9, v[1:2]
; GISEL-NEXT: v_mul_lo_u32 v4, v13, v9
; GISEL-NEXT: v_cndmask_b32_e32 v3, v5, v3, vcc
; GISEL-NEXT: v_mad_u64_u32 v[1:2], s[10:11], v12, v8, v[1:2]
; GISEL-NEXT: v_addc_co_u32_e64 v6, s[10:11], v7, v6, s[10:11]
; GISEL-NEXT: v_mad_u64_u32 v[1:2], s[10:11], v13, v8, v[6:7]
; GISEL-NEXT: v_addc_co_u32_e64 v6, s[10:11], v11, v10, s[10:11]
; GISEL-NEXT: v_addc_co_u32_e64 v4, s[8:9], v6, v4, s[8:9]
; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v10, v9, v[4:5]
; GISEL-NEXT: v_cndmask_b32_e64 v3, v3, 0, s[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[8:9], v14, v9, v[4:5]
; GISEL-NEXT: v_cndmask_b32_e64 v5, v3, 0, s[6:7]
; GISEL-NEXT: v_mad_u64_u32 v[3:4], s[6:7], v5, v8, v[6:7]
; GISEL-NEXT: ; implicit-def: $vgpr5
; GISEL-NEXT: v_mad_u64_u32 v[3:4], s[6:7], v3, v8, v[6:7]
; GISEL-NEXT: ; implicit-def: $vgpr6_vgpr7
; GISEL-NEXT: ; implicit-def: $vgpr8
; GISEL-NEXT: .LBB7_4: ; %Flow

View File

@ -5775,28 +5775,28 @@ define i64 @clpeak_imad_pat_i64(i64 %x, i64 %y) {
; GFX7-GISEL-LABEL: clpeak_imad_pat_i64:
; GFX7-GISEL: ; %bb.0: ; %entry
; GFX7-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-GISEL-NEXT: v_add_i32_e32 v6, vcc, 1, v0
; GFX7-GISEL-NEXT: v_addc_u32_e32 v7, vcc, 0, v1, vcc
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v6, v2, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v6, v3, v[1:2]
; GFX7-GISEL-NEXT: v_add_i32_e32 v8, vcc, v0, v6
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v7, v2, v[4:5]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v8, v2, 0
; GFX7-GISEL-NEXT: v_addc_u32_e32 v9, vcc, v4, v7, vcc
; GFX7-GISEL-NEXT: v_mov_b32_e32 v1, v6
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v8, v3, v[1:2]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v9, v2, v[6:7]
; GFX7-GISEL-NEXT: v_add_i32_e32 v6, vcc, 1, v0
; GFX7-GISEL-NEXT: v_addc_u32_e32 v1, vcc, 0, v4, vcc
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v5, v6, 0
; GFX7-GISEL-NEXT: v_add_i32_e32 v7, vcc, 1, v5
; GFX7-GISEL-NEXT: v_mov_b32_e32 v0, v4
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v5, v1, v[0:1]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v3, v7, 0
; GFX7-GISEL-NEXT: v_addc_u32_e32 v8, vcc, 0, v2, vcc
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v2, v6, v[4:5]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v3, v8, v[1:2]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v4, v7, v[1:2]
; GFX7-GISEL-NEXT: v_add_i32_e32 v8, vcc, 1, v0
; GFX7-GISEL-NEXT: v_addc_u32_e32 v9, vcc, 0, v1, vcc
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v8, v2, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v8, v3, v[1:2]
; GFX7-GISEL-NEXT: v_add_i32_e32 v10, vcc, v0, v8
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v9, v2, v[4:5]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v10, v2, 0
; GFX7-GISEL-NEXT: v_addc_u32_e32 v11, vcc, v6, v9, vcc
; GFX7-GISEL-NEXT: v_mov_b32_e32 v1, v5
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[7:8], s[4:5], v10, v3, v[1:2]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v11, v2, v[7:8]
; GFX7-GISEL-NEXT: v_add_i32_e32 v7, vcc, 1, v0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v4, v7, 0
; GFX7-GISEL-NEXT: v_addc_u32_e32 v1, vcc, 0, v6, vcc
; GFX7-GISEL-NEXT: v_mov_b32_e32 v0, v3
; GFX7-GISEL-NEXT: v_add_i32_e32 v8, vcc, 1, v4
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v4, v1, v[0:1]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v2, v8, 0
; GFX7-GISEL-NEXT: v_addc_u32_e32 v10, vcc, 0, v9, vcc
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v9, v7, v[5:6]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v2, v10, v[1:2]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v3, v8, v[4:5]
; GFX7-GISEL-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-SDAG-LABEL: clpeak_imad_pat_i64:
@ -5831,28 +5831,28 @@ define i64 @clpeak_imad_pat_i64(i64 %x, i64 %y) {
; GFX8-GISEL-LABEL: clpeak_imad_pat_i64:
; GFX8-GISEL: ; %bb.0: ; %entry
; GFX8-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-GISEL-NEXT: v_add_u32_e32 v6, vcc, 1, v0
; GFX8-GISEL-NEXT: v_addc_u32_e32 v7, vcc, 0, v1, vcc
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v6, v2, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v6, v3, v[1:2]
; GFX8-GISEL-NEXT: v_add_u32_e32 v8, vcc, v0, v6
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v7, v2, v[4:5]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v8, v2, 0
; GFX8-GISEL-NEXT: v_addc_u32_e32 v9, vcc, v4, v7, vcc
; GFX8-GISEL-NEXT: v_mov_b32_e32 v1, v6
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v8, v3, v[1:2]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v9, v2, v[6:7]
; GFX8-GISEL-NEXT: v_add_u32_e32 v6, vcc, 1, v0
; GFX8-GISEL-NEXT: v_addc_u32_e32 v1, vcc, 0, v4, vcc
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v5, v6, 0
; GFX8-GISEL-NEXT: v_add_u32_e32 v7, vcc, 1, v5
; GFX8-GISEL-NEXT: v_mov_b32_e32 v0, v4
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v5, v1, v[0:1]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v3, v7, 0
; GFX8-GISEL-NEXT: v_addc_u32_e32 v8, vcc, 0, v2, vcc
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v2, v6, v[4:5]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v3, v8, v[1:2]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v4, v7, v[1:2]
; GFX8-GISEL-NEXT: v_add_u32_e32 v8, vcc, 1, v0
; GFX8-GISEL-NEXT: v_addc_u32_e32 v9, vcc, 0, v1, vcc
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v8, v2, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v8, v3, v[1:2]
; GFX8-GISEL-NEXT: v_add_u32_e32 v10, vcc, v0, v8
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v9, v2, v[4:5]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v10, v2, 0
; GFX8-GISEL-NEXT: v_addc_u32_e32 v11, vcc, v6, v9, vcc
; GFX8-GISEL-NEXT: v_mov_b32_e32 v1, v5
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[7:8], s[4:5], v10, v3, v[1:2]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v11, v2, v[7:8]
; GFX8-GISEL-NEXT: v_add_u32_e32 v7, vcc, 1, v0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v4, v7, 0
; GFX8-GISEL-NEXT: v_addc_u32_e32 v1, vcc, 0, v6, vcc
; GFX8-GISEL-NEXT: v_mov_b32_e32 v0, v3
; GFX8-GISEL-NEXT: v_add_u32_e32 v8, vcc, 1, v4
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v4, v1, v[0:1]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v2, v8, 0
; GFX8-GISEL-NEXT: v_addc_u32_e32 v10, vcc, 0, v9, vcc
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v9, v7, v[5:6]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v2, v10, v[1:2]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v3, v8, v[4:5]
; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
;
; GFX900-SDAG-LABEL: clpeak_imad_pat_i64:
@ -5883,28 +5883,28 @@ define i64 @clpeak_imad_pat_i64(i64 %x, i64 %y) {
; GFX900-GISEL-LABEL: clpeak_imad_pat_i64:
; GFX900-GISEL: ; %bb.0: ; %entry
; GFX900-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX900-GISEL-NEXT: v_add_co_u32_e32 v6, vcc, 1, v0
; GFX900-GISEL-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v1, vcc
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v6, v2, 0
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v6, v3, v[1:2]
; GFX900-GISEL-NEXT: v_add_co_u32_e32 v8, vcc, v0, v6
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v7, v2, v[4:5]
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v8, v2, 0
; GFX900-GISEL-NEXT: v_addc_co_u32_e32 v9, vcc, v4, v7, vcc
; GFX900-GISEL-NEXT: v_mov_b32_e32 v1, v6
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v8, v3, v[1:2]
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v9, v2, v[6:7]
; GFX900-GISEL-NEXT: v_add_co_u32_e32 v6, vcc, 1, v0
; GFX900-GISEL-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v4, vcc
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v5, v6, 0
; GFX900-GISEL-NEXT: v_add_co_u32_e32 v7, vcc, 1, v5
; GFX900-GISEL-NEXT: v_mov_b32_e32 v0, v4
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v5, v1, v[0:1]
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v3, v7, 0
; GFX900-GISEL-NEXT: v_addc_co_u32_e32 v8, vcc, 0, v2, vcc
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v2, v6, v[4:5]
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v3, v8, v[1:2]
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v4, v7, v[1:2]
; GFX900-GISEL-NEXT: v_add_co_u32_e32 v8, vcc, 1, v0
; GFX900-GISEL-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v1, vcc
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v8, v2, 0
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v8, v3, v[1:2]
; GFX900-GISEL-NEXT: v_add_co_u32_e32 v10, vcc, v0, v8
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v9, v2, v[4:5]
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v10, v2, 0
; GFX900-GISEL-NEXT: v_addc_co_u32_e32 v11, vcc, v6, v9, vcc
; GFX900-GISEL-NEXT: v_mov_b32_e32 v1, v5
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[7:8], s[4:5], v10, v3, v[1:2]
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v11, v2, v[7:8]
; GFX900-GISEL-NEXT: v_add_co_u32_e32 v7, vcc, 1, v0
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v4, v7, 0
; GFX900-GISEL-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v6, vcc
; GFX900-GISEL-NEXT: v_mov_b32_e32 v0, v3
; GFX900-GISEL-NEXT: v_add_co_u32_e32 v8, vcc, 1, v4
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v4, v1, v[0:1]
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v2, v8, 0
; GFX900-GISEL-NEXT: v_addc_co_u32_e32 v10, vcc, 0, v9, vcc
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v9, v7, v[5:6]
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v2, v10, v[1:2]
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v3, v8, v[4:5]
; GFX900-GISEL-NEXT: s_setpc_b64 s[30:31]
;
; GFX90A-SDAG-LABEL: clpeak_imad_pat_i64:
@ -5935,29 +5935,29 @@ define i64 @clpeak_imad_pat_i64(i64 %x, i64 %y) {
; GFX90A-GISEL-LABEL: clpeak_imad_pat_i64:
; GFX90A-GISEL: ; %bb.0: ; %entry
; GFX90A-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-GISEL-NEXT: v_add_co_u32_e32 v6, vcc, 1, v0
; GFX90A-GISEL-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v1, vcc
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v6, v3, 0
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v6, v2, 0
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v7, v2, v[4:5]
; GFX90A-GISEL-NEXT: v_add_u32_e32 v1, v1, v4
; GFX90A-GISEL-NEXT: v_add_co_u32_e32 v6, vcc, v0, v6
; GFX90A-GISEL-NEXT: v_addc_co_u32_e32 v8, vcc, v1, v7, vcc
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v6, v2, 0
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v6, v3, 0
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v8, v2, v[6:7]
; GFX90A-GISEL-NEXT: v_add_co_u32_e32 v6, vcc, 1, v0
; GFX90A-GISEL-NEXT: v_addc_co_u32_e32 v0, vcc, 0, v1, vcc
; GFX90A-GISEL-NEXT: v_add_u32_e32 v5, v5, v2
; GFX90A-GISEL-NEXT: v_add_co_u32_e32 v7, vcc, 1, v4
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v4, v0, 0
; GFX90A-GISEL-NEXT: v_addc_co_u32_e32 v8, vcc, 0, v5, vcc
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v4, v6, 0
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v5, v6, v[0:1]
; GFX90A-GISEL-NEXT: v_add_u32_e32 v4, v3, v0
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v2, v7, 0
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v2, v8, 0
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v4, v7, v[2:3]
; GFX90A-GISEL-NEXT: v_add_co_u32_e32 v8, vcc, 1, v0
; GFX90A-GISEL-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v1, vcc
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v8, v3, 0
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v8, v2, 0
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v9, v2, v[4:5]
; GFX90A-GISEL-NEXT: v_add_u32_e32 v1, v1, v6
; GFX90A-GISEL-NEXT: v_add_co_u32_e32 v8, vcc, v0, v8
; GFX90A-GISEL-NEXT: v_addc_co_u32_e32 v10, vcc, v1, v9, vcc
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v8, v3, 0
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v8, v2, 0
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v10, v2, v[6:7]
; GFX90A-GISEL-NEXT: v_add_co_u32_e32 v7, vcc, 1, v0
; GFX90A-GISEL-NEXT: v_add_u32_e32 v6, v5, v8
; GFX90A-GISEL-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v1, vcc
; GFX90A-GISEL-NEXT: v_add_co_u32_e32 v8, vcc, 1, v4
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v4, v5, 0
; GFX90A-GISEL-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v6, vcc
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v4, v7, 0
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v6, v7, v[0:1]
; GFX90A-GISEL-NEXT: v_add_u32_e32 v6, v3, v4
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v2, v9, 0
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v2, v8, 0
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v6, v8, v[4:5]
; GFX90A-GISEL-NEXT: v_add_u32_e32 v1, v1, v2
; GFX90A-GISEL-NEXT: s_setpc_b64 s[30:31]
;
@ -6408,52 +6408,52 @@ define <2 x i64> @clpeak_imad_pat_v2i64(<2 x i64> %x, <2 x i64> %y) {
; GFX7-GISEL-LABEL: clpeak_imad_pat_v2i64:
; GFX7-GISEL: ; %bb.0: ; %entry
; GFX7-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-GISEL-NEXT: v_add_i32_e32 v12, vcc, 1, v0
; GFX7-GISEL-NEXT: v_addc_u32_e32 v13, vcc, 0, v1, vcc
; GFX7-GISEL-NEXT: v_add_i32_e32 v14, vcc, 1, v2
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v12, v4, 0
; GFX7-GISEL-NEXT: v_addc_u32_e32 v15, vcc, 0, v3, vcc
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v14, v6, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v12, v5, v[1:2]
; GFX7-GISEL-NEXT: v_add_i32_e32 v14, vcc, 1, v0
; GFX7-GISEL-NEXT: v_addc_u32_e32 v15, vcc, 0, v1, vcc
; GFX7-GISEL-NEXT: v_add_i32_e32 v16, vcc, 1, v2
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v14, v4, 0
; GFX7-GISEL-NEXT: v_addc_u32_e32 v17, vcc, 0, v3, vcc
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v16, v6, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v14, v5, v[1:2]
; GFX7-GISEL-NEXT: v_mov_b32_e32 v1, v3
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v14, v7, v[1:2]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v13, v4, v[8:9]
; GFX7-GISEL-NEXT: v_add_i32_e32 v3, vcc, v0, v12
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v15, v6, v[10:11]
; GFX7-GISEL-NEXT: v_addc_u32_e32 v16, vcc, v8, v13, vcc
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v3, v4, 0
; GFX7-GISEL-NEXT: v_add_i32_e32 v17, vcc, v2, v14
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[13:14], s[4:5], v17, v6, 0
; GFX7-GISEL-NEXT: v_mov_b32_e32 v1, v11
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[11:12], s[4:5], v3, v5, v[1:2]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v16, v7, v[1:2]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[12:13], s[4:5], v15, v4, v[8:9]
; GFX7-GISEL-NEXT: v_add_i32_e32 v3, vcc, v0, v14
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v17, v6, v[10:11]
; GFX7-GISEL-NEXT: v_addc_u32_e32 v18, vcc, v12, v15, vcc
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v3, v4, 0
; GFX7-GISEL-NEXT: v_add_i32_e32 v16, vcc, v2, v16
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[13:14], s[4:5], v16, v6, 0
; GFX7-GISEL-NEXT: v_mov_b32_e32 v1, v10
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v3, v5, v[1:2]
; GFX7-GISEL-NEXT: v_mov_b32_e32 v1, v14
; GFX7-GISEL-NEXT: v_addc_u32_e32 v18, vcc, v9, v15, vcc
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[14:15], s[4:5], v17, v7, v[1:2]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v16, v4, v[11:12]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v18, v6, v[14:15]
; GFX7-GISEL-NEXT: v_add_i32_e32 v11, vcc, 1, v0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v10, v11, 0
; GFX7-GISEL-NEXT: v_addc_u32_e32 v1, vcc, 0, v8, vcc
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[14:15], s[4:5], v16, v7, v[1:2]
; GFX7-GISEL-NEXT: v_addc_u32_e32 v19, vcc, v8, v17, vcc
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[16:17], s[4:5], v18, v4, v[10:11]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v19, v6, v[14:15]
; GFX7-GISEL-NEXT: v_add_i32_e32 v10, vcc, 1, v0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v9, v10, 0
; GFX7-GISEL-NEXT: v_addc_u32_e32 v6, vcc, 0, v12, vcc
; GFX7-GISEL-NEXT: v_add_i32_e32 v12, vcc, 1, v2
; GFX7-GISEL-NEXT: v_mov_b32_e32 v0, v5
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v9, v6, v[0:1]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v13, v12, 0
; GFX7-GISEL-NEXT: v_addc_u32_e32 v11, vcc, 0, v8, vcc
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[7:8], s[4:5], v16, v10, v[1:2]
; GFX7-GISEL-NEXT: v_mov_b32_e32 v0, v6
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v10, v1, v[0:1]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v13, v12, 0
; GFX7-GISEL-NEXT: v_addc_u32_e32 v2, vcc, 0, v9, vcc
; GFX7-GISEL-NEXT: v_add_i32_e32 v14, vcc, 1, v10
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v3, v11, v[0:1]
; GFX7-GISEL-NEXT: v_mov_b32_e32 v0, v7
; GFX7-GISEL-NEXT: v_addc_u32_e32 v15, vcc, 0, v3, vcc
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v13, v2, v[0:1]
; GFX7-GISEL-NEXT: v_add_i32_e32 v16, vcc, 1, v13
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v5, v14, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v4, v12, v[2:3]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v6, v16, 0
; GFX7-GISEL-NEXT: v_addc_u32_e32 v17, vcc, 0, v4, vcc
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v5, v15, v[1:2]
; GFX7-GISEL-NEXT: v_add_i32_e32 v14, vcc, 1, v9
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v13, v11, v[0:1]
; GFX7-GISEL-NEXT: v_addc_u32_e32 v15, vcc, 0, v16, vcc
; GFX7-GISEL-NEXT: v_add_i32_e32 v17, vcc, 1, v13
; GFX7-GISEL-NEXT: v_addc_u32_e32 v18, vcc, 0, v3, vcc
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v4, v14, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v3, v12, v[8:9]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v5, v17, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v4, v15, v[1:2]
; GFX7-GISEL-NEXT: v_mov_b32_e32 v1, v3
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v6, v17, v[1:2]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v8, v14, v[4:5]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v9, v16, v[10:11]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[11:12], s[4:5], v5, v18, v[1:2]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v7, v14, v[8:9]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v10, v17, v[11:12]
; GFX7-GISEL-NEXT: v_mov_b32_e32 v1, v5
; GFX7-GISEL-NEXT: s_setpc_b64 s[30:31]
;
@ -6513,52 +6513,52 @@ define <2 x i64> @clpeak_imad_pat_v2i64(<2 x i64> %x, <2 x i64> %y) {
; GFX8-GISEL-LABEL: clpeak_imad_pat_v2i64:
; GFX8-GISEL: ; %bb.0: ; %entry
; GFX8-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-GISEL-NEXT: v_add_u32_e32 v12, vcc, 1, v0
; GFX8-GISEL-NEXT: v_addc_u32_e32 v13, vcc, 0, v1, vcc
; GFX8-GISEL-NEXT: v_add_u32_e32 v14, vcc, 1, v2
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v12, v4, 0
; GFX8-GISEL-NEXT: v_addc_u32_e32 v15, vcc, 0, v3, vcc
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v14, v6, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v12, v5, v[1:2]
; GFX8-GISEL-NEXT: v_add_u32_e32 v14, vcc, 1, v0
; GFX8-GISEL-NEXT: v_addc_u32_e32 v15, vcc, 0, v1, vcc
; GFX8-GISEL-NEXT: v_add_u32_e32 v16, vcc, 1, v2
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v14, v4, 0
; GFX8-GISEL-NEXT: v_addc_u32_e32 v17, vcc, 0, v3, vcc
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v16, v6, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v14, v5, v[1:2]
; GFX8-GISEL-NEXT: v_mov_b32_e32 v1, v3
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v14, v7, v[1:2]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v13, v4, v[8:9]
; GFX8-GISEL-NEXT: v_add_u32_e32 v3, vcc, v0, v12
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v15, v6, v[10:11]
; GFX8-GISEL-NEXT: v_addc_u32_e32 v16, vcc, v8, v13, vcc
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v3, v4, 0
; GFX8-GISEL-NEXT: v_add_u32_e32 v17, vcc, v2, v14
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[13:14], s[4:5], v17, v6, 0
; GFX8-GISEL-NEXT: v_mov_b32_e32 v1, v11
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[11:12], s[4:5], v3, v5, v[1:2]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v16, v7, v[1:2]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[12:13], s[4:5], v15, v4, v[8:9]
; GFX8-GISEL-NEXT: v_add_u32_e32 v3, vcc, v0, v14
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v17, v6, v[10:11]
; GFX8-GISEL-NEXT: v_addc_u32_e32 v18, vcc, v12, v15, vcc
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v3, v4, 0
; GFX8-GISEL-NEXT: v_add_u32_e32 v16, vcc, v2, v16
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[13:14], s[4:5], v16, v6, 0
; GFX8-GISEL-NEXT: v_mov_b32_e32 v1, v10
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v3, v5, v[1:2]
; GFX8-GISEL-NEXT: v_mov_b32_e32 v1, v14
; GFX8-GISEL-NEXT: v_addc_u32_e32 v18, vcc, v9, v15, vcc
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[14:15], s[4:5], v17, v7, v[1:2]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v16, v4, v[11:12]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v18, v6, v[14:15]
; GFX8-GISEL-NEXT: v_add_u32_e32 v11, vcc, 1, v0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v10, v11, 0
; GFX8-GISEL-NEXT: v_addc_u32_e32 v1, vcc, 0, v8, vcc
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[14:15], s[4:5], v16, v7, v[1:2]
; GFX8-GISEL-NEXT: v_addc_u32_e32 v19, vcc, v8, v17, vcc
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[16:17], s[4:5], v18, v4, v[10:11]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v19, v6, v[14:15]
; GFX8-GISEL-NEXT: v_add_u32_e32 v10, vcc, 1, v0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v9, v10, 0
; GFX8-GISEL-NEXT: v_addc_u32_e32 v6, vcc, 0, v12, vcc
; GFX8-GISEL-NEXT: v_add_u32_e32 v12, vcc, 1, v2
; GFX8-GISEL-NEXT: v_mov_b32_e32 v0, v5
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v9, v6, v[0:1]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v13, v12, 0
; GFX8-GISEL-NEXT: v_addc_u32_e32 v11, vcc, 0, v8, vcc
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[7:8], s[4:5], v16, v10, v[1:2]
; GFX8-GISEL-NEXT: v_mov_b32_e32 v0, v6
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v10, v1, v[0:1]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v13, v12, 0
; GFX8-GISEL-NEXT: v_addc_u32_e32 v2, vcc, 0, v9, vcc
; GFX8-GISEL-NEXT: v_add_u32_e32 v14, vcc, 1, v10
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v3, v11, v[0:1]
; GFX8-GISEL-NEXT: v_mov_b32_e32 v0, v7
; GFX8-GISEL-NEXT: v_addc_u32_e32 v15, vcc, 0, v3, vcc
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v13, v2, v[0:1]
; GFX8-GISEL-NEXT: v_add_u32_e32 v16, vcc, 1, v13
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v5, v14, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v4, v12, v[2:3]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v6, v16, 0
; GFX8-GISEL-NEXT: v_addc_u32_e32 v17, vcc, 0, v4, vcc
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v5, v15, v[1:2]
; GFX8-GISEL-NEXT: v_add_u32_e32 v14, vcc, 1, v9
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v13, v11, v[0:1]
; GFX8-GISEL-NEXT: v_addc_u32_e32 v15, vcc, 0, v16, vcc
; GFX8-GISEL-NEXT: v_add_u32_e32 v17, vcc, 1, v13
; GFX8-GISEL-NEXT: v_addc_u32_e32 v18, vcc, 0, v3, vcc
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v4, v14, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v3, v12, v[8:9]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v5, v17, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v4, v15, v[1:2]
; GFX8-GISEL-NEXT: v_mov_b32_e32 v1, v3
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v6, v17, v[1:2]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v8, v14, v[4:5]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v9, v16, v[10:11]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[11:12], s[4:5], v5, v18, v[1:2]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v7, v14, v[8:9]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v10, v17, v[11:12]
; GFX8-GISEL-NEXT: v_mov_b32_e32 v1, v5
; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
;
@ -6610,52 +6610,52 @@ define <2 x i64> @clpeak_imad_pat_v2i64(<2 x i64> %x, <2 x i64> %y) {
; GFX900-GISEL-LABEL: clpeak_imad_pat_v2i64:
; GFX900-GISEL: ; %bb.0: ; %entry
; GFX900-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX900-GISEL-NEXT: v_add_co_u32_e32 v12, vcc, 1, v0
; GFX900-GISEL-NEXT: v_addc_co_u32_e32 v13, vcc, 0, v1, vcc
; GFX900-GISEL-NEXT: v_add_co_u32_e32 v14, vcc, 1, v2
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v12, v4, 0
; GFX900-GISEL-NEXT: v_addc_co_u32_e32 v15, vcc, 0, v3, vcc
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v14, v6, 0
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v12, v5, v[1:2]
; GFX900-GISEL-NEXT: v_add_co_u32_e32 v14, vcc, 1, v0
; GFX900-GISEL-NEXT: v_addc_co_u32_e32 v15, vcc, 0, v1, vcc
; GFX900-GISEL-NEXT: v_add_co_u32_e32 v16, vcc, 1, v2
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v14, v4, 0
; GFX900-GISEL-NEXT: v_addc_co_u32_e32 v17, vcc, 0, v3, vcc
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v16, v6, 0
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v14, v5, v[1:2]
; GFX900-GISEL-NEXT: v_mov_b32_e32 v1, v3
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v14, v7, v[1:2]
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v13, v4, v[8:9]
; GFX900-GISEL-NEXT: v_add_co_u32_e32 v3, vcc, v0, v12
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v15, v6, v[10:11]
; GFX900-GISEL-NEXT: v_addc_co_u32_e32 v16, vcc, v8, v13, vcc
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v3, v4, 0
; GFX900-GISEL-NEXT: v_add_co_u32_e32 v17, vcc, v2, v14
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[13:14], s[4:5], v17, v6, 0
; GFX900-GISEL-NEXT: v_mov_b32_e32 v1, v11
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[11:12], s[4:5], v3, v5, v[1:2]
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v16, v7, v[1:2]
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[12:13], s[4:5], v15, v4, v[8:9]
; GFX900-GISEL-NEXT: v_add_co_u32_e32 v3, vcc, v0, v14
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v17, v6, v[10:11]
; GFX900-GISEL-NEXT: v_addc_co_u32_e32 v18, vcc, v12, v15, vcc
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v3, v4, 0
; GFX900-GISEL-NEXT: v_add_co_u32_e32 v16, vcc, v2, v16
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[13:14], s[4:5], v16, v6, 0
; GFX900-GISEL-NEXT: v_mov_b32_e32 v1, v10
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v3, v5, v[1:2]
; GFX900-GISEL-NEXT: v_mov_b32_e32 v1, v14
; GFX900-GISEL-NEXT: v_addc_co_u32_e32 v18, vcc, v9, v15, vcc
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[14:15], s[4:5], v17, v7, v[1:2]
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v16, v4, v[11:12]
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v18, v6, v[14:15]
; GFX900-GISEL-NEXT: v_add_co_u32_e32 v11, vcc, 1, v0
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v10, v11, 0
; GFX900-GISEL-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v8, vcc
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[14:15], s[4:5], v16, v7, v[1:2]
; GFX900-GISEL-NEXT: v_addc_co_u32_e32 v19, vcc, v8, v17, vcc
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[16:17], s[4:5], v18, v4, v[10:11]
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v19, v6, v[14:15]
; GFX900-GISEL-NEXT: v_add_co_u32_e32 v10, vcc, 1, v0
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v9, v10, 0
; GFX900-GISEL-NEXT: v_addc_co_u32_e32 v6, vcc, 0, v12, vcc
; GFX900-GISEL-NEXT: v_add_co_u32_e32 v12, vcc, 1, v2
; GFX900-GISEL-NEXT: v_mov_b32_e32 v0, v5
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v9, v6, v[0:1]
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v13, v12, 0
; GFX900-GISEL-NEXT: v_addc_co_u32_e32 v11, vcc, 0, v8, vcc
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[7:8], s[4:5], v16, v10, v[1:2]
; GFX900-GISEL-NEXT: v_mov_b32_e32 v0, v6
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v10, v1, v[0:1]
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v13, v12, 0
; GFX900-GISEL-NEXT: v_addc_co_u32_e32 v2, vcc, 0, v9, vcc
; GFX900-GISEL-NEXT: v_add_co_u32_e32 v14, vcc, 1, v10
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v3, v11, v[0:1]
; GFX900-GISEL-NEXT: v_mov_b32_e32 v0, v7
; GFX900-GISEL-NEXT: v_addc_co_u32_e32 v15, vcc, 0, v3, vcc
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v13, v2, v[0:1]
; GFX900-GISEL-NEXT: v_add_co_u32_e32 v16, vcc, 1, v13
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v5, v14, 0
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v4, v12, v[2:3]
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v6, v16, 0
; GFX900-GISEL-NEXT: v_addc_co_u32_e32 v17, vcc, 0, v4, vcc
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v5, v15, v[1:2]
; GFX900-GISEL-NEXT: v_add_co_u32_e32 v14, vcc, 1, v9
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v13, v11, v[0:1]
; GFX900-GISEL-NEXT: v_addc_co_u32_e32 v15, vcc, 0, v16, vcc
; GFX900-GISEL-NEXT: v_add_co_u32_e32 v17, vcc, 1, v13
; GFX900-GISEL-NEXT: v_addc_co_u32_e32 v18, vcc, 0, v3, vcc
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v4, v14, 0
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v3, v12, v[8:9]
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v5, v17, 0
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v4, v15, v[1:2]
; GFX900-GISEL-NEXT: v_mov_b32_e32 v1, v3
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v6, v17, v[1:2]
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v8, v14, v[4:5]
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v9, v16, v[10:11]
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[11:12], s[4:5], v5, v18, v[1:2]
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v7, v14, v[8:9]
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v10, v17, v[11:12]
; GFX900-GISEL-NEXT: v_mov_b32_e32 v1, v5
; GFX900-GISEL-NEXT: s_setpc_b64 s[30:31]
;
@ -6707,54 +6707,54 @@ define <2 x i64> @clpeak_imad_pat_v2i64(<2 x i64> %x, <2 x i64> %y) {
; GFX90A-GISEL-LABEL: clpeak_imad_pat_v2i64:
; GFX90A-GISEL: ; %bb.0: ; %entry
; GFX90A-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-GISEL-NEXT: v_add_co_u32_e32 v10, vcc, 1, v0
; GFX90A-GISEL-NEXT: v_addc_co_u32_e32 v11, vcc, 0, v1, vcc
; GFX90A-GISEL-NEXT: v_add_co_u32_e32 v12, vcc, 1, v2
; GFX90A-GISEL-NEXT: v_addc_co_u32_e32 v13, vcc, 0, v3, vcc
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v10, v5, 0
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v10, v4, 0
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v11, v4, v[2:3]
; GFX90A-GISEL-NEXT: v_add_co_u32_e32 v12, vcc, 1, v0
; GFX90A-GISEL-NEXT: v_addc_co_u32_e32 v13, vcc, 0, v1, vcc
; GFX90A-GISEL-NEXT: v_add_co_u32_e32 v14, vcc, 1, v2
; GFX90A-GISEL-NEXT: v_addc_co_u32_e32 v15, vcc, 0, v3, vcc
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v12, v5, 0
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v12, v4, 0
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v13, v4, v[2:3]
; GFX90A-GISEL-NEXT: v_add_u32_e32 v1, v1, v8
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v14, v7, 0
; GFX90A-GISEL-NEXT: v_add_co_u32_e32 v12, vcc, v0, v12
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v14, v6, 0
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v15, v6, v[8:9]
; GFX90A-GISEL-NEXT: v_addc_co_u32_e32 v16, vcc, v1, v13, vcc
; GFX90A-GISEL-NEXT: v_add_u32_e32 v3, v3, v10
; GFX90A-GISEL-NEXT: v_add_co_u32_e32 v14, vcc, v2, v14
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v12, v5, 0
; GFX90A-GISEL-NEXT: v_addc_co_u32_e32 v15, vcc, v3, v15, vcc
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v12, v4, 0
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[12:13], s[4:5], v16, v4, v[10:11]
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v14, v7, 0
; GFX90A-GISEL-NEXT: v_add_u32_e32 v9, v9, v12
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v14, v6, 0
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[12:13], s[4:5], v15, v6, v[10:11]
; GFX90A-GISEL-NEXT: v_add_u32_e32 v10, v5, v12
; GFX90A-GISEL-NEXT: v_add_co_u32_e32 v5, vcc, 1, v0
; GFX90A-GISEL-NEXT: v_addc_co_u32_e32 v6, vcc, 0, v1, vcc
; GFX90A-GISEL-NEXT: v_add_co_u32_e32 v11, vcc, 1, v2
; GFX90A-GISEL-NEXT: v_addc_co_u32_e32 v12, vcc, 0, v3, vcc
; GFX90A-GISEL-NEXT: v_add_co_u32_e32 v13, vcc, 1, v8
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v8, v6, 0
; GFX90A-GISEL-NEXT: v_addc_co_u32_e32 v14, vcc, 0, v9, vcc
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v8, v5, 0
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v9, v5, v[0:1]
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v4, v12, 0
; GFX90A-GISEL-NEXT: v_add_co_u32_e32 v15, vcc, 1, v4
; GFX90A-GISEL-NEXT: v_add_u32_e32 v8, v3, v6
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v4, v11, 0
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v10, v11, v[0:1]
; GFX90A-GISEL-NEXT: v_addc_co_u32_e32 v16, vcc, 0, v10, vcc
; GFX90A-GISEL-NEXT: v_add_u32_e32 v9, v7, v4
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v2, v14, 0
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v2, v13, 0
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v8, v13, v[4:5]
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v6, v16, 0
; GFX90A-GISEL-NEXT: v_add_u32_e32 v1, v1, v2
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v12, v7, 0
; GFX90A-GISEL-NEXT: v_add_co_u32_e32 v10, vcc, v0, v10
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v12, v6, 0
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v13, v6, v[8:9]
; GFX90A-GISEL-NEXT: v_addc_co_u32_e32 v14, vcc, v1, v11, vcc
; GFX90A-GISEL-NEXT: v_add_u32_e32 v3, v3, v8
; GFX90A-GISEL-NEXT: v_add_co_u32_e32 v12, vcc, v2, v12
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v10, v4, 0
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v10, v5, 0
; GFX90A-GISEL-NEXT: v_addc_co_u32_e32 v13, vcc, v3, v13, vcc
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v14, v4, v[10:11]
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v12, v7, 0
; GFX90A-GISEL-NEXT: v_add_u32_e32 v9, v9, v4
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v12, v6, 0
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v13, v6, v[10:11]
; GFX90A-GISEL-NEXT: v_add_u32_e32 v5, v5, v6
; GFX90A-GISEL-NEXT: v_add_co_u32_e32 v6, vcc, 1, v0
; GFX90A-GISEL-NEXT: v_addc_co_u32_e32 v0, vcc, 0, v1, vcc
; GFX90A-GISEL-NEXT: v_add_co_u32_e32 v10, vcc, 1, v2
; GFX90A-GISEL-NEXT: v_addc_co_u32_e32 v11, vcc, 0, v3, vcc
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v8, v0, 0
; GFX90A-GISEL-NEXT: v_add_co_u32_e32 v12, vcc, 1, v8
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v8, v6, 0
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v9, v6, v[0:1]
; GFX90A-GISEL-NEXT: v_addc_co_u32_e32 v13, vcc, 0, v9, vcc
; GFX90A-GISEL-NEXT: v_add_u32_e32 v8, v3, v0
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v4, v11, 0
; GFX90A-GISEL-NEXT: v_add_co_u32_e32 v14, vcc, 1, v4
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v4, v10, 0
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v5, v10, v[0:1]
; GFX90A-GISEL-NEXT: v_addc_co_u32_e32 v15, vcc, 0, v5, vcc
; GFX90A-GISEL-NEXT: v_add_u32_e32 v7, v7, v0
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v2, v12, 0
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v2, v13, 0
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v8, v12, v[2:3]
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v6, v15, 0
; GFX90A-GISEL-NEXT: v_add_u32_e32 v1, v1, v2
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v6, v14, 0
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v7, v14, v[4:5]
; GFX90A-GISEL-NEXT: v_add_u32_e32 v3, v3, v4
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v6, v15, 0
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v9, v15, v[4:5]
; GFX90A-GISEL-NEXT: v_add_u32_e32 v3, v3, v6
; GFX90A-GISEL-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-SDAG-LABEL: clpeak_imad_pat_v2i64:

View File

@ -781,16 +781,23 @@ define amdgpu_cs_chain void @wwm_write_to_arg_reg(<3 x i32> inreg %sgpr, ptr inr
; GISEL12-NEXT: v_dual_mov_b32 v14, v38 :: v_dual_mov_b32 v15, v39
; GISEL12-NEXT: s_wait_kmcnt 0x0
; GISEL12-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GISEL12-NEXT: v_dual_mov_b32 v24, v0 :: v_dual_mov_b32 v25, v1
; GISEL12-NEXT: v_dual_mov_b32 v26, v2 :: v_dual_mov_b32 v27, v3
; GISEL12-NEXT: v_dual_mov_b32 v28, v4 :: v_dual_mov_b32 v29, v5
; GISEL12-NEXT: v_dual_mov_b32 v30, v6 :: v_dual_mov_b32 v31, v7
; GISEL12-NEXT: v_dual_mov_b32 v32, v8 :: v_dual_mov_b32 v33, v9
; GISEL12-NEXT: v_dual_mov_b32 v34, v10 :: v_dual_mov_b32 v35, v11
; GISEL12-NEXT: v_dual_mov_b32 v36, v12 :: v_dual_mov_b32 v37, v13
; GISEL12-NEXT: v_dual_mov_b32 v38, v14 :: v_dual_mov_b32 v39, v15
; GISEL12-NEXT: v_dual_mov_b32 v40, v0 :: v_dual_mov_b32 v41, v1
; GISEL12-NEXT: v_dual_mov_b32 v42, v2 :: v_dual_mov_b32 v43, v3
; GISEL12-NEXT: v_dual_mov_b32 v44, v4 :: v_dual_mov_b32 v45, v5
; GISEL12-NEXT: v_dual_mov_b32 v46, v6 :: v_dual_mov_b32 v47, v7
; GISEL12-NEXT: v_dual_mov_b32 v48, v8 :: v_dual_mov_b32 v49, v9
; GISEL12-NEXT: v_dual_mov_b32 v50, v10 :: v_dual_mov_b32 v51, v11
; GISEL12-NEXT: v_dual_mov_b32 v52, v12 :: v_dual_mov_b32 v53, v13
; GISEL12-NEXT: v_dual_mov_b32 v54, v14 :: v_dual_mov_b32 v55, v15
; GISEL12-NEXT: s_mov_b32 exec_lo, s9
; GISEL12-NEXT: ; kill: def $vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39 killed $vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39 killed $exec
; GISEL12-NEXT: v_dual_mov_b32 v24, v40 :: v_dual_mov_b32 v25, v41
; GISEL12-NEXT: v_dual_mov_b32 v26, v42 :: v_dual_mov_b32 v27, v43
; GISEL12-NEXT: v_dual_mov_b32 v28, v44 :: v_dual_mov_b32 v29, v45
; GISEL12-NEXT: v_dual_mov_b32 v30, v46 :: v_dual_mov_b32 v31, v47
; GISEL12-NEXT: v_dual_mov_b32 v32, v48 :: v_dual_mov_b32 v33, v49
; GISEL12-NEXT: v_dual_mov_b32 v34, v50 :: v_dual_mov_b32 v35, v51
; GISEL12-NEXT: v_dual_mov_b32 v36, v52 :: v_dual_mov_b32 v37, v53
; GISEL12-NEXT: v_dual_mov_b32 v38, v54 :: v_dual_mov_b32 v39, v55
; GISEL12-NEXT: .LBB5_2: ; %tail
; GISEL12-NEXT: s_wait_alu 0xfffe
; GISEL12-NEXT: s_or_b32 exec_lo, exec_lo, s4
@ -946,24 +953,39 @@ define amdgpu_cs_chain void @wwm_write_to_arg_reg(<3 x i32> inreg %sgpr, ptr inr
; GISEL10-NEXT: s_mov_b64 s[2:3], s[50:51]
; GISEL10-NEXT: s_waitcnt lgkmcnt(0)
; GISEL10-NEXT: s_swappc_b64 s[30:31], s[12:13]
; GISEL10-NEXT: v_mov_b32_e32 v24, v0
; GISEL10-NEXT: v_mov_b32_e32 v25, v1
; GISEL10-NEXT: v_mov_b32_e32 v26, v2
; GISEL10-NEXT: v_mov_b32_e32 v27, v3
; GISEL10-NEXT: v_mov_b32_e32 v28, v4
; GISEL10-NEXT: v_mov_b32_e32 v29, v5
; GISEL10-NEXT: v_mov_b32_e32 v30, v6
; GISEL10-NEXT: v_mov_b32_e32 v31, v7
; GISEL10-NEXT: v_mov_b32_e32 v32, v8
; GISEL10-NEXT: v_mov_b32_e32 v33, v9
; GISEL10-NEXT: v_mov_b32_e32 v34, v10
; GISEL10-NEXT: v_mov_b32_e32 v35, v11
; GISEL10-NEXT: v_mov_b32_e32 v36, v12
; GISEL10-NEXT: v_mov_b32_e32 v37, v13
; GISEL10-NEXT: v_mov_b32_e32 v38, v14
; GISEL10-NEXT: v_mov_b32_e32 v39, v15
; GISEL10-NEXT: v_mov_b32_e32 v40, v0
; GISEL10-NEXT: v_mov_b32_e32 v41, v1
; GISEL10-NEXT: v_mov_b32_e32 v42, v2
; GISEL10-NEXT: v_mov_b32_e32 v43, v3
; GISEL10-NEXT: v_mov_b32_e32 v44, v4
; GISEL10-NEXT: v_mov_b32_e32 v45, v5
; GISEL10-NEXT: v_mov_b32_e32 v46, v6
; GISEL10-NEXT: v_mov_b32_e32 v47, v7
; GISEL10-NEXT: v_mov_b32_e32 v48, v8
; GISEL10-NEXT: v_mov_b32_e32 v49, v9
; GISEL10-NEXT: v_mov_b32_e32 v50, v10
; GISEL10-NEXT: v_mov_b32_e32 v51, v11
; GISEL10-NEXT: v_mov_b32_e32 v52, v12
; GISEL10-NEXT: v_mov_b32_e32 v53, v13
; GISEL10-NEXT: v_mov_b32_e32 v54, v14
; GISEL10-NEXT: v_mov_b32_e32 v55, v15
; GISEL10-NEXT: s_mov_b32 exec_lo, s9
; GISEL10-NEXT: ; kill: def $vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39 killed $vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39 killed $exec
; GISEL10-NEXT: v_mov_b32_e32 v24, v40
; GISEL10-NEXT: v_mov_b32_e32 v25, v41
; GISEL10-NEXT: v_mov_b32_e32 v26, v42
; GISEL10-NEXT: v_mov_b32_e32 v27, v43
; GISEL10-NEXT: v_mov_b32_e32 v28, v44
; GISEL10-NEXT: v_mov_b32_e32 v29, v45
; GISEL10-NEXT: v_mov_b32_e32 v30, v46
; GISEL10-NEXT: v_mov_b32_e32 v31, v47
; GISEL10-NEXT: v_mov_b32_e32 v32, v48
; GISEL10-NEXT: v_mov_b32_e32 v33, v49
; GISEL10-NEXT: v_mov_b32_e32 v34, v50
; GISEL10-NEXT: v_mov_b32_e32 v35, v51
; GISEL10-NEXT: v_mov_b32_e32 v36, v52
; GISEL10-NEXT: v_mov_b32_e32 v37, v53
; GISEL10-NEXT: v_mov_b32_e32 v38, v54
; GISEL10-NEXT: v_mov_b32_e32 v39, v55
; GISEL10-NEXT: .LBB5_2: ; %tail
; GISEL10-NEXT: s_or_b32 exec_lo, exec_lo, s4
; GISEL10-NEXT: v_mov_b32_e32 v8, v24

View File

@ -2481,10 +2481,11 @@ define i64 @test_vector_reduce_mul_v2i64(<2 x i64> %v) {
; GFX7-GISEL: ; %bb.0: ; %entry
; GFX7-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-GISEL-NEXT: v_mov_b32_e32 v4, v0
; GFX7-GISEL-NEXT: v_mov_b32_e32 v5, v1
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v4, v2, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v4, v3, v[1:2]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v5, v2, v[3:4]
; GFX7-GISEL-NEXT: v_mov_b32_e32 v5, v2
; GFX7-GISEL-NEXT: v_mov_b32_e32 v6, v1
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v4, v5, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[7:8], s[4:5], v4, v3, v[1:2]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v6, v5, v[7:8]
; GFX7-GISEL-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-SDAG-LABEL: test_vector_reduce_mul_v2i64:
@ -2502,10 +2503,11 @@ define i64 @test_vector_reduce_mul_v2i64(<2 x i64> %v) {
; GFX8-GISEL: ; %bb.0: ; %entry
; GFX8-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-GISEL-NEXT: v_mov_b32_e32 v4, v0
; GFX8-GISEL-NEXT: v_mov_b32_e32 v5, v1
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v4, v2, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v4, v3, v[1:2]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v5, v2, v[3:4]
; GFX8-GISEL-NEXT: v_mov_b32_e32 v5, v2
; GFX8-GISEL-NEXT: v_mov_b32_e32 v6, v1
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v4, v5, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[7:8], s[4:5], v4, v3, v[1:2]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v6, v5, v[7:8]
; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-SDAG-LABEL: test_vector_reduce_mul_v2i64:
@ -2524,8 +2526,8 @@ define i64 @test_vector_reduce_mul_v2i64(<2 x i64> %v) {
; GFX9-GISEL-NEXT: v_mov_b32_e32 v5, v1
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[0:1], v4, v3, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[0:1], v4, v2, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[0:1], v5, v2, v[6:7]
; GFX9-GISEL-NEXT: v_add_u32_e32 v1, v1, v2
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[0:1], v5, v2, v[6:7]
; GFX9-GISEL-NEXT: v_add_u32_e32 v1, v1, v8
; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-SDAG-LABEL: test_vector_reduce_mul_v2i64:
@ -2626,9 +2628,9 @@ define i64 @test_vector_reduce_mul_v3i64(<3 x i64> %v) {
; GFX7-GISEL-NEXT: v_mov_b32_e32 v1, v8
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v0, v3, v[1:2]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v7, v4, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v6, v2, v[8:9]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v7, v5, v[1:2]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v2, v4, v[5:6]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v6, v2, v[8:9]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v7, v5, v[1:2]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v10, v4, v[8:9]
; GFX7-GISEL-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-SDAG-LABEL: test_vector_reduce_mul_v3i64:
@ -2654,9 +2656,9 @@ define i64 @test_vector_reduce_mul_v3i64(<3 x i64> %v) {
; GFX8-GISEL-NEXT: v_mov_b32_e32 v1, v8
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v0, v3, v[1:2]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v7, v4, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v6, v2, v[8:9]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v7, v5, v[1:2]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v2, v4, v[5:6]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v6, v2, v[8:9]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v7, v5, v[1:2]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v10, v4, v[8:9]
; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-SDAG-LABEL: test_vector_reduce_mul_v3i64:
@ -2677,12 +2679,12 @@ define i64 @test_vector_reduce_mul_v3i64(<3 x i64> %v) {
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[0:1], v0, v3, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[0:1], v0, v2, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[0:1], v1, v2, v[8:9]
; GFX9-GISEL-NEXT: v_add_u32_e32 v7, v7, v0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[0:1], v1, v2, v[8:9]
; GFX9-GISEL-NEXT: v_add_u32_e32 v8, v7, v10
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[0:1], v6, v5, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[0:1], v6, v4, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[0:1], v7, v4, v[2:3]
; GFX9-GISEL-NEXT: v_add_u32_e32 v1, v1, v2
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[0:1], v8, v4, v[2:3]
; GFX9-GISEL-NEXT: v_add_u32_e32 v1, v1, v6
; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-SDAG-LABEL: test_vector_reduce_mul_v3i64:
@ -2816,10 +2818,10 @@ define i64 @test_vector_reduce_mul_v4i64(<4 x i64> %v) {
; GFX7-GISEL-NEXT: v_mov_b32_e32 v0, v12
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[15:16], s[4:5], v2, v7, v[0:1]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v9, v11, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v3, v6, v[15:16]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v8, v4, v[13:14]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v9, v2, v[1:2]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v3, v11, v[1:2]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[17:18], s[4:5], v3, v6, v[15:16]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v8, v4, v[13:14]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v9, v17, v[1:2]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v5, v11, v[3:4]
; GFX7-GISEL-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-SDAG-LABEL: test_vector_reduce_mul_v4i64:
@ -2853,10 +2855,10 @@ define i64 @test_vector_reduce_mul_v4i64(<4 x i64> %v) {
; GFX8-GISEL-NEXT: v_mov_b32_e32 v0, v12
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[15:16], s[4:5], v2, v7, v[0:1]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v9, v11, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v3, v6, v[15:16]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v8, v4, v[13:14]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v9, v2, v[1:2]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v3, v11, v[1:2]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[17:18], s[4:5], v3, v6, v[15:16]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v8, v4, v[13:14]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v9, v17, v[1:2]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v5, v11, v[3:4]
; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-SDAG-LABEL: test_vector_reduce_mul_v4i64:
@ -2881,16 +2883,16 @@ define i64 @test_vector_reduce_mul_v4i64(<4 x i64> %v) {
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[0:1], v0, v5, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[0:1], v0, v4, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[0:1], v1, v4, v[10:11]
; GFX9-GISEL-NEXT: v_add_u32_e32 v9, v9, v0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[12:13], s[0:1], v1, v4, v[10:11]
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[0:1], v2, v7, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[0:1], v2, v6, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[0:1], v3, v6, v[0:1]
; GFX9-GISEL-NEXT: v_add_u32_e32 v2, v5, v0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[0:1], v8, v2, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[0:1], v3, v6, v[0:1]
; GFX9-GISEL-NEXT: v_add_u32_e32 v5, v5, v10
; GFX9-GISEL-NEXT: v_add_u32_e32 v9, v9, v12
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[0:1], v8, v5, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[0:1], v8, v4, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[0:1], v9, v4, v[2:3]
; GFX9-GISEL-NEXT: v_add_u32_e32 v1, v1, v2
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[0:1], v9, v4, v[2:3]
; GFX9-GISEL-NEXT: v_add_u32_e32 v1, v1, v6
; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-SDAG-LABEL: test_vector_reduce_mul_v4i64:
@ -3068,31 +3070,29 @@ define i64 @test_vector_reduce_mul_v8i64(<8 x i64> %v) {
; GFX7-GISEL: ; %bb.0: ; %entry
; GFX7-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[16:17], s[4:5], v4, v12, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[19:20], s[4:5], v0, v8, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[17:18], s[4:5], v4, v13, v[17:18]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v5, v12, v[17:18]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[12:13], s[4:5], v19, v16, 0
; GFX7-GISEL-NEXT: v_mov_b32_e32 v5, v13
; GFX7-GISEL-NEXT: v_mov_b32_e32 v13, v20
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[17:18], s[4:5], v0, v9, v[13:14]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v19, v4, v[5:6]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[20:21], s[4:5], v0, v8, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[18:19], s[4:5], v4, v13, v[17:18]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[22:23], s[4:5], v5, v12, v[18:19]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v20, v16, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[12:13], s[4:5], v20, v22, v[5:6]
; GFX7-GISEL-NEXT: v_mov_b32_e32 v5, v21
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[19:20], s[4:5], v6, v14, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v1, v8, v[17:18]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[17:18], s[4:5], v2, v10, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[17:18], s[4:5], v0, v9, v[5:6]
; GFX7-GISEL-NEXT: v_mov_b32_e32 v0, v20
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v6, v15, v[0:1]
; GFX7-GISEL-NEXT: v_mov_b32_e32 v6, v18
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[20:21], s[4:5], v2, v11, v[6:7]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v7, v14, v[0:1]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v17, v19, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v3, v10, v[20:21]
; GFX7-GISEL-NEXT: v_mov_b32_e32 v1, v7
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v17, v0, v[1:2]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v12, v6, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v2, v19, v[9:10]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v8, v16, v[4:5]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v12, v2, v[1:2]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v3, v6, v[1:2]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[21:22], s[4:5], v1, v8, v[17:18]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v6, v15, v[0:1]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v2, v10, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v2, v11, v[1:2]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v7, v14, v[8:9]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[7:8], s[4:5], v0, v19, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[14:15], s[4:5], v3, v10, v[5:6]
; GFX7-GISEL-NEXT: v_mov_b32_e32 v2, v8
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v0, v1, v[2:3]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v4, v7, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v14, v19, v[5:6]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v21, v16, v[12:13]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v4, v2, v[1:2]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v5, v7, v[8:9]
; GFX7-GISEL-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-SDAG-LABEL: test_vector_reduce_mul_v8i64:
@ -3139,31 +3139,29 @@ define i64 @test_vector_reduce_mul_v8i64(<8 x i64> %v) {
; GFX8-GISEL: ; %bb.0: ; %entry
; GFX8-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[16:17], s[4:5], v4, v12, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[19:20], s[4:5], v0, v8, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[17:18], s[4:5], v4, v13, v[17:18]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v5, v12, v[17:18]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[12:13], s[4:5], v19, v16, 0
; GFX8-GISEL-NEXT: v_mov_b32_e32 v5, v13
; GFX8-GISEL-NEXT: v_mov_b32_e32 v13, v20
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[17:18], s[4:5], v0, v9, v[13:14]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v19, v4, v[5:6]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[20:21], s[4:5], v0, v8, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[18:19], s[4:5], v4, v13, v[17:18]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[22:23], s[4:5], v5, v12, v[18:19]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v20, v16, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[12:13], s[4:5], v20, v22, v[5:6]
; GFX8-GISEL-NEXT: v_mov_b32_e32 v5, v21
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[19:20], s[4:5], v6, v14, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v1, v8, v[17:18]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[17:18], s[4:5], v2, v10, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[17:18], s[4:5], v0, v9, v[5:6]
; GFX8-GISEL-NEXT: v_mov_b32_e32 v0, v20
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v6, v15, v[0:1]
; GFX8-GISEL-NEXT: v_mov_b32_e32 v6, v18
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[20:21], s[4:5], v2, v11, v[6:7]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v7, v14, v[0:1]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v17, v19, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v3, v10, v[20:21]
; GFX8-GISEL-NEXT: v_mov_b32_e32 v1, v7
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v17, v0, v[1:2]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v12, v6, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v2, v19, v[9:10]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v8, v16, v[4:5]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v12, v2, v[1:2]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v3, v6, v[1:2]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[21:22], s[4:5], v1, v8, v[17:18]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v6, v15, v[0:1]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v2, v10, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v2, v11, v[1:2]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v7, v14, v[8:9]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[7:8], s[4:5], v0, v19, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[14:15], s[4:5], v3, v10, v[5:6]
; GFX8-GISEL-NEXT: v_mov_b32_e32 v2, v8
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v0, v1, v[2:3]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v4, v7, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v14, v19, v[5:6]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v21, v16, v[12:13]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v4, v2, v[1:2]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v5, v7, v[8:9]
; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-SDAG-LABEL: test_vector_reduce_mul_v8i64:
@ -3204,32 +3202,32 @@ define i64 @test_vector_reduce_mul_v8i64(<8 x i64> %v) {
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[18:19], s[0:1], v0, v9, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[16:17], s[0:1], v0, v8, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[0:1], v1, v8, v[18:19]
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[20:21], s[0:1], v1, v8, v[18:19]
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[0:1], v2, v11, 0
; GFX9-GISEL-NEXT: v_add_u32_e32 v17, v17, v0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[0:1], v2, v10, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[0:1], v3, v10, v[8:9]
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[18:19], s[0:1], v3, v10, v[8:9]
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[0:1], v4, v13, 0
; GFX9-GISEL-NEXT: v_add_u32_e32 v10, v1, v2
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[0:1], v2, v10, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[0:1], v4, v12, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[0:1], v5, v12, v[8:9]
; GFX9-GISEL-NEXT: v_add_u32_e32 v1, v3, v4
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[0:1], v5, v12, v[8:9]
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[0:1], v6, v15, 0
; GFX9-GISEL-NEXT: v_add_u32_e32 v18, v1, v18
; GFX9-GISEL-NEXT: v_add_u32_e32 v1, v3, v10
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[0:1], v6, v14, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[0:1], v7, v14, v[8:9]
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[0:1], v7, v14, v[8:9]
; GFX9-GISEL-NEXT: v_add_u32_e32 v17, v17, v20
; GFX9-GISEL-NEXT: v_add_u32_e32 v5, v5, v10
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[0:1], v16, v1, 0
; GFX9-GISEL-NEXT: v_add_u32_e32 v5, v5, v6
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[0:1], v17, v2, v[8:9]
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[0:1], v0, v5, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[0:1], v16, v2, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[0:1], v17, v2, v[8:9]
; GFX9-GISEL-NEXT: v_add_u32_e32 v7, v7, v2
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[0:1], v0, v4, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[0:1], v0, v5, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[0:1], v10, v4, v[0:1]
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[0:1], v18, v4, v[8:9]
; GFX9-GISEL-NEXT: v_add_u32_e32 v3, v3, v0
; GFX9-GISEL-NEXT: v_add_u32_e32 v10, v7, v10
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[0:1], v6, v3, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[0:1], v6, v2, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[0:1], v7, v2, v[4:5]
; GFX9-GISEL-NEXT: v_add_u32_e32 v1, v1, v2
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[0:1], v10, v2, v[4:5]
; GFX9-GISEL-NEXT: v_add_u32_e32 v1, v1, v6
; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-SDAG-LABEL: test_vector_reduce_mul_v8i64:
@ -3550,63 +3548,63 @@ define i64 @test_vector_reduce_mul_v16i64(<16 x i64> %v) {
; GFX7-GISEL: ; %bb.0: ; %entry
; GFX7-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[31:32], s[4:5], v8, v24, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[32:33], s[4:5], v8, v25, v[32:33]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v9, v24, v[32:33]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[24:25], s[4:5], v0, v16, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[32:33], s[4:5], v24, v31, 0
; GFX7-GISEL-NEXT: v_mov_b32_e32 v9, v33
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v24, v8, v[9:10]
; GFX7-GISEL-NEXT: v_mov_b32_e32 v24, v25
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[24:25], s[4:5], v0, v17, v[24:25]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v1, v16, v[24:25]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[16:17], s[4:5], v2, v18, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v0, v31, v[8:9]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[33:34], s[4:5], v8, v25, v[32:33]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[35:36], s[4:5], v9, v24, v[33:34]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v0, v16, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[24:25], s[4:5], v8, v31, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[32:33], s[4:5], v8, v35, v[25:26]
; GFX7-GISEL-NEXT: v_mov_b32_e32 v8, v9
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[34:35], s[4:5], v0, v17, v[8:9]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v1, v16, v[34:35]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v10, v26, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v10, v27, v[1:2]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v11, v26, v[9:10]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v16, v0, 0
; GFX7-GISEL-NEXT: v_mov_b32_e32 v1, v11
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[24:25], s[4:5], v16, v9, v[1:2]
; GFX7-GISEL-NEXT: buffer_load_dword v9, off, s[0:3], s32
; GFX7-GISEL-NEXT: v_mov_b32_e32 v1, v17
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v2, v19, v[1:2]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v3, v18, v[1:2]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[16:17], s[4:5], v8, v31, v[32:33]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v10, v27, v[1:2]
; GFX7-GISEL-NEXT: buffer_load_dword v27, off, s[0:3], s32
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[31:32], s[4:5], v2, v18, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[33:34], s[4:5], v11, v26, v[8:9]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v31, v0, 0
; GFX7-GISEL-NEXT: v_mov_b32_e32 v1, v9
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v31, v33, v[1:2]
; GFX7-GISEL-NEXT: v_mov_b32_e32 v1, v32
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[25:26], s[4:5], v2, v19, v[1:2]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v3, v18, v[25:26]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v12, v28, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[16:17], s[4:5], v1, v0, v[24:25]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[17:18], s[4:5], v4, v20, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[17:18], s[4:5], v1, v0, v[9:10]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v4, v20, 0
; GFX7-GISEL-NEXT: v_mov_b32_e32 v0, v3
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v12, v29, v[0:1]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[11:12], s[4:5], v17, v2, 0
; GFX7-GISEL-NEXT: v_mov_b32_e32 v3, v18
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v13, v28, v[0:1]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v4, v21, v[3:4]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[18:19], s[4:5], v12, v29, v[0:1]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[11:12], s[4:5], v9, v2, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v13, v28, v[18:19]
; GFX7-GISEL-NEXT: v_mov_b32_e32 v3, v10
; GFX7-GISEL-NEXT: v_mov_b32_e32 v1, v12
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[12:13], s[4:5], v14, v30, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v5, v20, v[3:4]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v17, v0, v[1:2]
; GFX7-GISEL-NEXT: v_mov_b32_e32 v4, v13
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[17:18], s[4:5], v6, v22, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v3, v2, v[0:1]
; GFX7-GISEL-NEXT: v_mov_b32_e32 v3, v18
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[12:13], s[4:5], v4, v21, v[3:4]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v14, v30, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[18:19], s[4:5], v9, v0, v[1:2]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v5, v20, v[12:13]
; GFX7-GISEL-NEXT: v_mov_b32_e32 v1, v4
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v6, v22, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[20:21], s[4:5], v0, v2, v[18:19]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[12:13], s[4:5], v4, v3, 0
; GFX7-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v14, v9, v[4:5]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[13:14], s[4:5], v17, v12, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v15, v30, v[4:5]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v6, v23, v[3:4]
; GFX7-GISEL-NEXT: v_mov_b32_e32 v2, v14
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v17, v1, v[2:3]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v7, v22, v[3:4]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v32, v11, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v3, v12, v[1:2]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v10, v13, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v32, v0, v[5:6]
; GFX7-GISEL-NEXT: v_mov_b32_e32 v0, v3
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v10, v1, v[0:1]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v4, v2, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v16, v13, v[9:10]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v8, v11, v[5:6]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v4, v9, v[1:2]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v5, v2, v[3:4]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v14, v27, v[1:2]
; GFX7-GISEL-NEXT: v_mov_b32_e32 v2, v5
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v15, v30, v[9:10]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v6, v23, v[2:3]
; GFX7-GISEL-NEXT: v_mov_b32_e32 v1, v13
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v4, v0, v[1:2]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v7, v22, v[9:10]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v24, v11, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v0, v3, v[5:6]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v8, v12, 0
; GFX7-GISEL-NEXT: v_mov_b32_e32 v0, v10
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v24, v20, v[0:1]
; GFX7-GISEL-NEXT: v_mov_b32_e32 v0, v4
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[13:14], s[4:5], v8, v1, v[0:1]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v9, v3, 0
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[7:8], s[4:5], v17, v12, v[13:14]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[12:13], s[4:5], v16, v11, v[5:6]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v9, v7, v[1:2]
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v12, v3, v[4:5]
; GFX7-GISEL-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-SDAG-LABEL: test_vector_reduce_mul_v16i64:
@ -3695,63 +3693,63 @@ define i64 @test_vector_reduce_mul_v16i64(<16 x i64> %v) {
; GFX8-GISEL: ; %bb.0: ; %entry
; GFX8-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[31:32], s[4:5], v8, v24, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[32:33], s[4:5], v8, v25, v[32:33]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v9, v24, v[32:33]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[24:25], s[4:5], v0, v16, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[32:33], s[4:5], v24, v31, 0
; GFX8-GISEL-NEXT: v_mov_b32_e32 v9, v33
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v24, v8, v[9:10]
; GFX8-GISEL-NEXT: v_mov_b32_e32 v24, v25
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[24:25], s[4:5], v0, v17, v[24:25]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v1, v16, v[24:25]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[16:17], s[4:5], v2, v18, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v0, v31, v[8:9]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[33:34], s[4:5], v8, v25, v[32:33]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[35:36], s[4:5], v9, v24, v[33:34]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v0, v16, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[24:25], s[4:5], v8, v31, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[32:33], s[4:5], v8, v35, v[25:26]
; GFX8-GISEL-NEXT: v_mov_b32_e32 v8, v9
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[34:35], s[4:5], v0, v17, v[8:9]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v1, v16, v[34:35]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v10, v26, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v10, v27, v[1:2]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v11, v26, v[9:10]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v16, v0, 0
; GFX8-GISEL-NEXT: v_mov_b32_e32 v1, v11
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[24:25], s[4:5], v16, v9, v[1:2]
; GFX8-GISEL-NEXT: buffer_load_dword v9, off, s[0:3], s32
; GFX8-GISEL-NEXT: v_mov_b32_e32 v1, v17
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v2, v19, v[1:2]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v3, v18, v[1:2]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[16:17], s[4:5], v8, v31, v[32:33]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v10, v27, v[1:2]
; GFX8-GISEL-NEXT: buffer_load_dword v27, off, s[0:3], s32
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[31:32], s[4:5], v2, v18, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[33:34], s[4:5], v11, v26, v[8:9]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v31, v0, 0
; GFX8-GISEL-NEXT: v_mov_b32_e32 v1, v9
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v31, v33, v[1:2]
; GFX8-GISEL-NEXT: v_mov_b32_e32 v1, v32
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[25:26], s[4:5], v2, v19, v[1:2]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v3, v18, v[25:26]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v12, v28, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[16:17], s[4:5], v1, v0, v[24:25]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[17:18], s[4:5], v4, v20, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[17:18], s[4:5], v1, v0, v[9:10]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v4, v20, 0
; GFX8-GISEL-NEXT: v_mov_b32_e32 v0, v3
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v12, v29, v[0:1]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[11:12], s[4:5], v17, v2, 0
; GFX8-GISEL-NEXT: v_mov_b32_e32 v3, v18
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v13, v28, v[0:1]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v4, v21, v[3:4]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[18:19], s[4:5], v12, v29, v[0:1]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[11:12], s[4:5], v9, v2, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v13, v28, v[18:19]
; GFX8-GISEL-NEXT: v_mov_b32_e32 v3, v10
; GFX8-GISEL-NEXT: v_mov_b32_e32 v1, v12
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[12:13], s[4:5], v14, v30, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v5, v20, v[3:4]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v17, v0, v[1:2]
; GFX8-GISEL-NEXT: v_mov_b32_e32 v4, v13
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[17:18], s[4:5], v6, v22, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v3, v2, v[0:1]
; GFX8-GISEL-NEXT: v_mov_b32_e32 v3, v18
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[12:13], s[4:5], v4, v21, v[3:4]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v14, v30, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[18:19], s[4:5], v9, v0, v[1:2]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v5, v20, v[12:13]
; GFX8-GISEL-NEXT: v_mov_b32_e32 v1, v4
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v6, v22, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[20:21], s[4:5], v0, v2, v[18:19]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[12:13], s[4:5], v4, v3, 0
; GFX8-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v14, v9, v[4:5]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[13:14], s[4:5], v17, v12, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v15, v30, v[4:5]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v6, v23, v[3:4]
; GFX8-GISEL-NEXT: v_mov_b32_e32 v2, v14
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v17, v1, v[2:3]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v7, v22, v[3:4]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v32, v11, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v3, v12, v[1:2]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v10, v13, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v32, v0, v[5:6]
; GFX8-GISEL-NEXT: v_mov_b32_e32 v0, v3
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v10, v1, v[0:1]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v4, v2, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v16, v13, v[9:10]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v8, v11, v[5:6]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v4, v9, v[1:2]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v5, v2, v[3:4]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v14, v27, v[1:2]
; GFX8-GISEL-NEXT: v_mov_b32_e32 v2, v5
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v15, v30, v[9:10]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v6, v23, v[2:3]
; GFX8-GISEL-NEXT: v_mov_b32_e32 v1, v13
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v4, v0, v[1:2]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v7, v22, v[9:10]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v24, v11, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v0, v3, v[5:6]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v8, v12, 0
; GFX8-GISEL-NEXT: v_mov_b32_e32 v0, v10
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v24, v20, v[0:1]
; GFX8-GISEL-NEXT: v_mov_b32_e32 v0, v4
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[13:14], s[4:5], v8, v1, v[0:1]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v9, v3, 0
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[7:8], s[4:5], v17, v12, v[13:14]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[12:13], s[4:5], v16, v11, v[5:6]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v9, v7, v[1:2]
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v12, v3, v[4:5]
; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-SDAG-LABEL: test_vector_reduce_mul_v16i64:
@ -3827,65 +3825,65 @@ define i64 @test_vector_reduce_mul_v16i64(<16 x i64> %v) {
; GFX9-GISEL-NEXT: scratch_load_dword v31, off, s32
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[34:35], s[0:1], v0, v17, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[32:33], s[0:1], v0, v16, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[0:1], v1, v16, v[34:35]
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[36:37], s[0:1], v1, v16, v[34:35]
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[16:17], s[0:1], v2, v19, 0
; GFX9-GISEL-NEXT: v_add_u32_e32 v33, v33, v0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[0:1], v2, v18, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[0:1], v3, v18, v[16:17]
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[34:35], s[0:1], v3, v18, v[16:17]
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[16:17], s[0:1], v4, v21, 0
; GFX9-GISEL-NEXT: v_add_u32_e32 v18, v1, v2
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[0:1], v2, v18, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[0:1], v4, v20, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[0:1], v5, v20, v[16:17]
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[18:19], s[0:1], v5, v20, v[16:17]
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[16:17], s[0:1], v6, v23, 0
; GFX9-GISEL-NEXT: v_add_u32_e32 v19, v3, v4
; GFX9-GISEL-NEXT: v_add_u32_e32 v20, v3, v18
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[0:1], v6, v22, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[0:1], v7, v22, v[16:17]
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[18:19], s[0:1], v7, v22, v[16:17]
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[16:17], s[0:1], v8, v25, 0
; GFX9-GISEL-NEXT: v_add_u32_e32 v20, v5, v6
; GFX9-GISEL-NEXT: v_add_u32_e32 v21, v5, v18
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[0:1], v8, v24, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[0:1], v9, v24, v[16:17]
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[18:19], s[0:1], v9, v24, v[16:17]
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[16:17], s[0:1], v10, v27, 0
; GFX9-GISEL-NEXT: v_add_u32_e32 v1, v7, v8
; GFX9-GISEL-NEXT: v_add_u32_e32 v34, v1, v34
; GFX9-GISEL-NEXT: v_add_u32_e32 v1, v7, v18
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[0:1], v10, v26, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[0:1], v11, v26, v[16:17]
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[18:19], s[0:1], v11, v26, v[16:17]
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[16:17], s[0:1], v12, v29, 0
; GFX9-GISEL-NEXT: v_add_u32_e32 v3, v9, v10
; GFX9-GISEL-NEXT: v_add_u32_e32 v3, v9, v18
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[0:1], v12, v28, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[12:13], s[0:1], v13, v28, v[16:17]
; GFX9-GISEL-NEXT: v_add_u32_e32 v5, v11, v12
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[18:19], s[0:1], v13, v28, v[16:17]
; GFX9-GISEL-NEXT: v_add_u32_e32 v33, v33, v36
; GFX9-GISEL-NEXT: v_add_u32_e32 v5, v11, v18
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[12:13], s[0:1], v14, v30, 0
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[16:17], s[0:1], v14, v31, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[14:15], s[0:1], v15, v30, v[16:17]
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[18:19], s[0:1], v15, v30, v[16:17]
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[16:17], s[0:1], v32, v1, 0
; GFX9-GISEL-NEXT: v_add_u32_e32 v9, v13, v14
; GFX9-GISEL-NEXT: v_add_u32_e32 v11, v13, v18
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[18:19], s[0:1], v33, v6, v[16:17]
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[16:17], s[0:1], v0, v3, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[14:15], s[0:1], v32, v6, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[0:1], v33, v6, v[16:17]
; GFX9-GISEL-NEXT: v_add_u32_e32 v11, v15, v6
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[0:1], v0, v8, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[0:1], v0, v3, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[0:1], v18, v8, v[0:1]
; GFX9-GISEL-NEXT: v_add_u32_e32 v7, v7, v0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[0:1], v34, v8, v[16:17]
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[0:1], v2, v5, 0
; GFX9-GISEL-NEXT: v_add_u32_e32 v13, v15, v18
; GFX9-GISEL-NEXT: v_add_u32_e32 v15, v7, v0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[0:1], v2, v10, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[0:1], v2, v5, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[0:1], v19, v10, v[2:3]
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[0:1], v20, v10, v[8:9]
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[0:1], v4, v11, 0
; GFX9-GISEL-NEXT: v_add_u32_e32 v1, v1, v2
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[0:1], v4, v12, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[0:1], v4, v9, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[0:1], v20, v12, v[4:5]
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[0:1], v14, v1, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[0:1], v21, v12, v[8:9]
; GFX9-GISEL-NEXT: v_add_u32_e32 v3, v3, v4
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[0:1], v14, v1, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[0:1], v14, v0, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[0:1], v11, v0, v[8:9]
; GFX9-GISEL-NEXT: v_add_u32_e32 v5, v5, v0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[0:1], v13, v0, v[8:9]
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[0:1], v6, v3, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[0:1], v6, v2, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[0:1], v7, v2, v[0:1]
; GFX9-GISEL-NEXT: v_add_u32_e32 v2, v9, v0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[0:1], v4, v2, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[0:1], v15, v2, v[0:1]
; GFX9-GISEL-NEXT: v_add_u32_e32 v10, v5, v10
; GFX9-GISEL-NEXT: v_add_u32_e32 v5, v9, v6
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[0:1], v4, v5, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[0:1], v4, v8, 0
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[0:1], v5, v8, v[2:3]
; GFX9-GISEL-NEXT: v_add_u32_e32 v1, v1, v2
; GFX9-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[0:1], v10, v8, v[2:3]
; GFX9-GISEL-NEXT: v_add_u32_e32 v1, v1, v4
; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-SDAG-LABEL: test_vector_reduce_mul_v16i64: