[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
@@ -264,7 +265,7 @@ def combine_extracted_vector_load : GICombineRule<
   (match (wip_match_opcode G_EXTRACT_VECTOR_ELT):$root,
         [{ return Helper.matchCombineExtractedVectorLoad(*${root}, ${matchinfo}); }]),
   (apply [{ Helper.applyBuildFn(*${root}, ${matchinfo}); }])>;
-  
+
 def combine_indexed_load_store : GICombineRule<
   (defs root:$root, indexed_load_store_matchdata:$matchinfo),
   (match (wip_match_opcode G_LOAD, G_SEXTLOAD, G_ZEXTLOAD, G_STORE):$root,
@@ -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),
@@ -1356,7 +1370,7 @@ def constant_fold_binops : GICombineGroup<[constant_fold_binop,
 
 def all_combines : GICombineGroup<[trivial_combines, insert_vec_elt_combines,
     extract_vec_elt_combines, combines_for_extload, combine_extracted_vector_load,
-    undef_combines, identity_combines, phi_combines, 
+    undef_combines, identity_combines, phi_combines,
     simplify_add_to_sub, hoist_logic_op_with_same_opcode_hands, shifts_too_big,
     reassocs, ptr_add_immed_chain,
     shl_ashr_to_sext_inreg, sext_inreg_of_load,
@@ -1373,7 +1387,7 @@ def all_combines : GICombineGroup<[trivial_combines, insert_vec_elt_combines,
     intdiv_combines, mulh_combines, redundant_neg_operands,
     and_or_disjoint_mask, fma_combines, fold_binop_into_select,
     sub_add_reg, select_to_minmax, redundant_binop_in_equality,
-    fsub_to_fneg, commute_constant_to_rhs, match_ands, match_ors, 
+    fsub_to_fneg, commute_constant_to_rhs, match_ands, match_ors,
     combine_concat_vector, double_icmp_zero_and_or_combine]>;
 
 // A combine group used to for prelegalizer combiners at -O0. The combines in

llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp

@@ -1490,7 +1490,7 @@ void CombinerHelper::applyOptBrCondByInvertingCond(MachineInstr &MI,
   Observer.changedInstr(*BrCond);
 }
 
- 
+
 bool CombinerHelper::tryEmitMemcpyInline(MachineInstr &MI) {
   MachineIRBuilder HelperBuilder(MI);
   GISelObserverWrapper DummyObserver;
@@ -5286,6 +5286,156 @@ 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) {
+    if (auto *CI = dyn_cast<ConstantInt>(C)) {
+      if (CI->getValue().isPowerOf2())
+        return true;
+      if (IsSigned && CI->getValue().isNegatedPowerOf2())
+        return true;
+    }
+    return false;
+  };
+  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);
+
+  Builder.setInstrAndDebugLoc(MI);
+
+  // Effectively we want to lower G_SDIV %lhs, %rhs, where %rhs is a power of 2,
+  // to the following version:
+  //
+  // %bits = G_CONSTANT $bitwidth
+  // %c1 = G_CTTZ %rhs
+  // %c1 = G_ZEXT %c1
+  // %inexact = G_SUB $bits, %c1
+  // %tmp = G_CONSTANT ($bitwidth - 1)
+  // %sign = %G_ASHR %lhs, %tmp
+  // %srl = G_SHL %sign, %inexact
+  // %add = G_ADD $lhs, $srl
+  // $sra = G_ASHR %add, %c1
+  // %one = G_CONSTANT $1
+  // %allones = G_CONSTANT %111..1
+  // %isone = G_ICMP EQ %rhs, $one
+  // %isallones = G_ICMP EQ %rhs, $allones
+  // %isoneorallones = G_OR %isone, $isallones
+  // %sra = G_SELECT, %isoneorallones, %lhs, %sra
+  // %zero = G_CONSTANT $0
+  // %sub = G_SUB %zero, %sra
+  // %isneg = G_ICMP SLT $lhs, %zero
+  // %res = G_SELECT %isneg, %sub, %sra
+  //
+  // When $rhs is a constant integer, or a splat vector, we can check its value
+  // at compile time such that the first two G_ICMP conditional statements, as
+  // well as the corresponding non-taken branches, can be eliminated. This can
+  // generate compact code even w/o any constant folding afterwards. When $rhs
+  // is not a splat vector, we have to generate those checks via instructions.
+
+  unsigned Bitwidth = Ty.getScalarSizeInBits();
+  auto Zero = Builder.buildConstant(Ty, 0);
+
+  if (auto RHSC = getConstantOrConstantSplatVector(RHS)) {
+    // Special case: (sdiv X, 1) -> X
+    if (RHSC->isOne()) {
+      replaceSingleDefInstWithReg(MI, LHS);
+      return;
+    }
+    // Special Case: (sdiv X, -1) -> 0-X
+    if (RHSC->isAllOnes()) {
+      auto Sub = Builder.buildSub(Ty, Zero, LHS);
+      replaceSingleDefInstWithReg(MI, Sub->getOperand(0).getReg());
+      return;
+    }
+
+    unsigned TrailingZeros = RHSC->countTrailingZeros();
+    auto C1 = Builder.buildConstant(ShiftAmtTy, TrailingZeros);
+    auto Inexact = Builder.buildConstant(ShiftAmtTy, Bitwidth - TrailingZeros);
+    auto Sign = Builder.buildAShr(
+        Ty, LHS, Builder.buildConstant(ShiftAmtTy, Bitwidth - 1));
+    // Add (LHS < 0) ? abs2 - 1 : 0;
+    auto Srl = Builder.buildShl(Ty, Sign, Inexact);
+    auto Add = Builder.buildAdd(Ty, LHS, Srl);
+    auto Sra = Builder.buildAShr(Ty, Add, C1);
+
+    // If dividing by a positive value, we're done. Otherwise, the result must
+    // be negated.
+    auto Res = RHSC->isNegative() ? Builder.buildSub(Ty, Zero, Sra) : Sra;
+    replaceSingleDefInstWithReg(MI, Res->getOperand(0).getReg());
+    return;
+  }
+
+  // RHS is not a splat vector. Build the above version with instructions.
+  auto Bits = Builder.buildConstant(ShiftAmtTy, Bitwidth);
+  auto C1 = Builder.buildCTTZ(Ty, RHS);
+  C1 = Builder.buildZExtOrTrunc(ShiftAmtTy, C1);
+  auto Inexact = Builder.buildSub(ShiftAmtTy, Bits, C1);
+  auto Sign = Builder.buildAShr(
+      Ty, LHS, Builder.buildConstant(ShiftAmtTy, Bitwidth - 1));
+
+  // Add (LHS < 0) ? abs2 - 1 : 0;
+  auto Srl = Builder.buildShl(Ty, Sign, Inexact);
+  auto Add = Builder.buildAdd(Ty, LHS, Srl);
+  auto Sra = Builder.buildAShr(Ty, Add, C1);
+
+  LLT CCVT = LLT::vector(Ty.getElementCount(), 1);
+
+  auto One = Builder.buildConstant(Ty, 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);
+  Sra = Builder.buildSelect(Ty, IsOneOrAllOnes, LHS, Sra);
+
+  // If dividing by a positive value, we're done. Otherwise, the result must
+  // be negated.
+  auto Sub = Builder.buildSub(Ty, Zero, Sra);
+  auto IsNeg = Builder.buildICmp(CmpInst::Predicate::ICMP_SLT, CCVT, LHS, Zero);
+  auto Res = Builder.buildSelect(Ty, IsNeg, Sub, Sra);
+  replaceSingleDefInstWithReg(MI, Res->getOperand(0).getReg());
+}
+
+void CombinerHelper::applyUDivByPow2(MachineInstr &MI) {
+  assert(MI.getOpcode() == TargetOpcode::G_UDIV && "Expected SDIV");
+  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 RHSC = getIConstantVRegValWithLookThrough(RHS, MRI);
+  assert(RHSC.has_value() && "RHS must be a constant");
+  auto RHSCV = RHSC->Value;
+
+  // Special case: (udiv X, 1) -> X
+  if (RHSCV.isOne()) {
+    replaceSingleDefInstWithReg(MI, LHS);
+    return;
+  }
+
+  unsigned TrailingZeros = RHSCV.countTrailingZeros();
+  auto C1 = Builder.buildConstant(ShiftAmtTy, TrailingZeros);
+  auto Res = Builder.buildLShr(Ty, LHS, C1);
+  replaceSingleDefInstWithReg(MI, Res->getOperand(0).getReg());
+}
+
 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_lshl_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)