1 //===-- AMDGPUISelLowering.cpp - AMDGPU Common DAG lowering functions -----===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
11 /// \brief This is the parent TargetLowering class for hardware code gen
14 //===----------------------------------------------------------------------===//
16 #include "AMDGPUISelLowering.h"
18 #include "AMDGPURegisterInfo.h"
19 #include "AMDGPUSubtarget.h"
20 #include "AMDILIntrinsicInfo.h"
21 #include "SIMachineFunctionInfo.h"
22 #include "llvm/CodeGen/CallingConvLower.h"
23 #include "llvm/CodeGen/MachineFunction.h"
24 #include "llvm/CodeGen/MachineRegisterInfo.h"
25 #include "llvm/CodeGen/SelectionDAG.h"
26 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
27 #include "llvm/IR/DataLayout.h"
31 #include "AMDGPUGenCallingConv.inc"
33 AMDGPUTargetLowering::AMDGPUTargetLowering(TargetMachine &TM) :
34 TargetLowering(TM, new TargetLoweringObjectFileELF()) {
36 // Initialize target lowering borrowed from AMDIL
39 // We need to custom lower some of the intrinsics
40 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
42 // Library functions. These default to Expand, but we have instructions
44 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
45 setOperationAction(ISD::FEXP2, MVT::f32, Legal);
46 setOperationAction(ISD::FPOW, MVT::f32, Legal);
47 setOperationAction(ISD::FLOG2, MVT::f32, Legal);
48 setOperationAction(ISD::FABS, MVT::f32, Legal);
49 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
50 setOperationAction(ISD::FRINT, MVT::f32, Legal);
52 // The hardware supports ROTR, but not ROTL
53 setOperationAction(ISD::ROTL, MVT::i32, Expand);
55 // Lower floating point store/load to integer store/load to reduce the number
56 // of patterns in tablegen.
57 setOperationAction(ISD::STORE, MVT::f32, Promote);
58 AddPromotedToType(ISD::STORE, MVT::f32, MVT::i32);
60 setOperationAction(ISD::STORE, MVT::v2f32, Promote);
61 AddPromotedToType(ISD::STORE, MVT::v2f32, MVT::v2i32);
63 setOperationAction(ISD::STORE, MVT::v4f32, Promote);
64 AddPromotedToType(ISD::STORE, MVT::v4f32, MVT::v4i32);
66 setOperationAction(ISD::STORE, MVT::f64, Promote);
67 AddPromotedToType(ISD::STORE, MVT::f64, MVT::i64);
69 setOperationAction(ISD::LOAD, MVT::f32, Promote);
70 AddPromotedToType(ISD::LOAD, MVT::f32, MVT::i32);
72 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
73 AddPromotedToType(ISD::LOAD, MVT::v2f32, MVT::v2i32);
75 setOperationAction(ISD::LOAD, MVT::v4f32, Promote);
76 AddPromotedToType(ISD::LOAD, MVT::v4f32, MVT::v4i32);
78 setOperationAction(ISD::LOAD, MVT::f64, Promote);
79 AddPromotedToType(ISD::LOAD, MVT::f64, MVT::i64);
81 setOperationAction(ISD::MUL, MVT::i64, Expand);
83 setOperationAction(ISD::UDIV, MVT::i32, Expand);
84 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
85 setOperationAction(ISD::UREM, MVT::i32, Expand);
86 setOperationAction(ISD::VSELECT, MVT::v2f32, Expand);
87 setOperationAction(ISD::VSELECT, MVT::v4f32, Expand);
89 static const int types[] = {
93 const size_t NumTypes = array_lengthof(types);
95 for (unsigned int x = 0; x < NumTypes; ++x) {
96 MVT::SimpleValueType VT = (MVT::SimpleValueType)types[x];
97 //Expand the following operations for the current type by default
98 setOperationAction(ISD::ADD, VT, Expand);
99 setOperationAction(ISD::AND, VT, Expand);
100 setOperationAction(ISD::MUL, VT, Expand);
101 setOperationAction(ISD::OR, VT, Expand);
102 setOperationAction(ISD::SHL, VT, Expand);
103 setOperationAction(ISD::SRL, VT, Expand);
104 setOperationAction(ISD::SRA, VT, Expand);
105 setOperationAction(ISD::SUB, VT, Expand);
106 setOperationAction(ISD::UDIV, VT, Expand);
107 setOperationAction(ISD::UREM, VT, Expand);
108 setOperationAction(ISD::VSELECT, VT, Expand);
109 setOperationAction(ISD::XOR, VT, Expand);
113 //===---------------------------------------------------------------------===//
114 // TargetLowering Callbacks
115 //===---------------------------------------------------------------------===//
117 void AMDGPUTargetLowering::AnalyzeFormalArguments(CCState &State,
118 const SmallVectorImpl<ISD::InputArg> &Ins) const {
120 State.AnalyzeFormalArguments(Ins, CC_AMDGPU);
123 SDValue AMDGPUTargetLowering::LowerReturn(
125 CallingConv::ID CallConv,
127 const SmallVectorImpl<ISD::OutputArg> &Outs,
128 const SmallVectorImpl<SDValue> &OutVals,
129 SDLoc DL, SelectionDAG &DAG) const {
130 return DAG.getNode(AMDGPUISD::RET_FLAG, DL, MVT::Other, Chain);
133 //===---------------------------------------------------------------------===//
134 // Target specific lowering
135 //===---------------------------------------------------------------------===//
137 SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG)
139 switch (Op.getOpcode()) {
141 Op.getNode()->dump();
142 assert(0 && "Custom lowering code for this"
143 "instruction is not implemented yet!");
145 // AMDIL DAG lowering
146 case ISD::SDIV: return LowerSDIV(Op, DAG);
147 case ISD::SREM: return LowerSREM(Op, DAG);
148 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
149 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
150 // AMDGPU DAG lowering
151 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
152 case ISD::UDIVREM: return LowerUDIVREM(Op, DAG);
157 SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
159 SelectionDAG &DAG) const {
161 const DataLayout *TD = getTargetMachine().getDataLayout();
162 GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
163 // XXX: What does the value of G->getOffset() mean?
164 assert(G->getOffset() == 0 &&
165 "Do not know what to do with an non-zero offset");
167 unsigned Offset = MFI->LDSSize;
168 const GlobalValue *GV = G->getGlobal();
169 uint64_t Size = TD->getTypeAllocSize(GV->getType()->getElementType());
171 // XXX: Account for alignment?
172 MFI->LDSSize += Size;
174 return DAG.getConstant(Offset, TD->getPointerSize() == 8 ? MVT::i64 : MVT::i32);
177 SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
178 SelectionDAG &DAG) const {
179 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
181 EVT VT = Op.getValueType();
183 switch (IntrinsicID) {
185 case AMDGPUIntrinsic::AMDIL_abs:
186 return LowerIntrinsicIABS(Op, DAG);
187 case AMDGPUIntrinsic::AMDIL_exp:
188 return DAG.getNode(ISD::FEXP2, DL, VT, Op.getOperand(1));
189 case AMDGPUIntrinsic::AMDGPU_lrp:
190 return LowerIntrinsicLRP(Op, DAG);
191 case AMDGPUIntrinsic::AMDIL_fraction:
192 return DAG.getNode(AMDGPUISD::FRACT, DL, VT, Op.getOperand(1));
193 case AMDGPUIntrinsic::AMDIL_max:
194 return DAG.getNode(AMDGPUISD::FMAX, DL, VT, Op.getOperand(1),
196 case AMDGPUIntrinsic::AMDGPU_imax:
197 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Op.getOperand(1),
199 case AMDGPUIntrinsic::AMDGPU_umax:
200 return DAG.getNode(AMDGPUISD::UMAX, DL, VT, Op.getOperand(1),
202 case AMDGPUIntrinsic::AMDIL_min:
203 return DAG.getNode(AMDGPUISD::FMIN, DL, VT, Op.getOperand(1),
205 case AMDGPUIntrinsic::AMDGPU_imin:
206 return DAG.getNode(AMDGPUISD::SMIN, DL, VT, Op.getOperand(1),
208 case AMDGPUIntrinsic::AMDGPU_umin:
209 return DAG.getNode(AMDGPUISD::UMIN, DL, VT, Op.getOperand(1),
211 case AMDGPUIntrinsic::AMDIL_round_nearest:
212 return DAG.getNode(ISD::FRINT, DL, VT, Op.getOperand(1));
216 ///IABS(a) = SMAX(sub(0, a), a)
217 SDValue AMDGPUTargetLowering::LowerIntrinsicIABS(SDValue Op,
218 SelectionDAG &DAG) const {
221 EVT VT = Op.getValueType();
222 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
225 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Neg, Op.getOperand(1));
228 /// Linear Interpolation
229 /// LRP(a, b, c) = muladd(a, b, (1 - a) * c)
230 SDValue AMDGPUTargetLowering::LowerIntrinsicLRP(SDValue Op,
231 SelectionDAG &DAG) const {
233 EVT VT = Op.getValueType();
234 SDValue OneSubA = DAG.getNode(ISD::FSUB, DL, VT,
235 DAG.getConstantFP(1.0f, MVT::f32),
237 SDValue OneSubAC = DAG.getNode(ISD::FMUL, DL, VT, OneSubA,
239 return DAG.getNode(ISD::FADD, DL, VT,
240 DAG.getNode(ISD::FMUL, DL, VT, Op.getOperand(1), Op.getOperand(2)),
244 /// \brief Generate Min/Max node
245 SDValue AMDGPUTargetLowering::LowerMinMax(SDValue Op,
246 SelectionDAG &DAG) const {
248 EVT VT = Op.getValueType();
250 SDValue LHS = Op.getOperand(0);
251 SDValue RHS = Op.getOperand(1);
252 SDValue True = Op.getOperand(2);
253 SDValue False = Op.getOperand(3);
254 SDValue CC = Op.getOperand(4);
256 if (VT != MVT::f32 ||
257 !((LHS == True && RHS == False) || (LHS == False && RHS == True))) {
261 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
275 assert(0 && "Operation should already be optimised !");
283 return DAG.getNode(AMDGPUISD::FMIN, DL, VT, LHS, RHS);
285 return DAG.getNode(AMDGPUISD::FMAX, DL, VT, LHS, RHS);
294 return DAG.getNode(AMDGPUISD::FMAX, DL, VT, LHS, RHS);
296 return DAG.getNode(AMDGPUISD::FMIN, DL, VT, LHS, RHS);
298 case ISD::SETCC_INVALID:
299 assert(0 && "Invalid setcc condcode !");
306 SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
307 SelectionDAG &DAG) const {
309 EVT VT = Op.getValueType();
311 SDValue Num = Op.getOperand(0);
312 SDValue Den = Op.getOperand(1);
314 SmallVector<SDValue, 8> Results;
316 // RCP = URECIP(Den) = 2^32 / Den + e
317 // e is rounding error.
318 SDValue RCP = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Den);
320 // RCP_LO = umulo(RCP, Den) */
321 SDValue RCP_LO = DAG.getNode(ISD::UMULO, DL, VT, RCP, Den);
323 // RCP_HI = mulhu (RCP, Den) */
324 SDValue RCP_HI = DAG.getNode(ISD::MULHU, DL, VT, RCP, Den);
326 // NEG_RCP_LO = -RCP_LO
327 SDValue NEG_RCP_LO = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
330 // ABS_RCP_LO = (RCP_HI == 0 ? NEG_RCP_LO : RCP_LO)
331 SDValue ABS_RCP_LO = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
334 // Calculate the rounding error from the URECIP instruction
335 // E = mulhu(ABS_RCP_LO, RCP)
336 SDValue E = DAG.getNode(ISD::MULHU, DL, VT, ABS_RCP_LO, RCP);
339 SDValue RCP_A_E = DAG.getNode(ISD::ADD, DL, VT, RCP, E);
342 SDValue RCP_S_E = DAG.getNode(ISD::SUB, DL, VT, RCP, E);
344 // Tmp0 = (RCP_HI == 0 ? RCP_A_E : RCP_SUB_E)
345 SDValue Tmp0 = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
348 // Quotient = mulhu(Tmp0, Num)
349 SDValue Quotient = DAG.getNode(ISD::MULHU, DL, VT, Tmp0, Num);
351 // Num_S_Remainder = Quotient * Den
352 SDValue Num_S_Remainder = DAG.getNode(ISD::UMULO, DL, VT, Quotient, Den);
354 // Remainder = Num - Num_S_Remainder
355 SDValue Remainder = DAG.getNode(ISD::SUB, DL, VT, Num, Num_S_Remainder);
357 // Remainder_GE_Den = (Remainder >= Den ? -1 : 0)
358 SDValue Remainder_GE_Den = DAG.getSelectCC(DL, Remainder, Den,
359 DAG.getConstant(-1, VT),
360 DAG.getConstant(0, VT),
362 // Remainder_GE_Zero = (Remainder >= 0 ? -1 : 0)
363 SDValue Remainder_GE_Zero = DAG.getSelectCC(DL, Remainder,
364 DAG.getConstant(0, VT),
365 DAG.getConstant(-1, VT),
366 DAG.getConstant(0, VT),
368 // Tmp1 = Remainder_GE_Den & Remainder_GE_Zero
369 SDValue Tmp1 = DAG.getNode(ISD::AND, DL, VT, Remainder_GE_Den,
372 // Calculate Division result:
374 // Quotient_A_One = Quotient + 1
375 SDValue Quotient_A_One = DAG.getNode(ISD::ADD, DL, VT, Quotient,
376 DAG.getConstant(1, VT));
378 // Quotient_S_One = Quotient - 1
379 SDValue Quotient_S_One = DAG.getNode(ISD::SUB, DL, VT, Quotient,
380 DAG.getConstant(1, VT));
382 // Div = (Tmp1 == 0 ? Quotient : Quotient_A_One)
383 SDValue Div = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
384 Quotient, Quotient_A_One, ISD::SETEQ);
386 // Div = (Remainder_GE_Zero == 0 ? Quotient_S_One : Div)
387 Div = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
388 Quotient_S_One, Div, ISD::SETEQ);
390 // Calculate Rem result:
392 // Remainder_S_Den = Remainder - Den
393 SDValue Remainder_S_Den = DAG.getNode(ISD::SUB, DL, VT, Remainder, Den);
395 // Remainder_A_Den = Remainder + Den
396 SDValue Remainder_A_Den = DAG.getNode(ISD::ADD, DL, VT, Remainder, Den);
398 // Rem = (Tmp1 == 0 ? Remainder : Remainder_S_Den)
399 SDValue Rem = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
400 Remainder, Remainder_S_Den, ISD::SETEQ);
402 // Rem = (Remainder_GE_Zero == 0 ? Remainder_A_Den : Rem)
403 Rem = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
404 Remainder_A_Den, Rem, ISD::SETEQ);
408 return DAG.getMergeValues(Ops, 2, DL);
411 //===----------------------------------------------------------------------===//
413 //===----------------------------------------------------------------------===//
415 bool AMDGPUTargetLowering::isHWTrueValue(SDValue Op) const {
416 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
417 return CFP->isExactlyValue(1.0);
419 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
420 return C->isAllOnesValue();
425 bool AMDGPUTargetLowering::isHWFalseValue(SDValue Op) const {
426 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
427 return CFP->getValueAPF().isZero();
429 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
430 return C->isNullValue();
435 SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
436 const TargetRegisterClass *RC,
437 unsigned Reg, EVT VT) const {
438 MachineFunction &MF = DAG.getMachineFunction();
439 MachineRegisterInfo &MRI = MF.getRegInfo();
440 unsigned VirtualRegister;
441 if (!MRI.isLiveIn(Reg)) {
442 VirtualRegister = MRI.createVirtualRegister(RC);
443 MRI.addLiveIn(Reg, VirtualRegister);
445 VirtualRegister = MRI.getLiveInVirtReg(Reg);
447 return DAG.getRegister(VirtualRegister, VT);
450 #define NODE_NAME_CASE(node) case AMDGPUISD::node: return #node;
452 const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
456 NODE_NAME_CASE(CALL);
457 NODE_NAME_CASE(UMUL);
458 NODE_NAME_CASE(DIV_INF);
459 NODE_NAME_CASE(RET_FLAG);
460 NODE_NAME_CASE(BRANCH_COND);
463 NODE_NAME_CASE(DWORDADDR)
464 NODE_NAME_CASE(FRACT)
471 NODE_NAME_CASE(URECIP)
472 NODE_NAME_CASE(EXPORT)
473 NODE_NAME_CASE(CONST_ADDRESS)
474 NODE_NAME_CASE(REGISTER_LOAD)
475 NODE_NAME_CASE(REGISTER_STORE)