1 | //===-- SparcISelLowering.cpp - Sparc DAG Lowering Implementation ---------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This file implements the interfaces that Sparc uses to lower LLVM code into a |
10 | // selection DAG. |
11 | // |
12 | //===----------------------------------------------------------------------===// |
13 | |
14 | #include "SparcISelLowering.h" |
15 | #include "MCTargetDesc/SparcMCTargetDesc.h" |
16 | #include "SparcMachineFunctionInfo.h" |
17 | #include "SparcRegisterInfo.h" |
18 | #include "SparcSelectionDAGInfo.h" |
19 | #include "SparcTargetMachine.h" |
20 | #include "SparcTargetObjectFile.h" |
21 | #include "llvm/ADT/StringExtras.h" |
22 | #include "llvm/ADT/StringSwitch.h" |
23 | #include "llvm/BinaryFormat/ELF.h" |
24 | #include "llvm/CodeGen/CallingConvLower.h" |
25 | #include "llvm/CodeGen/MachineFrameInfo.h" |
26 | #include "llvm/CodeGen/MachineFunction.h" |
27 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
28 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
29 | #include "llvm/CodeGen/SelectionDAG.h" |
30 | #include "llvm/CodeGen/SelectionDAGNodes.h" |
31 | #include "llvm/CodeGen/TargetLowering.h" |
32 | #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h" |
33 | #include "llvm/IR/DerivedTypes.h" |
34 | #include "llvm/IR/DiagnosticInfo.h" |
35 | #include "llvm/IR/Function.h" |
36 | #include "llvm/IR/Module.h" |
37 | #include "llvm/Support/ErrorHandling.h" |
38 | #include "llvm/Support/KnownBits.h" |
39 | using namespace llvm; |
40 | |
41 | |
42 | //===----------------------------------------------------------------------===// |
43 | // Calling Convention Implementation |
44 | //===----------------------------------------------------------------------===// |
45 | |
46 | static bool CC_Sparc_Assign_SRet(unsigned &ValNo, MVT &ValVT, |
47 | MVT &LocVT, CCValAssign::LocInfo &LocInfo, |
48 | ISD::ArgFlagsTy &ArgFlags, CCState &State) |
49 | { |
50 | assert (ArgFlags.isSRet()); |
51 | |
52 | // Assign SRet argument. |
53 | State.addLoc(V: CCValAssign::getCustomMem(ValNo, ValVT, |
54 | Offset: 0, |
55 | LocVT, HTP: LocInfo)); |
56 | return true; |
57 | } |
58 | |
59 | static bool CC_Sparc_Assign_Split_64(unsigned &ValNo, MVT &ValVT, |
60 | MVT &LocVT, CCValAssign::LocInfo &LocInfo, |
61 | ISD::ArgFlagsTy &ArgFlags, CCState &State) |
62 | { |
63 | static const MCPhysReg RegList[] = { |
64 | SP::I0, SP::I1, SP::I2, SP::I3, SP::I4, SP::I5 |
65 | }; |
66 | // Try to get first reg. |
67 | if (Register Reg = State.AllocateReg(Regs: RegList)) { |
68 | State.addLoc(V: CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, HTP: LocInfo)); |
69 | } else { |
70 | // Assign whole thing in stack. |
71 | State.addLoc(V: CCValAssign::getCustomMem( |
72 | ValNo, ValVT, Offset: State.AllocateStack(Size: 8, Alignment: Align(4)), LocVT, HTP: LocInfo)); |
73 | return true; |
74 | } |
75 | |
76 | // Try to get second reg. |
77 | if (Register Reg = State.AllocateReg(Regs: RegList)) |
78 | State.addLoc(V: CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, HTP: LocInfo)); |
79 | else |
80 | State.addLoc(V: CCValAssign::getCustomMem( |
81 | ValNo, ValVT, Offset: State.AllocateStack(Size: 4, Alignment: Align(4)), LocVT, HTP: LocInfo)); |
82 | return true; |
83 | } |
84 | |
85 | static bool CC_Sparc_Assign_Ret_Split_64(unsigned &ValNo, MVT &ValVT, |
86 | MVT &LocVT, CCValAssign::LocInfo &LocInfo, |
87 | ISD::ArgFlagsTy &ArgFlags, CCState &State) |
88 | { |
89 | static const MCPhysReg RegList[] = { |
90 | SP::I0, SP::I1, SP::I2, SP::I3, SP::I4, SP::I5 |
91 | }; |
92 | |
93 | // Try to get first reg. |
94 | if (Register Reg = State.AllocateReg(Regs: RegList)) |
95 | State.addLoc(V: CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, HTP: LocInfo)); |
96 | else |
97 | return false; |
98 | |
99 | // Try to get second reg. |
100 | if (Register Reg = State.AllocateReg(Regs: RegList)) |
101 | State.addLoc(V: CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, HTP: LocInfo)); |
102 | else |
103 | return false; |
104 | |
105 | return true; |
106 | } |
107 | |
108 | // Allocate a full-sized argument for the 64-bit ABI. |
109 | static bool Analyze_CC_Sparc64_Full(bool IsReturn, unsigned &ValNo, MVT &ValVT, |
110 | MVT &LocVT, CCValAssign::LocInfo &LocInfo, |
111 | ISD::ArgFlagsTy &ArgFlags, CCState &State) { |
112 | assert((LocVT == MVT::f32 || LocVT == MVT::f128 |
113 | || LocVT.getSizeInBits() == 64) && |
114 | "Can't handle non-64 bits locations" ); |
115 | |
116 | // Stack space is allocated for all arguments starting from [%fp+BIAS+128]. |
117 | unsigned size = (LocVT == MVT::f128) ? 16 : 8; |
118 | Align alignment = (LocVT == MVT::f128) ? Align(16) : Align(8); |
119 | unsigned Offset = State.AllocateStack(Size: size, Alignment: alignment); |
120 | unsigned Reg = 0; |
121 | |
122 | if (LocVT == MVT::i64 && Offset < 6*8) |
123 | // Promote integers to %i0-%i5. |
124 | Reg = SP::I0 + Offset/8; |
125 | else if (LocVT == MVT::f64 && Offset < 16*8) |
126 | // Promote doubles to %d0-%d30. (Which LLVM calls D0-D15). |
127 | Reg = SP::D0 + Offset/8; |
128 | else if (LocVT == MVT::f32 && Offset < 16*8) |
129 | // Promote floats to %f1, %f3, ... |
130 | Reg = SP::F1 + Offset/4; |
131 | else if (LocVT == MVT::f128 && Offset < 16*8) |
132 | // Promote long doubles to %q0-%q28. (Which LLVM calls Q0-Q7). |
133 | Reg = SP::Q0 + Offset/16; |
134 | |
135 | // Promote to register when possible, otherwise use the stack slot. |
136 | if (Reg) { |
137 | State.addLoc(V: CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, HTP: LocInfo)); |
138 | return true; |
139 | } |
140 | |
141 | // Bail out if this is a return CC and we run out of registers to place |
142 | // values into. |
143 | if (IsReturn) |
144 | return false; |
145 | |
146 | // This argument goes on the stack in an 8-byte slot. |
147 | // When passing floats, LocVT is smaller than 8 bytes. Adjust the offset to |
148 | // the right-aligned float. The first 4 bytes of the stack slot are undefined. |
149 | if (LocVT == MVT::f32) |
150 | Offset += 4; |
151 | |
152 | State.addLoc(V: CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, HTP: LocInfo)); |
153 | return true; |
154 | } |
155 | |
156 | // Allocate a half-sized argument for the 64-bit ABI. |
157 | // |
158 | // This is used when passing { float, int } structs by value in registers. |
159 | static bool Analyze_CC_Sparc64_Half(bool IsReturn, unsigned &ValNo, MVT &ValVT, |
160 | MVT &LocVT, CCValAssign::LocInfo &LocInfo, |
161 | ISD::ArgFlagsTy &ArgFlags, CCState &State) { |
162 | assert(LocVT.getSizeInBits() == 32 && "Can't handle non-32 bits locations" ); |
163 | unsigned Offset = State.AllocateStack(Size: 4, Alignment: Align(4)); |
164 | |
165 | if (LocVT == MVT::f32 && Offset < 16*8) { |
166 | // Promote floats to %f0-%f31. |
167 | State.addLoc(V: CCValAssign::getReg(ValNo, ValVT, Reg: SP::F0 + Offset/4, |
168 | LocVT, HTP: LocInfo)); |
169 | return true; |
170 | } |
171 | |
172 | if (LocVT == MVT::i32 && Offset < 6*8) { |
173 | // Promote integers to %i0-%i5, using half the register. |
174 | unsigned Reg = SP::I0 + Offset/8; |
175 | LocVT = MVT::i64; |
176 | LocInfo = CCValAssign::AExt; |
177 | |
178 | // Set the Custom bit if this i32 goes in the high bits of a register. |
179 | if (Offset % 8 == 0) |
180 | State.addLoc(V: CCValAssign::getCustomReg(ValNo, ValVT, Reg, |
181 | LocVT, HTP: LocInfo)); |
182 | else |
183 | State.addLoc(V: CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, HTP: LocInfo)); |
184 | return true; |
185 | } |
186 | |
187 | // Bail out if this is a return CC and we run out of registers to place |
188 | // values into. |
189 | if (IsReturn) |
190 | return false; |
191 | |
192 | State.addLoc(V: CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, HTP: LocInfo)); |
193 | return true; |
194 | } |
195 | |
196 | static bool CC_Sparc64_Full(unsigned &ValNo, MVT &ValVT, MVT &LocVT, |
197 | CCValAssign::LocInfo &LocInfo, |
198 | ISD::ArgFlagsTy &ArgFlags, CCState &State) { |
199 | return Analyze_CC_Sparc64_Full(IsReturn: false, ValNo, ValVT, LocVT, LocInfo, ArgFlags, |
200 | State); |
201 | } |
202 | |
203 | static bool CC_Sparc64_Half(unsigned &ValNo, MVT &ValVT, MVT &LocVT, |
204 | CCValAssign::LocInfo &LocInfo, |
205 | ISD::ArgFlagsTy &ArgFlags, CCState &State) { |
206 | return Analyze_CC_Sparc64_Half(IsReturn: false, ValNo, ValVT, LocVT, LocInfo, ArgFlags, |
207 | State); |
208 | } |
209 | |
210 | static bool RetCC_Sparc64_Full(unsigned &ValNo, MVT &ValVT, MVT &LocVT, |
211 | CCValAssign::LocInfo &LocInfo, |
212 | ISD::ArgFlagsTy &ArgFlags, CCState &State) { |
213 | return Analyze_CC_Sparc64_Full(IsReturn: true, ValNo, ValVT, LocVT, LocInfo, ArgFlags, |
214 | State); |
215 | } |
216 | |
217 | static bool RetCC_Sparc64_Half(unsigned &ValNo, MVT &ValVT, MVT &LocVT, |
218 | CCValAssign::LocInfo &LocInfo, |
219 | ISD::ArgFlagsTy &ArgFlags, CCState &State) { |
220 | return Analyze_CC_Sparc64_Half(IsReturn: true, ValNo, ValVT, LocVT, LocInfo, ArgFlags, |
221 | State); |
222 | } |
223 | |
224 | #include "SparcGenCallingConv.inc" |
225 | |
226 | // The calling conventions in SparcCallingConv.td are described in terms of the |
227 | // callee's register window. This function translates registers to the |
228 | // corresponding caller window %o register. |
229 | static unsigned toCallerWindow(unsigned Reg) { |
230 | static_assert(SP::I0 + 7 == SP::I7 && SP::O0 + 7 == SP::O7, |
231 | "Unexpected enum" ); |
232 | if (Reg >= SP::I0 && Reg <= SP::I7) |
233 | return Reg - SP::I0 + SP::O0; |
234 | return Reg; |
235 | } |
236 | |
237 | bool SparcTargetLowering::CanLowerReturn( |
238 | CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg, |
239 | const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context, |
240 | const Type *RetTy) const { |
241 | SmallVector<CCValAssign, 16> RVLocs; |
242 | CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context); |
243 | return CCInfo.CheckReturn(Outs, Fn: Subtarget->is64Bit() ? RetCC_Sparc64 |
244 | : RetCC_Sparc32); |
245 | } |
246 | |
247 | SDValue |
248 | SparcTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv, |
249 | bool IsVarArg, |
250 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
251 | const SmallVectorImpl<SDValue> &OutVals, |
252 | const SDLoc &DL, SelectionDAG &DAG) const { |
253 | if (Subtarget->is64Bit()) |
254 | return LowerReturn_64(Chain, CallConv, IsVarArg, Outs, OutVals, DL, DAG); |
255 | return LowerReturn_32(Chain, CallConv, IsVarArg, Outs, OutVals, DL, DAG); |
256 | } |
257 | |
258 | SDValue |
259 | SparcTargetLowering::LowerReturn_32(SDValue Chain, CallingConv::ID CallConv, |
260 | bool IsVarArg, |
261 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
262 | const SmallVectorImpl<SDValue> &OutVals, |
263 | const SDLoc &DL, SelectionDAG &DAG) const { |
264 | MachineFunction &MF = DAG.getMachineFunction(); |
265 | |
266 | // CCValAssign - represent the assignment of the return value to locations. |
267 | SmallVector<CCValAssign, 16> RVLocs; |
268 | |
269 | // CCState - Info about the registers and stack slot. |
270 | CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), RVLocs, |
271 | *DAG.getContext()); |
272 | |
273 | // Analyze return values. |
274 | CCInfo.AnalyzeReturn(Outs, Fn: RetCC_Sparc32); |
275 | |
276 | SDValue Glue; |
277 | SmallVector<SDValue, 4> RetOps(1, Chain); |
278 | // Make room for the return address offset. |
279 | RetOps.push_back(Elt: SDValue()); |
280 | |
281 | // Copy the result values into the output registers. |
282 | for (unsigned i = 0, realRVLocIdx = 0; |
283 | i != RVLocs.size(); |
284 | ++i, ++realRVLocIdx) { |
285 | CCValAssign &VA = RVLocs[i]; |
286 | assert(VA.isRegLoc() && "Can only return in registers!" ); |
287 | |
288 | SDValue Arg = OutVals[realRVLocIdx]; |
289 | |
290 | if (VA.needsCustom()) { |
291 | assert(VA.getLocVT() == MVT::v2i32); |
292 | // Legalize ret v2i32 -> ret 2 x i32 (Basically: do what would |
293 | // happen by default if this wasn't a legal type) |
294 | |
295 | SDValue Part0 = DAG.getNode(Opcode: ISD::EXTRACT_VECTOR_ELT, DL, VT: MVT::i32, |
296 | N1: Arg, |
297 | N2: DAG.getConstant(Val: 0, DL, VT: getVectorIdxTy(DL: DAG.getDataLayout()))); |
298 | SDValue Part1 = DAG.getNode(Opcode: ISD::EXTRACT_VECTOR_ELT, DL, VT: MVT::i32, |
299 | N1: Arg, |
300 | N2: DAG.getConstant(Val: 1, DL, VT: getVectorIdxTy(DL: DAG.getDataLayout()))); |
301 | |
302 | Chain = DAG.getCopyToReg(Chain, dl: DL, Reg: VA.getLocReg(), N: Part0, Glue); |
303 | Glue = Chain.getValue(R: 1); |
304 | RetOps.push_back(Elt: DAG.getRegister(Reg: VA.getLocReg(), VT: VA.getLocVT())); |
305 | VA = RVLocs[++i]; // skip ahead to next loc |
306 | Chain = DAG.getCopyToReg(Chain, dl: DL, Reg: VA.getLocReg(), N: Part1, |
307 | Glue); |
308 | } else |
309 | Chain = DAG.getCopyToReg(Chain, dl: DL, Reg: VA.getLocReg(), N: Arg, Glue); |
310 | |
311 | // Guarantee that all emitted copies are stuck together with flags. |
312 | Glue = Chain.getValue(R: 1); |
313 | RetOps.push_back(Elt: DAG.getRegister(Reg: VA.getLocReg(), VT: VA.getLocVT())); |
314 | } |
315 | |
316 | unsigned RetAddrOffset = 8; // Call Inst + Delay Slot |
317 | // If the function returns a struct, copy the SRetReturnReg to I0 |
318 | if (MF.getFunction().hasStructRetAttr()) { |
319 | SparcMachineFunctionInfo *SFI = MF.getInfo<SparcMachineFunctionInfo>(); |
320 | Register Reg = SFI->getSRetReturnReg(); |
321 | if (!Reg) |
322 | llvm_unreachable("sret virtual register not created in the entry block" ); |
323 | auto PtrVT = getPointerTy(DL: DAG.getDataLayout()); |
324 | SDValue Val = DAG.getCopyFromReg(Chain, dl: DL, Reg, VT: PtrVT); |
325 | Chain = DAG.getCopyToReg(Chain, dl: DL, Reg: SP::I0, N: Val, Glue); |
326 | Glue = Chain.getValue(R: 1); |
327 | RetOps.push_back(Elt: DAG.getRegister(Reg: SP::I0, VT: PtrVT)); |
328 | RetAddrOffset = 12; // CallInst + Delay Slot + Unimp |
329 | } |
330 | |
331 | RetOps[0] = Chain; // Update chain. |
332 | RetOps[1] = DAG.getConstant(Val: RetAddrOffset, DL, VT: MVT::i32); |
333 | |
334 | // Add the glue if we have it. |
335 | if (Glue.getNode()) |
336 | RetOps.push_back(Elt: Glue); |
337 | |
338 | return DAG.getNode(Opcode: SPISD::RET_GLUE, DL, VT: MVT::Other, Ops: RetOps); |
339 | } |
340 | |
341 | // Lower return values for the 64-bit ABI. |
342 | // Return values are passed the exactly the same way as function arguments. |
343 | SDValue |
344 | SparcTargetLowering::LowerReturn_64(SDValue Chain, CallingConv::ID CallConv, |
345 | bool IsVarArg, |
346 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
347 | const SmallVectorImpl<SDValue> &OutVals, |
348 | const SDLoc &DL, SelectionDAG &DAG) const { |
349 | // CCValAssign - represent the assignment of the return value to locations. |
350 | SmallVector<CCValAssign, 16> RVLocs; |
351 | |
352 | // CCState - Info about the registers and stack slot. |
353 | CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), RVLocs, |
354 | *DAG.getContext()); |
355 | |
356 | // Analyze return values. |
357 | CCInfo.AnalyzeReturn(Outs, Fn: RetCC_Sparc64); |
358 | |
359 | SDValue Glue; |
360 | SmallVector<SDValue, 4> RetOps(1, Chain); |
361 | |
362 | // The second operand on the return instruction is the return address offset. |
363 | // The return address is always %i7+8 with the 64-bit ABI. |
364 | RetOps.push_back(Elt: DAG.getConstant(Val: 8, DL, VT: MVT::i32)); |
365 | |
366 | // Copy the result values into the output registers. |
367 | for (unsigned i = 0; i != RVLocs.size(); ++i) { |
368 | CCValAssign &VA = RVLocs[i]; |
369 | assert(VA.isRegLoc() && "Can only return in registers!" ); |
370 | SDValue OutVal = OutVals[i]; |
371 | |
372 | // Integer return values must be sign or zero extended by the callee. |
373 | switch (VA.getLocInfo()) { |
374 | case CCValAssign::Full: break; |
375 | case CCValAssign::SExt: |
376 | OutVal = DAG.getNode(Opcode: ISD::SIGN_EXTEND, DL, VT: VA.getLocVT(), Operand: OutVal); |
377 | break; |
378 | case CCValAssign::ZExt: |
379 | OutVal = DAG.getNode(Opcode: ISD::ZERO_EXTEND, DL, VT: VA.getLocVT(), Operand: OutVal); |
380 | break; |
381 | case CCValAssign::AExt: |
382 | OutVal = DAG.getNode(Opcode: ISD::ANY_EXTEND, DL, VT: VA.getLocVT(), Operand: OutVal); |
383 | break; |
384 | default: |
385 | llvm_unreachable("Unknown loc info!" ); |
386 | } |
387 | |
388 | // The custom bit on an i32 return value indicates that it should be passed |
389 | // in the high bits of the register. |
390 | if (VA.getValVT() == MVT::i32 && VA.needsCustom()) { |
391 | OutVal = DAG.getNode(Opcode: ISD::SHL, DL, VT: MVT::i64, N1: OutVal, |
392 | N2: DAG.getConstant(Val: 32, DL, VT: MVT::i32)); |
393 | |
394 | // The next value may go in the low bits of the same register. |
395 | // Handle both at once. |
396 | if (i+1 < RVLocs.size() && RVLocs[i+1].getLocReg() == VA.getLocReg()) { |
397 | SDValue NV = DAG.getNode(Opcode: ISD::ZERO_EXTEND, DL, VT: MVT::i64, Operand: OutVals[i+1]); |
398 | OutVal = DAG.getNode(Opcode: ISD::OR, DL, VT: MVT::i64, N1: OutVal, N2: NV); |
399 | // Skip the next value, it's already done. |
400 | ++i; |
401 | } |
402 | } |
403 | |
404 | Chain = DAG.getCopyToReg(Chain, dl: DL, Reg: VA.getLocReg(), N: OutVal, Glue); |
405 | |
406 | // Guarantee that all emitted copies are stuck together with flags. |
407 | Glue = Chain.getValue(R: 1); |
408 | RetOps.push_back(Elt: DAG.getRegister(Reg: VA.getLocReg(), VT: VA.getLocVT())); |
409 | } |
410 | |
411 | RetOps[0] = Chain; // Update chain. |
412 | |
413 | // Add the flag if we have it. |
414 | if (Glue.getNode()) |
415 | RetOps.push_back(Elt: Glue); |
416 | |
417 | return DAG.getNode(Opcode: SPISD::RET_GLUE, DL, VT: MVT::Other, Ops: RetOps); |
418 | } |
419 | |
420 | SDValue SparcTargetLowering::LowerFormalArguments( |
421 | SDValue Chain, CallingConv::ID CallConv, bool IsVarArg, |
422 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL, |
423 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { |
424 | if (Subtarget->is64Bit()) |
425 | return LowerFormalArguments_64(Chain, CallConv, isVarArg: IsVarArg, Ins, |
426 | dl: DL, DAG, InVals); |
427 | return LowerFormalArguments_32(Chain, CallConv, isVarArg: IsVarArg, Ins, |
428 | dl: DL, DAG, InVals); |
429 | } |
430 | |
431 | /// LowerFormalArguments32 - V8 uses a very simple ABI, where all values are |
432 | /// passed in either one or two GPRs, including FP values. TODO: we should |
433 | /// pass FP values in FP registers for fastcc functions. |
434 | SDValue SparcTargetLowering::LowerFormalArguments_32( |
435 | SDValue Chain, CallingConv::ID CallConv, bool isVarArg, |
436 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl, |
437 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { |
438 | MachineFunction &MF = DAG.getMachineFunction(); |
439 | MachineRegisterInfo &RegInfo = MF.getRegInfo(); |
440 | SparcMachineFunctionInfo *FuncInfo = MF.getInfo<SparcMachineFunctionInfo>(); |
441 | |
442 | // Assign locations to all of the incoming arguments. |
443 | SmallVector<CCValAssign, 16> ArgLocs; |
444 | CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs, |
445 | *DAG.getContext()); |
446 | CCInfo.AnalyzeFormalArguments(Ins, Fn: CC_Sparc32); |
447 | |
448 | const unsigned StackOffset = 92; |
449 | bool IsLittleEndian = DAG.getDataLayout().isLittleEndian(); |
450 | |
451 | unsigned InIdx = 0; |
452 | for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i, ++InIdx) { |
453 | CCValAssign &VA = ArgLocs[i]; |
454 | |
455 | if (Ins[InIdx].Flags.isSRet()) { |
456 | if (InIdx != 0) |
457 | report_fatal_error(reason: "sparc only supports sret on the first parameter" ); |
458 | // Get SRet from [%fp+64]. |
459 | int FrameIdx = MF.getFrameInfo().CreateFixedObject(Size: 4, SPOffset: 64, IsImmutable: true); |
460 | SDValue FIPtr = DAG.getFrameIndex(FI: FrameIdx, VT: MVT::i32); |
461 | SDValue Arg = |
462 | DAG.getLoad(VT: MVT::i32, dl, Chain, Ptr: FIPtr, PtrInfo: MachinePointerInfo()); |
463 | InVals.push_back(Elt: Arg); |
464 | continue; |
465 | } |
466 | |
467 | if (VA.isRegLoc()) { |
468 | if (VA.needsCustom()) { |
469 | assert(VA.getLocVT() == MVT::f64 || VA.getLocVT() == MVT::v2i32); |
470 | |
471 | Register VRegHi = RegInfo.createVirtualRegister(RegClass: &SP::IntRegsRegClass); |
472 | MF.getRegInfo().addLiveIn(Reg: VA.getLocReg(), vreg: VRegHi); |
473 | SDValue HiVal = DAG.getCopyFromReg(Chain, dl, Reg: VRegHi, VT: MVT::i32); |
474 | |
475 | assert(i+1 < e); |
476 | CCValAssign &NextVA = ArgLocs[++i]; |
477 | |
478 | SDValue LoVal; |
479 | if (NextVA.isMemLoc()) { |
480 | int FrameIdx = MF.getFrameInfo(). |
481 | CreateFixedObject(Size: 4, SPOffset: StackOffset+NextVA.getLocMemOffset(),IsImmutable: true); |
482 | SDValue FIPtr = DAG.getFrameIndex(FI: FrameIdx, VT: MVT::i32); |
483 | LoVal = DAG.getLoad(VT: MVT::i32, dl, Chain, Ptr: FIPtr, PtrInfo: MachinePointerInfo()); |
484 | } else { |
485 | Register loReg = MF.addLiveIn(PReg: NextVA.getLocReg(), |
486 | RC: &SP::IntRegsRegClass); |
487 | LoVal = DAG.getCopyFromReg(Chain, dl, Reg: loReg, VT: MVT::i32); |
488 | } |
489 | |
490 | if (IsLittleEndian) |
491 | std::swap(a&: LoVal, b&: HiVal); |
492 | |
493 | SDValue WholeValue = |
494 | DAG.getNode(Opcode: ISD::BUILD_PAIR, DL: dl, VT: MVT::i64, N1: LoVal, N2: HiVal); |
495 | WholeValue = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: VA.getLocVT(), Operand: WholeValue); |
496 | InVals.push_back(Elt: WholeValue); |
497 | continue; |
498 | } |
499 | Register VReg = RegInfo.createVirtualRegister(RegClass: &SP::IntRegsRegClass); |
500 | MF.getRegInfo().addLiveIn(Reg: VA.getLocReg(), vreg: VReg); |
501 | SDValue Arg = DAG.getCopyFromReg(Chain, dl, Reg: VReg, VT: MVT::i32); |
502 | if (VA.getLocVT() == MVT::f32) |
503 | Arg = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: MVT::f32, Operand: Arg); |
504 | else if (VA.getLocVT() != MVT::i32) { |
505 | Arg = DAG.getNode(Opcode: ISD::AssertSext, DL: dl, VT: MVT::i32, N1: Arg, |
506 | N2: DAG.getValueType(VA.getLocVT())); |
507 | Arg = DAG.getNode(Opcode: ISD::TRUNCATE, DL: dl, VT: VA.getLocVT(), Operand: Arg); |
508 | } |
509 | InVals.push_back(Elt: Arg); |
510 | continue; |
511 | } |
512 | |
513 | assert(VA.isMemLoc()); |
514 | |
515 | unsigned Offset = VA.getLocMemOffset()+StackOffset; |
516 | auto PtrVT = getPointerTy(DL: DAG.getDataLayout()); |
517 | |
518 | if (VA.needsCustom()) { |
519 | assert(VA.getValVT() == MVT::f64 || VA.getValVT() == MVT::v2i32); |
520 | // If it is double-word aligned, just load. |
521 | if (Offset % 8 == 0) { |
522 | int FI = MF.getFrameInfo().CreateFixedObject(Size: 8, |
523 | SPOffset: Offset, |
524 | IsImmutable: true); |
525 | SDValue FIPtr = DAG.getFrameIndex(FI, VT: PtrVT); |
526 | SDValue Load = |
527 | DAG.getLoad(VT: VA.getValVT(), dl, Chain, Ptr: FIPtr, PtrInfo: MachinePointerInfo()); |
528 | InVals.push_back(Elt: Load); |
529 | continue; |
530 | } |
531 | |
532 | int FI = MF.getFrameInfo().CreateFixedObject(Size: 4, |
533 | SPOffset: Offset, |
534 | IsImmutable: true); |
535 | SDValue FIPtr = DAG.getFrameIndex(FI, VT: PtrVT); |
536 | SDValue HiVal = |
537 | DAG.getLoad(VT: MVT::i32, dl, Chain, Ptr: FIPtr, PtrInfo: MachinePointerInfo()); |
538 | int FI2 = MF.getFrameInfo().CreateFixedObject(Size: 4, |
539 | SPOffset: Offset+4, |
540 | IsImmutable: true); |
541 | SDValue FIPtr2 = DAG.getFrameIndex(FI: FI2, VT: PtrVT); |
542 | |
543 | SDValue LoVal = |
544 | DAG.getLoad(VT: MVT::i32, dl, Chain, Ptr: FIPtr2, PtrInfo: MachinePointerInfo()); |
545 | |
546 | if (IsLittleEndian) |
547 | std::swap(a&: LoVal, b&: HiVal); |
548 | |
549 | SDValue WholeValue = |
550 | DAG.getNode(Opcode: ISD::BUILD_PAIR, DL: dl, VT: MVT::i64, N1: LoVal, N2: HiVal); |
551 | WholeValue = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: VA.getValVT(), Operand: WholeValue); |
552 | InVals.push_back(Elt: WholeValue); |
553 | continue; |
554 | } |
555 | |
556 | int FI = MF.getFrameInfo().CreateFixedObject(Size: 4, |
557 | SPOffset: Offset, |
558 | IsImmutable: true); |
559 | SDValue FIPtr = DAG.getFrameIndex(FI, VT: PtrVT); |
560 | SDValue Load ; |
561 | if (VA.getValVT() == MVT::i32 || VA.getValVT() == MVT::f32) { |
562 | Load = DAG.getLoad(VT: VA.getValVT(), dl, Chain, Ptr: FIPtr, PtrInfo: MachinePointerInfo()); |
563 | } else if (VA.getValVT() == MVT::f128) { |
564 | report_fatal_error(reason: "SPARCv8 does not handle f128 in calls; " |
565 | "pass indirectly" ); |
566 | } else { |
567 | // We shouldn't see any other value types here. |
568 | llvm_unreachable("Unexpected ValVT encountered in frame lowering." ); |
569 | } |
570 | InVals.push_back(Elt: Load); |
571 | } |
572 | |
573 | if (MF.getFunction().hasStructRetAttr()) { |
574 | // Copy the SRet Argument to SRetReturnReg. |
575 | SparcMachineFunctionInfo *SFI = MF.getInfo<SparcMachineFunctionInfo>(); |
576 | Register Reg = SFI->getSRetReturnReg(); |
577 | if (!Reg) { |
578 | Reg = MF.getRegInfo().createVirtualRegister(RegClass: &SP::IntRegsRegClass); |
579 | SFI->setSRetReturnReg(Reg); |
580 | } |
581 | SDValue Copy = DAG.getCopyToReg(Chain: DAG.getEntryNode(), dl, Reg, N: InVals[0]); |
582 | Chain = DAG.getNode(Opcode: ISD::TokenFactor, DL: dl, VT: MVT::Other, N1: Copy, N2: Chain); |
583 | } |
584 | |
585 | // Store remaining ArgRegs to the stack if this is a varargs function. |
586 | if (isVarArg) { |
587 | static const MCPhysReg ArgRegs[] = { |
588 | SP::I0, SP::I1, SP::I2, SP::I3, SP::I4, SP::I5 |
589 | }; |
590 | unsigned NumAllocated = CCInfo.getFirstUnallocated(Regs: ArgRegs); |
591 | const MCPhysReg *CurArgReg = ArgRegs+NumAllocated, *ArgRegEnd = ArgRegs+6; |
592 | unsigned ArgOffset = CCInfo.getStackSize(); |
593 | if (NumAllocated == 6) |
594 | ArgOffset += StackOffset; |
595 | else { |
596 | assert(!ArgOffset); |
597 | ArgOffset = 68+4*NumAllocated; |
598 | } |
599 | |
600 | // Remember the vararg offset for the va_start implementation. |
601 | FuncInfo->setVarArgsFrameOffset(ArgOffset); |
602 | |
603 | std::vector<SDValue> OutChains; |
604 | |
605 | for (; CurArgReg != ArgRegEnd; ++CurArgReg) { |
606 | Register VReg = RegInfo.createVirtualRegister(RegClass: &SP::IntRegsRegClass); |
607 | MF.getRegInfo().addLiveIn(Reg: *CurArgReg, vreg: VReg); |
608 | SDValue Arg = DAG.getCopyFromReg(Chain: DAG.getRoot(), dl, Reg: VReg, VT: MVT::i32); |
609 | |
610 | int FrameIdx = MF.getFrameInfo().CreateFixedObject(Size: 4, SPOffset: ArgOffset, |
611 | IsImmutable: true); |
612 | SDValue FIPtr = DAG.getFrameIndex(FI: FrameIdx, VT: MVT::i32); |
613 | |
614 | OutChains.push_back( |
615 | x: DAG.getStore(Chain: DAG.getRoot(), dl, Val: Arg, Ptr: FIPtr, PtrInfo: MachinePointerInfo())); |
616 | ArgOffset += 4; |
617 | } |
618 | |
619 | if (!OutChains.empty()) { |
620 | OutChains.push_back(x: Chain); |
621 | Chain = DAG.getNode(Opcode: ISD::TokenFactor, DL: dl, VT: MVT::Other, Ops: OutChains); |
622 | } |
623 | } |
624 | |
625 | return Chain; |
626 | } |
627 | |
628 | // Lower formal arguments for the 64 bit ABI. |
629 | SDValue SparcTargetLowering::LowerFormalArguments_64( |
630 | SDValue Chain, CallingConv::ID CallConv, bool IsVarArg, |
631 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL, |
632 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { |
633 | MachineFunction &MF = DAG.getMachineFunction(); |
634 | |
635 | // Analyze arguments according to CC_Sparc64. |
636 | SmallVector<CCValAssign, 16> ArgLocs; |
637 | CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs, |
638 | *DAG.getContext()); |
639 | CCInfo.AnalyzeFormalArguments(Ins, Fn: CC_Sparc64); |
640 | |
641 | // The argument array begins at %fp+BIAS+128, after the register save area. |
642 | const unsigned ArgArea = 128; |
643 | |
644 | for (const CCValAssign &VA : ArgLocs) { |
645 | if (VA.isRegLoc()) { |
646 | // This argument is passed in a register. |
647 | // All integer register arguments are promoted by the caller to i64. |
648 | |
649 | // Create a virtual register for the promoted live-in value. |
650 | Register VReg = MF.addLiveIn(PReg: VA.getLocReg(), |
651 | RC: getRegClassFor(VT: VA.getLocVT())); |
652 | SDValue Arg = DAG.getCopyFromReg(Chain, dl: DL, Reg: VReg, VT: VA.getLocVT()); |
653 | |
654 | // Get the high bits for i32 struct elements. |
655 | if (VA.getValVT() == MVT::i32 && VA.needsCustom()) |
656 | Arg = DAG.getNode(Opcode: ISD::SRL, DL, VT: VA.getLocVT(), N1: Arg, |
657 | N2: DAG.getConstant(Val: 32, DL, VT: MVT::i32)); |
658 | |
659 | // The caller promoted the argument, so insert an Assert?ext SDNode so we |
660 | // won't promote the value again in this function. |
661 | switch (VA.getLocInfo()) { |
662 | case CCValAssign::SExt: |
663 | Arg = DAG.getNode(Opcode: ISD::AssertSext, DL, VT: VA.getLocVT(), N1: Arg, |
664 | N2: DAG.getValueType(VA.getValVT())); |
665 | break; |
666 | case CCValAssign::ZExt: |
667 | Arg = DAG.getNode(Opcode: ISD::AssertZext, DL, VT: VA.getLocVT(), N1: Arg, |
668 | N2: DAG.getValueType(VA.getValVT())); |
669 | break; |
670 | default: |
671 | break; |
672 | } |
673 | |
674 | // Truncate the register down to the argument type. |
675 | if (VA.isExtInLoc()) |
676 | Arg = DAG.getNode(Opcode: ISD::TRUNCATE, DL, VT: VA.getValVT(), Operand: Arg); |
677 | |
678 | InVals.push_back(Elt: Arg); |
679 | continue; |
680 | } |
681 | |
682 | // The registers are exhausted. This argument was passed on the stack. |
683 | assert(VA.isMemLoc()); |
684 | // The CC_Sparc64_Full/Half functions compute stack offsets relative to the |
685 | // beginning of the arguments area at %fp+BIAS+128. |
686 | unsigned Offset = VA.getLocMemOffset() + ArgArea; |
687 | unsigned ValSize = VA.getValVT().getSizeInBits() / 8; |
688 | // Adjust offset for extended arguments, SPARC is big-endian. |
689 | // The caller will have written the full slot with extended bytes, but we |
690 | // prefer our own extending loads. |
691 | if (VA.isExtInLoc()) |
692 | Offset += 8 - ValSize; |
693 | int FI = MF.getFrameInfo().CreateFixedObject(Size: ValSize, SPOffset: Offset, IsImmutable: true); |
694 | InVals.push_back( |
695 | Elt: DAG.getLoad(VT: VA.getValVT(), dl: DL, Chain, |
696 | Ptr: DAG.getFrameIndex(FI, VT: getPointerTy(DL: MF.getDataLayout())), |
697 | PtrInfo: MachinePointerInfo::getFixedStack(MF, FI))); |
698 | } |
699 | |
700 | if (!IsVarArg) |
701 | return Chain; |
702 | |
703 | // This function takes variable arguments, some of which may have been passed |
704 | // in registers %i0-%i5. Variable floating point arguments are never passed |
705 | // in floating point registers. They go on %i0-%i5 or on the stack like |
706 | // integer arguments. |
707 | // |
708 | // The va_start intrinsic needs to know the offset to the first variable |
709 | // argument. |
710 | unsigned ArgOffset = CCInfo.getStackSize(); |
711 | SparcMachineFunctionInfo *FuncInfo = MF.getInfo<SparcMachineFunctionInfo>(); |
712 | // Skip the 128 bytes of register save area. |
713 | FuncInfo->setVarArgsFrameOffset(ArgOffset + ArgArea + |
714 | Subtarget->getStackPointerBias()); |
715 | |
716 | // Save the variable arguments that were passed in registers. |
717 | // The caller is required to reserve stack space for 6 arguments regardless |
718 | // of how many arguments were actually passed. |
719 | SmallVector<SDValue, 8> OutChains; |
720 | for (; ArgOffset < 6*8; ArgOffset += 8) { |
721 | Register VReg = MF.addLiveIn(PReg: SP::I0 + ArgOffset/8, RC: &SP::I64RegsRegClass); |
722 | SDValue VArg = DAG.getCopyFromReg(Chain, dl: DL, Reg: VReg, VT: MVT::i64); |
723 | int FI = MF.getFrameInfo().CreateFixedObject(Size: 8, SPOffset: ArgOffset + ArgArea, IsImmutable: true); |
724 | auto PtrVT = getPointerTy(DL: MF.getDataLayout()); |
725 | OutChains.push_back( |
726 | Elt: DAG.getStore(Chain, dl: DL, Val: VArg, Ptr: DAG.getFrameIndex(FI, VT: PtrVT), |
727 | PtrInfo: MachinePointerInfo::getFixedStack(MF, FI))); |
728 | } |
729 | |
730 | if (!OutChains.empty()) |
731 | Chain = DAG.getNode(Opcode: ISD::TokenFactor, DL, VT: MVT::Other, Ops: OutChains); |
732 | |
733 | return Chain; |
734 | } |
735 | |
736 | // Check whether any of the argument registers are reserved |
737 | static bool isAnyArgRegReserved(const SparcRegisterInfo *TRI, |
738 | const MachineFunction &MF) { |
739 | // The register window design means that outgoing parameters at O* |
740 | // will appear in the callee as I*. |
741 | // Be conservative and check both sides of the register names. |
742 | bool Outgoing = |
743 | llvm::any_of(Range: SP::GPROutgoingArgRegClass, P: [TRI, &MF](MCPhysReg r) { |
744 | return TRI->isReservedReg(MF, Reg: r); |
745 | }); |
746 | bool Incoming = |
747 | llvm::any_of(Range: SP::GPRIncomingArgRegClass, P: [TRI, &MF](MCPhysReg r) { |
748 | return TRI->isReservedReg(MF, Reg: r); |
749 | }); |
750 | return Outgoing || Incoming; |
751 | } |
752 | |
753 | static void emitReservedArgRegCallError(const MachineFunction &MF) { |
754 | const Function &F = MF.getFunction(); |
755 | F.getContext().diagnose(DI: DiagnosticInfoUnsupported{ |
756 | F, ("SPARC doesn't support" |
757 | " function calls if any of the argument registers is reserved." )}); |
758 | } |
759 | |
760 | SDValue |
761 | SparcTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI, |
762 | SmallVectorImpl<SDValue> &InVals) const { |
763 | if (Subtarget->is64Bit()) |
764 | return LowerCall_64(CLI, InVals); |
765 | return LowerCall_32(CLI, InVals); |
766 | } |
767 | |
768 | static bool hasReturnsTwiceAttr(SelectionDAG &DAG, SDValue Callee, |
769 | const CallBase *Call) { |
770 | if (Call) |
771 | return Call->hasFnAttr(Kind: Attribute::ReturnsTwice); |
772 | |
773 | const Function *CalleeFn = nullptr; |
774 | if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Val&: Callee)) { |
775 | CalleeFn = dyn_cast<Function>(Val: G->getGlobal()); |
776 | } else if (ExternalSymbolSDNode *E = |
777 | dyn_cast<ExternalSymbolSDNode>(Val&: Callee)) { |
778 | const Function &Fn = DAG.getMachineFunction().getFunction(); |
779 | const Module *M = Fn.getParent(); |
780 | const char *CalleeName = E->getSymbol(); |
781 | CalleeFn = M->getFunction(Name: CalleeName); |
782 | } |
783 | |
784 | if (!CalleeFn) |
785 | return false; |
786 | return CalleeFn->hasFnAttribute(Kind: Attribute::ReturnsTwice); |
787 | } |
788 | |
789 | /// IsEligibleForTailCallOptimization - Check whether the call is eligible |
790 | /// for tail call optimization. |
791 | bool SparcTargetLowering::IsEligibleForTailCallOptimization( |
792 | CCState &CCInfo, CallLoweringInfo &CLI, MachineFunction &MF) const { |
793 | |
794 | auto &Outs = CLI.Outs; |
795 | auto &Caller = MF.getFunction(); |
796 | |
797 | // Do not tail call opt functions with "disable-tail-calls" attribute. |
798 | if (Caller.getFnAttribute(Kind: "disable-tail-calls" ).getValueAsString() == "true" ) |
799 | return false; |
800 | |
801 | // Do not tail call opt if the stack is used to pass parameters. |
802 | // 64-bit targets have a slightly higher limit since the ABI requires |
803 | // to allocate some space even when all the parameters fit inside registers. |
804 | unsigned StackSizeLimit = Subtarget->is64Bit() ? 48 : 0; |
805 | if (CCInfo.getStackSize() > StackSizeLimit) |
806 | return false; |
807 | |
808 | // Do not tail call opt if either the callee or caller returns |
809 | // a struct and the other does not. |
810 | if (!Outs.empty() && Caller.hasStructRetAttr() != Outs[0].Flags.isSRet()) |
811 | return false; |
812 | |
813 | // Byval parameters hand the function a pointer directly into the stack area |
814 | // we want to reuse during a tail call. |
815 | for (auto &Arg : Outs) |
816 | if (Arg.Flags.isByVal()) |
817 | return false; |
818 | |
819 | return true; |
820 | } |
821 | |
822 | // Lower a call for the 32-bit ABI. |
823 | SDValue |
824 | SparcTargetLowering::LowerCall_32(TargetLowering::CallLoweringInfo &CLI, |
825 | SmallVectorImpl<SDValue> &InVals) const { |
826 | SelectionDAG &DAG = CLI.DAG; |
827 | SDLoc &dl = CLI.DL; |
828 | SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs; |
829 | SmallVectorImpl<SDValue> &OutVals = CLI.OutVals; |
830 | SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins; |
831 | SDValue Chain = CLI.Chain; |
832 | SDValue Callee = CLI.Callee; |
833 | bool &isTailCall = CLI.IsTailCall; |
834 | CallingConv::ID CallConv = CLI.CallConv; |
835 | bool isVarArg = CLI.IsVarArg; |
836 | MachineFunction &MF = DAG.getMachineFunction(); |
837 | |
838 | // Analyze operands of the call, assigning locations to each operand. |
839 | SmallVector<CCValAssign, 16> ArgLocs; |
840 | CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs, |
841 | *DAG.getContext()); |
842 | CCInfo.AnalyzeCallOperands(Outs, Fn: CC_Sparc32); |
843 | |
844 | isTailCall = isTailCall && IsEligibleForTailCallOptimization( |
845 | CCInfo, CLI, MF&: DAG.getMachineFunction()); |
846 | |
847 | // Get the size of the outgoing arguments stack space requirement. |
848 | unsigned ArgsSize = CCInfo.getStackSize(); |
849 | |
850 | // Keep stack frames 8-byte aligned. |
851 | ArgsSize = (ArgsSize+7) & ~7; |
852 | |
853 | MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); |
854 | |
855 | // Create local copies for byval args. |
856 | SmallVector<SDValue, 8> ByValArgs; |
857 | for (unsigned i = 0, e = Outs.size(); i != e; ++i) { |
858 | ISD::ArgFlagsTy Flags = Outs[i].Flags; |
859 | if (!Flags.isByVal()) |
860 | continue; |
861 | |
862 | SDValue Arg = OutVals[i]; |
863 | unsigned Size = Flags.getByValSize(); |
864 | Align Alignment = Flags.getNonZeroByValAlign(); |
865 | |
866 | if (Size > 0U) { |
867 | int FI = MFI.CreateStackObject(Size, Alignment, isSpillSlot: false); |
868 | SDValue FIPtr = DAG.getFrameIndex(FI, VT: getPointerTy(DL: DAG.getDataLayout())); |
869 | SDValue SizeNode = DAG.getConstant(Val: Size, DL: dl, VT: MVT::i32); |
870 | |
871 | Chain = DAG.getMemcpy(Chain, dl, Dst: FIPtr, Src: Arg, Size: SizeNode, Alignment, |
872 | isVol: false, // isVolatile, |
873 | AlwaysInline: (Size <= 32), // AlwaysInline if size <= 32, |
874 | /*CI=*/nullptr, OverrideTailCall: std::nullopt, DstPtrInfo: MachinePointerInfo(), |
875 | SrcPtrInfo: MachinePointerInfo()); |
876 | ByValArgs.push_back(Elt: FIPtr); |
877 | } |
878 | else { |
879 | SDValue nullVal; |
880 | ByValArgs.push_back(Elt: nullVal); |
881 | } |
882 | } |
883 | |
884 | assert(!isTailCall || ArgsSize == 0); |
885 | |
886 | if (!isTailCall) |
887 | Chain = DAG.getCALLSEQ_START(Chain, InSize: ArgsSize, OutSize: 0, DL: dl); |
888 | |
889 | SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass; |
890 | SmallVector<SDValue, 8> MemOpChains; |
891 | |
892 | const unsigned StackOffset = 92; |
893 | bool hasStructRetAttr = false; |
894 | unsigned SRetArgSize = 0; |
895 | // Walk the register/memloc assignments, inserting copies/loads. |
896 | for (unsigned i = 0, realArgIdx = 0, byvalArgIdx = 0, e = ArgLocs.size(); |
897 | i != e; |
898 | ++i, ++realArgIdx) { |
899 | CCValAssign &VA = ArgLocs[i]; |
900 | SDValue Arg = OutVals[realArgIdx]; |
901 | |
902 | ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags; |
903 | |
904 | // Use local copy if it is a byval arg. |
905 | if (Flags.isByVal()) { |
906 | Arg = ByValArgs[byvalArgIdx++]; |
907 | if (!Arg) { |
908 | continue; |
909 | } |
910 | } |
911 | |
912 | // Promote the value if needed. |
913 | switch (VA.getLocInfo()) { |
914 | default: llvm_unreachable("Unknown loc info!" ); |
915 | case CCValAssign::Full: break; |
916 | case CCValAssign::SExt: |
917 | Arg = DAG.getNode(Opcode: ISD::SIGN_EXTEND, DL: dl, VT: VA.getLocVT(), Operand: Arg); |
918 | break; |
919 | case CCValAssign::ZExt: |
920 | Arg = DAG.getNode(Opcode: ISD::ZERO_EXTEND, DL: dl, VT: VA.getLocVT(), Operand: Arg); |
921 | break; |
922 | case CCValAssign::AExt: |
923 | Arg = DAG.getNode(Opcode: ISD::ANY_EXTEND, DL: dl, VT: VA.getLocVT(), Operand: Arg); |
924 | break; |
925 | case CCValAssign::BCvt: |
926 | Arg = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: VA.getLocVT(), Operand: Arg); |
927 | break; |
928 | } |
929 | |
930 | if (Flags.isSRet()) { |
931 | assert(VA.needsCustom()); |
932 | |
933 | if (isTailCall) |
934 | continue; |
935 | |
936 | // store SRet argument in %sp+64 |
937 | SDValue StackPtr = DAG.getRegister(Reg: SP::O6, VT: MVT::i32); |
938 | SDValue PtrOff = DAG.getIntPtrConstant(Val: 64, DL: dl); |
939 | PtrOff = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT: MVT::i32, N1: StackPtr, N2: PtrOff); |
940 | MemOpChains.push_back( |
941 | Elt: DAG.getStore(Chain, dl, Val: Arg, Ptr: PtrOff, PtrInfo: MachinePointerInfo())); |
942 | hasStructRetAttr = true; |
943 | // sret only allowed on first argument |
944 | assert(Outs[realArgIdx].OrigArgIndex == 0); |
945 | SRetArgSize = |
946 | DAG.getDataLayout().getTypeAllocSize(Ty: CLI.getArgs()[0].IndirectType); |
947 | continue; |
948 | } |
949 | |
950 | if (VA.needsCustom()) { |
951 | assert(VA.getLocVT() == MVT::f64 || VA.getLocVT() == MVT::v2i32); |
952 | |
953 | if (VA.isMemLoc()) { |
954 | unsigned Offset = VA.getLocMemOffset() + StackOffset; |
955 | // if it is double-word aligned, just store. |
956 | if (Offset % 8 == 0) { |
957 | SDValue StackPtr = DAG.getRegister(Reg: SP::O6, VT: MVT::i32); |
958 | SDValue PtrOff = DAG.getIntPtrConstant(Val: Offset, DL: dl); |
959 | PtrOff = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT: MVT::i32, N1: StackPtr, N2: PtrOff); |
960 | MemOpChains.push_back( |
961 | Elt: DAG.getStore(Chain, dl, Val: Arg, Ptr: PtrOff, PtrInfo: MachinePointerInfo())); |
962 | continue; |
963 | } |
964 | } |
965 | |
966 | if (VA.getLocVT() == MVT::f64) { |
967 | // Move from the float value from float registers into the |
968 | // integer registers. |
969 | if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Val&: Arg)) |
970 | Arg = bitcastConstantFPToInt(C, DL: dl, DAG); |
971 | else |
972 | Arg = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: MVT::v2i32, Operand: Arg); |
973 | } |
974 | |
975 | SDValue Part0 = DAG.getNode(Opcode: ISD::EXTRACT_VECTOR_ELT, DL: dl, VT: MVT::i32, |
976 | N1: Arg, |
977 | N2: DAG.getConstant(Val: 0, DL: dl, VT: getVectorIdxTy(DL: DAG.getDataLayout()))); |
978 | SDValue Part1 = DAG.getNode(Opcode: ISD::EXTRACT_VECTOR_ELT, DL: dl, VT: MVT::i32, |
979 | N1: Arg, |
980 | N2: DAG.getConstant(Val: 1, DL: dl, VT: getVectorIdxTy(DL: DAG.getDataLayout()))); |
981 | |
982 | if (VA.isRegLoc()) { |
983 | RegsToPass.push_back(Elt: std::make_pair(x: VA.getLocReg(), y&: Part0)); |
984 | assert(i+1 != e); |
985 | CCValAssign &NextVA = ArgLocs[++i]; |
986 | if (NextVA.isRegLoc()) { |
987 | RegsToPass.push_back(Elt: std::make_pair(x: NextVA.getLocReg(), y&: Part1)); |
988 | } else { |
989 | // Store the second part in stack. |
990 | unsigned Offset = NextVA.getLocMemOffset() + StackOffset; |
991 | SDValue StackPtr = DAG.getRegister(Reg: SP::O6, VT: MVT::i32); |
992 | SDValue PtrOff = DAG.getIntPtrConstant(Val: Offset, DL: dl); |
993 | PtrOff = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT: MVT::i32, N1: StackPtr, N2: PtrOff); |
994 | MemOpChains.push_back( |
995 | Elt: DAG.getStore(Chain, dl, Val: Part1, Ptr: PtrOff, PtrInfo: MachinePointerInfo())); |
996 | } |
997 | } else { |
998 | unsigned Offset = VA.getLocMemOffset() + StackOffset; |
999 | // Store the first part. |
1000 | SDValue StackPtr = DAG.getRegister(Reg: SP::O6, VT: MVT::i32); |
1001 | SDValue PtrOff = DAG.getIntPtrConstant(Val: Offset, DL: dl); |
1002 | PtrOff = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT: MVT::i32, N1: StackPtr, N2: PtrOff); |
1003 | MemOpChains.push_back( |
1004 | Elt: DAG.getStore(Chain, dl, Val: Part0, Ptr: PtrOff, PtrInfo: MachinePointerInfo())); |
1005 | // Store the second part. |
1006 | PtrOff = DAG.getIntPtrConstant(Val: Offset + 4, DL: dl); |
1007 | PtrOff = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT: MVT::i32, N1: StackPtr, N2: PtrOff); |
1008 | MemOpChains.push_back( |
1009 | Elt: DAG.getStore(Chain, dl, Val: Part1, Ptr: PtrOff, PtrInfo: MachinePointerInfo())); |
1010 | } |
1011 | continue; |
1012 | } |
1013 | |
1014 | // Arguments that can be passed on register must be kept at |
1015 | // RegsToPass vector |
1016 | if (VA.isRegLoc()) { |
1017 | if (VA.getLocVT() != MVT::f32) { |
1018 | RegsToPass.push_back(Elt: std::make_pair(x: VA.getLocReg(), y&: Arg)); |
1019 | continue; |
1020 | } |
1021 | Arg = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: MVT::i32, Operand: Arg); |
1022 | RegsToPass.push_back(Elt: std::make_pair(x: VA.getLocReg(), y&: Arg)); |
1023 | continue; |
1024 | } |
1025 | |
1026 | assert(VA.isMemLoc()); |
1027 | |
1028 | // Create a store off the stack pointer for this argument. |
1029 | SDValue StackPtr = DAG.getRegister(Reg: SP::O6, VT: MVT::i32); |
1030 | SDValue PtrOff = DAG.getIntPtrConstant(Val: VA.getLocMemOffset() + StackOffset, |
1031 | DL: dl); |
1032 | PtrOff = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT: MVT::i32, N1: StackPtr, N2: PtrOff); |
1033 | MemOpChains.push_back( |
1034 | Elt: DAG.getStore(Chain, dl, Val: Arg, Ptr: PtrOff, PtrInfo: MachinePointerInfo())); |
1035 | } |
1036 | |
1037 | |
1038 | // Emit all stores, make sure the occur before any copies into physregs. |
1039 | if (!MemOpChains.empty()) |
1040 | Chain = DAG.getNode(Opcode: ISD::TokenFactor, DL: dl, VT: MVT::Other, Ops: MemOpChains); |
1041 | |
1042 | // Build a sequence of copy-to-reg nodes chained together with token |
1043 | // chain and flag operands which copy the outgoing args into registers. |
1044 | // The InGlue in necessary since all emitted instructions must be |
1045 | // stuck together. |
1046 | SDValue InGlue; |
1047 | for (const auto &[OrigReg, N] : RegsToPass) { |
1048 | Register Reg = isTailCall ? OrigReg : toCallerWindow(Reg: OrigReg); |
1049 | Chain = DAG.getCopyToReg(Chain, dl, Reg, N, Glue: InGlue); |
1050 | InGlue = Chain.getValue(R: 1); |
1051 | } |
1052 | |
1053 | bool hasReturnsTwice = hasReturnsTwiceAttr(DAG, Callee, Call: CLI.CB); |
1054 | |
1055 | // If the callee is a GlobalAddress node (quite common, every direct call is) |
1056 | // turn it into a TargetGlobalAddress node so that legalize doesn't hack it. |
1057 | // Likewise ExternalSymbol -> TargetExternalSymbol. |
1058 | if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Val&: Callee)) |
1059 | Callee = DAG.getTargetGlobalAddress(GV: G->getGlobal(), DL: dl, VT: MVT::i32, offset: 0); |
1060 | else if (ExternalSymbolSDNode *E = dyn_cast<ExternalSymbolSDNode>(Val&: Callee)) |
1061 | Callee = DAG.getTargetExternalSymbol(Sym: E->getSymbol(), VT: MVT::i32); |
1062 | |
1063 | // Returns a chain & a flag for retval copy to use |
1064 | SDVTList NodeTys = DAG.getVTList(VT1: MVT::Other, VT2: MVT::Glue); |
1065 | SmallVector<SDValue, 8> Ops; |
1066 | Ops.push_back(Elt: Chain); |
1067 | Ops.push_back(Elt: Callee); |
1068 | if (hasStructRetAttr) |
1069 | Ops.push_back(Elt: DAG.getTargetConstant(Val: SRetArgSize, DL: dl, VT: MVT::i32)); |
1070 | for (const auto &[OrigReg, N] : RegsToPass) { |
1071 | Register Reg = isTailCall ? OrigReg : toCallerWindow(Reg: OrigReg); |
1072 | Ops.push_back(Elt: DAG.getRegister(Reg, VT: N.getValueType())); |
1073 | } |
1074 | |
1075 | // Add a register mask operand representing the call-preserved registers. |
1076 | const SparcRegisterInfo *TRI = Subtarget->getRegisterInfo(); |
1077 | const uint32_t *Mask = |
1078 | ((hasReturnsTwice) |
1079 | ? TRI->getRTCallPreservedMask(CC: CallConv) |
1080 | : TRI->getCallPreservedMask(MF: DAG.getMachineFunction(), CC: CallConv)); |
1081 | |
1082 | if (isAnyArgRegReserved(TRI, MF)) |
1083 | emitReservedArgRegCallError(MF); |
1084 | |
1085 | assert(Mask && "Missing call preserved mask for calling convention" ); |
1086 | Ops.push_back(Elt: DAG.getRegisterMask(RegMask: Mask)); |
1087 | |
1088 | if (InGlue.getNode()) |
1089 | Ops.push_back(Elt: InGlue); |
1090 | |
1091 | if (isTailCall) { |
1092 | DAG.getMachineFunction().getFrameInfo().setHasTailCall(); |
1093 | return DAG.getNode(Opcode: SPISD::TAIL_CALL, DL: dl, VT: MVT::Other, Ops); |
1094 | } |
1095 | |
1096 | Chain = DAG.getNode(Opcode: SPISD::CALL, DL: dl, VTList: NodeTys, Ops); |
1097 | InGlue = Chain.getValue(R: 1); |
1098 | |
1099 | Chain = DAG.getCALLSEQ_END(Chain, Size1: ArgsSize, Size2: 0, Glue: InGlue, DL: dl); |
1100 | InGlue = Chain.getValue(R: 1); |
1101 | |
1102 | // Assign locations to each value returned by this call. |
1103 | SmallVector<CCValAssign, 16> RVLocs; |
1104 | CCState RVInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs, |
1105 | *DAG.getContext()); |
1106 | |
1107 | RVInfo.AnalyzeCallResult(Ins, Fn: RetCC_Sparc32); |
1108 | |
1109 | // Copy all of the result registers out of their specified physreg. |
1110 | for (unsigned i = 0; i != RVLocs.size(); ++i) { |
1111 | assert(RVLocs[i].isRegLoc() && "Can only return in registers!" ); |
1112 | if (RVLocs[i].getLocVT() == MVT::v2i32) { |
1113 | SDValue Vec = DAG.getNode(Opcode: ISD::UNDEF, DL: dl, VT: MVT::v2i32); |
1114 | SDValue Lo = DAG.getCopyFromReg( |
1115 | Chain, dl, Reg: toCallerWindow(Reg: RVLocs[i++].getLocReg()), VT: MVT::i32, Glue: InGlue); |
1116 | Chain = Lo.getValue(R: 1); |
1117 | InGlue = Lo.getValue(R: 2); |
1118 | Vec = DAG.getNode(Opcode: ISD::INSERT_VECTOR_ELT, DL: dl, VT: MVT::v2i32, N1: Vec, N2: Lo, |
1119 | N3: DAG.getConstant(Val: 0, DL: dl, VT: MVT::i32)); |
1120 | SDValue Hi = DAG.getCopyFromReg( |
1121 | Chain, dl, Reg: toCallerWindow(Reg: RVLocs[i].getLocReg()), VT: MVT::i32, Glue: InGlue); |
1122 | Chain = Hi.getValue(R: 1); |
1123 | InGlue = Hi.getValue(R: 2); |
1124 | Vec = DAG.getNode(Opcode: ISD::INSERT_VECTOR_ELT, DL: dl, VT: MVT::v2i32, N1: Vec, N2: Hi, |
1125 | N3: DAG.getConstant(Val: 1, DL: dl, VT: MVT::i32)); |
1126 | InVals.push_back(Elt: Vec); |
1127 | } else { |
1128 | Chain = |
1129 | DAG.getCopyFromReg(Chain, dl, Reg: toCallerWindow(Reg: RVLocs[i].getLocReg()), |
1130 | VT: RVLocs[i].getValVT(), Glue: InGlue) |
1131 | .getValue(R: 1); |
1132 | InGlue = Chain.getValue(R: 2); |
1133 | InVals.push_back(Elt: Chain.getValue(R: 0)); |
1134 | } |
1135 | } |
1136 | |
1137 | return Chain; |
1138 | } |
1139 | |
1140 | // FIXME? Maybe this could be a TableGen attribute on some registers and |
1141 | // this table could be generated automatically from RegInfo. |
1142 | Register SparcTargetLowering::getRegisterByName(const char* RegName, LLT VT, |
1143 | const MachineFunction &MF) const { |
1144 | Register Reg = StringSwitch<Register>(RegName) |
1145 | .Case(S: "i0" , Value: SP::I0).Case(S: "i1" , Value: SP::I1).Case(S: "i2" , Value: SP::I2).Case(S: "i3" , Value: SP::I3) |
1146 | .Case(S: "i4" , Value: SP::I4).Case(S: "i5" , Value: SP::I5).Case(S: "i6" , Value: SP::I6).Case(S: "i7" , Value: SP::I7) |
1147 | .Case(S: "o0" , Value: SP::O0).Case(S: "o1" , Value: SP::O1).Case(S: "o2" , Value: SP::O2).Case(S: "o3" , Value: SP::O3) |
1148 | .Case(S: "o4" , Value: SP::O4).Case(S: "o5" , Value: SP::O5).Case(S: "o6" , Value: SP::O6).Case(S: "o7" , Value: SP::O7) |
1149 | .Case(S: "l0" , Value: SP::L0).Case(S: "l1" , Value: SP::L1).Case(S: "l2" , Value: SP::L2).Case(S: "l3" , Value: SP::L3) |
1150 | .Case(S: "l4" , Value: SP::L4).Case(S: "l5" , Value: SP::L5).Case(S: "l6" , Value: SP::L6).Case(S: "l7" , Value: SP::L7) |
1151 | .Case(S: "g0" , Value: SP::G0).Case(S: "g1" , Value: SP::G1).Case(S: "g2" , Value: SP::G2).Case(S: "g3" , Value: SP::G3) |
1152 | .Case(S: "g4" , Value: SP::G4).Case(S: "g5" , Value: SP::G5).Case(S: "g6" , Value: SP::G6).Case(S: "g7" , Value: SP::G7) |
1153 | .Default(Value: 0); |
1154 | |
1155 | // If we're directly referencing register names |
1156 | // (e.g in GCC C extension `register int r asm("g1");`), |
1157 | // make sure that said register is in the reserve list. |
1158 | const SparcRegisterInfo *TRI = Subtarget->getRegisterInfo(); |
1159 | if (!TRI->isReservedReg(MF, Reg)) |
1160 | Reg = Register(); |
1161 | |
1162 | return Reg; |
1163 | } |
1164 | |
1165 | // Fixup floating point arguments in the ... part of a varargs call. |
1166 | // |
1167 | // The SPARC v9 ABI requires that floating point arguments are treated the same |
1168 | // as integers when calling a varargs function. This does not apply to the |
1169 | // fixed arguments that are part of the function's prototype. |
1170 | // |
1171 | // This function post-processes a CCValAssign array created by |
1172 | // AnalyzeCallOperands(). |
1173 | static void fixupVariableFloatArgs(SmallVectorImpl<CCValAssign> &ArgLocs, |
1174 | ArrayRef<ISD::OutputArg> Outs) { |
1175 | for (CCValAssign &VA : ArgLocs) { |
1176 | MVT ValTy = VA.getLocVT(); |
1177 | // FIXME: What about f32 arguments? C promotes them to f64 when calling |
1178 | // varargs functions. |
1179 | if (!VA.isRegLoc() || (ValTy != MVT::f64 && ValTy != MVT::f128)) |
1180 | continue; |
1181 | // The fixed arguments to a varargs function still go in FP registers. |
1182 | if (Outs[VA.getValNo()].IsFixed) |
1183 | continue; |
1184 | |
1185 | // This floating point argument should be reassigned. |
1186 | // Determine the offset into the argument array. |
1187 | Register firstReg = (ValTy == MVT::f64) ? SP::D0 : SP::Q0; |
1188 | unsigned argSize = (ValTy == MVT::f64) ? 8 : 16; |
1189 | unsigned Offset = argSize * (VA.getLocReg() - firstReg); |
1190 | assert(Offset < 16*8 && "Offset out of range, bad register enum?" ); |
1191 | |
1192 | if (Offset < 6*8) { |
1193 | // This argument should go in %i0-%i5. |
1194 | unsigned IReg = SP::I0 + Offset/8; |
1195 | if (ValTy == MVT::f64) |
1196 | // Full register, just bitconvert into i64. |
1197 | VA = CCValAssign::getReg(ValNo: VA.getValNo(), ValVT: VA.getValVT(), Reg: IReg, LocVT: MVT::i64, |
1198 | HTP: CCValAssign::BCvt); |
1199 | else { |
1200 | assert(ValTy == MVT::f128 && "Unexpected type!" ); |
1201 | // Full register, just bitconvert into i128 -- We will lower this into |
1202 | // two i64s in LowerCall_64. |
1203 | VA = CCValAssign::getCustomReg(ValNo: VA.getValNo(), ValVT: VA.getValVT(), Reg: IReg, |
1204 | LocVT: MVT::i128, HTP: CCValAssign::BCvt); |
1205 | } |
1206 | } else { |
1207 | // This needs to go to memory, we're out of integer registers. |
1208 | VA = CCValAssign::getMem(ValNo: VA.getValNo(), ValVT: VA.getValVT(), Offset, |
1209 | LocVT: VA.getLocVT(), HTP: VA.getLocInfo()); |
1210 | } |
1211 | } |
1212 | } |
1213 | |
1214 | // Lower a call for the 64-bit ABI. |
1215 | SDValue |
1216 | SparcTargetLowering::LowerCall_64(TargetLowering::CallLoweringInfo &CLI, |
1217 | SmallVectorImpl<SDValue> &InVals) const { |
1218 | SelectionDAG &DAG = CLI.DAG; |
1219 | SDLoc DL = CLI.DL; |
1220 | SDValue Chain = CLI.Chain; |
1221 | auto PtrVT = getPointerTy(DL: DAG.getDataLayout()); |
1222 | MachineFunction &MF = DAG.getMachineFunction(); |
1223 | |
1224 | // Analyze operands of the call, assigning locations to each operand. |
1225 | SmallVector<CCValAssign, 16> ArgLocs; |
1226 | CCState CCInfo(CLI.CallConv, CLI.IsVarArg, DAG.getMachineFunction(), ArgLocs, |
1227 | *DAG.getContext()); |
1228 | CCInfo.AnalyzeCallOperands(Outs: CLI.Outs, Fn: CC_Sparc64); |
1229 | |
1230 | CLI.IsTailCall = CLI.IsTailCall && IsEligibleForTailCallOptimization( |
1231 | CCInfo, CLI, MF&: DAG.getMachineFunction()); |
1232 | |
1233 | // Get the size of the outgoing arguments stack space requirement. |
1234 | // The stack offset computed by CC_Sparc64 includes all arguments. |
1235 | // Called functions expect 6 argument words to exist in the stack frame, used |
1236 | // or not. |
1237 | unsigned StackReserved = 6 * 8u; |
1238 | unsigned ArgsSize = std::max<unsigned>(a: StackReserved, b: CCInfo.getStackSize()); |
1239 | |
1240 | // Keep stack frames 16-byte aligned. |
1241 | ArgsSize = alignTo(Value: ArgsSize, Align: 16); |
1242 | |
1243 | // Varargs calls require special treatment. |
1244 | if (CLI.IsVarArg) |
1245 | fixupVariableFloatArgs(ArgLocs, Outs: CLI.Outs); |
1246 | |
1247 | assert(!CLI.IsTailCall || ArgsSize == StackReserved); |
1248 | |
1249 | // Adjust the stack pointer to make room for the arguments. |
1250 | // FIXME: Use hasReservedCallFrame to avoid %sp adjustments around all calls |
1251 | // with more than 6 arguments. |
1252 | if (!CLI.IsTailCall) |
1253 | Chain = DAG.getCALLSEQ_START(Chain, InSize: ArgsSize, OutSize: 0, DL); |
1254 | |
1255 | // Collect the set of registers to pass to the function and their values. |
1256 | // This will be emitted as a sequence of CopyToReg nodes glued to the call |
1257 | // instruction. |
1258 | SmallVector<std::pair<Register, SDValue>, 8> RegsToPass; |
1259 | |
1260 | // Collect chains from all the memory opeations that copy arguments to the |
1261 | // stack. They must follow the stack pointer adjustment above and precede the |
1262 | // call instruction itself. |
1263 | SmallVector<SDValue, 8> MemOpChains; |
1264 | |
1265 | for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { |
1266 | const CCValAssign &VA = ArgLocs[i]; |
1267 | SDValue Arg = CLI.OutVals[i]; |
1268 | |
1269 | // Promote the value if needed. |
1270 | switch (VA.getLocInfo()) { |
1271 | default: |
1272 | llvm_unreachable("Unknown location info!" ); |
1273 | case CCValAssign::Full: |
1274 | break; |
1275 | case CCValAssign::SExt: |
1276 | Arg = DAG.getNode(Opcode: ISD::SIGN_EXTEND, DL, VT: VA.getLocVT(), Operand: Arg); |
1277 | break; |
1278 | case CCValAssign::ZExt: |
1279 | Arg = DAG.getNode(Opcode: ISD::ZERO_EXTEND, DL, VT: VA.getLocVT(), Operand: Arg); |
1280 | break; |
1281 | case CCValAssign::AExt: |
1282 | Arg = DAG.getNode(Opcode: ISD::ANY_EXTEND, DL, VT: VA.getLocVT(), Operand: Arg); |
1283 | break; |
1284 | case CCValAssign::BCvt: |
1285 | // fixupVariableFloatArgs() may create bitcasts from f128 to i128. But |
1286 | // SPARC does not support i128 natively. Lower it into two i64, see below. |
1287 | if (!VA.needsCustom() || VA.getValVT() != MVT::f128 |
1288 | || VA.getLocVT() != MVT::i128) |
1289 | Arg = DAG.getNode(Opcode: ISD::BITCAST, DL, VT: VA.getLocVT(), Operand: Arg); |
1290 | break; |
1291 | } |
1292 | |
1293 | if (VA.isRegLoc()) { |
1294 | if (VA.needsCustom() && VA.getValVT() == MVT::f128 |
1295 | && VA.getLocVT() == MVT::i128) { |
1296 | // Store and reload into the integer register reg and reg+1. |
1297 | unsigned Offset = 8 * (VA.getLocReg() - SP::I0); |
1298 | unsigned StackOffset = Offset + Subtarget->getStackPointerBias() + 128; |
1299 | SDValue StackPtr = DAG.getRegister(Reg: SP::O6, VT: PtrVT); |
1300 | SDValue HiPtrOff = DAG.getIntPtrConstant(Val: StackOffset, DL); |
1301 | HiPtrOff = DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrVT, N1: StackPtr, N2: HiPtrOff); |
1302 | SDValue LoPtrOff = DAG.getIntPtrConstant(Val: StackOffset + 8, DL); |
1303 | LoPtrOff = DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrVT, N1: StackPtr, N2: LoPtrOff); |
1304 | |
1305 | // Store to %sp+BIAS+128+Offset |
1306 | SDValue Store = |
1307 | DAG.getStore(Chain, dl: DL, Val: Arg, Ptr: HiPtrOff, PtrInfo: MachinePointerInfo()); |
1308 | // Load into Reg and Reg+1 |
1309 | SDValue Hi64 = |
1310 | DAG.getLoad(VT: MVT::i64, dl: DL, Chain: Store, Ptr: HiPtrOff, PtrInfo: MachinePointerInfo()); |
1311 | SDValue Lo64 = |
1312 | DAG.getLoad(VT: MVT::i64, dl: DL, Chain: Store, Ptr: LoPtrOff, PtrInfo: MachinePointerInfo()); |
1313 | |
1314 | Register HiReg = VA.getLocReg(); |
1315 | Register LoReg = VA.getLocReg() + 1; |
1316 | if (!CLI.IsTailCall) { |
1317 | HiReg = toCallerWindow(Reg: HiReg); |
1318 | LoReg = toCallerWindow(Reg: LoReg); |
1319 | } |
1320 | |
1321 | RegsToPass.push_back(Elt: std::make_pair(x&: HiReg, y&: Hi64)); |
1322 | RegsToPass.push_back(Elt: std::make_pair(x&: LoReg, y&: Lo64)); |
1323 | continue; |
1324 | } |
1325 | |
1326 | // The custom bit on an i32 return value indicates that it should be |
1327 | // passed in the high bits of the register. |
1328 | if (VA.getValVT() == MVT::i32 && VA.needsCustom()) { |
1329 | Arg = DAG.getNode(Opcode: ISD::SHL, DL, VT: MVT::i64, N1: Arg, |
1330 | N2: DAG.getConstant(Val: 32, DL, VT: MVT::i32)); |
1331 | |
1332 | // The next value may go in the low bits of the same register. |
1333 | // Handle both at once. |
1334 | if (i+1 < ArgLocs.size() && ArgLocs[i+1].isRegLoc() && |
1335 | ArgLocs[i+1].getLocReg() == VA.getLocReg()) { |
1336 | SDValue NV = DAG.getNode(Opcode: ISD::ZERO_EXTEND, DL, VT: MVT::i64, |
1337 | Operand: CLI.OutVals[i+1]); |
1338 | Arg = DAG.getNode(Opcode: ISD::OR, DL, VT: MVT::i64, N1: Arg, N2: NV); |
1339 | // Skip the next value, it's already done. |
1340 | ++i; |
1341 | } |
1342 | } |
1343 | |
1344 | Register Reg = VA.getLocReg(); |
1345 | if (!CLI.IsTailCall) |
1346 | Reg = toCallerWindow(Reg); |
1347 | RegsToPass.push_back(Elt: std::make_pair(x&: Reg, y&: Arg)); |
1348 | continue; |
1349 | } |
1350 | |
1351 | assert(VA.isMemLoc()); |
1352 | |
1353 | // Create a store off the stack pointer for this argument. |
1354 | SDValue StackPtr = DAG.getRegister(Reg: SP::O6, VT: PtrVT); |
1355 | // The argument area starts at %fp+BIAS+128 in the callee frame, |
1356 | // %sp+BIAS+128 in ours. |
1357 | SDValue PtrOff = DAG.getIntPtrConstant(Val: VA.getLocMemOffset() + |
1358 | Subtarget->getStackPointerBias() + |
1359 | 128, DL); |
1360 | PtrOff = DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrVT, N1: StackPtr, N2: PtrOff); |
1361 | MemOpChains.push_back( |
1362 | Elt: DAG.getStore(Chain, dl: DL, Val: Arg, Ptr: PtrOff, PtrInfo: MachinePointerInfo())); |
1363 | } |
1364 | |
1365 | // Emit all stores, make sure they occur before the call. |
1366 | if (!MemOpChains.empty()) |
1367 | Chain = DAG.getNode(Opcode: ISD::TokenFactor, DL, VT: MVT::Other, Ops: MemOpChains); |
1368 | |
1369 | // Build a sequence of CopyToReg nodes glued together with token chain and |
1370 | // glue operands which copy the outgoing args into registers. The InGlue is |
1371 | // necessary since all emitted instructions must be stuck together in order |
1372 | // to pass the live physical registers. |
1373 | SDValue InGlue; |
1374 | for (const auto &[Reg, N] : RegsToPass) { |
1375 | Chain = DAG.getCopyToReg(Chain, dl: DL, Reg, N, Glue: InGlue); |
1376 | InGlue = Chain.getValue(R: 1); |
1377 | } |
1378 | |
1379 | // If the callee is a GlobalAddress node (quite common, every direct call is) |
1380 | // turn it into a TargetGlobalAddress node so that legalize doesn't hack it. |
1381 | // Likewise ExternalSymbol -> TargetExternalSymbol. |
1382 | SDValue Callee = CLI.Callee; |
1383 | bool hasReturnsTwice = hasReturnsTwiceAttr(DAG, Callee, Call: CLI.CB); |
1384 | if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Val&: Callee)) |
1385 | Callee = DAG.getTargetGlobalAddress(GV: G->getGlobal(), DL, VT: PtrVT, offset: 0); |
1386 | else if (ExternalSymbolSDNode *E = dyn_cast<ExternalSymbolSDNode>(Val&: Callee)) |
1387 | Callee = DAG.getTargetExternalSymbol(Sym: E->getSymbol(), VT: PtrVT); |
1388 | |
1389 | // Build the operands for the call instruction itself. |
1390 | SmallVector<SDValue, 8> Ops; |
1391 | Ops.push_back(Elt: Chain); |
1392 | Ops.push_back(Elt: Callee); |
1393 | for (const auto &[Reg, N] : RegsToPass) |
1394 | Ops.push_back(Elt: DAG.getRegister(Reg, VT: N.getValueType())); |
1395 | |
1396 | // Add a register mask operand representing the call-preserved registers. |
1397 | const SparcRegisterInfo *TRI = Subtarget->getRegisterInfo(); |
1398 | const uint32_t *Mask = |
1399 | ((hasReturnsTwice) ? TRI->getRTCallPreservedMask(CC: CLI.CallConv) |
1400 | : TRI->getCallPreservedMask(MF: DAG.getMachineFunction(), |
1401 | CC: CLI.CallConv)); |
1402 | |
1403 | if (isAnyArgRegReserved(TRI, MF)) |
1404 | emitReservedArgRegCallError(MF); |
1405 | |
1406 | assert(Mask && "Missing call preserved mask for calling convention" ); |
1407 | Ops.push_back(Elt: DAG.getRegisterMask(RegMask: Mask)); |
1408 | |
1409 | // Make sure the CopyToReg nodes are glued to the call instruction which |
1410 | // consumes the registers. |
1411 | if (InGlue.getNode()) |
1412 | Ops.push_back(Elt: InGlue); |
1413 | |
1414 | // Now the call itself. |
1415 | if (CLI.IsTailCall) { |
1416 | DAG.getMachineFunction().getFrameInfo().setHasTailCall(); |
1417 | return DAG.getNode(Opcode: SPISD::TAIL_CALL, DL, VT: MVT::Other, Ops); |
1418 | } |
1419 | SDVTList NodeTys = DAG.getVTList(VT1: MVT::Other, VT2: MVT::Glue); |
1420 | Chain = DAG.getNode(Opcode: SPISD::CALL, DL, VTList: NodeTys, Ops); |
1421 | InGlue = Chain.getValue(R: 1); |
1422 | |
1423 | // Revert the stack pointer immediately after the call. |
1424 | Chain = DAG.getCALLSEQ_END(Chain, Size1: ArgsSize, Size2: 0, Glue: InGlue, DL); |
1425 | InGlue = Chain.getValue(R: 1); |
1426 | |
1427 | // Now extract the return values. This is more or less the same as |
1428 | // LowerFormalArguments_64. |
1429 | |
1430 | // Assign locations to each value returned by this call. |
1431 | SmallVector<CCValAssign, 16> RVLocs; |
1432 | CCState RVInfo(CLI.CallConv, CLI.IsVarArg, DAG.getMachineFunction(), RVLocs, |
1433 | *DAG.getContext()); |
1434 | |
1435 | // Set inreg flag manually for codegen generated library calls that |
1436 | // return float. |
1437 | if (CLI.Ins.size() == 1 && CLI.Ins[0].VT == MVT::f32 && !CLI.CB) |
1438 | CLI.Ins[0].Flags.setInReg(); |
1439 | |
1440 | RVInfo.AnalyzeCallResult(Ins: CLI.Ins, Fn: RetCC_Sparc64); |
1441 | |
1442 | // Copy all of the result registers out of their specified physreg. |
1443 | for (unsigned i = 0; i != RVLocs.size(); ++i) { |
1444 | CCValAssign &VA = RVLocs[i]; |
1445 | assert(VA.isRegLoc() && "Can only return in registers!" ); |
1446 | unsigned Reg = toCallerWindow(Reg: VA.getLocReg()); |
1447 | |
1448 | // When returning 'inreg {i32, i32 }', two consecutive i32 arguments can |
1449 | // reside in the same register in the high and low bits. Reuse the |
1450 | // CopyFromReg previous node to avoid duplicate copies. |
1451 | SDValue RV; |
1452 | if (RegisterSDNode *SrcReg = dyn_cast<RegisterSDNode>(Val: Chain.getOperand(i: 1))) |
1453 | if (SrcReg->getReg() == Reg && Chain->getOpcode() == ISD::CopyFromReg) |
1454 | RV = Chain.getValue(R: 0); |
1455 | |
1456 | // But usually we'll create a new CopyFromReg for a different register. |
1457 | if (!RV.getNode()) { |
1458 | RV = DAG.getCopyFromReg(Chain, dl: DL, Reg, VT: RVLocs[i].getLocVT(), Glue: InGlue); |
1459 | Chain = RV.getValue(R: 1); |
1460 | InGlue = Chain.getValue(R: 2); |
1461 | } |
1462 | |
1463 | // Get the high bits for i32 struct elements. |
1464 | if (VA.getValVT() == MVT::i32 && VA.needsCustom()) |
1465 | RV = DAG.getNode(Opcode: ISD::SRL, DL, VT: VA.getLocVT(), N1: RV, |
1466 | N2: DAG.getConstant(Val: 32, DL, VT: MVT::i32)); |
1467 | |
1468 | // The callee promoted the return value, so insert an Assert?ext SDNode so |
1469 | // we won't promote the value again in this function. |
1470 | switch (VA.getLocInfo()) { |
1471 | case CCValAssign::SExt: |
1472 | RV = DAG.getNode(Opcode: ISD::AssertSext, DL, VT: VA.getLocVT(), N1: RV, |
1473 | N2: DAG.getValueType(VA.getValVT())); |
1474 | break; |
1475 | case CCValAssign::ZExt: |
1476 | RV = DAG.getNode(Opcode: ISD::AssertZext, DL, VT: VA.getLocVT(), N1: RV, |
1477 | N2: DAG.getValueType(VA.getValVT())); |
1478 | break; |
1479 | default: |
1480 | break; |
1481 | } |
1482 | |
1483 | // Truncate the register down to the return value type. |
1484 | if (VA.isExtInLoc()) |
1485 | RV = DAG.getNode(Opcode: ISD::TRUNCATE, DL, VT: VA.getValVT(), Operand: RV); |
1486 | |
1487 | InVals.push_back(Elt: RV); |
1488 | } |
1489 | |
1490 | return Chain; |
1491 | } |
1492 | |
1493 | //===----------------------------------------------------------------------===// |
1494 | // TargetLowering Implementation |
1495 | //===----------------------------------------------------------------------===// |
1496 | |
1497 | TargetLowering::AtomicExpansionKind SparcTargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const { |
1498 | if (AI->getOperation() == AtomicRMWInst::Xchg && |
1499 | AI->getType()->getPrimitiveSizeInBits() == 32) |
1500 | return AtomicExpansionKind::None; // Uses xchg instruction |
1501 | |
1502 | return AtomicExpansionKind::CmpXChg; |
1503 | } |
1504 | |
1505 | /// intCondCCodeToRcond - Convert a DAG integer condition code to a SPARC |
1506 | /// rcond condition. |
1507 | static SPCC::CondCodes intCondCCodeToRcond(ISD::CondCode CC) { |
1508 | switch (CC) { |
1509 | default: |
1510 | llvm_unreachable("Unknown/unsigned integer condition code!" ); |
1511 | case ISD::SETEQ: |
1512 | return SPCC::REG_Z; |
1513 | case ISD::SETNE: |
1514 | return SPCC::REG_NZ; |
1515 | case ISD::SETLT: |
1516 | return SPCC::REG_LZ; |
1517 | case ISD::SETGT: |
1518 | return SPCC::REG_GZ; |
1519 | case ISD::SETLE: |
1520 | return SPCC::REG_LEZ; |
1521 | case ISD::SETGE: |
1522 | return SPCC::REG_GEZ; |
1523 | } |
1524 | } |
1525 | |
1526 | /// IntCondCCodeToICC - Convert a DAG integer condition code to a SPARC ICC |
1527 | /// condition. |
1528 | static SPCC::CondCodes IntCondCCodeToICC(ISD::CondCode CC) { |
1529 | switch (CC) { |
1530 | default: llvm_unreachable("Unknown integer condition code!" ); |
1531 | case ISD::SETEQ: return SPCC::ICC_E; |
1532 | case ISD::SETNE: return SPCC::ICC_NE; |
1533 | case ISD::SETLT: return SPCC::ICC_L; |
1534 | case ISD::SETGT: return SPCC::ICC_G; |
1535 | case ISD::SETLE: return SPCC::ICC_LE; |
1536 | case ISD::SETGE: return SPCC::ICC_GE; |
1537 | case ISD::SETULT: return SPCC::ICC_CS; |
1538 | case ISD::SETULE: return SPCC::ICC_LEU; |
1539 | case ISD::SETUGT: return SPCC::ICC_GU; |
1540 | case ISD::SETUGE: return SPCC::ICC_CC; |
1541 | } |
1542 | } |
1543 | |
1544 | /// FPCondCCodeToFCC - Convert a DAG floatingp oint condition code to a SPARC |
1545 | /// FCC condition. |
1546 | static SPCC::CondCodes FPCondCCodeToFCC(ISD::CondCode CC) { |
1547 | switch (CC) { |
1548 | default: llvm_unreachable("Unknown fp condition code!" ); |
1549 | case ISD::SETEQ: |
1550 | case ISD::SETOEQ: return SPCC::FCC_E; |
1551 | case ISD::SETNE: |
1552 | case ISD::SETUNE: return SPCC::FCC_NE; |
1553 | case ISD::SETLT: |
1554 | case ISD::SETOLT: return SPCC::FCC_L; |
1555 | case ISD::SETGT: |
1556 | case ISD::SETOGT: return SPCC::FCC_G; |
1557 | case ISD::SETLE: |
1558 | case ISD::SETOLE: return SPCC::FCC_LE; |
1559 | case ISD::SETGE: |
1560 | case ISD::SETOGE: return SPCC::FCC_GE; |
1561 | case ISD::SETULT: return SPCC::FCC_UL; |
1562 | case ISD::SETULE: return SPCC::FCC_ULE; |
1563 | case ISD::SETUGT: return SPCC::FCC_UG; |
1564 | case ISD::SETUGE: return SPCC::FCC_UGE; |
1565 | case ISD::SETUO: return SPCC::FCC_U; |
1566 | case ISD::SETO: return SPCC::FCC_O; |
1567 | case ISD::SETONE: return SPCC::FCC_LG; |
1568 | case ISD::SETUEQ: return SPCC::FCC_UE; |
1569 | } |
1570 | } |
1571 | |
1572 | SparcTargetLowering::SparcTargetLowering(const TargetMachine &TM, |
1573 | const SparcSubtarget &STI) |
1574 | : TargetLowering(TM), Subtarget(&STI) { |
1575 | MVT PtrVT = MVT::getIntegerVT(BitWidth: TM.getPointerSizeInBits(AS: 0)); |
1576 | |
1577 | // Instructions which use registers as conditionals examine all the |
1578 | // bits (as does the pseudo SELECT_CC expansion). I don't think it |
1579 | // matters much whether it's ZeroOrOneBooleanContent, or |
1580 | // ZeroOrNegativeOneBooleanContent, so, arbitrarily choose the |
1581 | // former. |
1582 | setBooleanContents(ZeroOrOneBooleanContent); |
1583 | setBooleanVectorContents(ZeroOrOneBooleanContent); |
1584 | |
1585 | // Set up the register classes. |
1586 | addRegisterClass(VT: MVT::i32, RC: &SP::IntRegsRegClass); |
1587 | if (!Subtarget->useSoftFloat()) { |
1588 | addRegisterClass(VT: MVT::f32, RC: &SP::FPRegsRegClass); |
1589 | addRegisterClass(VT: MVT::f64, RC: &SP::DFPRegsRegClass); |
1590 | addRegisterClass(VT: MVT::f128, RC: &SP::QFPRegsRegClass); |
1591 | } |
1592 | if (Subtarget->is64Bit()) { |
1593 | addRegisterClass(VT: MVT::i64, RC: &SP::I64RegsRegClass); |
1594 | } else { |
1595 | // On 32bit sparc, we define a double-register 32bit register |
1596 | // class, as well. This is modeled in LLVM as a 2-vector of i32. |
1597 | addRegisterClass(VT: MVT::v2i32, RC: &SP::IntPairRegClass); |
1598 | |
1599 | // ...but almost all operations must be expanded, so set that as |
1600 | // the default. |
1601 | for (unsigned Op = 0; Op < ISD::BUILTIN_OP_END; ++Op) { |
1602 | setOperationAction(Op, VT: MVT::v2i32, Action: Expand); |
1603 | } |
1604 | // Truncating/extending stores/loads are also not supported. |
1605 | for (MVT VT : MVT::integer_fixedlen_vector_valuetypes()) { |
1606 | setLoadExtAction(ExtType: ISD::SEXTLOAD, ValVT: VT, MemVT: MVT::v2i32, Action: Expand); |
1607 | setLoadExtAction(ExtType: ISD::ZEXTLOAD, ValVT: VT, MemVT: MVT::v2i32, Action: Expand); |
1608 | setLoadExtAction(ExtType: ISD::EXTLOAD, ValVT: VT, MemVT: MVT::v2i32, Action: Expand); |
1609 | |
1610 | setLoadExtAction(ExtType: ISD::SEXTLOAD, ValVT: MVT::v2i32, MemVT: VT, Action: Expand); |
1611 | setLoadExtAction(ExtType: ISD::ZEXTLOAD, ValVT: MVT::v2i32, MemVT: VT, Action: Expand); |
1612 | setLoadExtAction(ExtType: ISD::EXTLOAD, ValVT: MVT::v2i32, MemVT: VT, Action: Expand); |
1613 | |
1614 | setTruncStoreAction(ValVT: VT, MemVT: MVT::v2i32, Action: Expand); |
1615 | setTruncStoreAction(ValVT: MVT::v2i32, MemVT: VT, Action: Expand); |
1616 | } |
1617 | // However, load and store *are* legal. |
1618 | setOperationAction(Op: ISD::LOAD, VT: MVT::v2i32, Action: Legal); |
1619 | setOperationAction(Op: ISD::STORE, VT: MVT::v2i32, Action: Legal); |
1620 | setOperationAction(Op: ISD::EXTRACT_VECTOR_ELT, VT: MVT::v2i32, Action: Legal); |
1621 | setOperationAction(Op: ISD::BUILD_VECTOR, VT: MVT::v2i32, Action: Legal); |
1622 | |
1623 | // And we need to promote i64 loads/stores into vector load/store |
1624 | setOperationAction(Op: ISD::LOAD, VT: MVT::i64, Action: Custom); |
1625 | setOperationAction(Op: ISD::STORE, VT: MVT::i64, Action: Custom); |
1626 | |
1627 | // Sadly, this doesn't work: |
1628 | // AddPromotedToType(ISD::LOAD, MVT::i64, MVT::v2i32); |
1629 | // AddPromotedToType(ISD::STORE, MVT::i64, MVT::v2i32); |
1630 | } |
1631 | |
1632 | // Turn FP extload into load/fpextend |
1633 | for (MVT VT : MVT::fp_valuetypes()) { |
1634 | setLoadExtAction(ExtType: ISD::EXTLOAD, ValVT: VT, MemVT: MVT::f16, Action: Expand); |
1635 | setLoadExtAction(ExtType: ISD::EXTLOAD, ValVT: VT, MemVT: MVT::f32, Action: Expand); |
1636 | setLoadExtAction(ExtType: ISD::EXTLOAD, ValVT: VT, MemVT: MVT::f64, Action: Expand); |
1637 | } |
1638 | |
1639 | // Sparc doesn't have i1 sign extending load |
1640 | for (MVT VT : MVT::integer_valuetypes()) |
1641 | setLoadExtAction(ExtType: ISD::SEXTLOAD, ValVT: VT, MemVT: MVT::i1, Action: Promote); |
1642 | |
1643 | // Turn FP truncstore into trunc + store. |
1644 | setTruncStoreAction(ValVT: MVT::f32, MemVT: MVT::f16, Action: Expand); |
1645 | setTruncStoreAction(ValVT: MVT::f64, MemVT: MVT::f16, Action: Expand); |
1646 | setTruncStoreAction(ValVT: MVT::f64, MemVT: MVT::f32, Action: Expand); |
1647 | setTruncStoreAction(ValVT: MVT::f128, MemVT: MVT::f16, Action: Expand); |
1648 | setTruncStoreAction(ValVT: MVT::f128, MemVT: MVT::f32, Action: Expand); |
1649 | setTruncStoreAction(ValVT: MVT::f128, MemVT: MVT::f64, Action: Expand); |
1650 | |
1651 | // Custom legalize GlobalAddress nodes into LO/HI parts. |
1652 | setOperationAction(Op: ISD::GlobalAddress, VT: PtrVT, Action: Custom); |
1653 | setOperationAction(Op: ISD::GlobalTLSAddress, VT: PtrVT, Action: Custom); |
1654 | setOperationAction(Op: ISD::ConstantPool, VT: PtrVT, Action: Custom); |
1655 | setOperationAction(Op: ISD::BlockAddress, VT: PtrVT, Action: Custom); |
1656 | |
1657 | // Sparc doesn't have sext_inreg, replace them with shl/sra |
1658 | setOperationAction(Op: ISD::SIGN_EXTEND_INREG, VT: MVT::i16, Action: Expand); |
1659 | setOperationAction(Op: ISD::SIGN_EXTEND_INREG, VT: MVT::i8 , Action: Expand); |
1660 | setOperationAction(Op: ISD::SIGN_EXTEND_INREG, VT: MVT::i1 , Action: Expand); |
1661 | |
1662 | // Sparc has no REM or DIVREM operations. |
1663 | setOperationAction(Op: ISD::UREM, VT: MVT::i32, Action: Expand); |
1664 | setOperationAction(Op: ISD::SREM, VT: MVT::i32, Action: Expand); |
1665 | setOperationAction(Op: ISD::SDIVREM, VT: MVT::i32, Action: Expand); |
1666 | setOperationAction(Op: ISD::UDIVREM, VT: MVT::i32, Action: Expand); |
1667 | |
1668 | // ... nor does SparcV9. |
1669 | if (Subtarget->is64Bit()) { |
1670 | setOperationAction(Op: ISD::UREM, VT: MVT::i64, Action: Expand); |
1671 | setOperationAction(Op: ISD::SREM, VT: MVT::i64, Action: Expand); |
1672 | setOperationAction(Op: ISD::SDIVREM, VT: MVT::i64, Action: Expand); |
1673 | setOperationAction(Op: ISD::UDIVREM, VT: MVT::i64, Action: Expand); |
1674 | } |
1675 | |
1676 | // Custom expand fp<->sint |
1677 | setOperationAction(Op: ISD::FP_TO_SINT, VT: MVT::i32, Action: Custom); |
1678 | setOperationAction(Op: ISD::SINT_TO_FP, VT: MVT::i32, Action: Custom); |
1679 | setOperationAction(Op: ISD::FP_TO_SINT, VT: MVT::i64, Action: Custom); |
1680 | setOperationAction(Op: ISD::SINT_TO_FP, VT: MVT::i64, Action: Custom); |
1681 | |
1682 | // Custom Expand fp<->uint |
1683 | setOperationAction(Op: ISD::FP_TO_UINT, VT: MVT::i32, Action: Custom); |
1684 | setOperationAction(Op: ISD::UINT_TO_FP, VT: MVT::i32, Action: Custom); |
1685 | setOperationAction(Op: ISD::FP_TO_UINT, VT: MVT::i64, Action: Custom); |
1686 | setOperationAction(Op: ISD::UINT_TO_FP, VT: MVT::i64, Action: Custom); |
1687 | |
1688 | // Lower f16 conversion operations into library calls |
1689 | setOperationAction(Op: ISD::FP16_TO_FP, VT: MVT::f32, Action: Expand); |
1690 | setOperationAction(Op: ISD::FP_TO_FP16, VT: MVT::f32, Action: Expand); |
1691 | setOperationAction(Op: ISD::FP16_TO_FP, VT: MVT::f64, Action: Expand); |
1692 | setOperationAction(Op: ISD::FP_TO_FP16, VT: MVT::f64, Action: Expand); |
1693 | setOperationAction(Op: ISD::FP16_TO_FP, VT: MVT::f128, Action: Expand); |
1694 | setOperationAction(Op: ISD::FP_TO_FP16, VT: MVT::f128, Action: Expand); |
1695 | |
1696 | setOperationAction(Op: ISD::BITCAST, VT: MVT::f32, |
1697 | Action: Subtarget->isVIS3() ? Legal : Expand); |
1698 | setOperationAction(Op: ISD::BITCAST, VT: MVT::i32, |
1699 | Action: Subtarget->isVIS3() ? Legal : Expand); |
1700 | |
1701 | // Sparc has no select or setcc: expand to SELECT_CC. |
1702 | setOperationAction(Op: ISD::SELECT, VT: MVT::i32, Action: Expand); |
1703 | setOperationAction(Op: ISD::SELECT, VT: MVT::f32, Action: Expand); |
1704 | setOperationAction(Op: ISD::SELECT, VT: MVT::f64, Action: Expand); |
1705 | setOperationAction(Op: ISD::SELECT, VT: MVT::f128, Action: Expand); |
1706 | |
1707 | setOperationAction(Op: ISD::SETCC, VT: MVT::i32, Action: Expand); |
1708 | setOperationAction(Op: ISD::SETCC, VT: MVT::f32, Action: Expand); |
1709 | setOperationAction(Op: ISD::SETCC, VT: MVT::f64, Action: Expand); |
1710 | setOperationAction(Op: ISD::SETCC, VT: MVT::f128, Action: Expand); |
1711 | |
1712 | // Sparc doesn't have BRCOND either, it has BR_CC. |
1713 | setOperationAction(Op: ISD::BRCOND, VT: MVT::Other, Action: Expand); |
1714 | setOperationAction(Op: ISD::BRIND, VT: MVT::Other, Action: Expand); |
1715 | setOperationAction(Op: ISD::BR_JT, VT: MVT::Other, Action: Expand); |
1716 | setOperationAction(Op: ISD::BR_CC, VT: MVT::i32, Action: Custom); |
1717 | setOperationAction(Op: ISD::BR_CC, VT: MVT::f32, Action: Custom); |
1718 | setOperationAction(Op: ISD::BR_CC, VT: MVT::f64, Action: Custom); |
1719 | setOperationAction(Op: ISD::BR_CC, VT: MVT::f128, Action: Custom); |
1720 | |
1721 | setOperationAction(Op: ISD::SELECT_CC, VT: MVT::i32, Action: Custom); |
1722 | setOperationAction(Op: ISD::SELECT_CC, VT: MVT::f32, Action: Custom); |
1723 | setOperationAction(Op: ISD::SELECT_CC, VT: MVT::f64, Action: Custom); |
1724 | setOperationAction(Op: ISD::SELECT_CC, VT: MVT::f128, Action: Custom); |
1725 | |
1726 | setOperationAction(Op: ISD::ADDC, VT: MVT::i32, Action: Legal); |
1727 | setOperationAction(Op: ISD::ADDE, VT: MVT::i32, Action: Legal); |
1728 | setOperationAction(Op: ISD::SUBC, VT: MVT::i32, Action: Legal); |
1729 | setOperationAction(Op: ISD::SUBE, VT: MVT::i32, Action: Legal); |
1730 | |
1731 | if (Subtarget->isVIS3()) { |
1732 | setOperationAction(Op: ISD::ADDC, VT: MVT::i64, Action: Legal); |
1733 | setOperationAction(Op: ISD::ADDE, VT: MVT::i64, Action: Legal); |
1734 | } |
1735 | |
1736 | if (Subtarget->is64Bit()) { |
1737 | setOperationAction(Op: ISD::BITCAST, VT: MVT::f64, |
1738 | Action: Subtarget->isVIS3() ? Legal : Expand); |
1739 | setOperationAction(Op: ISD::BITCAST, VT: MVT::i64, |
1740 | Action: Subtarget->isVIS3() ? Legal : Expand); |
1741 | setOperationAction(Op: ISD::SELECT, VT: MVT::i64, Action: Expand); |
1742 | setOperationAction(Op: ISD::SETCC, VT: MVT::i64, Action: Expand); |
1743 | setOperationAction(Op: ISD::BR_CC, VT: MVT::i64, Action: Custom); |
1744 | setOperationAction(Op: ISD::SELECT_CC, VT: MVT::i64, Action: Custom); |
1745 | |
1746 | setOperationAction(Op: ISD::CTPOP, VT: MVT::i64, |
1747 | Action: Subtarget->usePopc() ? Legal : Expand); |
1748 | setOperationAction(Op: ISD::BSWAP, VT: MVT::i64, Action: Expand); |
1749 | setOperationAction(Op: ISD::ROTL , VT: MVT::i64, Action: Expand); |
1750 | setOperationAction(Op: ISD::ROTR , VT: MVT::i64, Action: Expand); |
1751 | setOperationAction(Op: ISD::DYNAMIC_STACKALLOC, VT: MVT::i64, Action: Custom); |
1752 | } |
1753 | |
1754 | // ATOMICs. |
1755 | // Atomics are supported on SparcV9. 32-bit atomics are also |
1756 | // supported by some Leon SparcV8 variants. Otherwise, atomics |
1757 | // are unsupported. |
1758 | if (Subtarget->isV9()) { |
1759 | // TODO: we _ought_ to be able to support 64-bit atomics on 32-bit sparcv9, |
1760 | // but it hasn't been implemented in the backend yet. |
1761 | if (Subtarget->is64Bit()) |
1762 | setMaxAtomicSizeInBitsSupported(64); |
1763 | else |
1764 | setMaxAtomicSizeInBitsSupported(32); |
1765 | } else if (Subtarget->hasLeonCasa()) |
1766 | setMaxAtomicSizeInBitsSupported(32); |
1767 | else |
1768 | setMaxAtomicSizeInBitsSupported(0); |
1769 | |
1770 | setMinCmpXchgSizeInBits(32); |
1771 | |
1772 | setOperationAction(Op: ISD::ATOMIC_SWAP, VT: MVT::i32, Action: Legal); |
1773 | |
1774 | setOperationAction(Op: ISD::ATOMIC_FENCE, VT: MVT::Other, Action: Legal); |
1775 | |
1776 | // Custom Lower Atomic LOAD/STORE |
1777 | setOperationAction(Op: ISD::ATOMIC_LOAD, VT: MVT::i32, Action: Custom); |
1778 | setOperationAction(Op: ISD::ATOMIC_STORE, VT: MVT::i32, Action: Custom); |
1779 | |
1780 | if (Subtarget->is64Bit()) { |
1781 | setOperationAction(Op: ISD::ATOMIC_CMP_SWAP, VT: MVT::i64, Action: Legal); |
1782 | setOperationAction(Op: ISD::ATOMIC_SWAP, VT: MVT::i64, Action: Legal); |
1783 | setOperationAction(Op: ISD::ATOMIC_LOAD, VT: MVT::i64, Action: Custom); |
1784 | setOperationAction(Op: ISD::ATOMIC_STORE, VT: MVT::i64, Action: Custom); |
1785 | } |
1786 | |
1787 | if (!Subtarget->isV9()) { |
1788 | // SparcV8 does not have FNEGD and FABSD. |
1789 | setOperationAction(Op: ISD::FNEG, VT: MVT::f64, Action: Custom); |
1790 | setOperationAction(Op: ISD::FABS, VT: MVT::f64, Action: Custom); |
1791 | } |
1792 | |
1793 | setOperationAction(Op: ISD::FSIN , VT: MVT::f128, Action: Expand); |
1794 | setOperationAction(Op: ISD::FCOS , VT: MVT::f128, Action: Expand); |
1795 | setOperationAction(Op: ISD::FSINCOS, VT: MVT::f128, Action: Expand); |
1796 | setOperationAction(Op: ISD::FREM , VT: MVT::f128, Action: Expand); |
1797 | setOperationAction(Op: ISD::FMA , VT: MVT::f128, Action: Expand); |
1798 | setOperationAction(Op: ISD::FSIN , VT: MVT::f64, Action: Expand); |
1799 | setOperationAction(Op: ISD::FCOS , VT: MVT::f64, Action: Expand); |
1800 | setOperationAction(Op: ISD::FSINCOS, VT: MVT::f64, Action: Expand); |
1801 | setOperationAction(Op: ISD::FREM , VT: MVT::f64, Action: Expand); |
1802 | setOperationAction(Op: ISD::FMA , VT: MVT::f64, Action: Expand); |
1803 | setOperationAction(Op: ISD::FSIN , VT: MVT::f32, Action: Expand); |
1804 | setOperationAction(Op: ISD::FCOS , VT: MVT::f32, Action: Expand); |
1805 | setOperationAction(Op: ISD::FSINCOS, VT: MVT::f32, Action: Expand); |
1806 | setOperationAction(Op: ISD::FREM , VT: MVT::f32, Action: Expand); |
1807 | setOperationAction(Op: ISD::FMA, VT: MVT::f32, Action: Expand); |
1808 | setOperationAction(Op: ISD::ROTL , VT: MVT::i32, Action: Expand); |
1809 | setOperationAction(Op: ISD::ROTR , VT: MVT::i32, Action: Expand); |
1810 | setOperationAction(Op: ISD::BSWAP, VT: MVT::i32, Action: Expand); |
1811 | setOperationAction(Op: ISD::FCOPYSIGN, VT: MVT::f128, Action: Expand); |
1812 | setOperationAction(Op: ISD::FCOPYSIGN, VT: MVT::f64, Action: Expand); |
1813 | setOperationAction(Op: ISD::FCOPYSIGN, VT: MVT::f32, Action: Expand); |
1814 | setOperationAction(Op: ISD::FPOW , VT: MVT::f128, Action: Expand); |
1815 | setOperationAction(Op: ISD::FPOW , VT: MVT::f64, Action: Expand); |
1816 | setOperationAction(Op: ISD::FPOW , VT: MVT::f32, Action: Expand); |
1817 | |
1818 | setOperationAction(Op: ISD::SHL_PARTS, VT: MVT::i32, Action: Expand); |
1819 | setOperationAction(Op: ISD::SRA_PARTS, VT: MVT::i32, Action: Expand); |
1820 | setOperationAction(Op: ISD::SRL_PARTS, VT: MVT::i32, Action: Expand); |
1821 | |
1822 | // Expands to [SU]MUL_LOHI. |
1823 | setOperationAction(Op: ISD::MULHU, VT: MVT::i32, Action: Expand); |
1824 | setOperationAction(Op: ISD::MULHS, VT: MVT::i32, Action: Expand); |
1825 | setOperationAction(Op: ISD::MUL, VT: MVT::i32, Action: Expand); |
1826 | |
1827 | if (Subtarget->useSoftMulDiv()) { |
1828 | // .umul works for both signed and unsigned |
1829 | setOperationAction(Op: ISD::SMUL_LOHI, VT: MVT::i32, Action: Expand); |
1830 | setOperationAction(Op: ISD::UMUL_LOHI, VT: MVT::i32, Action: Expand); |
1831 | setLibcallImpl(Call: RTLIB::MUL_I32, Impl: RTLIB::sparc_umul); |
1832 | |
1833 | setOperationAction(Op: ISD::SDIV, VT: MVT::i32, Action: Expand); |
1834 | setLibcallImpl(Call: RTLIB::SDIV_I32, Impl: RTLIB::sparc_div); |
1835 | |
1836 | setOperationAction(Op: ISD::UDIV, VT: MVT::i32, Action: Expand); |
1837 | setLibcallImpl(Call: RTLIB::UDIV_I32, Impl: RTLIB::sparc_udiv); |
1838 | |
1839 | setLibcallImpl(Call: RTLIB::SREM_I32, Impl: RTLIB::sparc_rem); |
1840 | setLibcallImpl(Call: RTLIB::UREM_I32, Impl: RTLIB::sparc_urem); |
1841 | } |
1842 | |
1843 | if (Subtarget->is64Bit()) { |
1844 | setOperationAction(Op: ISD::UMUL_LOHI, VT: MVT::i64, Action: Expand); |
1845 | setOperationAction(Op: ISD::SMUL_LOHI, VT: MVT::i64, Action: Expand); |
1846 | setOperationAction(Op: ISD::MULHU, VT: MVT::i64, |
1847 | Action: Subtarget->isVIS3() ? Legal : Expand); |
1848 | setOperationAction(Op: ISD::MULHS, VT: MVT::i64, |
1849 | Action: Subtarget->isVIS3() ? Legal : Expand); |
1850 | |
1851 | setOperationAction(Op: ISD::SHL_PARTS, VT: MVT::i64, Action: Expand); |
1852 | setOperationAction(Op: ISD::SRA_PARTS, VT: MVT::i64, Action: Expand); |
1853 | setOperationAction(Op: ISD::SRL_PARTS, VT: MVT::i64, Action: Expand); |
1854 | } |
1855 | |
1856 | // VASTART needs to be custom lowered to use the VarArgsFrameIndex. |
1857 | setOperationAction(Op: ISD::VASTART , VT: MVT::Other, Action: Custom); |
1858 | // VAARG needs to be lowered to not do unaligned accesses for doubles. |
1859 | setOperationAction(Op: ISD::VAARG , VT: MVT::Other, Action: Custom); |
1860 | |
1861 | setOperationAction(Op: ISD::TRAP , VT: MVT::Other, Action: Legal); |
1862 | setOperationAction(Op: ISD::DEBUGTRAP , VT: MVT::Other, Action: Legal); |
1863 | |
1864 | // Use the default implementation. |
1865 | setOperationAction(Op: ISD::VACOPY , VT: MVT::Other, Action: Expand); |
1866 | setOperationAction(Op: ISD::VAEND , VT: MVT::Other, Action: Expand); |
1867 | setOperationAction(Op: ISD::STACKSAVE , VT: MVT::Other, Action: Expand); |
1868 | setOperationAction(Op: ISD::STACKRESTORE , VT: MVT::Other, Action: Expand); |
1869 | setOperationAction(Op: ISD::DYNAMIC_STACKALLOC, VT: MVT::i32 , Action: Custom); |
1870 | |
1871 | setStackPointerRegisterToSaveRestore(SP::O6); |
1872 | |
1873 | setOperationAction(Op: ISD::CTPOP, VT: MVT::i32, |
1874 | Action: Subtarget->usePopc() ? Legal : Expand); |
1875 | |
1876 | if (Subtarget->isV9() && Subtarget->hasHardQuad()) { |
1877 | setOperationAction(Op: ISD::LOAD, VT: MVT::f128, Action: Legal); |
1878 | setOperationAction(Op: ISD::STORE, VT: MVT::f128, Action: Legal); |
1879 | } else { |
1880 | setOperationAction(Op: ISD::LOAD, VT: MVT::f128, Action: Custom); |
1881 | setOperationAction(Op: ISD::STORE, VT: MVT::f128, Action: Custom); |
1882 | } |
1883 | |
1884 | if (Subtarget->hasHardQuad()) { |
1885 | setOperationAction(Op: ISD::FADD, VT: MVT::f128, Action: Legal); |
1886 | setOperationAction(Op: ISD::FSUB, VT: MVT::f128, Action: Legal); |
1887 | setOperationAction(Op: ISD::FMUL, VT: MVT::f128, Action: Legal); |
1888 | setOperationAction(Op: ISD::FDIV, VT: MVT::f128, Action: Legal); |
1889 | setOperationAction(Op: ISD::FSQRT, VT: MVT::f128, Action: Legal); |
1890 | setOperationAction(Op: ISD::FP_EXTEND, VT: MVT::f128, Action: Legal); |
1891 | setOperationAction(Op: ISD::FP_ROUND, VT: MVT::f64, Action: Legal); |
1892 | if (Subtarget->isV9()) { |
1893 | setOperationAction(Op: ISD::FNEG, VT: MVT::f128, Action: Legal); |
1894 | setOperationAction(Op: ISD::FABS, VT: MVT::f128, Action: Legal); |
1895 | } else { |
1896 | setOperationAction(Op: ISD::FNEG, VT: MVT::f128, Action: Custom); |
1897 | setOperationAction(Op: ISD::FABS, VT: MVT::f128, Action: Custom); |
1898 | } |
1899 | |
1900 | if (!Subtarget->is64Bit()) { |
1901 | setLibcallImpl(Call: RTLIB::FPTOSINT_F128_I64, Impl: RTLIB::_Q_qtoll); |
1902 | setLibcallImpl(Call: RTLIB::FPTOUINT_F128_I64, Impl: RTLIB::_Q_qtoull); |
1903 | setLibcallImpl(Call: RTLIB::SINTTOFP_I64_F128, Impl: RTLIB::_Q_lltoq); |
1904 | setLibcallImpl(Call: RTLIB::UINTTOFP_I64_F128, Impl: RTLIB::_Q_ulltoq); |
1905 | } |
1906 | |
1907 | } else { |
1908 | // Custom legalize f128 operations. |
1909 | |
1910 | setOperationAction(Op: ISD::FADD, VT: MVT::f128, Action: Custom); |
1911 | setOperationAction(Op: ISD::FSUB, VT: MVT::f128, Action: Custom); |
1912 | setOperationAction(Op: ISD::FMUL, VT: MVT::f128, Action: Custom); |
1913 | setOperationAction(Op: ISD::FDIV, VT: MVT::f128, Action: Custom); |
1914 | setOperationAction(Op: ISD::FSQRT, VT: MVT::f128, Action: Custom); |
1915 | setOperationAction(Op: ISD::FNEG, VT: MVT::f128, Action: Custom); |
1916 | setOperationAction(Op: ISD::FABS, VT: MVT::f128, Action: Custom); |
1917 | |
1918 | setOperationAction(Op: ISD::FP_EXTEND, VT: MVT::f128, Action: Custom); |
1919 | setOperationAction(Op: ISD::FP_ROUND, VT: MVT::f64, Action: Custom); |
1920 | setOperationAction(Op: ISD::FP_ROUND, VT: MVT::f32, Action: Custom); |
1921 | |
1922 | // Setup Runtime library names. |
1923 | if (Subtarget->is64Bit() && !Subtarget->useSoftFloat()) { |
1924 | setLibcallImpl(Call: RTLIB::ADD_F128, Impl: RTLIB::_Qp_add); |
1925 | setLibcallImpl(Call: RTLIB::SUB_F128, Impl: RTLIB::_Qp_sub); |
1926 | setLibcallImpl(Call: RTLIB::MUL_F128, Impl: RTLIB::_Qp_mul); |
1927 | setLibcallImpl(Call: RTLIB::DIV_F128, Impl: RTLIB::_Qp_div); |
1928 | setLibcallImpl(Call: RTLIB::SQRT_F128, Impl: RTLIB::_Qp_sqrt); |
1929 | setLibcallImpl(Call: RTLIB::FPTOSINT_F128_I32, Impl: RTLIB::_Qp_qtoi); |
1930 | setLibcallImpl(Call: RTLIB::FPTOUINT_F128_I32, Impl: RTLIB::_Qp_qtoui); |
1931 | setLibcallImpl(Call: RTLIB::SINTTOFP_I32_F128, Impl: RTLIB::_Qp_itoq); |
1932 | setLibcallImpl(Call: RTLIB::UINTTOFP_I32_F128, Impl: RTLIB::_Qp_uitoq); |
1933 | setLibcallImpl(Call: RTLIB::FPTOSINT_F128_I64, Impl: RTLIB::_Qp_qtox); |
1934 | setLibcallImpl(Call: RTLIB::FPTOUINT_F128_I64, Impl: RTLIB::_Qp_qtoux); |
1935 | setLibcallImpl(Call: RTLIB::SINTTOFP_I64_F128, Impl: RTLIB::_Qp_xtoq); |
1936 | setLibcallImpl(Call: RTLIB::UINTTOFP_I64_F128, Impl: RTLIB::_Qp_uxtoq); |
1937 | setLibcallImpl(Call: RTLIB::FPEXT_F32_F128, Impl: RTLIB::_Qp_stoq); |
1938 | setLibcallImpl(Call: RTLIB::FPEXT_F64_F128, Impl: RTLIB::_Qp_dtoq); |
1939 | setLibcallImpl(Call: RTLIB::FPROUND_F128_F32, Impl: RTLIB::_Qp_qtos); |
1940 | setLibcallImpl(Call: RTLIB::FPROUND_F128_F64, Impl: RTLIB::_Qp_qtod); |
1941 | } else if (!Subtarget->useSoftFloat()) { |
1942 | setLibcallImpl(Call: RTLIB::ADD_F128, Impl: RTLIB::_Q_add); |
1943 | setLibcallImpl(Call: RTLIB::SUB_F128, Impl: RTLIB::_Q_sub); |
1944 | setLibcallImpl(Call: RTLIB::MUL_F128, Impl: RTLIB::_Q_mul); |
1945 | setLibcallImpl(Call: RTLIB::DIV_F128, Impl: RTLIB::_Q_div); |
1946 | setLibcallImpl(Call: RTLIB::SQRT_F128, Impl: RTLIB::_Q_sqrt); |
1947 | setLibcallImpl(Call: RTLIB::FPTOSINT_F128_I32, Impl: RTLIB::_Q_qtoi); |
1948 | setLibcallImpl(Call: RTLIB::FPTOUINT_F128_I32, Impl: RTLIB::_Q_qtou); |
1949 | setLibcallImpl(Call: RTLIB::SINTTOFP_I32_F128, Impl: RTLIB::_Q_itoq); |
1950 | setLibcallImpl(Call: RTLIB::UINTTOFP_I32_F128, Impl: RTLIB::_Q_utoq); |
1951 | setLibcallImpl(Call: RTLIB::FPTOSINT_F128_I64, Impl: RTLIB::_Q_qtoll); |
1952 | setLibcallImpl(Call: RTLIB::FPTOUINT_F128_I64, Impl: RTLIB::_Q_qtoull); |
1953 | setLibcallImpl(Call: RTLIB::SINTTOFP_I64_F128, Impl: RTLIB::_Q_lltoq); |
1954 | setLibcallImpl(Call: RTLIB::UINTTOFP_I64_F128, Impl: RTLIB::_Q_ulltoq); |
1955 | setLibcallImpl(Call: RTLIB::FPEXT_F32_F128, Impl: RTLIB::_Q_stoq); |
1956 | setLibcallImpl(Call: RTLIB::FPEXT_F64_F128, Impl: RTLIB::_Q_dtoq); |
1957 | setLibcallImpl(Call: RTLIB::FPROUND_F128_F32, Impl: RTLIB::_Q_qtos); |
1958 | setLibcallImpl(Call: RTLIB::FPROUND_F128_F64, Impl: RTLIB::_Q_qtod); |
1959 | } |
1960 | } |
1961 | |
1962 | if (Subtarget->fixAllFDIVSQRT()) { |
1963 | // Promote FDIVS and FSQRTS to FDIVD and FSQRTD instructions instead as |
1964 | // the former instructions generate errata on LEON processors. |
1965 | setOperationAction(Op: ISD::FDIV, VT: MVT::f32, Action: Promote); |
1966 | setOperationAction(Op: ISD::FSQRT, VT: MVT::f32, Action: Promote); |
1967 | } |
1968 | |
1969 | if (Subtarget->hasNoFMULS()) { |
1970 | setOperationAction(Op: ISD::FMUL, VT: MVT::f32, Action: Promote); |
1971 | } |
1972 | |
1973 | // Custom combine bitcast between f64 and v2i32 |
1974 | if (!Subtarget->is64Bit()) |
1975 | setTargetDAGCombine(ISD::BITCAST); |
1976 | |
1977 | if (Subtarget->hasLeonCycleCounter()) |
1978 | setOperationAction(Op: ISD::READCYCLECOUNTER, VT: MVT::i64, Action: Custom); |
1979 | |
1980 | if (Subtarget->isVIS3()) { |
1981 | setOperationAction(Op: ISD::CTLZ, VT: MVT::i32, Action: Legal); |
1982 | setOperationAction(Op: ISD::CTLZ, VT: MVT::i64, Action: Legal); |
1983 | setOperationAction(Op: ISD::CTLZ_ZERO_UNDEF, VT: MVT::i32, Action: Legal); |
1984 | setOperationAction(Op: ISD::CTLZ_ZERO_UNDEF, VT: MVT::i64, Action: Legal); |
1985 | |
1986 | setOperationAction(Op: ISD::CTTZ, VT: MVT::i32, |
1987 | Action: Subtarget->is64Bit() ? Promote : Expand); |
1988 | setOperationAction(Op: ISD::CTTZ, VT: MVT::i64, Action: Expand); |
1989 | setOperationAction(Op: ISD::CTTZ_ZERO_UNDEF, VT: MVT::i32, |
1990 | Action: Subtarget->is64Bit() ? Promote : Expand); |
1991 | setOperationAction(Op: ISD::CTTZ_ZERO_UNDEF, VT: MVT::i64, Action: Expand); |
1992 | } else if (Subtarget->usePopc()) { |
1993 | setOperationAction(Op: ISD::CTLZ, VT: MVT::i32, Action: Expand); |
1994 | setOperationAction(Op: ISD::CTLZ, VT: MVT::i64, Action: Expand); |
1995 | setOperationAction(Op: ISD::CTLZ_ZERO_UNDEF, VT: MVT::i32, Action: Expand); |
1996 | setOperationAction(Op: ISD::CTLZ_ZERO_UNDEF, VT: MVT::i64, Action: Expand); |
1997 | |
1998 | setOperationAction(Op: ISD::CTTZ, VT: MVT::i32, Action: Expand); |
1999 | setOperationAction(Op: ISD::CTTZ, VT: MVT::i64, Action: Expand); |
2000 | setOperationAction(Op: ISD::CTTZ_ZERO_UNDEF, VT: MVT::i32, Action: Expand); |
2001 | setOperationAction(Op: ISD::CTTZ_ZERO_UNDEF, VT: MVT::i64, Action: Expand); |
2002 | } else { |
2003 | setOperationAction(Op: ISD::CTLZ, VT: MVT::i32, Action: Expand); |
2004 | setOperationAction(Op: ISD::CTLZ, VT: MVT::i64, Action: Expand); |
2005 | setOperationAction(Op: ISD::CTLZ_ZERO_UNDEF, VT: MVT::i32, |
2006 | Action: Subtarget->is64Bit() ? Promote : LibCall); |
2007 | setOperationAction(Op: ISD::CTLZ_ZERO_UNDEF, VT: MVT::i64, Action: LibCall); |
2008 | |
2009 | // FIXME here we don't have any ISA extensions that could help us, so to |
2010 | // prevent large expansions those should be made into LibCalls. |
2011 | setOperationAction(Op: ISD::CTTZ, VT: MVT::i32, Action: Expand); |
2012 | setOperationAction(Op: ISD::CTTZ, VT: MVT::i64, Action: Expand); |
2013 | setOperationAction(Op: ISD::CTTZ_ZERO_UNDEF, VT: MVT::i32, Action: Expand); |
2014 | setOperationAction(Op: ISD::CTTZ_ZERO_UNDEF, VT: MVT::i64, Action: Expand); |
2015 | } |
2016 | |
2017 | setOperationAction(Op: ISD::INTRINSIC_WO_CHAIN, VT: MVT::Other, Action: Custom); |
2018 | |
2019 | setMinFunctionAlignment(Align(4)); |
2020 | |
2021 | computeRegisterProperties(TRI: Subtarget->getRegisterInfo()); |
2022 | } |
2023 | |
2024 | bool SparcTargetLowering::useSoftFloat() const { |
2025 | return Subtarget->useSoftFloat(); |
2026 | } |
2027 | |
2028 | EVT SparcTargetLowering::getSetCCResultType(const DataLayout &, LLVMContext &, |
2029 | EVT VT) const { |
2030 | if (!VT.isVector()) |
2031 | return MVT::i32; |
2032 | return VT.changeVectorElementTypeToInteger(); |
2033 | } |
2034 | |
2035 | /// isMaskedValueZeroForTargetNode - Return true if 'Op & Mask' is known to |
2036 | /// be zero. Op is expected to be a target specific node. Used by DAG |
2037 | /// combiner. |
2038 | void SparcTargetLowering::computeKnownBitsForTargetNode |
2039 | (const SDValue Op, |
2040 | KnownBits &Known, |
2041 | const APInt &DemandedElts, |
2042 | const SelectionDAG &DAG, |
2043 | unsigned Depth) const { |
2044 | KnownBits Known2; |
2045 | Known.resetAll(); |
2046 | |
2047 | switch (Op.getOpcode()) { |
2048 | default: break; |
2049 | case SPISD::SELECT_ICC: |
2050 | case SPISD::SELECT_XCC: |
2051 | case SPISD::SELECT_FCC: |
2052 | Known = DAG.computeKnownBits(Op: Op.getOperand(i: 1), Depth: Depth + 1); |
2053 | Known2 = DAG.computeKnownBits(Op: Op.getOperand(i: 0), Depth: Depth + 1); |
2054 | |
2055 | // Only known if known in both the LHS and RHS. |
2056 | Known = Known.intersectWith(RHS: Known2); |
2057 | break; |
2058 | } |
2059 | } |
2060 | |
2061 | // Look at LHS/RHS/CC and see if they are a lowered setcc instruction. If so |
2062 | // set LHS/RHS and SPCC to the LHS/RHS of the setcc and SPCC to the condition. |
2063 | static void LookThroughSetCC(SDValue &LHS, SDValue &RHS, |
2064 | ISD::CondCode CC, unsigned &SPCC) { |
2065 | if (isNullConstant(V: RHS) && CC == ISD::SETNE && |
2066 | (((LHS.getOpcode() == SPISD::SELECT_ICC || |
2067 | LHS.getOpcode() == SPISD::SELECT_XCC) && |
2068 | LHS.getOperand(i: 3).getOpcode() == SPISD::CMPICC) || |
2069 | (LHS.getOpcode() == SPISD::SELECT_FCC && |
2070 | (LHS.getOperand(i: 3).getOpcode() == SPISD::CMPFCC || |
2071 | LHS.getOperand(i: 3).getOpcode() == SPISD::CMPFCC_V9))) && |
2072 | isOneConstant(V: LHS.getOperand(i: 0)) && isNullConstant(V: LHS.getOperand(i: 1))) { |
2073 | SDValue CMPCC = LHS.getOperand(i: 3); |
2074 | SPCC = LHS.getConstantOperandVal(i: 2); |
2075 | LHS = CMPCC.getOperand(i: 0); |
2076 | RHS = CMPCC.getOperand(i: 1); |
2077 | } |
2078 | } |
2079 | |
2080 | // Convert to a target node and set target flags. |
2081 | SDValue SparcTargetLowering::withTargetFlags(SDValue Op, unsigned TF, |
2082 | SelectionDAG &DAG) const { |
2083 | if (const GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Val&: Op)) |
2084 | return DAG.getTargetGlobalAddress(GV: GA->getGlobal(), |
2085 | DL: SDLoc(GA), |
2086 | VT: GA->getValueType(ResNo: 0), |
2087 | offset: GA->getOffset(), TargetFlags: TF); |
2088 | |
2089 | if (const ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Val&: Op)) |
2090 | return DAG.getTargetConstantPool(C: CP->getConstVal(), VT: CP->getValueType(ResNo: 0), |
2091 | Align: CP->getAlign(), Offset: CP->getOffset(), TargetFlags: TF); |
2092 | |
2093 | if (const BlockAddressSDNode *BA = dyn_cast<BlockAddressSDNode>(Val&: Op)) |
2094 | return DAG.getTargetBlockAddress(BA: BA->getBlockAddress(), |
2095 | VT: Op.getValueType(), |
2096 | Offset: 0, |
2097 | TargetFlags: TF); |
2098 | |
2099 | if (const ExternalSymbolSDNode *ES = dyn_cast<ExternalSymbolSDNode>(Val&: Op)) |
2100 | return DAG.getTargetExternalSymbol(Sym: ES->getSymbol(), |
2101 | VT: ES->getValueType(ResNo: 0), TargetFlags: TF); |
2102 | |
2103 | llvm_unreachable("Unhandled address SDNode" ); |
2104 | } |
2105 | |
2106 | // Split Op into high and low parts according to HiTF and LoTF. |
2107 | // Return an ADD node combining the parts. |
2108 | SDValue SparcTargetLowering::makeHiLoPair(SDValue Op, |
2109 | unsigned HiTF, unsigned LoTF, |
2110 | SelectionDAG &DAG) const { |
2111 | SDLoc DL(Op); |
2112 | EVT VT = Op.getValueType(); |
2113 | SDValue Hi = DAG.getNode(Opcode: SPISD::Hi, DL, VT, Operand: withTargetFlags(Op, TF: HiTF, DAG)); |
2114 | SDValue Lo = DAG.getNode(Opcode: SPISD::Lo, DL, VT, Operand: withTargetFlags(Op, TF: LoTF, DAG)); |
2115 | return DAG.getNode(Opcode: ISD::ADD, DL, VT, N1: Hi, N2: Lo); |
2116 | } |
2117 | |
2118 | // Build SDNodes for producing an address from a GlobalAddress, ConstantPool, |
2119 | // or ExternalSymbol SDNode. |
2120 | SDValue SparcTargetLowering::makeAddress(SDValue Op, SelectionDAG &DAG) const { |
2121 | SDLoc DL(Op); |
2122 | EVT VT = getPointerTy(DL: DAG.getDataLayout()); |
2123 | |
2124 | // Handle PIC mode first. SPARC needs a got load for every variable! |
2125 | if (isPositionIndependent()) { |
2126 | const Module *M = DAG.getMachineFunction().getFunction().getParent(); |
2127 | PICLevel::Level picLevel = M->getPICLevel(); |
2128 | SDValue Idx; |
2129 | |
2130 | if (picLevel == PICLevel::SmallPIC) { |
2131 | // This is the pic13 code model, the GOT is known to be smaller than 8KiB. |
2132 | Idx = DAG.getNode(Opcode: SPISD::Lo, DL, VT: Op.getValueType(), |
2133 | Operand: withTargetFlags(Op, TF: ELF::R_SPARC_GOT13, DAG)); |
2134 | } else { |
2135 | // This is the pic32 code model, the GOT is known to be smaller than 4GB. |
2136 | Idx = makeHiLoPair(Op, HiTF: ELF::R_SPARC_GOT22, LoTF: ELF::R_SPARC_GOT10, DAG); |
2137 | } |
2138 | |
2139 | SDValue GlobalBase = DAG.getNode(Opcode: SPISD::GLOBAL_BASE_REG, DL, VT); |
2140 | SDValue AbsAddr = DAG.getNode(Opcode: ISD::ADD, DL, VT, N1: GlobalBase, N2: Idx); |
2141 | // GLOBAL_BASE_REG codegen'ed with call. Inform MFI that this |
2142 | // function has calls. |
2143 | MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); |
2144 | MFI.setHasCalls(true); |
2145 | return DAG.getLoad(VT, dl: DL, Chain: DAG.getEntryNode(), Ptr: AbsAddr, |
2146 | PtrInfo: MachinePointerInfo::getGOT(MF&: DAG.getMachineFunction())); |
2147 | } |
2148 | |
2149 | // This is one of the absolute code models. |
2150 | switch(getTargetMachine().getCodeModel()) { |
2151 | default: |
2152 | llvm_unreachable("Unsupported absolute code model" ); |
2153 | case CodeModel::Small: |
2154 | // abs32. |
2155 | return makeHiLoPair(Op, HiTF: ELF::R_SPARC_HI22, LoTF: ELF::R_SPARC_LO10, DAG); |
2156 | case CodeModel::Medium: { |
2157 | // abs44. |
2158 | SDValue H44 = makeHiLoPair(Op, HiTF: ELF::R_SPARC_H44, LoTF: ELF::R_SPARC_M44, DAG); |
2159 | H44 = DAG.getNode(Opcode: ISD::SHL, DL, VT, N1: H44, N2: DAG.getConstant(Val: 12, DL, VT: MVT::i32)); |
2160 | SDValue L44 = withTargetFlags(Op, TF: ELF::R_SPARC_L44, DAG); |
2161 | L44 = DAG.getNode(Opcode: SPISD::Lo, DL, VT, Operand: L44); |
2162 | return DAG.getNode(Opcode: ISD::ADD, DL, VT, N1: H44, N2: L44); |
2163 | } |
2164 | case CodeModel::Large: { |
2165 | // abs64. |
2166 | SDValue Hi = makeHiLoPair(Op, HiTF: ELF::R_SPARC_HH22, LoTF: ELF::R_SPARC_HM10, DAG); |
2167 | Hi = DAG.getNode(Opcode: ISD::SHL, DL, VT, N1: Hi, N2: DAG.getConstant(Val: 32, DL, VT: MVT::i32)); |
2168 | SDValue Lo = makeHiLoPair(Op, HiTF: ELF::R_SPARC_HI22, LoTF: ELF::R_SPARC_LO10, DAG); |
2169 | return DAG.getNode(Opcode: ISD::ADD, DL, VT, N1: Hi, N2: Lo); |
2170 | } |
2171 | } |
2172 | } |
2173 | |
2174 | SDValue SparcTargetLowering::LowerGlobalAddress(SDValue Op, |
2175 | SelectionDAG &DAG) const { |
2176 | return makeAddress(Op, DAG); |
2177 | } |
2178 | |
2179 | SDValue SparcTargetLowering::LowerConstantPool(SDValue Op, |
2180 | SelectionDAG &DAG) const { |
2181 | return makeAddress(Op, DAG); |
2182 | } |
2183 | |
2184 | SDValue SparcTargetLowering::LowerBlockAddress(SDValue Op, |
2185 | SelectionDAG &DAG) const { |
2186 | return makeAddress(Op, DAG); |
2187 | } |
2188 | |
2189 | SDValue SparcTargetLowering::LowerGlobalTLSAddress(SDValue Op, |
2190 | SelectionDAG &DAG) const { |
2191 | |
2192 | GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Val&: Op); |
2193 | if (DAG.getTarget().useEmulatedTLS()) |
2194 | return LowerToTLSEmulatedModel(GA, DAG); |
2195 | |
2196 | SDLoc DL(GA); |
2197 | const GlobalValue *GV = GA->getGlobal(); |
2198 | EVT PtrVT = getPointerTy(DL: DAG.getDataLayout()); |
2199 | |
2200 | TLSModel::Model model = getTargetMachine().getTLSModel(GV); |
2201 | |
2202 | if (model == TLSModel::GeneralDynamic || model == TLSModel::LocalDynamic) { |
2203 | unsigned HiTF = |
2204 | ((model == TLSModel::GeneralDynamic) ? ELF::R_SPARC_TLS_GD_HI22 |
2205 | : ELF::R_SPARC_TLS_LDM_HI22); |
2206 | unsigned LoTF = |
2207 | ((model == TLSModel::GeneralDynamic) ? ELF::R_SPARC_TLS_GD_LO10 |
2208 | : ELF::R_SPARC_TLS_LDM_LO10); |
2209 | unsigned addTF = |
2210 | ((model == TLSModel::GeneralDynamic) ? ELF::R_SPARC_TLS_GD_ADD |
2211 | : ELF::R_SPARC_TLS_LDM_ADD); |
2212 | unsigned callTF = |
2213 | ((model == TLSModel::GeneralDynamic) ? ELF::R_SPARC_TLS_GD_CALL |
2214 | : ELF::R_SPARC_TLS_LDM_CALL); |
2215 | |
2216 | SDValue HiLo = makeHiLoPair(Op, HiTF, LoTF, DAG); |
2217 | SDValue Base = DAG.getNode(Opcode: SPISD::GLOBAL_BASE_REG, DL, VT: PtrVT); |
2218 | SDValue Argument = DAG.getNode(Opcode: SPISD::TLS_ADD, DL, VT: PtrVT, N1: Base, N2: HiLo, |
2219 | N3: withTargetFlags(Op, TF: addTF, DAG)); |
2220 | |
2221 | SDValue Chain = DAG.getEntryNode(); |
2222 | SDValue InGlue; |
2223 | |
2224 | Chain = DAG.getCALLSEQ_START(Chain, InSize: 1, OutSize: 0, DL); |
2225 | Chain = DAG.getCopyToReg(Chain, dl: DL, Reg: SP::O0, N: Argument, Glue: InGlue); |
2226 | InGlue = Chain.getValue(R: 1); |
2227 | SDValue Callee = DAG.getTargetExternalSymbol(Sym: "__tls_get_addr" , VT: PtrVT); |
2228 | SDValue Symbol = withTargetFlags(Op, TF: callTF, DAG); |
2229 | |
2230 | SDVTList NodeTys = DAG.getVTList(VT1: MVT::Other, VT2: MVT::Glue); |
2231 | const uint32_t *Mask = Subtarget->getRegisterInfo()->getCallPreservedMask( |
2232 | MF: DAG.getMachineFunction(), CC: CallingConv::C); |
2233 | assert(Mask && "Missing call preserved mask for calling convention" ); |
2234 | SDValue Ops[] = {Chain, |
2235 | Callee, |
2236 | Symbol, |
2237 | DAG.getRegister(Reg: SP::O0, VT: PtrVT), |
2238 | DAG.getRegisterMask(RegMask: Mask), |
2239 | InGlue}; |
2240 | Chain = DAG.getNode(Opcode: SPISD::TLS_CALL, DL, VTList: NodeTys, Ops); |
2241 | InGlue = Chain.getValue(R: 1); |
2242 | Chain = DAG.getCALLSEQ_END(Chain, Size1: 1, Size2: 0, Glue: InGlue, DL); |
2243 | InGlue = Chain.getValue(R: 1); |
2244 | SDValue Ret = DAG.getCopyFromReg(Chain, dl: DL, Reg: SP::O0, VT: PtrVT, Glue: InGlue); |
2245 | |
2246 | if (model != TLSModel::LocalDynamic) |
2247 | return Ret; |
2248 | |
2249 | SDValue Hi = |
2250 | DAG.getNode(Opcode: SPISD::Hi, DL, VT: PtrVT, |
2251 | Operand: withTargetFlags(Op, TF: ELF::R_SPARC_TLS_LDO_HIX22, DAG)); |
2252 | SDValue Lo = |
2253 | DAG.getNode(Opcode: SPISD::Lo, DL, VT: PtrVT, |
2254 | Operand: withTargetFlags(Op, TF: ELF::R_SPARC_TLS_LDO_LOX10, DAG)); |
2255 | HiLo = DAG.getNode(Opcode: ISD::XOR, DL, VT: PtrVT, N1: Hi, N2: Lo); |
2256 | return DAG.getNode(Opcode: SPISD::TLS_ADD, DL, VT: PtrVT, N1: Ret, N2: HiLo, |
2257 | N3: withTargetFlags(Op, TF: ELF::R_SPARC_TLS_LDO_ADD, DAG)); |
2258 | } |
2259 | |
2260 | if (model == TLSModel::InitialExec) { |
2261 | unsigned ldTF = ((PtrVT == MVT::i64) ? ELF::R_SPARC_TLS_IE_LDX |
2262 | : ELF::R_SPARC_TLS_IE_LD); |
2263 | |
2264 | SDValue Base = DAG.getNode(Opcode: SPISD::GLOBAL_BASE_REG, DL, VT: PtrVT); |
2265 | |
2266 | // GLOBAL_BASE_REG codegen'ed with call. Inform MFI that this |
2267 | // function has calls. |
2268 | MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); |
2269 | MFI.setHasCalls(true); |
2270 | |
2271 | SDValue TGA = makeHiLoPair(Op, HiTF: ELF::R_SPARC_TLS_IE_HI22, |
2272 | LoTF: ELF::R_SPARC_TLS_IE_LO10, DAG); |
2273 | SDValue Ptr = DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrVT, N1: Base, N2: TGA); |
2274 | SDValue Offset = DAG.getNode(Opcode: SPISD::TLS_LD, |
2275 | DL, VT: PtrVT, N1: Ptr, |
2276 | N2: withTargetFlags(Op, TF: ldTF, DAG)); |
2277 | return DAG.getNode(Opcode: SPISD::TLS_ADD, DL, VT: PtrVT, |
2278 | N1: DAG.getRegister(Reg: SP::G7, VT: PtrVT), N2: Offset, |
2279 | N3: withTargetFlags(Op, TF: ELF::R_SPARC_TLS_IE_ADD, DAG)); |
2280 | } |
2281 | |
2282 | assert(model == TLSModel::LocalExec); |
2283 | SDValue Hi = DAG.getNode(Opcode: SPISD::Hi, DL, VT: PtrVT, |
2284 | Operand: withTargetFlags(Op, TF: ELF::R_SPARC_TLS_LE_HIX22, DAG)); |
2285 | SDValue Lo = DAG.getNode(Opcode: SPISD::Lo, DL, VT: PtrVT, |
2286 | Operand: withTargetFlags(Op, TF: ELF::R_SPARC_TLS_LE_LOX10, DAG)); |
2287 | SDValue Offset = DAG.getNode(Opcode: ISD::XOR, DL, VT: PtrVT, N1: Hi, N2: Lo); |
2288 | |
2289 | return DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrVT, |
2290 | N1: DAG.getRegister(Reg: SP::G7, VT: PtrVT), N2: Offset); |
2291 | } |
2292 | |
2293 | SDValue SparcTargetLowering::LowerF128_LibCallArg(SDValue Chain, |
2294 | ArgListTy &Args, SDValue Arg, |
2295 | const SDLoc &DL, |
2296 | SelectionDAG &DAG) const { |
2297 | MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); |
2298 | EVT ArgVT = Arg.getValueType(); |
2299 | Type *ArgTy = ArgVT.getTypeForEVT(Context&: *DAG.getContext()); |
2300 | |
2301 | ArgListEntry Entry; |
2302 | Entry.Node = Arg; |
2303 | Entry.Ty = ArgTy; |
2304 | |
2305 | if (ArgTy->isFP128Ty()) { |
2306 | // Create a stack object and pass the pointer to the library function. |
2307 | int FI = MFI.CreateStackObject(Size: 16, Alignment: Align(8), isSpillSlot: false); |
2308 | SDValue FIPtr = DAG.getFrameIndex(FI, VT: getPointerTy(DL: DAG.getDataLayout())); |
2309 | Chain = DAG.getStore(Chain, dl: DL, Val: Entry.Node, Ptr: FIPtr, PtrInfo: MachinePointerInfo(), |
2310 | Alignment: Align(8)); |
2311 | |
2312 | Entry.Node = FIPtr; |
2313 | Entry.Ty = PointerType::getUnqual(C&: ArgTy->getContext()); |
2314 | } |
2315 | Args.push_back(x: Entry); |
2316 | return Chain; |
2317 | } |
2318 | |
2319 | SDValue |
2320 | SparcTargetLowering::LowerF128Op(SDValue Op, SelectionDAG &DAG, |
2321 | const char *LibFuncName, |
2322 | unsigned numArgs) const { |
2323 | |
2324 | ArgListTy Args; |
2325 | |
2326 | MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); |
2327 | auto PtrVT = getPointerTy(DL: DAG.getDataLayout()); |
2328 | |
2329 | SDValue Callee = DAG.getExternalSymbol(Sym: LibFuncName, VT: PtrVT); |
2330 | Type *RetTy = Op.getValueType().getTypeForEVT(Context&: *DAG.getContext()); |
2331 | Type *RetTyABI = RetTy; |
2332 | SDValue Chain = DAG.getEntryNode(); |
2333 | SDValue RetPtr; |
2334 | |
2335 | if (RetTy->isFP128Ty()) { |
2336 | // Create a Stack Object to receive the return value of type f128. |
2337 | ArgListEntry Entry; |
2338 | int RetFI = MFI.CreateStackObject(Size: 16, Alignment: Align(8), isSpillSlot: false); |
2339 | RetPtr = DAG.getFrameIndex(FI: RetFI, VT: PtrVT); |
2340 | Entry.Node = RetPtr; |
2341 | Entry.Ty = PointerType::getUnqual(C&: RetTy->getContext()); |
2342 | if (!Subtarget->is64Bit()) { |
2343 | Entry.IsSRet = true; |
2344 | Entry.IndirectType = RetTy; |
2345 | } |
2346 | Entry.IsReturned = false; |
2347 | Args.push_back(x: Entry); |
2348 | RetTyABI = Type::getVoidTy(C&: *DAG.getContext()); |
2349 | } |
2350 | |
2351 | assert(Op->getNumOperands() >= numArgs && "Not enough operands!" ); |
2352 | for (unsigned i = 0, e = numArgs; i != e; ++i) { |
2353 | Chain = LowerF128_LibCallArg(Chain, Args, Arg: Op.getOperand(i), DL: SDLoc(Op), DAG); |
2354 | } |
2355 | TargetLowering::CallLoweringInfo CLI(DAG); |
2356 | CLI.setDebugLoc(SDLoc(Op)).setChain(Chain) |
2357 | .setCallee(CC: CallingConv::C, ResultType: RetTyABI, Target: Callee, ArgsList: std::move(Args)); |
2358 | |
2359 | std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI); |
2360 | |
2361 | // chain is in second result. |
2362 | if (RetTyABI == RetTy) |
2363 | return CallInfo.first; |
2364 | |
2365 | assert (RetTy->isFP128Ty() && "Unexpected return type!" ); |
2366 | |
2367 | Chain = CallInfo.second; |
2368 | |
2369 | // Load RetPtr to get the return value. |
2370 | return DAG.getLoad(VT: Op.getValueType(), dl: SDLoc(Op), Chain, Ptr: RetPtr, |
2371 | PtrInfo: MachinePointerInfo(), Alignment: Align(8)); |
2372 | } |
2373 | |
2374 | SDValue SparcTargetLowering::LowerF128Compare(SDValue LHS, SDValue RHS, |
2375 | unsigned &SPCC, const SDLoc &DL, |
2376 | SelectionDAG &DAG) const { |
2377 | |
2378 | const char *LibCall = nullptr; |
2379 | bool is64Bit = Subtarget->is64Bit(); |
2380 | switch(SPCC) { |
2381 | default: llvm_unreachable("Unhandled conditional code!" ); |
2382 | case SPCC::FCC_E : LibCall = is64Bit? "_Qp_feq" : "_Q_feq" ; break; |
2383 | case SPCC::FCC_NE : LibCall = is64Bit? "_Qp_fne" : "_Q_fne" ; break; |
2384 | case SPCC::FCC_L : LibCall = is64Bit? "_Qp_flt" : "_Q_flt" ; break; |
2385 | case SPCC::FCC_G : LibCall = is64Bit? "_Qp_fgt" : "_Q_fgt" ; break; |
2386 | case SPCC::FCC_LE : LibCall = is64Bit? "_Qp_fle" : "_Q_fle" ; break; |
2387 | case SPCC::FCC_GE : LibCall = is64Bit? "_Qp_fge" : "_Q_fge" ; break; |
2388 | case SPCC::FCC_UL : |
2389 | case SPCC::FCC_ULE: |
2390 | case SPCC::FCC_UG : |
2391 | case SPCC::FCC_UGE: |
2392 | case SPCC::FCC_U : |
2393 | case SPCC::FCC_O : |
2394 | case SPCC::FCC_LG : |
2395 | case SPCC::FCC_UE : LibCall = is64Bit? "_Qp_cmp" : "_Q_cmp" ; break; |
2396 | } |
2397 | |
2398 | auto PtrVT = getPointerTy(DL: DAG.getDataLayout()); |
2399 | SDValue Callee = DAG.getExternalSymbol(Sym: LibCall, VT: PtrVT); |
2400 | Type *RetTy = Type::getInt32Ty(C&: *DAG.getContext()); |
2401 | ArgListTy Args; |
2402 | SDValue Chain = DAG.getEntryNode(); |
2403 | Chain = LowerF128_LibCallArg(Chain, Args, Arg: LHS, DL, DAG); |
2404 | Chain = LowerF128_LibCallArg(Chain, Args, Arg: RHS, DL, DAG); |
2405 | |
2406 | TargetLowering::CallLoweringInfo CLI(DAG); |
2407 | CLI.setDebugLoc(DL).setChain(Chain) |
2408 | .setCallee(CC: CallingConv::C, ResultType: RetTy, Target: Callee, ArgsList: std::move(Args)); |
2409 | |
2410 | std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI); |
2411 | |
2412 | // result is in first, and chain is in second result. |
2413 | SDValue Result = CallInfo.first; |
2414 | |
2415 | switch(SPCC) { |
2416 | default: { |
2417 | SDValue RHS = DAG.getConstant(Val: 0, DL, VT: Result.getValueType()); |
2418 | SPCC = SPCC::ICC_NE; |
2419 | return DAG.getNode(Opcode: SPISD::CMPICC, DL, VT: MVT::Glue, N1: Result, N2: RHS); |
2420 | } |
2421 | case SPCC::FCC_UL : { |
2422 | SDValue Mask = DAG.getConstant(Val: 1, DL, VT: Result.getValueType()); |
2423 | Result = DAG.getNode(Opcode: ISD::AND, DL, VT: Result.getValueType(), N1: Result, N2: Mask); |
2424 | SDValue RHS = DAG.getConstant(Val: 0, DL, VT: Result.getValueType()); |
2425 | SPCC = SPCC::ICC_NE; |
2426 | return DAG.getNode(Opcode: SPISD::CMPICC, DL, VT: MVT::Glue, N1: Result, N2: RHS); |
2427 | } |
2428 | case SPCC::FCC_ULE: { |
2429 | SDValue RHS = DAG.getConstant(Val: 2, DL, VT: Result.getValueType()); |
2430 | SPCC = SPCC::ICC_NE; |
2431 | return DAG.getNode(Opcode: SPISD::CMPICC, DL, VT: MVT::Glue, N1: Result, N2: RHS); |
2432 | } |
2433 | case SPCC::FCC_UG : { |
2434 | SDValue RHS = DAG.getConstant(Val: 1, DL, VT: Result.getValueType()); |
2435 | SPCC = SPCC::ICC_G; |
2436 | return DAG.getNode(Opcode: SPISD::CMPICC, DL, VT: MVT::Glue, N1: Result, N2: RHS); |
2437 | } |
2438 | case SPCC::FCC_UGE: { |
2439 | SDValue RHS = DAG.getConstant(Val: 1, DL, VT: Result.getValueType()); |
2440 | SPCC = SPCC::ICC_NE; |
2441 | return DAG.getNode(Opcode: SPISD::CMPICC, DL, VT: MVT::Glue, N1: Result, N2: RHS); |
2442 | } |
2443 | |
2444 | case SPCC::FCC_U : { |
2445 | SDValue RHS = DAG.getConstant(Val: 3, DL, VT: Result.getValueType()); |
2446 | SPCC = SPCC::ICC_E; |
2447 | return DAG.getNode(Opcode: SPISD::CMPICC, DL, VT: MVT::Glue, N1: Result, N2: RHS); |
2448 | } |
2449 | case SPCC::FCC_O : { |
2450 | SDValue RHS = DAG.getConstant(Val: 3, DL, VT: Result.getValueType()); |
2451 | SPCC = SPCC::ICC_NE; |
2452 | return DAG.getNode(Opcode: SPISD::CMPICC, DL, VT: MVT::Glue, N1: Result, N2: RHS); |
2453 | } |
2454 | case SPCC::FCC_LG : { |
2455 | SDValue Mask = DAG.getConstant(Val: 3, DL, VT: Result.getValueType()); |
2456 | Result = DAG.getNode(Opcode: ISD::AND, DL, VT: Result.getValueType(), N1: Result, N2: Mask); |
2457 | SDValue RHS = DAG.getConstant(Val: 0, DL, VT: Result.getValueType()); |
2458 | SPCC = SPCC::ICC_NE; |
2459 | return DAG.getNode(Opcode: SPISD::CMPICC, DL, VT: MVT::Glue, N1: Result, N2: RHS); |
2460 | } |
2461 | case SPCC::FCC_UE : { |
2462 | SDValue Mask = DAG.getConstant(Val: 3, DL, VT: Result.getValueType()); |
2463 | Result = DAG.getNode(Opcode: ISD::AND, DL, VT: Result.getValueType(), N1: Result, N2: Mask); |
2464 | SDValue RHS = DAG.getConstant(Val: 0, DL, VT: Result.getValueType()); |
2465 | SPCC = SPCC::ICC_E; |
2466 | return DAG.getNode(Opcode: SPISD::CMPICC, DL, VT: MVT::Glue, N1: Result, N2: RHS); |
2467 | } |
2468 | } |
2469 | } |
2470 | |
2471 | static SDValue |
2472 | LowerF128_FPEXTEND(SDValue Op, SelectionDAG &DAG, |
2473 | const SparcTargetLowering &TLI) { |
2474 | |
2475 | if (Op.getOperand(i: 0).getValueType() == MVT::f64) |
2476 | return TLI.LowerF128Op(Op, DAG, |
2477 | LibFuncName: TLI.getLibcallName(Call: RTLIB::FPEXT_F64_F128), numArgs: 1); |
2478 | |
2479 | if (Op.getOperand(i: 0).getValueType() == MVT::f32) |
2480 | return TLI.LowerF128Op(Op, DAG, |
2481 | LibFuncName: TLI.getLibcallName(Call: RTLIB::FPEXT_F32_F128), numArgs: 1); |
2482 | |
2483 | llvm_unreachable("fpextend with non-float operand!" ); |
2484 | return SDValue(); |
2485 | } |
2486 | |
2487 | static SDValue |
2488 | LowerF128_FPROUND(SDValue Op, SelectionDAG &DAG, |
2489 | const SparcTargetLowering &TLI) { |
2490 | // FP_ROUND on f64 and f32 are legal. |
2491 | if (Op.getOperand(i: 0).getValueType() != MVT::f128) |
2492 | return Op; |
2493 | |
2494 | if (Op.getValueType() == MVT::f64) |
2495 | return TLI.LowerF128Op(Op, DAG, |
2496 | LibFuncName: TLI.getLibcallName(Call: RTLIB::FPROUND_F128_F64), numArgs: 1); |
2497 | if (Op.getValueType() == MVT::f32) |
2498 | return TLI.LowerF128Op(Op, DAG, |
2499 | LibFuncName: TLI.getLibcallName(Call: RTLIB::FPROUND_F128_F32), numArgs: 1); |
2500 | |
2501 | llvm_unreachable("fpround to non-float!" ); |
2502 | return SDValue(); |
2503 | } |
2504 | |
2505 | static SDValue LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG, |
2506 | const SparcTargetLowering &TLI, |
2507 | bool hasHardQuad) { |
2508 | SDLoc dl(Op); |
2509 | EVT VT = Op.getValueType(); |
2510 | assert(VT == MVT::i32 || VT == MVT::i64); |
2511 | |
2512 | // Expand f128 operations to fp128 abi calls. |
2513 | if (Op.getOperand(i: 0).getValueType() == MVT::f128 |
2514 | && (!hasHardQuad || !TLI.isTypeLegal(VT))) { |
2515 | const char *libName = TLI.getLibcallName(Call: VT == MVT::i32 |
2516 | ? RTLIB::FPTOSINT_F128_I32 |
2517 | : RTLIB::FPTOSINT_F128_I64); |
2518 | return TLI.LowerF128Op(Op, DAG, LibFuncName: libName, numArgs: 1); |
2519 | } |
2520 | |
2521 | // Expand if the resulting type is illegal. |
2522 | if (!TLI.isTypeLegal(VT)) |
2523 | return SDValue(); |
2524 | |
2525 | // Otherwise, Convert the fp value to integer in an FP register. |
2526 | if (VT == MVT::i32) |
2527 | Op = DAG.getNode(Opcode: SPISD::FTOI, DL: dl, VT: MVT::f32, Operand: Op.getOperand(i: 0)); |
2528 | else |
2529 | Op = DAG.getNode(Opcode: SPISD::FTOX, DL: dl, VT: MVT::f64, Operand: Op.getOperand(i: 0)); |
2530 | |
2531 | return DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT, Operand: Op); |
2532 | } |
2533 | |
2534 | static SDValue LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG, |
2535 | const SparcTargetLowering &TLI, |
2536 | bool hasHardQuad) { |
2537 | SDLoc dl(Op); |
2538 | EVT OpVT = Op.getOperand(i: 0).getValueType(); |
2539 | assert(OpVT == MVT::i32 || (OpVT == MVT::i64)); |
2540 | |
2541 | EVT floatVT = (OpVT == MVT::i32) ? MVT::f32 : MVT::f64; |
2542 | |
2543 | // Expand f128 operations to fp128 ABI calls. |
2544 | if (Op.getValueType() == MVT::f128 |
2545 | && (!hasHardQuad || !TLI.isTypeLegal(VT: OpVT))) { |
2546 | const char *libName = TLI.getLibcallName(Call: OpVT == MVT::i32 |
2547 | ? RTLIB::SINTTOFP_I32_F128 |
2548 | : RTLIB::SINTTOFP_I64_F128); |
2549 | return TLI.LowerF128Op(Op, DAG, LibFuncName: libName, numArgs: 1); |
2550 | } |
2551 | |
2552 | // Expand if the operand type is illegal. |
2553 | if (!TLI.isTypeLegal(VT: OpVT)) |
2554 | return SDValue(); |
2555 | |
2556 | // Otherwise, Convert the int value to FP in an FP register. |
2557 | SDValue Tmp = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: floatVT, Operand: Op.getOperand(i: 0)); |
2558 | unsigned opcode = (OpVT == MVT::i32)? SPISD::ITOF : SPISD::XTOF; |
2559 | return DAG.getNode(Opcode: opcode, DL: dl, VT: Op.getValueType(), Operand: Tmp); |
2560 | } |
2561 | |
2562 | static SDValue LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG, |
2563 | const SparcTargetLowering &TLI, |
2564 | bool hasHardQuad) { |
2565 | EVT VT = Op.getValueType(); |
2566 | |
2567 | // Expand if it does not involve f128 or the target has support for |
2568 | // quad floating point instructions and the resulting type is legal. |
2569 | if (Op.getOperand(i: 0).getValueType() != MVT::f128 || |
2570 | (hasHardQuad && TLI.isTypeLegal(VT))) |
2571 | return SDValue(); |
2572 | |
2573 | assert(VT == MVT::i32 || VT == MVT::i64); |
2574 | |
2575 | return TLI.LowerF128Op(Op, DAG, |
2576 | LibFuncName: TLI.getLibcallName(Call: VT == MVT::i32 |
2577 | ? RTLIB::FPTOUINT_F128_I32 |
2578 | : RTLIB::FPTOUINT_F128_I64), |
2579 | numArgs: 1); |
2580 | } |
2581 | |
2582 | static SDValue LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG, |
2583 | const SparcTargetLowering &TLI, |
2584 | bool hasHardQuad) { |
2585 | EVT OpVT = Op.getOperand(i: 0).getValueType(); |
2586 | assert(OpVT == MVT::i32 || OpVT == MVT::i64); |
2587 | |
2588 | // Expand if it does not involve f128 or the target has support for |
2589 | // quad floating point instructions and the operand type is legal. |
2590 | if (Op.getValueType() != MVT::f128 || (hasHardQuad && TLI.isTypeLegal(VT: OpVT))) |
2591 | return SDValue(); |
2592 | |
2593 | return TLI.LowerF128Op(Op, DAG, |
2594 | LibFuncName: TLI.getLibcallName(Call: OpVT == MVT::i32 |
2595 | ? RTLIB::UINTTOFP_I32_F128 |
2596 | : RTLIB::UINTTOFP_I64_F128), |
2597 | numArgs: 1); |
2598 | } |
2599 | |
2600 | static SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG, |
2601 | const SparcTargetLowering &TLI, bool hasHardQuad, |
2602 | bool isV9, bool is64Bit) { |
2603 | SDValue Chain = Op.getOperand(i: 0); |
2604 | ISD::CondCode CC = cast<CondCodeSDNode>(Val: Op.getOperand(i: 1))->get(); |
2605 | SDValue LHS = Op.getOperand(i: 2); |
2606 | SDValue RHS = Op.getOperand(i: 3); |
2607 | SDValue Dest = Op.getOperand(i: 4); |
2608 | SDLoc dl(Op); |
2609 | unsigned Opc, SPCC = ~0U; |
2610 | |
2611 | // If this is a br_cc of a "setcc", and if the setcc got lowered into |
2612 | // an CMP[IF]CC/SELECT_[IF]CC pair, find the original compared values. |
2613 | LookThroughSetCC(LHS, RHS, CC, SPCC); |
2614 | assert(LHS.getValueType() == RHS.getValueType()); |
2615 | |
2616 | // Get the condition flag. |
2617 | SDValue CompareFlag; |
2618 | if (LHS.getValueType().isInteger()) { |
2619 | // On V9 processors running in 64-bit mode, if CC compares two `i64`s |
2620 | // and the RHS is zero we might be able to use a specialized branch. |
2621 | if (is64Bit && isV9 && LHS.getValueType() == MVT::i64 && |
2622 | isNullConstant(V: RHS) && !ISD::isUnsignedIntSetCC(Code: CC)) |
2623 | return DAG.getNode(Opcode: SPISD::BR_REG, DL: dl, VT: MVT::Other, N1: Chain, N2: Dest, |
2624 | N3: DAG.getConstant(Val: intCondCCodeToRcond(CC), DL: dl, VT: MVT::i32), |
2625 | N4: LHS); |
2626 | |
2627 | CompareFlag = DAG.getNode(Opcode: SPISD::CMPICC, DL: dl, VT: MVT::Glue, N1: LHS, N2: RHS); |
2628 | if (SPCC == ~0U) SPCC = IntCondCCodeToICC(CC); |
2629 | if (isV9) |
2630 | // 32-bit compares use the icc flags, 64-bit uses the xcc flags. |
2631 | Opc = LHS.getValueType() == MVT::i32 ? SPISD::BPICC : SPISD::BPXCC; |
2632 | else |
2633 | // Non-v9 targets don't have xcc. |
2634 | Opc = SPISD::BRICC; |
2635 | } else { |
2636 | if (!hasHardQuad && LHS.getValueType() == MVT::f128) { |
2637 | if (SPCC == ~0U) SPCC = FPCondCCodeToFCC(CC); |
2638 | CompareFlag = TLI.LowerF128Compare(LHS, RHS, SPCC, DL: dl, DAG); |
2639 | Opc = isV9 ? SPISD::BPICC : SPISD::BRICC; |
2640 | } else { |
2641 | unsigned CmpOpc = isV9 ? SPISD::CMPFCC_V9 : SPISD::CMPFCC; |
2642 | CompareFlag = DAG.getNode(Opcode: CmpOpc, DL: dl, VT: MVT::Glue, N1: LHS, N2: RHS); |
2643 | if (SPCC == ~0U) SPCC = FPCondCCodeToFCC(CC); |
2644 | Opc = isV9 ? SPISD::BRFCC_V9 : SPISD::BRFCC; |
2645 | } |
2646 | } |
2647 | return DAG.getNode(Opcode: Opc, DL: dl, VT: MVT::Other, N1: Chain, N2: Dest, |
2648 | N3: DAG.getConstant(Val: SPCC, DL: dl, VT: MVT::i32), N4: CompareFlag); |
2649 | } |
2650 | |
2651 | static SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG, |
2652 | const SparcTargetLowering &TLI, bool hasHardQuad, |
2653 | bool isV9, bool is64Bit) { |
2654 | SDValue LHS = Op.getOperand(i: 0); |
2655 | SDValue RHS = Op.getOperand(i: 1); |
2656 | ISD::CondCode CC = cast<CondCodeSDNode>(Val: Op.getOperand(i: 4))->get(); |
2657 | SDValue TrueVal = Op.getOperand(i: 2); |
2658 | SDValue FalseVal = Op.getOperand(i: 3); |
2659 | SDLoc dl(Op); |
2660 | unsigned Opc, SPCC = ~0U; |
2661 | |
2662 | // If this is a select_cc of a "setcc", and if the setcc got lowered into |
2663 | // an CMP[IF]CC/SELECT_[IF]CC pair, find the original compared values. |
2664 | LookThroughSetCC(LHS, RHS, CC, SPCC); |
2665 | assert(LHS.getValueType() == RHS.getValueType()); |
2666 | |
2667 | SDValue CompareFlag; |
2668 | if (LHS.getValueType().isInteger()) { |
2669 | // On V9 processors running in 64-bit mode, if CC compares two `i64`s |
2670 | // and the RHS is zero we might be able to use a specialized select. |
2671 | // All SELECT_CC between any two scalar integer types are eligible for |
2672 | // lowering to specialized instructions. Additionally, f32 and f64 types |
2673 | // are also eligible, but for f128 we can only use the specialized |
2674 | // instruction when we have hardquad. |
2675 | EVT ValType = TrueVal.getValueType(); |
2676 | bool IsEligibleType = ValType.isScalarInteger() || ValType == MVT::f32 || |
2677 | ValType == MVT::f64 || |
2678 | (ValType == MVT::f128 && hasHardQuad); |
2679 | if (is64Bit && isV9 && LHS.getValueType() == MVT::i64 && |
2680 | isNullConstant(V: RHS) && !ISD::isUnsignedIntSetCC(Code: CC) && IsEligibleType) |
2681 | return DAG.getNode( |
2682 | Opcode: SPISD::SELECT_REG, DL: dl, VT: TrueVal.getValueType(), N1: TrueVal, N2: FalseVal, |
2683 | N3: DAG.getConstant(Val: intCondCCodeToRcond(CC), DL: dl, VT: MVT::i32), N4: LHS); |
2684 | |
2685 | CompareFlag = DAG.getNode(Opcode: SPISD::CMPICC, DL: dl, VT: MVT::Glue, N1: LHS, N2: RHS); |
2686 | Opc = LHS.getValueType() == MVT::i32 ? |
2687 | SPISD::SELECT_ICC : SPISD::SELECT_XCC; |
2688 | if (SPCC == ~0U) SPCC = IntCondCCodeToICC(CC); |
2689 | } else { |
2690 | if (!hasHardQuad && LHS.getValueType() == MVT::f128) { |
2691 | if (SPCC == ~0U) SPCC = FPCondCCodeToFCC(CC); |
2692 | CompareFlag = TLI.LowerF128Compare(LHS, RHS, SPCC, DL: dl, DAG); |
2693 | Opc = SPISD::SELECT_ICC; |
2694 | } else { |
2695 | unsigned CmpOpc = isV9 ? SPISD::CMPFCC_V9 : SPISD::CMPFCC; |
2696 | CompareFlag = DAG.getNode(Opcode: CmpOpc, DL: dl, VT: MVT::Glue, N1: LHS, N2: RHS); |
2697 | Opc = SPISD::SELECT_FCC; |
2698 | if (SPCC == ~0U) SPCC = FPCondCCodeToFCC(CC); |
2699 | } |
2700 | } |
2701 | return DAG.getNode(Opcode: Opc, DL: dl, VT: TrueVal.getValueType(), N1: TrueVal, N2: FalseVal, |
2702 | N3: DAG.getConstant(Val: SPCC, DL: dl, VT: MVT::i32), N4: CompareFlag); |
2703 | } |
2704 | |
2705 | static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG, |
2706 | const SparcTargetLowering &TLI) { |
2707 | MachineFunction &MF = DAG.getMachineFunction(); |
2708 | SparcMachineFunctionInfo *FuncInfo = MF.getInfo<SparcMachineFunctionInfo>(); |
2709 | auto PtrVT = TLI.getPointerTy(DL: DAG.getDataLayout()); |
2710 | |
2711 | // Need frame address to find the address of VarArgsFrameIndex. |
2712 | MF.getFrameInfo().setFrameAddressIsTaken(true); |
2713 | |
2714 | // vastart just stores the address of the VarArgsFrameIndex slot into the |
2715 | // memory location argument. |
2716 | SDLoc DL(Op); |
2717 | SDValue Offset = |
2718 | DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrVT, N1: DAG.getRegister(Reg: SP::I6, VT: PtrVT), |
2719 | N2: DAG.getIntPtrConstant(Val: FuncInfo->getVarArgsFrameOffset(), DL)); |
2720 | const Value *SV = cast<SrcValueSDNode>(Val: Op.getOperand(i: 2))->getValue(); |
2721 | return DAG.getStore(Chain: Op.getOperand(i: 0), dl: DL, Val: Offset, Ptr: Op.getOperand(i: 1), |
2722 | PtrInfo: MachinePointerInfo(SV)); |
2723 | } |
2724 | |
2725 | static SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) { |
2726 | SDNode *Node = Op.getNode(); |
2727 | EVT VT = Node->getValueType(ResNo: 0); |
2728 | SDValue InChain = Node->getOperand(Num: 0); |
2729 | SDValue VAListPtr = Node->getOperand(Num: 1); |
2730 | EVT PtrVT = VAListPtr.getValueType(); |
2731 | const Value *SV = cast<SrcValueSDNode>(Val: Node->getOperand(Num: 2))->getValue(); |
2732 | SDLoc DL(Node); |
2733 | SDValue VAList = |
2734 | DAG.getLoad(VT: PtrVT, dl: DL, Chain: InChain, Ptr: VAListPtr, PtrInfo: MachinePointerInfo(SV)); |
2735 | // Increment the pointer, VAList, to the next vaarg. |
2736 | SDValue NextPtr = DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrVT, N1: VAList, |
2737 | N2: DAG.getIntPtrConstant(Val: VT.getSizeInBits()/8, |
2738 | DL)); |
2739 | // Store the incremented VAList to the legalized pointer. |
2740 | InChain = DAG.getStore(Chain: VAList.getValue(R: 1), dl: DL, Val: NextPtr, Ptr: VAListPtr, |
2741 | PtrInfo: MachinePointerInfo(SV)); |
2742 | // Load the actual argument out of the pointer VAList. |
2743 | // We can't count on greater alignment than the word size. |
2744 | return DAG.getLoad( |
2745 | VT, dl: DL, Chain: InChain, Ptr: VAList, PtrInfo: MachinePointerInfo(), |
2746 | Alignment: Align(std::min(a: PtrVT.getFixedSizeInBits(), b: VT.getFixedSizeInBits()) / 8)); |
2747 | } |
2748 | |
2749 | static SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG, |
2750 | const SparcSubtarget *Subtarget) { |
2751 | SDValue Chain = Op.getOperand(i: 0); |
2752 | SDValue Size = Op.getOperand(i: 1); |
2753 | SDValue Alignment = Op.getOperand(i: 2); |
2754 | MaybeAlign MaybeAlignment = |
2755 | cast<ConstantSDNode>(Val&: Alignment)->getMaybeAlignValue(); |
2756 | EVT VT = Size->getValueType(ResNo: 0); |
2757 | SDLoc dl(Op); |
2758 | |
2759 | unsigned SPReg = SP::O6; |
2760 | SDValue SP = DAG.getCopyFromReg(Chain, dl, Reg: SPReg, VT); |
2761 | |
2762 | // The resultant pointer needs to be above the register spill area |
2763 | // at the bottom of the stack. |
2764 | unsigned regSpillArea; |
2765 | if (Subtarget->is64Bit()) { |
2766 | regSpillArea = 128; |
2767 | } else { |
2768 | // On Sparc32, the size of the spill area is 92. Unfortunately, |
2769 | // that's only 4-byte aligned, not 8-byte aligned (the stack |
2770 | // pointer is 8-byte aligned). So, if the user asked for an 8-byte |
2771 | // aligned dynamic allocation, we actually need to add 96 to the |
2772 | // bottom of the stack, instead of 92, to ensure 8-byte alignment. |
2773 | |
2774 | // That also means adding 4 to the size of the allocation -- |
2775 | // before applying the 8-byte rounding. Unfortunately, we the |
2776 | // value we get here has already had rounding applied. So, we need |
2777 | // to add 8, instead, wasting a bit more memory. |
2778 | |
2779 | // Further, this only actually needs to be done if the required |
2780 | // alignment is > 4, but, we've lost that info by this point, too, |
2781 | // so we always apply it. |
2782 | |
2783 | // (An alternative approach would be to always reserve 96 bytes |
2784 | // instead of the required 92, but then we'd waste 4 extra bytes |
2785 | // in every frame, not just those with dynamic stack allocations) |
2786 | |
2787 | // TODO: modify code in SelectionDAGBuilder to make this less sad. |
2788 | |
2789 | Size = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT, N1: Size, |
2790 | N2: DAG.getConstant(Val: 8, DL: dl, VT)); |
2791 | regSpillArea = 96; |
2792 | } |
2793 | |
2794 | int64_t Bias = Subtarget->getStackPointerBias(); |
2795 | |
2796 | // Debias and increment SP past the reserved spill area. |
2797 | // We need the SP to point to the first usable region before calculating |
2798 | // anything to prevent any of the pointers from becoming out of alignment when |
2799 | // we rebias the SP later on. |
2800 | SDValue StartOfUsableStack = DAG.getNode( |
2801 | Opcode: ISD::ADD, DL: dl, VT, N1: SP, N2: DAG.getConstant(Val: regSpillArea + Bias, DL: dl, VT)); |
2802 | SDValue AllocatedPtr = |
2803 | DAG.getNode(Opcode: ISD::SUB, DL: dl, VT, N1: StartOfUsableStack, N2: Size); |
2804 | |
2805 | bool IsOveraligned = MaybeAlignment.has_value(); |
2806 | SDValue AlignedPtr = |
2807 | IsOveraligned |
2808 | ? DAG.getNode(Opcode: ISD::AND, DL: dl, VT, N1: AllocatedPtr, |
2809 | N2: DAG.getSignedConstant(Val: -MaybeAlignment->value(), DL: dl, VT)) |
2810 | : AllocatedPtr; |
2811 | |
2812 | // Now that we are done, restore the bias and reserved spill area. |
2813 | SDValue NewSP = DAG.getNode(Opcode: ISD::SUB, DL: dl, VT, N1: AlignedPtr, |
2814 | N2: DAG.getConstant(Val: regSpillArea + Bias, DL: dl, VT)); |
2815 | Chain = DAG.getCopyToReg(Chain: SP.getValue(R: 1), dl, Reg: SPReg, N: NewSP); |
2816 | SDValue Ops[2] = {AlignedPtr, Chain}; |
2817 | return DAG.getMergeValues(Ops, dl); |
2818 | } |
2819 | |
2820 | |
2821 | static SDValue getFLUSHW(SDValue Op, SelectionDAG &DAG) { |
2822 | SDLoc dl(Op); |
2823 | SDValue Chain = DAG.getNode(Opcode: SPISD::FLUSHW, |
2824 | DL: dl, VT: MVT::Other, Operand: DAG.getEntryNode()); |
2825 | return Chain; |
2826 | } |
2827 | |
2828 | static SDValue getFRAMEADDR(uint64_t depth, SDValue Op, SelectionDAG &DAG, |
2829 | const SparcSubtarget *Subtarget, |
2830 | bool AlwaysFlush = false) { |
2831 | MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); |
2832 | MFI.setFrameAddressIsTaken(true); |
2833 | |
2834 | EVT VT = Op.getValueType(); |
2835 | SDLoc dl(Op); |
2836 | unsigned FrameReg = SP::I6; |
2837 | unsigned stackBias = Subtarget->getStackPointerBias(); |
2838 | |
2839 | SDValue FrameAddr; |
2840 | SDValue Chain; |
2841 | |
2842 | // flush first to make sure the windowed registers' values are in stack |
2843 | Chain = (depth || AlwaysFlush) ? getFLUSHW(Op, DAG) : DAG.getEntryNode(); |
2844 | |
2845 | FrameAddr = DAG.getCopyFromReg(Chain, dl, Reg: FrameReg, VT); |
2846 | |
2847 | unsigned Offset = (Subtarget->is64Bit()) ? (stackBias + 112) : 56; |
2848 | |
2849 | while (depth--) { |
2850 | SDValue Ptr = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT, N1: FrameAddr, |
2851 | N2: DAG.getIntPtrConstant(Val: Offset, DL: dl)); |
2852 | FrameAddr = DAG.getLoad(VT, dl, Chain, Ptr, PtrInfo: MachinePointerInfo()); |
2853 | } |
2854 | if (Subtarget->is64Bit()) |
2855 | FrameAddr = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT, N1: FrameAddr, |
2856 | N2: DAG.getIntPtrConstant(Val: stackBias, DL: dl)); |
2857 | return FrameAddr; |
2858 | } |
2859 | |
2860 | |
2861 | static SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG, |
2862 | const SparcSubtarget *Subtarget) { |
2863 | |
2864 | uint64_t depth = Op.getConstantOperandVal(i: 0); |
2865 | |
2866 | return getFRAMEADDR(depth, Op, DAG, Subtarget); |
2867 | |
2868 | } |
2869 | |
2870 | static SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG, |
2871 | const SparcTargetLowering &TLI, |
2872 | const SparcSubtarget *Subtarget) { |
2873 | MachineFunction &MF = DAG.getMachineFunction(); |
2874 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
2875 | MFI.setReturnAddressIsTaken(true); |
2876 | |
2877 | if (TLI.verifyReturnAddressArgumentIsConstant(Op, DAG)) |
2878 | return SDValue(); |
2879 | |
2880 | EVT VT = Op.getValueType(); |
2881 | SDLoc dl(Op); |
2882 | uint64_t depth = Op.getConstantOperandVal(i: 0); |
2883 | |
2884 | SDValue RetAddr; |
2885 | if (depth == 0) { |
2886 | auto PtrVT = TLI.getPointerTy(DL: DAG.getDataLayout()); |
2887 | Register RetReg = MF.addLiveIn(PReg: SP::I7, RC: TLI.getRegClassFor(VT: PtrVT)); |
2888 | RetAddr = DAG.getCopyFromReg(Chain: DAG.getEntryNode(), dl, Reg: RetReg, VT); |
2889 | return RetAddr; |
2890 | } |
2891 | |
2892 | // Need frame address to find return address of the caller. |
2893 | SDValue FrameAddr = getFRAMEADDR(depth: depth - 1, Op, DAG, Subtarget, AlwaysFlush: true); |
2894 | |
2895 | unsigned Offset = (Subtarget->is64Bit()) ? 120 : 60; |
2896 | SDValue Ptr = DAG.getNode(Opcode: ISD::ADD, |
2897 | DL: dl, VT, |
2898 | N1: FrameAddr, |
2899 | N2: DAG.getIntPtrConstant(Val: Offset, DL: dl)); |
2900 | RetAddr = DAG.getLoad(VT, dl, Chain: DAG.getEntryNode(), Ptr, PtrInfo: MachinePointerInfo()); |
2901 | |
2902 | return RetAddr; |
2903 | } |
2904 | |
2905 | static SDValue LowerF64Op(SDValue SrcReg64, const SDLoc &dl, SelectionDAG &DAG, |
2906 | unsigned opcode) { |
2907 | assert(SrcReg64.getValueType() == MVT::f64 && "LowerF64Op called on non-double!" ); |
2908 | assert(opcode == ISD::FNEG || opcode == ISD::FABS); |
2909 | |
2910 | // Lower fneg/fabs on f64 to fneg/fabs on f32. |
2911 | // fneg f64 => fneg f32:sub_even, fmov f32:sub_odd. |
2912 | // fabs f64 => fabs f32:sub_even, fmov f32:sub_odd. |
2913 | |
2914 | // Note: in little-endian, the floating-point value is stored in the |
2915 | // registers are in the opposite order, so the subreg with the sign |
2916 | // bit is the highest-numbered (odd), rather than the |
2917 | // lowest-numbered (even). |
2918 | |
2919 | SDValue Hi32 = DAG.getTargetExtractSubreg(SRIdx: SP::sub_even, DL: dl, VT: MVT::f32, |
2920 | Operand: SrcReg64); |
2921 | SDValue Lo32 = DAG.getTargetExtractSubreg(SRIdx: SP::sub_odd, DL: dl, VT: MVT::f32, |
2922 | Operand: SrcReg64); |
2923 | |
2924 | if (DAG.getDataLayout().isLittleEndian()) |
2925 | Lo32 = DAG.getNode(Opcode: opcode, DL: dl, VT: MVT::f32, Operand: Lo32); |
2926 | else |
2927 | Hi32 = DAG.getNode(Opcode: opcode, DL: dl, VT: MVT::f32, Operand: Hi32); |
2928 | |
2929 | SDValue DstReg64 = SDValue(DAG.getMachineNode(Opcode: TargetOpcode::IMPLICIT_DEF, |
2930 | dl, VT: MVT::f64), 0); |
2931 | DstReg64 = DAG.getTargetInsertSubreg(SRIdx: SP::sub_even, DL: dl, VT: MVT::f64, |
2932 | Operand: DstReg64, Subreg: Hi32); |
2933 | DstReg64 = DAG.getTargetInsertSubreg(SRIdx: SP::sub_odd, DL: dl, VT: MVT::f64, |
2934 | Operand: DstReg64, Subreg: Lo32); |
2935 | return DstReg64; |
2936 | } |
2937 | |
2938 | // Lower a f128 load into two f64 loads. |
2939 | static SDValue LowerF128Load(SDValue Op, SelectionDAG &DAG) |
2940 | { |
2941 | SDLoc dl(Op); |
2942 | LoadSDNode *LdNode = cast<LoadSDNode>(Val: Op.getNode()); |
2943 | assert(LdNode->getOffset().isUndef() && "Unexpected node type" ); |
2944 | |
2945 | Align Alignment = commonAlignment(A: LdNode->getBaseAlign(), Offset: 8); |
2946 | |
2947 | SDValue Hi64 = |
2948 | DAG.getLoad(VT: MVT::f64, dl, Chain: LdNode->getChain(), Ptr: LdNode->getBasePtr(), |
2949 | PtrInfo: LdNode->getPointerInfo(), Alignment); |
2950 | EVT addrVT = LdNode->getBasePtr().getValueType(); |
2951 | SDValue LoPtr = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT: addrVT, |
2952 | N1: LdNode->getBasePtr(), |
2953 | N2: DAG.getConstant(Val: 8, DL: dl, VT: addrVT)); |
2954 | SDValue Lo64 = DAG.getLoad(VT: MVT::f64, dl, Chain: LdNode->getChain(), Ptr: LoPtr, |
2955 | PtrInfo: LdNode->getPointerInfo().getWithOffset(O: 8), |
2956 | Alignment); |
2957 | |
2958 | SDValue SubRegEven = DAG.getTargetConstant(Val: SP::sub_even64, DL: dl, VT: MVT::i32); |
2959 | SDValue SubRegOdd = DAG.getTargetConstant(Val: SP::sub_odd64, DL: dl, VT: MVT::i32); |
2960 | |
2961 | SDNode *InFP128 = DAG.getMachineNode(Opcode: TargetOpcode::IMPLICIT_DEF, |
2962 | dl, VT: MVT::f128); |
2963 | InFP128 = DAG.getMachineNode(Opcode: TargetOpcode::INSERT_SUBREG, dl, |
2964 | VT: MVT::f128, |
2965 | Op1: SDValue(InFP128, 0), |
2966 | Op2: Hi64, |
2967 | Op3: SubRegEven); |
2968 | InFP128 = DAG.getMachineNode(Opcode: TargetOpcode::INSERT_SUBREG, dl, |
2969 | VT: MVT::f128, |
2970 | Op1: SDValue(InFP128, 0), |
2971 | Op2: Lo64, |
2972 | Op3: SubRegOdd); |
2973 | SDValue OutChains[2] = { SDValue(Hi64.getNode(), 1), |
2974 | SDValue(Lo64.getNode(), 1) }; |
2975 | SDValue OutChain = DAG.getNode(Opcode: ISD::TokenFactor, DL: dl, VT: MVT::Other, Ops: OutChains); |
2976 | SDValue Ops[2] = {SDValue(InFP128,0), OutChain}; |
2977 | return DAG.getMergeValues(Ops, dl); |
2978 | } |
2979 | |
2980 | static SDValue LowerLOAD(SDValue Op, SelectionDAG &DAG) |
2981 | { |
2982 | LoadSDNode *LdNode = cast<LoadSDNode>(Val: Op.getNode()); |
2983 | |
2984 | EVT MemVT = LdNode->getMemoryVT(); |
2985 | if (MemVT == MVT::f128) |
2986 | return LowerF128Load(Op, DAG); |
2987 | |
2988 | return Op; |
2989 | } |
2990 | |
2991 | // Lower a f128 store into two f64 stores. |
2992 | static SDValue LowerF128Store(SDValue Op, SelectionDAG &DAG) { |
2993 | SDLoc dl(Op); |
2994 | StoreSDNode *StNode = cast<StoreSDNode>(Val: Op.getNode()); |
2995 | assert(StNode->getOffset().isUndef() && "Unexpected node type" ); |
2996 | |
2997 | SDValue SubRegEven = DAG.getTargetConstant(Val: SP::sub_even64, DL: dl, VT: MVT::i32); |
2998 | SDValue SubRegOdd = DAG.getTargetConstant(Val: SP::sub_odd64, DL: dl, VT: MVT::i32); |
2999 | |
3000 | SDNode *Hi64 = DAG.getMachineNode(Opcode: TargetOpcode::EXTRACT_SUBREG, |
3001 | dl, |
3002 | VT: MVT::f64, |
3003 | Op1: StNode->getValue(), |
3004 | Op2: SubRegEven); |
3005 | SDNode *Lo64 = DAG.getMachineNode(Opcode: TargetOpcode::EXTRACT_SUBREG, |
3006 | dl, |
3007 | VT: MVT::f64, |
3008 | Op1: StNode->getValue(), |
3009 | Op2: SubRegOdd); |
3010 | |
3011 | Align Alignment = commonAlignment(A: StNode->getBaseAlign(), Offset: 8); |
3012 | |
3013 | SDValue OutChains[2]; |
3014 | OutChains[0] = |
3015 | DAG.getStore(Chain: StNode->getChain(), dl, Val: SDValue(Hi64, 0), |
3016 | Ptr: StNode->getBasePtr(), PtrInfo: StNode->getPointerInfo(), |
3017 | Alignment); |
3018 | EVT addrVT = StNode->getBasePtr().getValueType(); |
3019 | SDValue LoPtr = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT: addrVT, |
3020 | N1: StNode->getBasePtr(), |
3021 | N2: DAG.getConstant(Val: 8, DL: dl, VT: addrVT)); |
3022 | OutChains[1] = DAG.getStore(Chain: StNode->getChain(), dl, Val: SDValue(Lo64, 0), Ptr: LoPtr, |
3023 | PtrInfo: StNode->getPointerInfo().getWithOffset(O: 8), |
3024 | Alignment); |
3025 | return DAG.getNode(Opcode: ISD::TokenFactor, DL: dl, VT: MVT::Other, Ops: OutChains); |
3026 | } |
3027 | |
3028 | static SDValue LowerSTORE(SDValue Op, SelectionDAG &DAG) |
3029 | { |
3030 | SDLoc dl(Op); |
3031 | StoreSDNode *St = cast<StoreSDNode>(Val: Op.getNode()); |
3032 | |
3033 | EVT MemVT = St->getMemoryVT(); |
3034 | if (MemVT == MVT::f128) |
3035 | return LowerF128Store(Op, DAG); |
3036 | |
3037 | if (MemVT == MVT::i64) { |
3038 | // Custom handling for i64 stores: turn it into a bitcast and a |
3039 | // v2i32 store. |
3040 | SDValue Val = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: MVT::v2i32, Operand: St->getValue()); |
3041 | SDValue Chain = DAG.getStore( |
3042 | Chain: St->getChain(), dl, Val, Ptr: St->getBasePtr(), PtrInfo: St->getPointerInfo(), |
3043 | Alignment: St->getBaseAlign(), MMOFlags: St->getMemOperand()->getFlags(), AAInfo: St->getAAInfo()); |
3044 | return Chain; |
3045 | } |
3046 | |
3047 | return SDValue(); |
3048 | } |
3049 | |
3050 | static SDValue LowerFNEGorFABS(SDValue Op, SelectionDAG &DAG, bool isV9) { |
3051 | assert((Op.getOpcode() == ISD::FNEG || Op.getOpcode() == ISD::FABS) |
3052 | && "invalid opcode" ); |
3053 | |
3054 | SDLoc dl(Op); |
3055 | |
3056 | if (Op.getValueType() == MVT::f64) |
3057 | return LowerF64Op(SrcReg64: Op.getOperand(i: 0), dl, DAG, opcode: Op.getOpcode()); |
3058 | if (Op.getValueType() != MVT::f128) |
3059 | return Op; |
3060 | |
3061 | // Lower fabs/fneg on f128 to fabs/fneg on f64 |
3062 | // fabs/fneg f128 => fabs/fneg f64:sub_even64, fmov f64:sub_odd64 |
3063 | // (As with LowerF64Op, on little-endian, we need to negate the odd |
3064 | // subreg) |
3065 | |
3066 | SDValue SrcReg128 = Op.getOperand(i: 0); |
3067 | SDValue Hi64 = DAG.getTargetExtractSubreg(SRIdx: SP::sub_even64, DL: dl, VT: MVT::f64, |
3068 | Operand: SrcReg128); |
3069 | SDValue Lo64 = DAG.getTargetExtractSubreg(SRIdx: SP::sub_odd64, DL: dl, VT: MVT::f64, |
3070 | Operand: SrcReg128); |
3071 | |
3072 | if (DAG.getDataLayout().isLittleEndian()) { |
3073 | if (isV9) |
3074 | Lo64 = DAG.getNode(Opcode: Op.getOpcode(), DL: dl, VT: MVT::f64, Operand: Lo64); |
3075 | else |
3076 | Lo64 = LowerF64Op(SrcReg64: Lo64, dl, DAG, opcode: Op.getOpcode()); |
3077 | } else { |
3078 | if (isV9) |
3079 | Hi64 = DAG.getNode(Opcode: Op.getOpcode(), DL: dl, VT: MVT::f64, Operand: Hi64); |
3080 | else |
3081 | Hi64 = LowerF64Op(SrcReg64: Hi64, dl, DAG, opcode: Op.getOpcode()); |
3082 | } |
3083 | |
3084 | SDValue DstReg128 = SDValue(DAG.getMachineNode(Opcode: TargetOpcode::IMPLICIT_DEF, |
3085 | dl, VT: MVT::f128), 0); |
3086 | DstReg128 = DAG.getTargetInsertSubreg(SRIdx: SP::sub_even64, DL: dl, VT: MVT::f128, |
3087 | Operand: DstReg128, Subreg: Hi64); |
3088 | DstReg128 = DAG.getTargetInsertSubreg(SRIdx: SP::sub_odd64, DL: dl, VT: MVT::f128, |
3089 | Operand: DstReg128, Subreg: Lo64); |
3090 | return DstReg128; |
3091 | } |
3092 | |
3093 | static SDValue LowerATOMIC_LOAD_STORE(SDValue Op, SelectionDAG &DAG) { |
3094 | if (isStrongerThanMonotonic(AO: cast<AtomicSDNode>(Val&: Op)->getSuccessOrdering())) { |
3095 | // Expand with a fence. |
3096 | return SDValue(); |
3097 | } |
3098 | |
3099 | // Monotonic load/stores are legal. |
3100 | return Op; |
3101 | } |
3102 | |
3103 | SDValue SparcTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, |
3104 | SelectionDAG &DAG) const { |
3105 | unsigned IntNo = Op.getConstantOperandVal(i: 0); |
3106 | switch (IntNo) { |
3107 | default: return SDValue(); // Don't custom lower most intrinsics. |
3108 | case Intrinsic::thread_pointer: { |
3109 | EVT PtrVT = getPointerTy(DL: DAG.getDataLayout()); |
3110 | return DAG.getRegister(Reg: SP::G7, VT: PtrVT); |
3111 | } |
3112 | } |
3113 | } |
3114 | |
3115 | SDValue SparcTargetLowering:: |
3116 | LowerOperation(SDValue Op, SelectionDAG &DAG) const { |
3117 | |
3118 | bool hasHardQuad = Subtarget->hasHardQuad(); |
3119 | bool isV9 = Subtarget->isV9(); |
3120 | bool is64Bit = Subtarget->is64Bit(); |
3121 | |
3122 | switch (Op.getOpcode()) { |
3123 | default: llvm_unreachable("Should not custom lower this!" ); |
3124 | |
3125 | case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG, TLI: *this, |
3126 | Subtarget); |
3127 | case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG, |
3128 | Subtarget); |
3129 | case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG); |
3130 | case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG); |
3131 | case ISD::BlockAddress: return LowerBlockAddress(Op, DAG); |
3132 | case ISD::ConstantPool: return LowerConstantPool(Op, DAG); |
3133 | case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG, TLI: *this, |
3134 | hasHardQuad); |
3135 | case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG, TLI: *this, |
3136 | hasHardQuad); |
3137 | case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG, TLI: *this, |
3138 | hasHardQuad); |
3139 | case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG, TLI: *this, |
3140 | hasHardQuad); |
3141 | case ISD::BR_CC: |
3142 | return LowerBR_CC(Op, DAG, TLI: *this, hasHardQuad, isV9, is64Bit); |
3143 | case ISD::SELECT_CC: |
3144 | return LowerSELECT_CC(Op, DAG, TLI: *this, hasHardQuad, isV9, is64Bit); |
3145 | case ISD::VASTART: return LowerVASTART(Op, DAG, TLI: *this); |
3146 | case ISD::VAARG: return LowerVAARG(Op, DAG); |
3147 | case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG, |
3148 | Subtarget); |
3149 | |
3150 | case ISD::LOAD: return LowerLOAD(Op, DAG); |
3151 | case ISD::STORE: return LowerSTORE(Op, DAG); |
3152 | case ISD::FADD: return LowerF128Op(Op, DAG, |
3153 | LibFuncName: getLibcallName(Call: RTLIB::ADD_F128), numArgs: 2); |
3154 | case ISD::FSUB: return LowerF128Op(Op, DAG, |
3155 | LibFuncName: getLibcallName(Call: RTLIB::SUB_F128), numArgs: 2); |
3156 | case ISD::FMUL: return LowerF128Op(Op, DAG, |
3157 | LibFuncName: getLibcallName(Call: RTLIB::MUL_F128), numArgs: 2); |
3158 | case ISD::FDIV: return LowerF128Op(Op, DAG, |
3159 | LibFuncName: getLibcallName(Call: RTLIB::DIV_F128), numArgs: 2); |
3160 | case ISD::FSQRT: return LowerF128Op(Op, DAG, |
3161 | LibFuncName: getLibcallName(Call: RTLIB::SQRT_F128),numArgs: 1); |
3162 | case ISD::FABS: |
3163 | case ISD::FNEG: return LowerFNEGorFABS(Op, DAG, isV9); |
3164 | case ISD::FP_EXTEND: return LowerF128_FPEXTEND(Op, DAG, TLI: *this); |
3165 | case ISD::FP_ROUND: return LowerF128_FPROUND(Op, DAG, TLI: *this); |
3166 | case ISD::ATOMIC_LOAD: |
3167 | case ISD::ATOMIC_STORE: return LowerATOMIC_LOAD_STORE(Op, DAG); |
3168 | case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG); |
3169 | } |
3170 | } |
3171 | |
3172 | SDValue SparcTargetLowering::bitcastConstantFPToInt(ConstantFPSDNode *C, |
3173 | const SDLoc &DL, |
3174 | SelectionDAG &DAG) const { |
3175 | APInt V = C->getValueAPF().bitcastToAPInt(); |
3176 | SDValue Lo = DAG.getConstant(Val: V.zextOrTrunc(width: 32), DL, VT: MVT::i32); |
3177 | SDValue Hi = DAG.getConstant(Val: V.lshr(shiftAmt: 32).zextOrTrunc(width: 32), DL, VT: MVT::i32); |
3178 | if (DAG.getDataLayout().isLittleEndian()) |
3179 | std::swap(a&: Lo, b&: Hi); |
3180 | return DAG.getBuildVector(VT: MVT::v2i32, DL, Ops: {Hi, Lo}); |
3181 | } |
3182 | |
3183 | SDValue SparcTargetLowering::PerformBITCASTCombine(SDNode *N, |
3184 | DAGCombinerInfo &DCI) const { |
3185 | SDLoc dl(N); |
3186 | SDValue Src = N->getOperand(Num: 0); |
3187 | |
3188 | if (isa<ConstantFPSDNode>(Val: Src) && N->getSimpleValueType(ResNo: 0) == MVT::v2i32 && |
3189 | Src.getSimpleValueType() == MVT::f64) |
3190 | return bitcastConstantFPToInt(C: cast<ConstantFPSDNode>(Val&: Src), DL: dl, DAG&: DCI.DAG); |
3191 | |
3192 | return SDValue(); |
3193 | } |
3194 | |
3195 | SDValue SparcTargetLowering::PerformDAGCombine(SDNode *N, |
3196 | DAGCombinerInfo &DCI) const { |
3197 | switch (N->getOpcode()) { |
3198 | default: |
3199 | break; |
3200 | case ISD::BITCAST: |
3201 | return PerformBITCASTCombine(N, DCI); |
3202 | } |
3203 | return SDValue(); |
3204 | } |
3205 | |
3206 | MachineBasicBlock * |
3207 | SparcTargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI, |
3208 | MachineBasicBlock *BB) const { |
3209 | switch (MI.getOpcode()) { |
3210 | default: llvm_unreachable("Unknown SELECT_CC!" ); |
3211 | case SP::SELECT_CC_Int_ICC: |
3212 | case SP::SELECT_CC_FP_ICC: |
3213 | case SP::SELECT_CC_DFP_ICC: |
3214 | case SP::SELECT_CC_QFP_ICC: |
3215 | if (Subtarget->isV9()) |
3216 | return expandSelectCC(MI, BB, BROpcode: SP::BPICC); |
3217 | return expandSelectCC(MI, BB, BROpcode: SP::BCOND); |
3218 | case SP::SELECT_CC_Int_XCC: |
3219 | case SP::SELECT_CC_FP_XCC: |
3220 | case SP::SELECT_CC_DFP_XCC: |
3221 | case SP::SELECT_CC_QFP_XCC: |
3222 | return expandSelectCC(MI, BB, BROpcode: SP::BPXCC); |
3223 | case SP::SELECT_CC_Int_FCC: |
3224 | case SP::SELECT_CC_FP_FCC: |
3225 | case SP::SELECT_CC_DFP_FCC: |
3226 | case SP::SELECT_CC_QFP_FCC: |
3227 | if (Subtarget->isV9()) |
3228 | return expandSelectCC(MI, BB, BROpcode: SP::FBCOND_V9); |
3229 | return expandSelectCC(MI, BB, BROpcode: SP::FBCOND); |
3230 | } |
3231 | } |
3232 | |
3233 | MachineBasicBlock * |
3234 | SparcTargetLowering::expandSelectCC(MachineInstr &MI, MachineBasicBlock *BB, |
3235 | unsigned BROpcode) const { |
3236 | const TargetInstrInfo &TII = *Subtarget->getInstrInfo(); |
3237 | DebugLoc dl = MI.getDebugLoc(); |
3238 | unsigned CC = (SPCC::CondCodes)MI.getOperand(i: 3).getImm(); |
3239 | |
3240 | // To "insert" a SELECT_CC instruction, we actually have to insert the |
3241 | // triangle control-flow pattern. The incoming instruction knows the |
3242 | // destination vreg to set, the condition code register to branch on, the |
3243 | // true/false values to select between, and the condition code for the branch. |
3244 | // |
3245 | // We produce the following control flow: |
3246 | // ThisMBB |
3247 | // | \ |
3248 | // | IfFalseMBB |
3249 | // | / |
3250 | // SinkMBB |
3251 | const BasicBlock *LLVM_BB = BB->getBasicBlock(); |
3252 | MachineFunction::iterator It = ++BB->getIterator(); |
3253 | |
3254 | MachineBasicBlock *ThisMBB = BB; |
3255 | MachineFunction *F = BB->getParent(); |
3256 | MachineBasicBlock *IfFalseMBB = F->CreateMachineBasicBlock(BB: LLVM_BB); |
3257 | MachineBasicBlock *SinkMBB = F->CreateMachineBasicBlock(BB: LLVM_BB); |
3258 | F->insert(MBBI: It, MBB: IfFalseMBB); |
3259 | F->insert(MBBI: It, MBB: SinkMBB); |
3260 | |
3261 | // Transfer the remainder of ThisMBB and its successor edges to SinkMBB. |
3262 | SinkMBB->splice(Where: SinkMBB->begin(), Other: ThisMBB, |
3263 | From: std::next(x: MachineBasicBlock::iterator(MI)), To: ThisMBB->end()); |
3264 | SinkMBB->transferSuccessorsAndUpdatePHIs(FromMBB: ThisMBB); |
3265 | |
3266 | // Set the new successors for ThisMBB. |
3267 | ThisMBB->addSuccessor(Succ: IfFalseMBB); |
3268 | ThisMBB->addSuccessor(Succ: SinkMBB); |
3269 | |
3270 | BuildMI(BB: ThisMBB, MIMD: dl, MCID: TII.get(Opcode: BROpcode)) |
3271 | .addMBB(MBB: SinkMBB) |
3272 | .addImm(Val: CC); |
3273 | |
3274 | // IfFalseMBB just falls through to SinkMBB. |
3275 | IfFalseMBB->addSuccessor(Succ: SinkMBB); |
3276 | |
3277 | // %Result = phi [ %TrueValue, ThisMBB ], [ %FalseValue, IfFalseMBB ] |
3278 | BuildMI(BB&: *SinkMBB, I: SinkMBB->begin(), MIMD: dl, MCID: TII.get(Opcode: SP::PHI), |
3279 | DestReg: MI.getOperand(i: 0).getReg()) |
3280 | .addReg(RegNo: MI.getOperand(i: 1).getReg()) |
3281 | .addMBB(MBB: ThisMBB) |
3282 | .addReg(RegNo: MI.getOperand(i: 2).getReg()) |
3283 | .addMBB(MBB: IfFalseMBB); |
3284 | |
3285 | MI.eraseFromParent(); // The pseudo instruction is gone now. |
3286 | return SinkMBB; |
3287 | } |
3288 | |
3289 | //===----------------------------------------------------------------------===// |
3290 | // Sparc Inline Assembly Support |
3291 | //===----------------------------------------------------------------------===// |
3292 | |
3293 | /// getConstraintType - Given a constraint letter, return the type of |
3294 | /// constraint it is for this target. |
3295 | SparcTargetLowering::ConstraintType |
3296 | SparcTargetLowering::getConstraintType(StringRef Constraint) const { |
3297 | if (Constraint.size() == 1) { |
3298 | switch (Constraint[0]) { |
3299 | default: break; |
3300 | case 'r': |
3301 | case 'f': |
3302 | case 'e': |
3303 | return C_RegisterClass; |
3304 | case 'I': // SIMM13 |
3305 | return C_Immediate; |
3306 | } |
3307 | } |
3308 | |
3309 | return TargetLowering::getConstraintType(Constraint); |
3310 | } |
3311 | |
3312 | TargetLowering::ConstraintWeight SparcTargetLowering:: |
3313 | getSingleConstraintMatchWeight(AsmOperandInfo &info, |
3314 | const char *constraint) const { |
3315 | ConstraintWeight weight = CW_Invalid; |
3316 | Value *CallOperandVal = info.CallOperandVal; |
3317 | // If we don't have a value, we can't do a match, |
3318 | // but allow it at the lowest weight. |
3319 | if (!CallOperandVal) |
3320 | return CW_Default; |
3321 | |
3322 | // Look at the constraint type. |
3323 | switch (*constraint) { |
3324 | default: |
3325 | weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint); |
3326 | break; |
3327 | case 'I': // SIMM13 |
3328 | if (ConstantInt *C = dyn_cast<ConstantInt>(Val: info.CallOperandVal)) { |
3329 | if (isInt<13>(x: C->getSExtValue())) |
3330 | weight = CW_Constant; |
3331 | } |
3332 | break; |
3333 | } |
3334 | return weight; |
3335 | } |
3336 | |
3337 | /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops |
3338 | /// vector. If it is invalid, don't add anything to Ops. |
3339 | void SparcTargetLowering::LowerAsmOperandForConstraint( |
3340 | SDValue Op, StringRef Constraint, std::vector<SDValue> &Ops, |
3341 | SelectionDAG &DAG) const { |
3342 | SDValue Result; |
3343 | |
3344 | // Only support length 1 constraints for now. |
3345 | if (Constraint.size() > 1) |
3346 | return; |
3347 | |
3348 | char ConstraintLetter = Constraint[0]; |
3349 | switch (ConstraintLetter) { |
3350 | default: break; |
3351 | case 'I': |
3352 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val&: Op)) { |
3353 | if (isInt<13>(x: C->getSExtValue())) { |
3354 | Result = DAG.getSignedTargetConstant(Val: C->getSExtValue(), DL: SDLoc(Op), |
3355 | VT: Op.getValueType()); |
3356 | break; |
3357 | } |
3358 | return; |
3359 | } |
3360 | } |
3361 | |
3362 | if (Result.getNode()) { |
3363 | Ops.push_back(x: Result); |
3364 | return; |
3365 | } |
3366 | TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG); |
3367 | } |
3368 | |
3369 | std::pair<unsigned, const TargetRegisterClass *> |
3370 | SparcTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI, |
3371 | StringRef Constraint, |
3372 | MVT VT) const { |
3373 | if (Constraint.empty()) |
3374 | return std::make_pair(x: 0U, y: nullptr); |
3375 | |
3376 | if (Constraint.size() == 1) { |
3377 | switch (Constraint[0]) { |
3378 | case 'r': |
3379 | if (VT == MVT::v2i32) |
3380 | return std::make_pair(x: 0U, y: &SP::IntPairRegClass); |
3381 | else if (Subtarget->is64Bit()) |
3382 | return std::make_pair(x: 0U, y: &SP::I64RegsRegClass); |
3383 | else |
3384 | return std::make_pair(x: 0U, y: &SP::IntRegsRegClass); |
3385 | case 'f': |
3386 | if (VT == MVT::f32 || VT == MVT::i32) |
3387 | return std::make_pair(x: 0U, y: &SP::FPRegsRegClass); |
3388 | else if (VT == MVT::f64 || VT == MVT::i64) |
3389 | return std::make_pair(x: 0U, y: &SP::LowDFPRegsRegClass); |
3390 | else if (VT == MVT::f128) |
3391 | return std::make_pair(x: 0U, y: &SP::LowQFPRegsRegClass); |
3392 | // This will generate an error message |
3393 | return std::make_pair(x: 0U, y: nullptr); |
3394 | case 'e': |
3395 | if (VT == MVT::f32 || VT == MVT::i32) |
3396 | return std::make_pair(x: 0U, y: &SP::FPRegsRegClass); |
3397 | else if (VT == MVT::f64 || VT == MVT::i64 ) |
3398 | return std::make_pair(x: 0U, y: &SP::DFPRegsRegClass); |
3399 | else if (VT == MVT::f128) |
3400 | return std::make_pair(x: 0U, y: &SP::QFPRegsRegClass); |
3401 | // This will generate an error message |
3402 | return std::make_pair(x: 0U, y: nullptr); |
3403 | } |
3404 | } |
3405 | |
3406 | if (Constraint.front() != '{') |
3407 | return std::make_pair(x: 0U, y: nullptr); |
3408 | |
3409 | assert(Constraint.back() == '}' && "Not a brace enclosed constraint?" ); |
3410 | StringRef RegName(Constraint.data() + 1, Constraint.size() - 2); |
3411 | if (RegName.empty()) |
3412 | return std::make_pair(x: 0U, y: nullptr); |
3413 | |
3414 | unsigned long long RegNo; |
3415 | // Handle numbered register aliases. |
3416 | if (RegName[0] == 'r' && |
3417 | getAsUnsignedInteger(Str: RegName.begin() + 1, Radix: 10, Result&: RegNo)) { |
3418 | // r0-r7 -> g0-g7 |
3419 | // r8-r15 -> o0-o7 |
3420 | // r16-r23 -> l0-l7 |
3421 | // r24-r31 -> i0-i7 |
3422 | if (RegNo > 31) |
3423 | return std::make_pair(x: 0U, y: nullptr); |
3424 | const char RegTypes[] = {'g', 'o', 'l', 'i'}; |
3425 | char RegType = RegTypes[RegNo / 8]; |
3426 | char RegIndex = '0' + (RegNo % 8); |
3427 | char Tmp[] = {'{', RegType, RegIndex, '}', 0}; |
3428 | return getRegForInlineAsmConstraint(TRI, Constraint: Tmp, VT); |
3429 | } |
3430 | |
3431 | // Rewrite the fN constraint according to the value type if needed. |
3432 | if (VT != MVT::f32 && VT != MVT::Other && RegName[0] == 'f' && |
3433 | getAsUnsignedInteger(Str: RegName.begin() + 1, Radix: 10, Result&: RegNo)) { |
3434 | if (VT == MVT::f64 && (RegNo % 2 == 0)) { |
3435 | return getRegForInlineAsmConstraint( |
3436 | TRI, Constraint: StringRef("{d" + utostr(X: RegNo / 2) + "}" ), VT); |
3437 | } else if (VT == MVT::f128 && (RegNo % 4 == 0)) { |
3438 | return getRegForInlineAsmConstraint( |
3439 | TRI, Constraint: StringRef("{q" + utostr(X: RegNo / 4) + "}" ), VT); |
3440 | } else { |
3441 | return std::make_pair(x: 0U, y: nullptr); |
3442 | } |
3443 | } |
3444 | |
3445 | auto ResultPair = |
3446 | TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT); |
3447 | if (!ResultPair.second) |
3448 | return std::make_pair(x: 0U, y: nullptr); |
3449 | |
3450 | // Force the use of I64Regs over IntRegs for 64-bit values. |
3451 | if (Subtarget->is64Bit() && VT == MVT::i64) { |
3452 | assert(ResultPair.second == &SP::IntRegsRegClass && |
3453 | "Unexpected register class" ); |
3454 | return std::make_pair(x&: ResultPair.first, y: &SP::I64RegsRegClass); |
3455 | } |
3456 | |
3457 | return ResultPair; |
3458 | } |
3459 | |
3460 | bool |
3461 | SparcTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const { |
3462 | // The Sparc target isn't yet aware of offsets. |
3463 | return false; |
3464 | } |
3465 | |
3466 | void SparcTargetLowering::ReplaceNodeResults(SDNode *N, |
3467 | SmallVectorImpl<SDValue>& Results, |
3468 | SelectionDAG &DAG) const { |
3469 | |
3470 | SDLoc dl(N); |
3471 | |
3472 | RTLIB::Libcall libCall = RTLIB::UNKNOWN_LIBCALL; |
3473 | |
3474 | switch (N->getOpcode()) { |
3475 | default: |
3476 | llvm_unreachable("Do not know how to custom type legalize this operation!" ); |
3477 | |
3478 | case ISD::FP_TO_SINT: |
3479 | case ISD::FP_TO_UINT: |
3480 | // Custom lower only if it involves f128 or i64. |
3481 | if (N->getOperand(Num: 0).getValueType() != MVT::f128 |
3482 | || N->getValueType(ResNo: 0) != MVT::i64) |
3483 | return; |
3484 | libCall = ((N->getOpcode() == ISD::FP_TO_SINT) |
3485 | ? RTLIB::FPTOSINT_F128_I64 |
3486 | : RTLIB::FPTOUINT_F128_I64); |
3487 | |
3488 | Results.push_back(Elt: LowerF128Op(Op: SDValue(N, 0), |
3489 | DAG, |
3490 | LibFuncName: getLibcallName(Call: libCall), |
3491 | numArgs: 1)); |
3492 | return; |
3493 | case ISD::READCYCLECOUNTER: { |
3494 | assert(Subtarget->hasLeonCycleCounter()); |
3495 | SDValue Lo = DAG.getCopyFromReg(Chain: N->getOperand(Num: 0), dl, Reg: SP::ASR23, VT: MVT::i32); |
3496 | SDValue Hi = DAG.getCopyFromReg(Chain: Lo, dl, Reg: SP::G0, VT: MVT::i32); |
3497 | SDValue Ops[] = { Lo, Hi }; |
3498 | SDValue Pair = DAG.getNode(Opcode: ISD::BUILD_PAIR, DL: dl, VT: MVT::i64, Ops); |
3499 | Results.push_back(Elt: Pair); |
3500 | Results.push_back(Elt: N->getOperand(Num: 0)); |
3501 | return; |
3502 | } |
3503 | case ISD::SINT_TO_FP: |
3504 | case ISD::UINT_TO_FP: |
3505 | // Custom lower only if it involves f128 or i64. |
3506 | if (N->getValueType(ResNo: 0) != MVT::f128 |
3507 | || N->getOperand(Num: 0).getValueType() != MVT::i64) |
3508 | return; |
3509 | |
3510 | libCall = ((N->getOpcode() == ISD::SINT_TO_FP) |
3511 | ? RTLIB::SINTTOFP_I64_F128 |
3512 | : RTLIB::UINTTOFP_I64_F128); |
3513 | |
3514 | Results.push_back(Elt: LowerF128Op(Op: SDValue(N, 0), |
3515 | DAG, |
3516 | LibFuncName: getLibcallName(Call: libCall), |
3517 | numArgs: 1)); |
3518 | return; |
3519 | case ISD::LOAD: { |
3520 | LoadSDNode *Ld = cast<LoadSDNode>(Val: N); |
3521 | // Custom handling only for i64: turn i64 load into a v2i32 load, |
3522 | // and a bitcast. |
3523 | if (Ld->getValueType(ResNo: 0) != MVT::i64 || Ld->getMemoryVT() != MVT::i64) |
3524 | return; |
3525 | |
3526 | SDLoc dl(N); |
3527 | SDValue LoadRes = DAG.getExtLoad( |
3528 | ExtType: Ld->getExtensionType(), dl, VT: MVT::v2i32, Chain: Ld->getChain(), |
3529 | Ptr: Ld->getBasePtr(), PtrInfo: Ld->getPointerInfo(), MemVT: MVT::v2i32, Alignment: Ld->getBaseAlign(), |
3530 | MMOFlags: Ld->getMemOperand()->getFlags(), AAInfo: Ld->getAAInfo()); |
3531 | |
3532 | SDValue Res = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: MVT::i64, Operand: LoadRes); |
3533 | Results.push_back(Elt: Res); |
3534 | Results.push_back(Elt: LoadRes.getValue(R: 1)); |
3535 | return; |
3536 | } |
3537 | } |
3538 | } |
3539 | |
3540 | // Override to enable LOAD_STACK_GUARD lowering on Linux. |
3541 | bool SparcTargetLowering::useLoadStackGuardNode(const Module &M) const { |
3542 | if (!Subtarget->isTargetLinux()) |
3543 | return TargetLowering::useLoadStackGuardNode(M); |
3544 | return true; |
3545 | } |
3546 | |
3547 | bool SparcTargetLowering::isFNegFree(EVT VT) const { |
3548 | if (Subtarget->isVIS3()) |
3549 | return VT == MVT::f32 || VT == MVT::f64; |
3550 | return false; |
3551 | } |
3552 | |
3553 | bool SparcTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT, |
3554 | bool ForCodeSize) const { |
3555 | if (VT != MVT::f32 && VT != MVT::f64) |
3556 | return false; |
3557 | if (Subtarget->isVIS() && Imm.isZero()) |
3558 | return true; |
3559 | if (Subtarget->isVIS3()) |
3560 | return Imm.isExactlyValue(V: +0.5) || Imm.isExactlyValue(V: -0.5) || |
3561 | Imm.getExactLog2Abs() == -1; |
3562 | return false; |
3563 | } |
3564 | |
3565 | bool SparcTargetLowering::isCtlzFast() const { return Subtarget->isVIS3(); } |
3566 | |
3567 | bool SparcTargetLowering::isCheapToSpeculateCttz(Type *Ty) const { |
3568 | // We lack native cttz, however, |
3569 | // On 64-bit targets it is cheap to implement it in terms of popc. |
3570 | if (Subtarget->is64Bit() && Subtarget->usePopc()) |
3571 | return true; |
3572 | // Otherwise, implementing cttz in terms of ctlz is still cheap. |
3573 | return isCheapToSpeculateCtlz(Ty); |
3574 | } |
3575 | |
3576 | // Override to disable global variable loading on Linux. |
3577 | void SparcTargetLowering::insertSSPDeclarations(Module &M) const { |
3578 | if (!Subtarget->isTargetLinux()) |
3579 | return TargetLowering::insertSSPDeclarations(M); |
3580 | } |
3581 | |
3582 | void SparcTargetLowering::AdjustInstrPostInstrSelection(MachineInstr &MI, |
3583 | SDNode *Node) const { |
3584 | assert(MI.getOpcode() == SP::SUBCCrr || MI.getOpcode() == SP::SUBCCri); |
3585 | // If the result is dead, replace it with %g0. |
3586 | if (!Node->hasAnyUseOfValue(Value: 0)) |
3587 | MI.getOperand(i: 0).setReg(SP::G0); |
3588 | } |
3589 | |