[GlobalISel] Handle div-by-pow2

This patch adds similar handling of div-by-pow2 as in `SelectionDAG`.

Author: shiltian

llvm/include/llvm/CodeGen/GlobalISel/CombinerHelper.h

@@ -673,6 +673,14 @@ class CombinerHelper {
   bool matchSDivByConst(MachineInstr &MI);
   void applySDivByConst(MachineInstr &MI);
 
+  /// Given an G_SDIV \p MI expressing a signed divided by a pow2 constant,
+  /// return expressions that implements it by shifting.
+  bool matchDivByPow2(MachineInstr &MI, bool IsSigned);
+  void applySDivByPow2(MachineInstr &MI);
+  /// Given an G_UDIV \p MI expressing an unsigned divided by a pow2 constant,
+  /// return expressions that implements it by shifting.
+  void applyUDivByPow2(MachineInstr &MI);
+
   // G_UMULH x, (1 << c)) -> x >> (bitwidth - c)
   bool matchUMulHToLShr(MachineInstr &MI);
   void applyUMulHToLShr(MachineInstr &MI);

llvm/include/llvm/Target/GlobalISel/Combine.td

@@ -179,6 +179,7 @@ def FmArcp      : MIFlagEnum<"FmArcp">;
 def FmContract  : MIFlagEnum<"FmContract">;
 def FmAfn       : MIFlagEnum<"FmAfn">;
 def FmReassoc   : MIFlagEnum<"FmReassoc">;
+def IsExact     : MIFlagEnum<"IsExact">;
 
 def MIFlags;
 // def not; -> Already defined as a SDNode
@@ -1036,7 +1037,20 @@ def sdiv_by_const : GICombineRule<
    [{ return Helper.matchSDivByConst(*${root}); }]),
   (apply [{ Helper.applySDivByConst(*${root}); }])>;
 
-def intdiv_combines : GICombineGroup<[udiv_by_const, sdiv_by_const]>;
+def sdiv_by_pow2 : GICombineRule<
+  (defs root:$root),
+  (match (G_SDIV $dst, $x, $y, (MIFlags (not IsExact))):$root,
+   [{ return Helper.matchDivByPow2(*${root}, /*IsSigned=*/true); }]),
+  (apply [{ Helper.applySDivByPow2(*${root}); }])>;
+
+def udiv_by_pow2 : GICombineRule<
+  (defs root:$root),
+  (match (G_UDIV $dst, $x, $y, (MIFlags (not IsExact))):$root,
+   [{ return Helper.matchDivByPow2(*${root}, /*IsSigned=*/false); }]),
+  (apply [{ Helper.applyUDivByPow2(*${root}); }])>;
+
+def intdiv_combines : GICombineGroup<[udiv_by_const, sdiv_by_const,
+                                      sdiv_by_pow2, udiv_by_pow2]>;
 
 def reassoc_ptradd : GICombineRule<
   (defs root:$root, build_fn_matchinfo:$matchinfo),

llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp

@@ -5270,6 +5270,98 @@ MachineInstr *CombinerHelper::buildSDivUsingMul(MachineInstr &MI) {
   return MIB.buildMul(Ty, Res, Factor);
 }
 
+bool CombinerHelper::matchDivByPow2(MachineInstr &MI, bool IsSigned) {
+  assert((MI.getOpcode() == TargetOpcode::G_SDIV ||
+          MI.getOpcode() == TargetOpcode::G_UDIV) &&
+         "Expected SDIV or UDIV");
+  auto &Div = cast<GenericMachineInstr>(MI);
+  Register RHS = Div.getReg(2);
+  auto MatchPow2 = [&](const Constant *C) {
+    auto *CI = dyn_cast<ConstantInt>(C);
+    return CI && (CI->getValue().isPowerOf2() ||
+                  (IsSigned && CI->getValue().isNegatedPowerOf2()));
+  };
+  return matchUnaryPredicate(MRI, RHS, MatchPow2, /*AllowUndefs=*/false);
+}
+
+void CombinerHelper::applySDivByPow2(MachineInstr &MI) {
+  assert(MI.getOpcode() == TargetOpcode::G_SDIV && "Expected SDIV");
+  auto &SDiv = cast<GenericMachineInstr>(MI);
+  Register Dst = SDiv.getReg(0);
+  Register LHS = SDiv.getReg(1);
+  Register RHS = SDiv.getReg(2);
+  LLT Ty = MRI.getType(Dst);
+  LLT ShiftAmtTy = getTargetLowering().getPreferredShiftAmountTy(Ty);
+  LLT CCVT =
+      Ty.isVector() ? LLT::vector(Ty.getElementCount(), 1) : LLT::scalar(1);
+
+  Builder.setInstrAndDebugLoc(MI);
+
+  // Effectively we want to lower G_SDIV %lhs, %rhs, where %rhs is a power of 2,
+  // to the following version:
+  //
+  // %c1 = G_CTTZ %rhs
+  // %inexact = G_SUB $bitwidth, %c1
+  // %sign = %G_ASHR %lhs, $(bitwidth - 1)
+  // %lshr = G_LSHR %sign, %inexact
+  // %add = G_ADD %lhs, %lshr
+  // %ashr = G_ASHR %add, %c1
+  // %ashr = G_SELECT, %isoneorallones, %lhs, %ashr
+  // %zero = G_CONSTANT $0
+  // %neg = G_NEG %ashr
+  // %isneg = G_ICMP SLT %rhs, %zero
+  // %res = G_SELECT %isneg, %neg, %ashr
+
+  unsigned BitWidth = Ty.getScalarSizeInBits();
+  auto Zero = Builder.buildConstant(Ty, 0);
+
+  auto Bits = Builder.buildConstant(ShiftAmtTy, BitWidth);
+  auto C1 = Builder.buildCTTZ(ShiftAmtTy, RHS);
+  auto Inexact = Builder.buildSub(ShiftAmtTy, Bits, C1);
+  // Splat the sign bit into the register
+  auto Sign = Builder.buildAShr(
+      Ty, LHS, Builder.buildConstant(ShiftAmtTy, BitWidth - 1));
+
+  // Add (LHS < 0) ? abs2 - 1 : 0;
+  auto LSrl = Builder.buildLShr(Ty, Sign, Inexact);
+  auto Add = Builder.buildAdd(Ty, LHS, LSrl);
+  auto AShr = Builder.buildAShr(Ty, Add, C1);
+
+  // Special case: (sdiv X, 1) -> X
+  // Special Case: (sdiv X, -1) -> 0-X
+  auto One = Builder.buildConstant(Ty, /*Val=*/1);
+  auto AllOnes =
+      Builder.buildConstant(Ty, APInt::getAllOnes(Ty.getScalarSizeInBits()));
+  auto IsOne = Builder.buildICmp(CmpInst::Predicate::ICMP_EQ, CCVT, RHS, One);
+  auto IsAllOnes =
+      Builder.buildICmp(CmpInst::Predicate::ICMP_EQ, CCVT, RHS, AllOnes);
+  auto IsOneOrAllOnes = Builder.buildOr(CCVT, IsOne, IsAllOnes);
+  AShr = Builder.buildSelect(Ty, IsOneOrAllOnes, LHS, AShr);
+
+  // If divided by a positive value, we're done. Otherwise, the result must be
+  // negated.
+  auto Neg = Builder.buildNeg(Ty, AShr);
+  auto IsNeg = Builder.buildICmp(CmpInst::Predicate::ICMP_SLT, CCVT, RHS, Zero);
+  Builder.buildSelect(MI.getOperand(0).getReg(), IsNeg, Neg, AShr);
+  MI.eraseFromParent();
+}
+
+void CombinerHelper::applyUDivByPow2(MachineInstr &MI) {
+  assert(MI.getOpcode() == TargetOpcode::G_UDIV && "Expected UDIV");
+  auto &UDiv = cast<GenericMachineInstr>(MI);
+  Register Dst = UDiv.getReg(0);
+  Register LHS = UDiv.getReg(1);
+  Register RHS = UDiv.getReg(2);
+  LLT Ty = MRI.getType(Dst);
+  LLT ShiftAmtTy = getTargetLowering().getPreferredShiftAmountTy(Ty);
+
+  Builder.setInstrAndDebugLoc(MI);
+
+  auto C1 = Builder.buildCTTZ(ShiftAmtTy, RHS);
+  Builder.buildLShr(MI.getOperand(0).getReg(), LHS, C1);
+  MI.eraseFromParent();
+}
+
 bool CombinerHelper::matchUMulHToLShr(MachineInstr &MI) {
   assert(MI.getOpcode() == TargetOpcode::G_UMULH);
   Register RHS = MI.getOperand(2).getReg();

llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.sbfe.ll

@@ -670,36 +670,19 @@ define amdgpu_kernel void @bfe_sext_in_reg_i24(ptr addrspace(1) %out, ptr addrsp
 define amdgpu_kernel void @simplify_demanded_bfe_sdiv(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 {
 ; GFX6-LABEL: simplify_demanded_bfe_sdiv:
 ; GFX6:       ; %bb.0:
-; GFX6-NEXT:    v_rcp_iflag_f32_e32 v0, 2.0
-; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x0
-; GFX6-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
-; GFX6-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX6-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x0
 ; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX6-NEXT:    s_load_dword s0, s[6:7], 0x0
-; GFX6-NEXT:    s_mov_b32 s6, -1
-; GFX6-NEXT:    s_mov_b32 s7, 0xf000
-; GFX6-NEXT:    v_mul_lo_u32 v1, v0, -2
-; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX6-NEXT:    s_bfe_i32 s0, s0, 0x100001
-; GFX6-NEXT:    s_ashr_i32 s2, s0, 31
-; GFX6-NEXT:    v_mul_hi_u32 v1, v0, v1
-; GFX6-NEXT:    s_add_i32 s0, s0, s2
-; GFX6-NEXT:    s_xor_b32 s0, s0, s2
-; GFX6-NEXT:    v_add_i32_e32 v0, vcc, v0, v1
-; GFX6-NEXT:    v_mul_hi_u32 v0, s0, v0
-; GFX6-NEXT:    v_lshlrev_b32_e32 v1, 1, v0
-; GFX6-NEXT:    v_add_i32_e32 v2, vcc, 1, v0
-; GFX6-NEXT:    v_sub_i32_e32 v1, vcc, s0, v1
-; GFX6-NEXT:    v_cmp_le_u32_e32 vcc, 2, v1
-; GFX6-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
-; GFX6-NEXT:    v_subrev_i32_e64 v2, s[0:1], 2, v1
-; GFX6-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc
-; GFX6-NEXT:    v_add_i32_e32 v2, vcc, 1, v0
-; GFX6-NEXT:    v_cmp_le_u32_e32 vcc, 2, v1
-; GFX6-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
-; GFX6-NEXT:    v_xor_b32_e32 v0, s2, v0
-; GFX6-NEXT:    v_subrev_i32_e32 v0, vcc, s2, v0
-; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_load_dword s3, s[2:3], 0x0
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_bfe_i32 s3, s3, 0x100001
+; GFX6-NEXT:    s_ashr_i32 s4, s3, 31
+; GFX6-NEXT:    s_lshr_b32 s4, s4, 31
+; GFX6-NEXT:    s_add_i32 s3, s3, s4
+; GFX6-NEXT:    s_ashr_i32 s3, s3, 1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s3
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; GFX6-NEXT:    s_endpgm
   %src = load i32, ptr addrspace(1) %in, align 4
   %bfe = call i32 @llvm.amdgcn.sbfe.i32(i32 %src, i32 1, i32 16)

llvm/test/CodeGen/AMDGPU/GlobalISel/sdiv.i32.ll

@@ -279,125 +279,27 @@ define i32 @v_sdiv_i32_pow2k_denom(i32 %num) {
 ; CHECK:       ; %bb.0:
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; CHECK-NEXT:    v_ashrrev_i32_e32 v1, 31, v0
-; CHECK-NEXT:    v_rcp_iflag_f32_e32 v2, 0x45800000
-; CHECK-NEXT:    v_mov_b32_e32 v3, 0xfffff000
-; CHECK-NEXT:    v_mov_b32_e32 v4, 0x1000
+; CHECK-NEXT:    v_lshrrev_b32_e32 v1, 20, v1
 ; CHECK-NEXT:    v_add_i32_e32 v0, vcc, v0, v1
-; CHECK-NEXT:    v_mul_f32_e32 v2, 0x4f7ffffe, v2
-; CHECK-NEXT:    v_xor_b32_e32 v0, v0, v1
-; CHECK-NEXT:    v_cvt_u32_f32_e32 v2, v2
-; CHECK-NEXT:    v_mul_lo_u32 v3, v2, v3
-; CHECK-NEXT:    v_mul_hi_u32 v3, v2, v3
-; CHECK-NEXT:    v_add_i32_e32 v2, vcc, v2, v3
-; CHECK-NEXT:    v_mul_hi_u32 v2, v0, v2
-; CHECK-NEXT:    v_lshlrev_b32_e32 v3, 12, v2
-; CHECK-NEXT:    v_add_i32_e32 v5, vcc, 1, v2
-; CHECK-NEXT:    v_sub_i32_e32 v0, vcc, v0, v3
-; CHECK-NEXT:    v_cmp_ge_u32_e64 s[4:5], v0, v4
-; CHECK-NEXT:    v_cndmask_b32_e64 v2, v2, v5, s[4:5]
-; CHECK-NEXT:    v_subrev_i32_e32 v3, vcc, 0x1000, v0
-; CHECK-NEXT:    v_cndmask_b32_e64 v0, v0, v3, s[4:5]
-; CHECK-NEXT:    v_add_i32_e32 v3, vcc, 1, v2
-; CHECK-NEXT:    v_cmp_ge_u32_e32 vcc, v0, v4
-; CHECK-NEXT:    v_cndmask_b32_e32 v0, v2, v3, vcc
-; CHECK-NEXT:    v_xor_b32_e32 v0, v0, v1
-; CHECK-NEXT:    v_sub_i32_e32 v0, vcc, v0, v1
+; CHECK-NEXT:    v_ashrrev_i32_e32 v0, 12, v0
 ; CHECK-NEXT:    s_setpc_b64 s[30:31]
   %result = sdiv i32 %num, 4096
   ret i32 %result
 }
 
 define <2 x i32> @v_sdiv_v2i32_pow2k_denom(<2 x i32> %num) {
-; GISEL-LABEL: v_sdiv_v2i32_pow2k_denom:
-; GISEL:       ; %bb.0:
-; GISEL-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GISEL-NEXT:    v_ashrrev_i32_e32 v2, 31, v0
-; GISEL-NEXT:    v_mov_b32_e32 v3, 0x1000
-; GISEL-NEXT:    v_cvt_f32_u32_e32 v4, 0x1000
-; GISEL-NEXT:    v_mov_b32_e32 v5, 0xfffff000
-; GISEL-NEXT:    v_ashrrev_i32_e32 v6, 31, v1
-; GISEL-NEXT:    v_add_i32_e32 v0, vcc, v0, v2
-; GISEL-NEXT:    v_rcp_iflag_f32_e32 v4, v4
-; GISEL-NEXT:    v_add_i32_e32 v1, vcc, v1, v6
-; GISEL-NEXT:    v_xor_b32_e32 v0, v0, v2
-; GISEL-NEXT:    v_mul_f32_e32 v4, 0x4f7ffffe, v4
-; GISEL-NEXT:    v_xor_b32_e32 v1, v1, v6
-; GISEL-NEXT:    v_cvt_u32_f32_e32 v4, v4
-; GISEL-NEXT:    v_mul_lo_u32 v5, v4, v5
-; GISEL-NEXT:    v_mul_hi_u32 v5, v4, v5
-; GISEL-NEXT:    v_add_i32_e32 v4, vcc, v4, v5
-; GISEL-NEXT:    v_mul_hi_u32 v5, v0, v4
-; GISEL-NEXT:    v_mul_hi_u32 v4, v1, v4
-; GISEL-NEXT:    v_lshlrev_b32_e32 v7, 12, v5
-; GISEL-NEXT:    v_add_i32_e32 v8, vcc, 1, v5
-; GISEL-NEXT:    v_lshlrev_b32_e32 v9, 12, v4
-; GISEL-NEXT:    v_add_i32_e32 v10, vcc, 1, v4
-; GISEL-NEXT:    v_sub_i32_e32 v0, vcc, v0, v7
-; GISEL-NEXT:    v_sub_i32_e32 v1, vcc, v1, v9
-; GISEL-NEXT:    v_cmp_ge_u32_e64 s[4:5], v0, v3
-; GISEL-NEXT:    v_cndmask_b32_e64 v5, v5, v8, s[4:5]
-; GISEL-NEXT:    v_sub_i32_e32 v7, vcc, v0, v3
-; GISEL-NEXT:    v_cmp_ge_u32_e64 s[6:7], v1, v3
-; GISEL-NEXT:    v_cndmask_b32_e64 v4, v4, v10, s[6:7]
-; GISEL-NEXT:    v_subrev_i32_e32 v8, vcc, 0x1000, v1
-; GISEL-NEXT:    v_cndmask_b32_e64 v0, v0, v7, s[4:5]
-; GISEL-NEXT:    v_add_i32_e32 v7, vcc, 1, v5
-; GISEL-NEXT:    v_cndmask_b32_e64 v1, v1, v8, s[6:7]
-; GISEL-NEXT:    v_add_i32_e32 v8, vcc, 1, v4
-; GISEL-NEXT:    v_cmp_ge_u32_e32 vcc, v0, v3
-; GISEL-NEXT:    v_cndmask_b32_e32 v0, v5, v7, vcc
-; GISEL-NEXT:    v_cmp_ge_u32_e32 vcc, v1, v3
-; GISEL-NEXT:    v_cndmask_b32_e32 v1, v4, v8, vcc
-; GISEL-NEXT:    v_xor_b32_e32 v0, v0, v2
-; GISEL-NEXT:    v_xor_b32_e32 v1, v1, v6
-; GISEL-NEXT:    v_sub_i32_e32 v0, vcc, v0, v2
-; GISEL-NEXT:    v_sub_i32_e32 v1, vcc, v1, v6
-; GISEL-NEXT:    s_setpc_b64 s[30:31]
-;
-; CGP-LABEL: v_sdiv_v2i32_pow2k_denom:
-; CGP:       ; %bb.0:
-; CGP-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; CGP-NEXT:    v_ashrrev_i32_e32 v2, 31, v0
-; CGP-NEXT:    v_rcp_iflag_f32_e32 v3, 0x45800000
-; CGP-NEXT:    v_mov_b32_e32 v4, 0xfffff000
-; CGP-NEXT:    v_mov_b32_e32 v5, 0x1000
-; CGP-NEXT:    v_ashrrev_i32_e32 v6, 31, v1
-; CGP-NEXT:    v_add_i32_e32 v0, vcc, v0, v2
-; CGP-NEXT:    v_mul_f32_e32 v3, 0x4f7ffffe, v3
-; CGP-NEXT:    v_add_i32_e32 v1, vcc, v1, v6
-; CGP-NEXT:    v_xor_b32_e32 v0, v0, v2
-; CGP-NEXT:    v_cvt_u32_f32_e32 v3, v3
-; CGP-NEXT:    v_xor_b32_e32 v1, v1, v6
-; CGP-NEXT:    v_mul_lo_u32 v4, v3, v4
-; CGP-NEXT:    v_mul_hi_u32 v4, v3, v4
-; CGP-NEXT:    v_add_i32_e32 v3, vcc, v3, v4
-; CGP-NEXT:    v_mul_hi_u32 v4, v0, v3
-; CGP-NEXT:    v_mul_hi_u32 v3, v1, v3
-; CGP-NEXT:    v_lshlrev_b32_e32 v7, 12, v4
-; CGP-NEXT:    v_add_i32_e32 v8, vcc, 1, v4
-; CGP-NEXT:    v_lshlrev_b32_e32 v9, 12, v3
-; CGP-NEXT:    v_add_i32_e32 v10, vcc, 1, v3
-; CGP-NEXT:    v_sub_i32_e32 v0, vcc, v0, v7
-; CGP-NEXT:    v_sub_i32_e32 v1, vcc, v1, v9
-; CGP-NEXT:    v_cmp_ge_u32_e64 s[4:5], v0, v5
-; CGP-NEXT:    v_cndmask_b32_e64 v4, v4, v8, s[4:5]
-; CGP-NEXT:    v_sub_i32_e32 v7, vcc, v0, v5
-; CGP-NEXT:    v_cmp_ge_u32_e64 s[6:7], v1, v5
-; CGP-NEXT:    v_cndmask_b32_e64 v3, v3, v10, s[6:7]
-; CGP-NEXT:    v_subrev_i32_e32 v8, vcc, 0x1000, v1
-; CGP-NEXT:    v_cndmask_b32_e64 v0, v0, v7, s[4:5]
-; CGP-NEXT:    v_add_i32_e32 v7, vcc, 1, v4
-; CGP-NEXT:    v_cndmask_b32_e64 v1, v1, v8, s[6:7]
-; CGP-NEXT:    v_add_i32_e32 v8, vcc, 1, v3
-; CGP-NEXT:    v_cmp_ge_u32_e32 vcc, v0, v5
-; CGP-NEXT:    v_cndmask_b32_e32 v0, v4, v7, vcc
-; CGP-NEXT:    v_cmp_ge_u32_e32 vcc, v1, v5
-; CGP-NEXT:    v_cndmask_b32_e32 v1, v3, v8, vcc
-; CGP-NEXT:    v_xor_b32_e32 v0, v0, v2
-; CGP-NEXT:    v_xor_b32_e32 v1, v1, v6
-; CGP-NEXT:    v_sub_i32_e32 v0, vcc, v0, v2
-; CGP-NEXT:    v_sub_i32_e32 v1, vcc, v1, v6
-; CGP-NEXT:    s_setpc_b64 s[30:31]
+; CHECK-LABEL: v_sdiv_v2i32_pow2k_denom:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_ashrrev_i32_e32 v2, 31, v0
+; CHECK-NEXT:    v_ashrrev_i32_e32 v3, 31, v1
+; CHECK-NEXT:    v_lshrrev_b32_e32 v2, 20, v2
+; CHECK-NEXT:    v_lshrrev_b32_e32 v3, 20, v3
+; CHECK-NEXT:    v_add_i32_e32 v0, vcc, v0, v2
+; CHECK-NEXT:    v_add_i32_e32 v1, vcc, v1, v3
+; CHECK-NEXT:    v_ashrrev_i32_e32 v0, 12, v0
+; CHECK-NEXT:    v_ashrrev_i32_e32 v1, 12, v1
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
   %result = sdiv <2 x i32> %num, <i32 4096, i32 4096>
   ret <2 x i32> %result
 }
@@ -884,3 +786,27 @@ define <2 x i32> @v_sdiv_v2i32_24bit(<2 x i32> %num, <2 x i32> %den) {
   %result = sdiv <2 x i32> %num.mask, %den.mask
   ret <2 x i32> %result
 }
+
+define <2 x i32> @v_sdiv_i32_exact(<2 x i32> %num) {
+; GISEL-LABEL: v_sdiv_i32_exact:
+; GISEL:       ; %bb.0:
+; GISEL-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GISEL-NEXT:    v_ashrrev_i32_e32 v0, 12, v0
+; GISEL-NEXT:    v_ashrrev_i32_e32 v1, 10, v1
+; GISEL-NEXT:    s_setpc_b64 s[30:31]
+;
+; CGP-LABEL: v_sdiv_i32_exact:
+; CGP:       ; %bb.0:
+; CGP-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CGP-NEXT:    v_ashrrev_i32_e32 v2, 31, v0
+; CGP-NEXT:    v_ashrrev_i32_e32 v3, 31, v1
+; CGP-NEXT:    v_lshrrev_b32_e32 v2, 20, v2
+; CGP-NEXT:    v_lshrrev_b32_e32 v3, 22, v3
+; CGP-NEXT:    v_add_i32_e32 v0, vcc, v0, v2
+; CGP-NEXT:    v_add_i32_e32 v1, vcc, v1, v3
+; CGP-NEXT:    v_ashrrev_i32_e32 v0, 12, v0
+; CGP-NEXT:    v_ashrrev_i32_e32 v1, 10, v1
+; CGP-NEXT:    s_setpc_b64 s[30:31]
+  %result = sdiv exact <2 x i32> %num, <i32 4096, i32 1024>
+  ret <2 x i32> %result
+}