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::v4f32, Promote);
61 AddPromotedToType(ISD::STORE, MVT::v4f32, MVT::v4i32);
63 setOperationAction(ISD::STORE, MVT::f64, Promote);
64 AddPromotedToType(ISD::STORE, MVT::f64, MVT::i64);
66 setOperationAction(ISD::LOAD, MVT::f32, Promote);
67 AddPromotedToType(ISD::LOAD, MVT::f32, MVT::i32);
69 setOperationAction(ISD::LOAD, MVT::v4f32, Promote);
70 AddPromotedToType(ISD::LOAD, MVT::v4f32, MVT::v4i32);
72 setOperationAction(ISD::LOAD, MVT::f64, Promote);
73 AddPromotedToType(ISD::LOAD, MVT::f64, MVT::i64);
75 setOperationAction(ISD::MUL, MVT::i64, Expand);
77 setOperationAction(ISD::UDIV, MVT::i32, Expand);
78 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
79 setOperationAction(ISD::UREM, MVT::i32, Expand);
81 static const int types[] = {
85 const size_t NumTypes = array_lengthof(types);
87 for (unsigned int x = 0; x < NumTypes; ++x) {
88 MVT::SimpleValueType VT = (MVT::SimpleValueType)types[x];
89 //Expand the following operations for the current type by default
90 setOperationAction(ISD::ADD, VT, Expand);
91 setOperationAction(ISD::AND, VT, Expand);
92 setOperationAction(ISD::MUL, VT, Expand);
93 setOperationAction(ISD::OR, VT, Expand);
94 setOperationAction(ISD::SHL, VT, Expand);
95 setOperationAction(ISD::SRL, VT, Expand);
96 setOperationAction(ISD::SRA, VT, Expand);
97 setOperationAction(ISD::SUB, VT, Expand);
98 setOperationAction(ISD::UDIV, VT, Expand);
99 setOperationAction(ISD::UREM, VT, Expand);
100 setOperationAction(ISD::XOR, VT, Expand);
104 //===---------------------------------------------------------------------===//
105 // TargetLowering Callbacks
106 //===---------------------------------------------------------------------===//
108 void AMDGPUTargetLowering::AnalyzeFormalArguments(CCState &State,
109 const SmallVectorImpl<ISD::InputArg> &Ins) const {
111 State.AnalyzeFormalArguments(Ins, CC_AMDGPU);
114 SDValue AMDGPUTargetLowering::LowerReturn(
116 CallingConv::ID CallConv,
118 const SmallVectorImpl<ISD::OutputArg> &Outs,
119 const SmallVectorImpl<SDValue> &OutVals,
120 SDLoc DL, SelectionDAG &DAG) const {
121 return DAG.getNode(AMDGPUISD::RET_FLAG, DL, MVT::Other, Chain);
124 //===---------------------------------------------------------------------===//
125 // Target specific lowering
126 //===---------------------------------------------------------------------===//
128 SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG)
130 switch (Op.getOpcode()) {
132 Op.getNode()->dump();
133 assert(0 && "Custom lowering code for this"
134 "instruction is not implemented yet!");
136 // AMDIL DAG lowering
137 case ISD::SDIV: return LowerSDIV(Op, DAG);
138 case ISD::SREM: return LowerSREM(Op, DAG);
139 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
140 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
141 // AMDGPU DAG lowering
142 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
143 case ISD::UDIVREM: return LowerUDIVREM(Op, DAG);
148 SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
150 SelectionDAG &DAG) const {
152 const DataLayout *TD = getTargetMachine().getDataLayout();
153 GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
154 // XXX: What does the value of G->getOffset() mean?
155 assert(G->getOffset() == 0 &&
156 "Do not know what to do with an non-zero offset");
158 unsigned Offset = MFI->LDSSize;
159 const GlobalValue *GV = G->getGlobal();
160 uint64_t Size = TD->getTypeAllocSize(GV->getType()->getElementType());
162 // XXX: Account for alignment?
163 MFI->LDSSize += Size;
165 return DAG.getConstant(Offset, TD->getPointerSize() == 8 ? MVT::i64 : MVT::i32);
168 SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
169 SelectionDAG &DAG) const {
170 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
172 EVT VT = Op.getValueType();
174 switch (IntrinsicID) {
176 case AMDGPUIntrinsic::AMDIL_abs:
177 return LowerIntrinsicIABS(Op, DAG);
178 case AMDGPUIntrinsic::AMDIL_exp:
179 return DAG.getNode(ISD::FEXP2, DL, VT, Op.getOperand(1));
180 case AMDGPUIntrinsic::AMDGPU_lrp:
181 return LowerIntrinsicLRP(Op, DAG);
182 case AMDGPUIntrinsic::AMDIL_fraction:
183 return DAG.getNode(AMDGPUISD::FRACT, DL, VT, Op.getOperand(1));
184 case AMDGPUIntrinsic::AMDIL_max:
185 return DAG.getNode(AMDGPUISD::FMAX, DL, VT, Op.getOperand(1),
187 case AMDGPUIntrinsic::AMDGPU_imax:
188 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Op.getOperand(1),
190 case AMDGPUIntrinsic::AMDGPU_umax:
191 return DAG.getNode(AMDGPUISD::UMAX, DL, VT, Op.getOperand(1),
193 case AMDGPUIntrinsic::AMDIL_min:
194 return DAG.getNode(AMDGPUISD::FMIN, DL, VT, Op.getOperand(1),
196 case AMDGPUIntrinsic::AMDGPU_imin:
197 return DAG.getNode(AMDGPUISD::SMIN, DL, VT, Op.getOperand(1),
199 case AMDGPUIntrinsic::AMDGPU_umin:
200 return DAG.getNode(AMDGPUISD::UMIN, DL, VT, Op.getOperand(1),
202 case AMDGPUIntrinsic::AMDIL_round_nearest:
203 return DAG.getNode(ISD::FRINT, DL, VT, Op.getOperand(1));
207 ///IABS(a) = SMAX(sub(0, a), a)
208 SDValue AMDGPUTargetLowering::LowerIntrinsicIABS(SDValue Op,
209 SelectionDAG &DAG) const {
212 EVT VT = Op.getValueType();
213 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
216 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Neg, Op.getOperand(1));
219 /// Linear Interpolation
220 /// LRP(a, b, c) = muladd(a, b, (1 - a) * c)
221 SDValue AMDGPUTargetLowering::LowerIntrinsicLRP(SDValue Op,
222 SelectionDAG &DAG) const {
224 EVT VT = Op.getValueType();
225 SDValue OneSubA = DAG.getNode(ISD::FSUB, DL, VT,
226 DAG.getConstantFP(1.0f, MVT::f32),
228 SDValue OneSubAC = DAG.getNode(ISD::FMUL, DL, VT, OneSubA,
230 return DAG.getNode(ISD::FADD, DL, VT,
231 DAG.getNode(ISD::FMUL, DL, VT, Op.getOperand(1), Op.getOperand(2)),
235 /// \brief Generate Min/Max node
236 SDValue AMDGPUTargetLowering::LowerMinMax(SDValue Op,
237 SelectionDAG &DAG) const {
239 EVT VT = Op.getValueType();
241 SDValue LHS = Op.getOperand(0);
242 SDValue RHS = Op.getOperand(1);
243 SDValue True = Op.getOperand(2);
244 SDValue False = Op.getOperand(3);
245 SDValue CC = Op.getOperand(4);
247 if (VT != MVT::f32 ||
248 !((LHS == True && RHS == False) || (LHS == False && RHS == True))) {
252 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
266 assert(0 && "Operation should already be optimised !");
274 return DAG.getNode(AMDGPUISD::FMIN, DL, VT, LHS, RHS);
276 return DAG.getNode(AMDGPUISD::FMAX, DL, VT, LHS, RHS);
285 return DAG.getNode(AMDGPUISD::FMAX, DL, VT, LHS, RHS);
287 return DAG.getNode(AMDGPUISD::FMIN, DL, VT, LHS, RHS);
289 case ISD::SETCC_INVALID:
290 assert(0 && "Invalid setcc condcode !");
297 SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
298 SelectionDAG &DAG) const {
300 EVT VT = Op.getValueType();
302 SDValue Num = Op.getOperand(0);
303 SDValue Den = Op.getOperand(1);
305 SmallVector<SDValue, 8> Results;
307 // RCP = URECIP(Den) = 2^32 / Den + e
308 // e is rounding error.
309 SDValue RCP = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Den);
311 // RCP_LO = umulo(RCP, Den) */
312 SDValue RCP_LO = DAG.getNode(ISD::UMULO, DL, VT, RCP, Den);
314 // RCP_HI = mulhu (RCP, Den) */
315 SDValue RCP_HI = DAG.getNode(ISD::MULHU, DL, VT, RCP, Den);
317 // NEG_RCP_LO = -RCP_LO
318 SDValue NEG_RCP_LO = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
321 // ABS_RCP_LO = (RCP_HI == 0 ? NEG_RCP_LO : RCP_LO)
322 SDValue ABS_RCP_LO = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
325 // Calculate the rounding error from the URECIP instruction
326 // E = mulhu(ABS_RCP_LO, RCP)
327 SDValue E = DAG.getNode(ISD::MULHU, DL, VT, ABS_RCP_LO, RCP);
330 SDValue RCP_A_E = DAG.getNode(ISD::ADD, DL, VT, RCP, E);
333 SDValue RCP_S_E = DAG.getNode(ISD::SUB, DL, VT, RCP, E);
335 // Tmp0 = (RCP_HI == 0 ? RCP_A_E : RCP_SUB_E)
336 SDValue Tmp0 = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
339 // Quotient = mulhu(Tmp0, Num)
340 SDValue Quotient = DAG.getNode(ISD::MULHU, DL, VT, Tmp0, Num);
342 // Num_S_Remainder = Quotient * Den
343 SDValue Num_S_Remainder = DAG.getNode(ISD::UMULO, DL, VT, Quotient, Den);
345 // Remainder = Num - Num_S_Remainder
346 SDValue Remainder = DAG.getNode(ISD::SUB, DL, VT, Num, Num_S_Remainder);
348 // Remainder_GE_Den = (Remainder >= Den ? -1 : 0)
349 SDValue Remainder_GE_Den = DAG.getSelectCC(DL, Remainder, Den,
350 DAG.getConstant(-1, VT),
351 DAG.getConstant(0, VT),
353 // Remainder_GE_Zero = (Remainder >= 0 ? -1 : 0)
354 SDValue Remainder_GE_Zero = DAG.getSelectCC(DL, Remainder,
355 DAG.getConstant(0, VT),
356 DAG.getConstant(-1, VT),
357 DAG.getConstant(0, VT),
359 // Tmp1 = Remainder_GE_Den & Remainder_GE_Zero
360 SDValue Tmp1 = DAG.getNode(ISD::AND, DL, VT, Remainder_GE_Den,
363 // Calculate Division result:
365 // Quotient_A_One = Quotient + 1
366 SDValue Quotient_A_One = DAG.getNode(ISD::ADD, DL, VT, Quotient,
367 DAG.getConstant(1, VT));
369 // Quotient_S_One = Quotient - 1
370 SDValue Quotient_S_One = DAG.getNode(ISD::SUB, DL, VT, Quotient,
371 DAG.getConstant(1, VT));
373 // Div = (Tmp1 == 0 ? Quotient : Quotient_A_One)
374 SDValue Div = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
375 Quotient, Quotient_A_One, ISD::SETEQ);
377 // Div = (Remainder_GE_Zero == 0 ? Quotient_S_One : Div)
378 Div = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
379 Quotient_S_One, Div, ISD::SETEQ);
381 // Calculate Rem result:
383 // Remainder_S_Den = Remainder - Den
384 SDValue Remainder_S_Den = DAG.getNode(ISD::SUB, DL, VT, Remainder, Den);
386 // Remainder_A_Den = Remainder + Den
387 SDValue Remainder_A_Den = DAG.getNode(ISD::ADD, DL, VT, Remainder, Den);
389 // Rem = (Tmp1 == 0 ? Remainder : Remainder_S_Den)
390 SDValue Rem = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
391 Remainder, Remainder_S_Den, ISD::SETEQ);
393 // Rem = (Remainder_GE_Zero == 0 ? Remainder_A_Den : Rem)
394 Rem = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
395 Remainder_A_Den, Rem, ISD::SETEQ);
399 return DAG.getMergeValues(Ops, 2, DL);
402 //===----------------------------------------------------------------------===//
404 //===----------------------------------------------------------------------===//
406 bool AMDGPUTargetLowering::isHWTrueValue(SDValue Op) const {
407 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
408 return CFP->isExactlyValue(1.0);
410 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
411 return C->isAllOnesValue();
416 bool AMDGPUTargetLowering::isHWFalseValue(SDValue Op) const {
417 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
418 return CFP->getValueAPF().isZero();
420 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
421 return C->isNullValue();
426 SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
427 const TargetRegisterClass *RC,
428 unsigned Reg, EVT VT) const {
429 MachineFunction &MF = DAG.getMachineFunction();
430 MachineRegisterInfo &MRI = MF.getRegInfo();
431 unsigned VirtualRegister;
432 if (!MRI.isLiveIn(Reg)) {
433 VirtualRegister = MRI.createVirtualRegister(RC);
434 MRI.addLiveIn(Reg, VirtualRegister);
436 VirtualRegister = MRI.getLiveInVirtReg(Reg);
438 return DAG.getRegister(VirtualRegister, VT);
441 #define NODE_NAME_CASE(node) case AMDGPUISD::node: return #node;
443 const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
447 NODE_NAME_CASE(CALL);
448 NODE_NAME_CASE(UMUL);
449 NODE_NAME_CASE(DIV_INF);
450 NODE_NAME_CASE(RET_FLAG);
451 NODE_NAME_CASE(BRANCH_COND);
454 NODE_NAME_CASE(DWORDADDR)
455 NODE_NAME_CASE(FRACT)
462 NODE_NAME_CASE(URECIP)
463 NODE_NAME_CASE(EXPORT)
464 NODE_NAME_CASE(CONST_ADDRESS)
465 NODE_NAME_CASE(REGISTER_LOAD)
466 NODE_NAME_CASE(REGISTER_STORE)