1 | //===- HexagonBitTracker.cpp ----------------------------------------------===// |
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 | #include "HexagonBitTracker.h" |
10 | #include "Hexagon.h" |
11 | #include "HexagonInstrInfo.h" |
12 | #include "HexagonRegisterInfo.h" |
13 | #include "HexagonSubtarget.h" |
14 | #include "llvm/CodeGen/MachineFrameInfo.h" |
15 | #include "llvm/CodeGen/MachineFunction.h" |
16 | #include "llvm/CodeGen/MachineInstr.h" |
17 | #include "llvm/CodeGen/MachineOperand.h" |
18 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
19 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
20 | #include "llvm/IR/Argument.h" |
21 | #include "llvm/IR/Attributes.h" |
22 | #include "llvm/IR/Function.h" |
23 | #include "llvm/IR/Type.h" |
24 | #include "llvm/Support/Compiler.h" |
25 | #include "llvm/Support/Debug.h" |
26 | #include "llvm/Support/ErrorHandling.h" |
27 | #include "llvm/Support/MathExtras.h" |
28 | #include "llvm/Support/raw_ostream.h" |
29 | #include <cassert> |
30 | #include <cstddef> |
31 | #include <cstdint> |
32 | #include <cstdlib> |
33 | #include <utility> |
34 | #include <vector> |
35 | |
36 | using namespace llvm; |
37 | |
38 | using BT = BitTracker; |
39 | |
40 | HexagonEvaluator::HexagonEvaluator(const HexagonRegisterInfo &tri, |
41 | MachineRegisterInfo &mri, |
42 | const HexagonInstrInfo &tii, |
43 | MachineFunction &mf) |
44 | : MachineEvaluator(tri, mri), MF(mf), MFI(mf.getFrameInfo()), TII(tii) { |
45 | // Populate the VRX map (VR to extension-type). |
46 | // Go over all the formal parameters of the function. If a given parameter |
47 | // P is sign- or zero-extended, locate the virtual register holding that |
48 | // parameter and create an entry in the VRX map indicating the type of ex- |
49 | // tension (and the source type). |
50 | // This is a bit complicated to do accurately, since the memory layout in- |
51 | // formation is necessary to precisely determine whether an aggregate para- |
52 | // meter will be passed in a register or in memory. What is given in MRI |
53 | // is the association between the physical register that is live-in (i.e. |
54 | // holds an argument), and the virtual register that this value will be |
55 | // copied into. This, by itself, is not sufficient to map back the virtual |
56 | // register to a formal parameter from Function (since consecutive live-ins |
57 | // from MRI may not correspond to consecutive formal parameters from Func- |
58 | // tion). To avoid the complications with in-memory arguments, only consi- |
59 | // der the initial sequence of formal parameters that are known to be |
60 | // passed via registers. |
61 | unsigned InVirtReg, InPhysReg = 0; |
62 | |
63 | for (const Argument &Arg : MF.getFunction().args()) { |
64 | Type *ATy = Arg.getType(); |
65 | unsigned Width = 0; |
66 | if (ATy->isIntegerTy()) |
67 | Width = ATy->getIntegerBitWidth(); |
68 | else if (ATy->isPointerTy()) |
69 | Width = 32; |
70 | // If pointer size is not set through target data, it will default to |
71 | // Module::AnyPointerSize. |
72 | if (Width == 0 || Width > 64) |
73 | break; |
74 | if (Arg.hasAttribute(Kind: Attribute::ByVal)) |
75 | continue; |
76 | InPhysReg = getNextPhysReg(PReg: InPhysReg, Width); |
77 | if (!InPhysReg) |
78 | break; |
79 | InVirtReg = getVirtRegFor(PReg: InPhysReg); |
80 | if (!InVirtReg) |
81 | continue; |
82 | if (Arg.hasAttribute(Kind: Attribute::SExt)) |
83 | VRX.insert(KV: std::make_pair(x&: InVirtReg, y: ExtType(ExtType::SExt, Width))); |
84 | else if (Arg.hasAttribute(Kind: Attribute::ZExt)) |
85 | VRX.insert(KV: std::make_pair(x&: InVirtReg, y: ExtType(ExtType::ZExt, Width))); |
86 | } |
87 | } |
88 | |
89 | BT::BitMask HexagonEvaluator::mask(Register Reg, unsigned Sub) const { |
90 | if (Sub == 0) |
91 | return MachineEvaluator::mask(Reg, Sub: 0); |
92 | const TargetRegisterClass &RC = *MRI.getRegClass(Reg); |
93 | unsigned ID = RC.getID(); |
94 | uint16_t RW = getRegBitWidth(RR: RegisterRef(Reg, Sub)); |
95 | const auto &HRI = static_cast<const HexagonRegisterInfo&>(TRI); |
96 | bool IsSubLo = (Sub == HRI.getHexagonSubRegIndex(RC, GenIdx: Hexagon::ps_sub_lo)); |
97 | switch (ID) { |
98 | case Hexagon::DoubleRegsRegClassID: |
99 | case Hexagon::HvxWRRegClassID: |
100 | case Hexagon::HvxVQRRegClassID: |
101 | return IsSubLo ? BT::BitMask(0, RW-1) |
102 | : BT::BitMask(RW, 2*RW-1); |
103 | default: |
104 | break; |
105 | } |
106 | #ifndef NDEBUG |
107 | dbgs() << printReg(Reg, &TRI, Sub) << " in reg class " |
108 | << TRI.getRegClassName(&RC) << '\n'; |
109 | #endif |
110 | llvm_unreachable("Unexpected register/subregister" ); |
111 | } |
112 | |
113 | uint16_t HexagonEvaluator::getPhysRegBitWidth(MCRegister Reg) const { |
114 | using namespace Hexagon; |
115 | const auto &HST = MF.getSubtarget<HexagonSubtarget>(); |
116 | if (HST.useHVXOps()) { |
117 | for (auto &RC : {HvxVRRegClass, HvxWRRegClass, HvxQRRegClass, |
118 | HvxVQRRegClass}) |
119 | if (RC.contains(Reg)) |
120 | return TRI.getRegSizeInBits(RC); |
121 | } |
122 | // Default treatment for other physical registers. |
123 | if (const TargetRegisterClass *RC = TRI.getMinimalPhysRegClass(Reg)) |
124 | return TRI.getRegSizeInBits(RC: *RC); |
125 | |
126 | llvm_unreachable( |
127 | (Twine("Unhandled physical register" ) + TRI.getName(Reg)).str().c_str()); |
128 | } |
129 | |
130 | const TargetRegisterClass &HexagonEvaluator::composeWithSubRegIndex( |
131 | const TargetRegisterClass &RC, unsigned Idx) const { |
132 | if (Idx == 0) |
133 | return RC; |
134 | |
135 | #ifndef NDEBUG |
136 | const auto &HRI = static_cast<const HexagonRegisterInfo&>(TRI); |
137 | bool IsSubLo = (Idx == HRI.getHexagonSubRegIndex(RC, Hexagon::ps_sub_lo)); |
138 | bool IsSubHi = (Idx == HRI.getHexagonSubRegIndex(RC, Hexagon::ps_sub_hi)); |
139 | assert(IsSubLo != IsSubHi && "Must refer to either low or high subreg" ); |
140 | #endif |
141 | |
142 | switch (RC.getID()) { |
143 | case Hexagon::DoubleRegsRegClassID: |
144 | return Hexagon::IntRegsRegClass; |
145 | case Hexagon::HvxWRRegClassID: |
146 | return Hexagon::HvxVRRegClass; |
147 | case Hexagon::HvxVQRRegClassID: |
148 | return Hexagon::HvxWRRegClass; |
149 | default: |
150 | break; |
151 | } |
152 | #ifndef NDEBUG |
153 | dbgs() << "Reg class id: " << RC.getID() << " idx: " << Idx << '\n'; |
154 | #endif |
155 | llvm_unreachable("Unimplemented combination of reg class/subreg idx" ); |
156 | } |
157 | |
158 | namespace { |
159 | |
160 | class RegisterRefs { |
161 | std::vector<BT::RegisterRef> Vector; |
162 | |
163 | public: |
164 | RegisterRefs(const MachineInstr &MI) : Vector(MI.getNumOperands()) { |
165 | for (unsigned i = 0, n = Vector.size(); i < n; ++i) { |
166 | const MachineOperand &MO = MI.getOperand(i); |
167 | if (MO.isReg()) |
168 | Vector[i] = BT::RegisterRef(MO); |
169 | // For indices that don't correspond to registers, the entry will |
170 | // remain constructed via the default constructor. |
171 | } |
172 | } |
173 | |
174 | size_t size() const { return Vector.size(); } |
175 | |
176 | const BT::RegisterRef &operator[](unsigned n) const { |
177 | // The main purpose of this operator is to assert with bad argument. |
178 | assert(n < Vector.size()); |
179 | return Vector[n]; |
180 | } |
181 | }; |
182 | |
183 | } // end anonymous namespace |
184 | |
185 | bool HexagonEvaluator::evaluate(const MachineInstr &MI, |
186 | const CellMapType &Inputs, |
187 | CellMapType &Outputs) const { |
188 | using namespace Hexagon; |
189 | |
190 | unsigned NumDefs = 0; |
191 | |
192 | // Basic correctness check: there should not be any defs with subregisters. |
193 | for (const MachineOperand &MO : MI.operands()) { |
194 | if (!MO.isReg() || !MO.isDef()) |
195 | continue; |
196 | NumDefs++; |
197 | assert(MO.getSubReg() == 0); |
198 | } |
199 | |
200 | if (NumDefs == 0) |
201 | return false; |
202 | |
203 | unsigned Opc = MI.getOpcode(); |
204 | |
205 | if (MI.mayLoad()) { |
206 | switch (Opc) { |
207 | // These instructions may be marked as mayLoad, but they are generating |
208 | // immediate values, so skip them. |
209 | case CONST32: |
210 | case CONST64: |
211 | break; |
212 | default: |
213 | return evaluateLoad(MI, Inputs, Outputs); |
214 | } |
215 | } |
216 | |
217 | // Check COPY instructions that copy formal parameters into virtual |
218 | // registers. Such parameters can be sign- or zero-extended at the |
219 | // call site, and we should take advantage of this knowledge. The MRI |
220 | // keeps a list of pairs of live-in physical and virtual registers, |
221 | // which provides information about which virtual registers will hold |
222 | // the argument values. The function will still contain instructions |
223 | // defining those virtual registers, and in practice those are COPY |
224 | // instructions from a physical to a virtual register. In such cases, |
225 | // applying the argument extension to the virtual register can be seen |
226 | // as simply mirroring the extension that had already been applied to |
227 | // the physical register at the call site. If the defining instruction |
228 | // was not a COPY, it would not be clear how to mirror that extension |
229 | // on the callee's side. For that reason, only check COPY instructions |
230 | // for potential extensions. |
231 | if (MI.isCopy()) { |
232 | if (evaluateFormalCopy(MI, Inputs, Outputs)) |
233 | return true; |
234 | } |
235 | |
236 | // Beyond this point, if any operand is a global, skip that instruction. |
237 | // The reason is that certain instructions that can take an immediate |
238 | // operand can also have a global symbol in that operand. To avoid |
239 | // checking what kind of operand a given instruction has individually |
240 | // for each instruction, do it here. Global symbols as operands gene- |
241 | // rally do not provide any useful information. |
242 | for (const MachineOperand &MO : MI.operands()) { |
243 | if (MO.isGlobal() || MO.isBlockAddress() || MO.isSymbol() || MO.isJTI() || |
244 | MO.isCPI()) |
245 | return false; |
246 | } |
247 | |
248 | RegisterRefs Reg(MI); |
249 | #define op(i) MI.getOperand(i) |
250 | #define rc(i) RegisterCell::ref(getCell(Reg[i], Inputs)) |
251 | #define im(i) MI.getOperand(i).getImm() |
252 | |
253 | // If the instruction has no register operands, skip it. |
254 | if (Reg.size() == 0) |
255 | return false; |
256 | |
257 | // Record result for register in operand 0. |
258 | auto rr0 = [this,Reg] (const BT::RegisterCell &Val, CellMapType &Outputs) |
259 | -> bool { |
260 | putCell(RR: Reg[0], RC: Val, M&: Outputs); |
261 | return true; |
262 | }; |
263 | // Get the cell corresponding to the N-th operand. |
264 | auto cop = [this, &Reg, &MI, &Inputs](unsigned N, |
265 | uint16_t W) -> BT::RegisterCell { |
266 | const MachineOperand &Op = MI.getOperand(i: N); |
267 | if (Op.isImm()) |
268 | return eIMM(V: Op.getImm(), W); |
269 | if (!Op.isReg()) |
270 | return RegisterCell::self(Reg: 0, Width: W); |
271 | assert(getRegBitWidth(Reg[N]) == W && "Register width mismatch" ); |
272 | return rc(N); |
273 | }; |
274 | // Extract RW low bits of the cell. |
275 | auto lo = [this] (const BT::RegisterCell &RC, uint16_t RW) |
276 | -> BT::RegisterCell { |
277 | assert(RW <= RC.width()); |
278 | return eXTR(A1: RC, B: 0, E: RW); |
279 | }; |
280 | // Extract RW high bits of the cell. |
281 | auto hi = [this] (const BT::RegisterCell &RC, uint16_t RW) |
282 | -> BT::RegisterCell { |
283 | uint16_t W = RC.width(); |
284 | assert(RW <= W); |
285 | return eXTR(A1: RC, B: W-RW, E: W); |
286 | }; |
287 | // Extract N-th halfword (counting from the least significant position). |
288 | auto half = [this] (const BT::RegisterCell &RC, unsigned N) |
289 | -> BT::RegisterCell { |
290 | assert(N*16+16 <= RC.width()); |
291 | return eXTR(A1: RC, B: N*16, E: N*16+16); |
292 | }; |
293 | // Shuffle bits (pick even/odd from cells and merge into result). |
294 | auto shuffle = [this] (const BT::RegisterCell &Rs, const BT::RegisterCell &Rt, |
295 | uint16_t BW, bool Odd) -> BT::RegisterCell { |
296 | uint16_t I = Odd, Ws = Rs.width(); |
297 | assert(Ws == Rt.width()); |
298 | RegisterCell RC = eXTR(A1: Rt, B: I*BW, E: I*BW+BW).cat(RC: eXTR(A1: Rs, B: I*BW, E: I*BW+BW)); |
299 | I += 2; |
300 | while (I*BW < Ws) { |
301 | RC.cat(RC: eXTR(A1: Rt, B: I*BW, E: I*BW+BW)).cat(RC: eXTR(A1: Rs, B: I*BW, E: I*BW+BW)); |
302 | I += 2; |
303 | } |
304 | return RC; |
305 | }; |
306 | |
307 | // The bitwidth of the 0th operand. In most (if not all) of the |
308 | // instructions below, the 0th operand is the defined register. |
309 | // Pre-compute the bitwidth here, because it is needed in many cases |
310 | // cases below. |
311 | uint16_t W0 = (Reg[0].Reg != 0) ? getRegBitWidth(RR: Reg[0]) : 0; |
312 | |
313 | // Register id of the 0th operand. It can be 0. |
314 | unsigned Reg0 = Reg[0].Reg; |
315 | |
316 | switch (Opc) { |
317 | // Transfer immediate: |
318 | |
319 | case A2_tfrsi: |
320 | case A2_tfrpi: |
321 | case CONST32: |
322 | case CONST64: |
323 | return rr0(eIMM(im(1), W: W0), Outputs); |
324 | case PS_false: |
325 | return rr0(RegisterCell(W0).fill(B: 0, E: W0, V: BT::BitValue::Zero), Outputs); |
326 | case PS_true: |
327 | return rr0(RegisterCell(W0).fill(B: 0, E: W0, V: BT::BitValue::One), Outputs); |
328 | case PS_fi: { |
329 | int FI = op(1).getIndex(); |
330 | int Off = op(2).getImm(); |
331 | unsigned A = MFI.getObjectAlign(ObjectIdx: FI).value() + std::abs(x: Off); |
332 | unsigned L = llvm::countr_zero(Val: A); |
333 | RegisterCell RC = RegisterCell::self(Reg: Reg[0].Reg, Width: W0); |
334 | RC.fill(B: 0, E: L, V: BT::BitValue::Zero); |
335 | return rr0(RC, Outputs); |
336 | } |
337 | |
338 | // Transfer register: |
339 | |
340 | case A2_tfr: |
341 | case A2_tfrp: |
342 | case C2_pxfer_map: |
343 | return rr0(rc(1), Outputs); |
344 | case C2_tfrpr: { |
345 | uint16_t RW = W0; |
346 | uint16_t PW = 8; // XXX Pred size: getRegBitWidth(Reg[1]); |
347 | assert(PW <= RW); |
348 | RegisterCell PC = eXTR(rc(1), B: 0, E: PW); |
349 | RegisterCell RC = RegisterCell(RW).insert(RC: PC, M: BT::BitMask(0, PW-1)); |
350 | RC.fill(B: PW, E: RW, V: BT::BitValue::Zero); |
351 | return rr0(RC, Outputs); |
352 | } |
353 | case C2_tfrrp: { |
354 | uint16_t RW = W0; |
355 | uint16_t PW = 8; // XXX Pred size: getRegBitWidth(Reg[1]); |
356 | RegisterCell RC = RegisterCell::self(Reg: Reg[0].Reg, Width: RW); |
357 | RC.fill(B: PW, E: RW, V: BT::BitValue::Zero); |
358 | return rr0(eINS(A1: RC, A2: eXTR(rc(1), B: 0, E: PW), AtN: 0), Outputs); |
359 | } |
360 | |
361 | // Arithmetic: |
362 | |
363 | case A2_abs: |
364 | case A2_absp: |
365 | // TODO |
366 | break; |
367 | |
368 | case A2_addsp: { |
369 | uint16_t W1 = getRegBitWidth(RR: Reg[1]); |
370 | assert(W0 == 64 && W1 == 32); |
371 | RegisterCell CW = RegisterCell(W0).insert(rc(1), M: BT::BitMask(0, W1-1)); |
372 | RegisterCell RC = eADD(A1: eSXT(A1: CW, FromN: W1), rc(2)); |
373 | return rr0(RC, Outputs); |
374 | } |
375 | case A2_add: |
376 | case A2_addp: |
377 | return rr0(eADD(rc(1), rc(2)), Outputs); |
378 | case A2_addi: |
379 | return rr0(eADD(rc(1), A2: eIMM(im(2), W: W0)), Outputs); |
380 | case S4_addi_asl_ri: { |
381 | RegisterCell RC = eADD(A1: eIMM(im(1), W: W0), A2: eASL(rc(2), im(3))); |
382 | return rr0(RC, Outputs); |
383 | } |
384 | case S4_addi_lsr_ri: { |
385 | RegisterCell RC = eADD(A1: eIMM(im(1), W: W0), A2: eLSR(rc(2), im(3))); |
386 | return rr0(RC, Outputs); |
387 | } |
388 | case S4_addaddi: { |
389 | RegisterCell RC = eADD(rc(1), A2: eADD(rc(2), A2: eIMM(im(3), W: W0))); |
390 | return rr0(RC, Outputs); |
391 | } |
392 | case M4_mpyri_addi: { |
393 | RegisterCell M = eMLS(rc(2), A2: eIMM(im(3), W: W0)); |
394 | RegisterCell RC = eADD(A1: eIMM(im(1), W: W0), A2: lo(M, W0)); |
395 | return rr0(RC, Outputs); |
396 | } |
397 | case M4_mpyrr_addi: { |
398 | RegisterCell M = eMLS(rc(2), rc(3)); |
399 | RegisterCell RC = eADD(A1: eIMM(im(1), W: W0), A2: lo(M, W0)); |
400 | return rr0(RC, Outputs); |
401 | } |
402 | case M4_mpyri_addr_u2: { |
403 | RegisterCell M = eMLS(A1: eIMM(im(2), W: W0), rc(3)); |
404 | RegisterCell RC = eADD(rc(1), A2: lo(M, W0)); |
405 | return rr0(RC, Outputs); |
406 | } |
407 | case M4_mpyri_addr: { |
408 | RegisterCell M = eMLS(rc(2), A2: eIMM(im(3), W: W0)); |
409 | RegisterCell RC = eADD(rc(1), A2: lo(M, W0)); |
410 | return rr0(RC, Outputs); |
411 | } |
412 | case M4_mpyrr_addr: { |
413 | RegisterCell M = eMLS(rc(2), rc(3)); |
414 | RegisterCell RC = eADD(rc(1), A2: lo(M, W0)); |
415 | return rr0(RC, Outputs); |
416 | } |
417 | case S4_subaddi: { |
418 | RegisterCell RC = eADD(rc(1), A2: eSUB(A1: eIMM(im(2), W: W0), rc(3))); |
419 | return rr0(RC, Outputs); |
420 | } |
421 | case M2_accii: { |
422 | RegisterCell RC = eADD(rc(1), A2: eADD(rc(2), A2: eIMM(im(3), W: W0))); |
423 | return rr0(RC, Outputs); |
424 | } |
425 | case M2_acci: { |
426 | RegisterCell RC = eADD(rc(1), A2: eADD(rc(2), rc(3))); |
427 | return rr0(RC, Outputs); |
428 | } |
429 | case M2_subacc: { |
430 | RegisterCell RC = eADD(rc(1), A2: eSUB(rc(2), rc(3))); |
431 | return rr0(RC, Outputs); |
432 | } |
433 | case S2_addasl_rrri: { |
434 | RegisterCell RC = eADD(rc(1), A2: eASL(rc(2), im(3))); |
435 | return rr0(RC, Outputs); |
436 | } |
437 | case C4_addipc: { |
438 | RegisterCell RPC = RegisterCell::self(Reg: Reg[0].Reg, Width: W0); |
439 | RPC.fill(B: 0, E: 2, V: BT::BitValue::Zero); |
440 | return rr0(eADD(A1: RPC, A2: eIMM(im(2), W: W0)), Outputs); |
441 | } |
442 | case A2_sub: |
443 | case A2_subp: |
444 | return rr0(eSUB(rc(1), rc(2)), Outputs); |
445 | case A2_subri: |
446 | return rr0(eSUB(A1: eIMM(im(1), W: W0), rc(2)), Outputs); |
447 | case S4_subi_asl_ri: { |
448 | RegisterCell RC = eSUB(A1: eIMM(im(1), W: W0), A2: eASL(rc(2), im(3))); |
449 | return rr0(RC, Outputs); |
450 | } |
451 | case S4_subi_lsr_ri: { |
452 | RegisterCell RC = eSUB(A1: eIMM(im(1), W: W0), A2: eLSR(rc(2), im(3))); |
453 | return rr0(RC, Outputs); |
454 | } |
455 | case M2_naccii: { |
456 | RegisterCell RC = eSUB(rc(1), A2: eADD(rc(2), A2: eIMM(im(3), W: W0))); |
457 | return rr0(RC, Outputs); |
458 | } |
459 | case M2_nacci: { |
460 | RegisterCell RC = eSUB(rc(1), A2: eADD(rc(2), rc(3))); |
461 | return rr0(RC, Outputs); |
462 | } |
463 | // 32-bit negation is done by "Rd = A2_subri 0, Rs" |
464 | case A2_negp: |
465 | return rr0(eSUB(A1: eIMM(V: 0, W: W0), rc(1)), Outputs); |
466 | |
467 | case M2_mpy_up: { |
468 | RegisterCell M = eMLS(rc(1), rc(2)); |
469 | return rr0(hi(M, W0), Outputs); |
470 | } |
471 | case M2_dpmpyss_s0: |
472 | return rr0(eMLS(rc(1), rc(2)), Outputs); |
473 | case M2_dpmpyss_acc_s0: |
474 | return rr0(eADD(rc(1), A2: eMLS(rc(2), rc(3))), Outputs); |
475 | case M2_dpmpyss_nac_s0: |
476 | return rr0(eSUB(rc(1), A2: eMLS(rc(2), rc(3))), Outputs); |
477 | case M2_mpyi: { |
478 | RegisterCell M = eMLS(rc(1), rc(2)); |
479 | return rr0(lo(M, W0), Outputs); |
480 | } |
481 | case M2_macsip: { |
482 | RegisterCell M = eMLS(rc(2), A2: eIMM(im(3), W: W0)); |
483 | RegisterCell RC = eADD(rc(1), A2: lo(M, W0)); |
484 | return rr0(RC, Outputs); |
485 | } |
486 | case M2_macsin: { |
487 | RegisterCell M = eMLS(rc(2), A2: eIMM(im(3), W: W0)); |
488 | RegisterCell RC = eSUB(rc(1), A2: lo(M, W0)); |
489 | return rr0(RC, Outputs); |
490 | } |
491 | case M2_maci: { |
492 | RegisterCell M = eMLS(rc(2), rc(3)); |
493 | RegisterCell RC = eADD(rc(1), A2: lo(M, W0)); |
494 | return rr0(RC, Outputs); |
495 | } |
496 | case M2_mnaci: { |
497 | RegisterCell M = eMLS(rc(2), rc(3)); |
498 | RegisterCell RC = eSUB(rc(1), A2: lo(M, W0)); |
499 | return rr0(RC, Outputs); |
500 | } |
501 | case M2_mpysmi: { |
502 | RegisterCell M = eMLS(rc(1), A2: eIMM(im(2), W: W0)); |
503 | return rr0(lo(M, 32), Outputs); |
504 | } |
505 | case M2_mpysin: { |
506 | RegisterCell M = eMLS(rc(1), A2: eIMM(V: -im(2), W: W0)); |
507 | return rr0(lo(M, 32), Outputs); |
508 | } |
509 | case M2_mpysip: { |
510 | RegisterCell M = eMLS(rc(1), A2: eIMM(im(2), W: W0)); |
511 | return rr0(lo(M, 32), Outputs); |
512 | } |
513 | case M2_mpyu_up: { |
514 | RegisterCell M = eMLU(rc(1), rc(2)); |
515 | return rr0(hi(M, W0), Outputs); |
516 | } |
517 | case M2_dpmpyuu_s0: |
518 | return rr0(eMLU(rc(1), rc(2)), Outputs); |
519 | case M2_dpmpyuu_acc_s0: |
520 | return rr0(eADD(rc(1), A2: eMLU(rc(2), rc(3))), Outputs); |
521 | case M2_dpmpyuu_nac_s0: |
522 | return rr0(eSUB(rc(1), A2: eMLU(rc(2), rc(3))), Outputs); |
523 | //case M2_mpysu_up: |
524 | |
525 | // Logical/bitwise: |
526 | |
527 | case A2_andir: |
528 | return rr0(eAND(rc(1), A2: eIMM(im(2), W: W0)), Outputs); |
529 | case A2_and: |
530 | case A2_andp: |
531 | return rr0(eAND(rc(1), rc(2)), Outputs); |
532 | case A4_andn: |
533 | case A4_andnp: |
534 | return rr0(eAND(rc(1), A2: eNOT(rc(2))), Outputs); |
535 | case S4_andi_asl_ri: { |
536 | RegisterCell RC = eAND(A1: eIMM(im(1), W: W0), A2: eASL(rc(2), im(3))); |
537 | return rr0(RC, Outputs); |
538 | } |
539 | case S4_andi_lsr_ri: { |
540 | RegisterCell RC = eAND(A1: eIMM(im(1), W: W0), A2: eLSR(rc(2), im(3))); |
541 | return rr0(RC, Outputs); |
542 | } |
543 | case M4_and_and: |
544 | return rr0(eAND(rc(1), A2: eAND(rc(2), rc(3))), Outputs); |
545 | case M4_and_andn: |
546 | return rr0(eAND(rc(1), A2: eAND(rc(2), A2: eNOT(rc(3)))), Outputs); |
547 | case M4_and_or: |
548 | return rr0(eAND(rc(1), A2: eORL(rc(2), rc(3))), Outputs); |
549 | case M4_and_xor: |
550 | return rr0(eAND(rc(1), A2: eXOR(rc(2), rc(3))), Outputs); |
551 | case A2_orir: |
552 | return rr0(eORL(rc(1), A2: eIMM(im(2), W: W0)), Outputs); |
553 | case A2_or: |
554 | case A2_orp: |
555 | return rr0(eORL(rc(1), rc(2)), Outputs); |
556 | case A4_orn: |
557 | case A4_ornp: |
558 | return rr0(eORL(rc(1), A2: eNOT(rc(2))), Outputs); |
559 | case S4_ori_asl_ri: { |
560 | RegisterCell RC = eORL(A1: eIMM(im(1), W: W0), A2: eASL(rc(2), im(3))); |
561 | return rr0(RC, Outputs); |
562 | } |
563 | case S4_ori_lsr_ri: { |
564 | RegisterCell RC = eORL(A1: eIMM(im(1), W: W0), A2: eLSR(rc(2), im(3))); |
565 | return rr0(RC, Outputs); |
566 | } |
567 | case M4_or_and: |
568 | return rr0(eORL(rc(1), A2: eAND(rc(2), rc(3))), Outputs); |
569 | case M4_or_andn: |
570 | return rr0(eORL(rc(1), A2: eAND(rc(2), A2: eNOT(rc(3)))), Outputs); |
571 | case S4_or_andi: |
572 | case S4_or_andix: { |
573 | RegisterCell RC = eORL(rc(1), A2: eAND(rc(2), A2: eIMM(im(3), W: W0))); |
574 | return rr0(RC, Outputs); |
575 | } |
576 | case S4_or_ori: { |
577 | RegisterCell RC = eORL(rc(1), A2: eORL(rc(2), A2: eIMM(im(3), W: W0))); |
578 | return rr0(RC, Outputs); |
579 | } |
580 | case M4_or_or: |
581 | return rr0(eORL(rc(1), A2: eORL(rc(2), rc(3))), Outputs); |
582 | case M4_or_xor: |
583 | return rr0(eORL(rc(1), A2: eXOR(rc(2), rc(3))), Outputs); |
584 | case A2_xor: |
585 | case A2_xorp: |
586 | return rr0(eXOR(rc(1), rc(2)), Outputs); |
587 | case M4_xor_and: |
588 | return rr0(eXOR(rc(1), A2: eAND(rc(2), rc(3))), Outputs); |
589 | case M4_xor_andn: |
590 | return rr0(eXOR(rc(1), A2: eAND(rc(2), A2: eNOT(rc(3)))), Outputs); |
591 | case M4_xor_or: |
592 | return rr0(eXOR(rc(1), A2: eORL(rc(2), rc(3))), Outputs); |
593 | case M4_xor_xacc: |
594 | return rr0(eXOR(rc(1), A2: eXOR(rc(2), rc(3))), Outputs); |
595 | case A2_not: |
596 | case A2_notp: |
597 | return rr0(eNOT(rc(1)), Outputs); |
598 | |
599 | case S2_asl_i_r: |
600 | case S2_asl_i_p: |
601 | return rr0(eASL(rc(1), im(2)), Outputs); |
602 | case A2_aslh: |
603 | return rr0(eASL(rc(1), Sh: 16), Outputs); |
604 | case S2_asl_i_r_acc: |
605 | case S2_asl_i_p_acc: |
606 | return rr0(eADD(rc(1), A2: eASL(rc(2), im(3))), Outputs); |
607 | case S2_asl_i_r_nac: |
608 | case S2_asl_i_p_nac: |
609 | return rr0(eSUB(rc(1), A2: eASL(rc(2), im(3))), Outputs); |
610 | case S2_asl_i_r_and: |
611 | case S2_asl_i_p_and: |
612 | return rr0(eAND(rc(1), A2: eASL(rc(2), im(3))), Outputs); |
613 | case S2_asl_i_r_or: |
614 | case S2_asl_i_p_or: |
615 | return rr0(eORL(rc(1), A2: eASL(rc(2), im(3))), Outputs); |
616 | case S2_asl_i_r_xacc: |
617 | case S2_asl_i_p_xacc: |
618 | return rr0(eXOR(rc(1), A2: eASL(rc(2), im(3))), Outputs); |
619 | case S2_asl_i_vh: |
620 | case S2_asl_i_vw: |
621 | // TODO |
622 | break; |
623 | |
624 | case S2_asr_i_r: |
625 | case S2_asr_i_p: |
626 | return rr0(eASR(rc(1), im(2)), Outputs); |
627 | case A2_asrh: |
628 | return rr0(eASR(rc(1), Sh: 16), Outputs); |
629 | case S2_asr_i_r_acc: |
630 | case S2_asr_i_p_acc: |
631 | return rr0(eADD(rc(1), A2: eASR(rc(2), im(3))), Outputs); |
632 | case S2_asr_i_r_nac: |
633 | case S2_asr_i_p_nac: |
634 | return rr0(eSUB(rc(1), A2: eASR(rc(2), im(3))), Outputs); |
635 | case S2_asr_i_r_and: |
636 | case S2_asr_i_p_and: |
637 | return rr0(eAND(rc(1), A2: eASR(rc(2), im(3))), Outputs); |
638 | case S2_asr_i_r_or: |
639 | case S2_asr_i_p_or: |
640 | return rr0(eORL(rc(1), A2: eASR(rc(2), im(3))), Outputs); |
641 | case S2_asr_i_r_rnd: { |
642 | // The input is first sign-extended to 64 bits, then the output |
643 | // is truncated back to 32 bits. |
644 | assert(W0 == 32); |
645 | RegisterCell XC = eSXT(rc(1).cat(RC: eIMM(V: 0, W: W0)), FromN: W0); |
646 | RegisterCell RC = eASR(A1: eADD(A1: eASR(A1: XC, im(2)), A2: eIMM(V: 1, W: 2*W0)), Sh: 1); |
647 | return rr0(eXTR(A1: RC, B: 0, E: W0), Outputs); |
648 | } |
649 | case S2_asr_i_r_rnd_goodsyntax: { |
650 | int64_t S = im(2); |
651 | if (S == 0) |
652 | return rr0(rc(1), Outputs); |
653 | // Result: S2_asr_i_r_rnd Rs, u5-1 |
654 | RegisterCell XC = eSXT(rc(1).cat(RC: eIMM(V: 0, W: W0)), FromN: W0); |
655 | RegisterCell RC = eLSR(A1: eADD(A1: eASR(A1: XC, Sh: S-1), A2: eIMM(V: 1, W: 2*W0)), Sh: 1); |
656 | return rr0(eXTR(A1: RC, B: 0, E: W0), Outputs); |
657 | } |
658 | case S2_asr_r_vh: |
659 | case S2_asr_i_vw: |
660 | case S2_asr_i_svw_trun: |
661 | // TODO |
662 | break; |
663 | |
664 | case S2_lsr_i_r: |
665 | case S2_lsr_i_p: |
666 | return rr0(eLSR(rc(1), im(2)), Outputs); |
667 | case S2_lsr_i_r_acc: |
668 | case S2_lsr_i_p_acc: |
669 | return rr0(eADD(rc(1), A2: eLSR(rc(2), im(3))), Outputs); |
670 | case S2_lsr_i_r_nac: |
671 | case S2_lsr_i_p_nac: |
672 | return rr0(eSUB(rc(1), A2: eLSR(rc(2), im(3))), Outputs); |
673 | case S2_lsr_i_r_and: |
674 | case S2_lsr_i_p_and: |
675 | return rr0(eAND(rc(1), A2: eLSR(rc(2), im(3))), Outputs); |
676 | case S2_lsr_i_r_or: |
677 | case S2_lsr_i_p_or: |
678 | return rr0(eORL(rc(1), A2: eLSR(rc(2), im(3))), Outputs); |
679 | case S2_lsr_i_r_xacc: |
680 | case S2_lsr_i_p_xacc: |
681 | return rr0(eXOR(rc(1), A2: eLSR(rc(2), im(3))), Outputs); |
682 | |
683 | case S2_clrbit_i: { |
684 | RegisterCell RC = rc(1); |
685 | RC[im(2)] = BT::BitValue::Zero; |
686 | return rr0(RC, Outputs); |
687 | } |
688 | case S2_setbit_i: { |
689 | RegisterCell RC = rc(1); |
690 | RC[im(2)] = BT::BitValue::One; |
691 | return rr0(RC, Outputs); |
692 | } |
693 | case S2_togglebit_i: { |
694 | RegisterCell RC = rc(1); |
695 | uint16_t BX = im(2); |
696 | RC[BX] = RC[BX].is(T: 0) ? BT::BitValue::One |
697 | : RC[BX].is(T: 1) ? BT::BitValue::Zero |
698 | : BT::BitValue::self(); |
699 | return rr0(RC, Outputs); |
700 | } |
701 | |
702 | case A4_bitspliti: { |
703 | uint16_t W1 = getRegBitWidth(RR: Reg[1]); |
704 | uint16_t BX = im(2); |
705 | // Res.uw[1] = Rs[bx+1:], Res.uw[0] = Rs[0:bx] |
706 | const BT::BitValue Zero = BT::BitValue::Zero; |
707 | RegisterCell RZ = RegisterCell(W0).fill(B: BX, E: W1, V: Zero) |
708 | .fill(B: W1+(W1-BX), E: W0, V: Zero); |
709 | RegisterCell BF1 = eXTR(rc(1), B: 0, E: BX), BF2 = eXTR(rc(1), B: BX, E: W1); |
710 | RegisterCell RC = eINS(A1: eINS(A1: RZ, A2: BF1, AtN: 0), A2: BF2, AtN: W1); |
711 | return rr0(RC, Outputs); |
712 | } |
713 | case S4_extract: |
714 | case S4_extractp: |
715 | case S2_extractu: |
716 | case S2_extractup: { |
717 | uint16_t Wd = im(2), Of = im(3); |
718 | assert(Wd <= W0); |
719 | if (Wd == 0) |
720 | return rr0(eIMM(V: 0, W: W0), Outputs); |
721 | // If the width extends beyond the register size, pad the register |
722 | // with 0 bits. |
723 | RegisterCell Pad = (Wd+Of > W0) ? rc(1).cat(RC: eIMM(V: 0, W: Wd+Of-W0)) : rc(1); |
724 | RegisterCell Ext = eXTR(A1: Pad, B: Of, E: Wd+Of); |
725 | // Ext is short, need to extend it with 0s or sign bit. |
726 | RegisterCell RC = RegisterCell(W0).insert(RC: Ext, M: BT::BitMask(0, Wd-1)); |
727 | if (Opc == S2_extractu || Opc == S2_extractup) |
728 | return rr0(eZXT(A1: RC, FromN: Wd), Outputs); |
729 | return rr0(eSXT(A1: RC, FromN: Wd), Outputs); |
730 | } |
731 | case S2_insert: |
732 | case S2_insertp: { |
733 | uint16_t Wd = im(3), Of = im(4); |
734 | assert(Wd < W0 && Of < W0); |
735 | // If Wd+Of exceeds W0, the inserted bits are truncated. |
736 | if (Wd+Of > W0) |
737 | Wd = W0-Of; |
738 | if (Wd == 0) |
739 | return rr0(rc(1), Outputs); |
740 | return rr0(eINS(rc(1), A2: eXTR(rc(2), B: 0, E: Wd), AtN: Of), Outputs); |
741 | } |
742 | |
743 | // Bit permutations: |
744 | |
745 | case A2_combineii: |
746 | case A4_combineii: |
747 | case A4_combineir: |
748 | case A4_combineri: |
749 | case A2_combinew: |
750 | case V6_vcombine: |
751 | assert(W0 % 2 == 0); |
752 | return rr0(cop(2, W0/2).cat(RC: cop(1, W0/2)), Outputs); |
753 | case A2_combine_ll: |
754 | case A2_combine_lh: |
755 | case A2_combine_hl: |
756 | case A2_combine_hh: { |
757 | assert(W0 == 32); |
758 | assert(getRegBitWidth(Reg[1]) == 32 && getRegBitWidth(Reg[2]) == 32); |
759 | // Low half in the output is 0 for _ll and _hl, 1 otherwise: |
760 | unsigned LoH = !(Opc == A2_combine_ll || Opc == A2_combine_hl); |
761 | // High half in the output is 0 for _ll and _lh, 1 otherwise: |
762 | unsigned HiH = !(Opc == A2_combine_ll || Opc == A2_combine_lh); |
763 | RegisterCell R1 = rc(1); |
764 | RegisterCell R2 = rc(2); |
765 | RegisterCell RC = half(R2, LoH).cat(RC: half(R1, HiH)); |
766 | return rr0(RC, Outputs); |
767 | } |
768 | case S2_packhl: { |
769 | assert(W0 == 64); |
770 | assert(getRegBitWidth(Reg[1]) == 32 && getRegBitWidth(Reg[2]) == 32); |
771 | RegisterCell R1 = rc(1); |
772 | RegisterCell R2 = rc(2); |
773 | RegisterCell RC = half(R2, 0).cat(RC: half(R1, 0)).cat(RC: half(R2, 1)) |
774 | .cat(RC: half(R1, 1)); |
775 | return rr0(RC, Outputs); |
776 | } |
777 | case S2_shuffeb: { |
778 | RegisterCell RC = shuffle(rc(1), rc(2), 8, false); |
779 | return rr0(RC, Outputs); |
780 | } |
781 | case S2_shuffeh: { |
782 | RegisterCell RC = shuffle(rc(1), rc(2), 16, false); |
783 | return rr0(RC, Outputs); |
784 | } |
785 | case S2_shuffob: { |
786 | RegisterCell RC = shuffle(rc(1), rc(2), 8, true); |
787 | return rr0(RC, Outputs); |
788 | } |
789 | case S2_shuffoh: { |
790 | RegisterCell RC = shuffle(rc(1), rc(2), 16, true); |
791 | return rr0(RC, Outputs); |
792 | } |
793 | case C2_mask: { |
794 | uint16_t WR = W0; |
795 | uint16_t WP = 8; // XXX Pred size: getRegBitWidth(Reg[1]); |
796 | assert(WR == 64 && WP == 8); |
797 | RegisterCell R1 = rc(1); |
798 | RegisterCell RC(WR); |
799 | for (uint16_t i = 0; i < WP; ++i) { |
800 | const BT::BitValue &V = R1[i]; |
801 | BT::BitValue F = (V.is(T: 0) || V.is(T: 1)) ? V : BT::BitValue::self(); |
802 | RC.fill(B: i*8, E: i*8+8, V: F); |
803 | } |
804 | return rr0(RC, Outputs); |
805 | } |
806 | |
807 | // Mux: |
808 | |
809 | case C2_muxii: |
810 | case C2_muxir: |
811 | case C2_muxri: |
812 | case C2_mux: { |
813 | BT::BitValue PC0 = rc(1)[0]; |
814 | RegisterCell R2 = cop(2, W0); |
815 | RegisterCell R3 = cop(3, W0); |
816 | if (PC0.is(T: 0) || PC0.is(T: 1)) |
817 | return rr0(RegisterCell::ref(C: PC0 ? R2 : R3), Outputs); |
818 | R2.meet(RC: R3, SelfR: Reg[0].Reg); |
819 | return rr0(R2, Outputs); |
820 | } |
821 | case C2_vmux: |
822 | // TODO |
823 | break; |
824 | |
825 | // Sign- and zero-extension: |
826 | |
827 | case A2_sxtb: |
828 | return rr0(eSXT(rc(1), FromN: 8), Outputs); |
829 | case A2_sxth: |
830 | return rr0(eSXT(rc(1), FromN: 16), Outputs); |
831 | case A2_sxtw: { |
832 | uint16_t W1 = getRegBitWidth(RR: Reg[1]); |
833 | assert(W0 == 64 && W1 == 32); |
834 | RegisterCell RC = eSXT(rc(1).cat(RC: eIMM(V: 0, W: W1)), FromN: W1); |
835 | return rr0(RC, Outputs); |
836 | } |
837 | case A2_zxtb: |
838 | return rr0(eZXT(rc(1), FromN: 8), Outputs); |
839 | case A2_zxth: |
840 | return rr0(eZXT(rc(1), FromN: 16), Outputs); |
841 | |
842 | // Saturations |
843 | |
844 | case A2_satb: |
845 | return rr0(eSXT(A1: RegisterCell::self(Reg: 0, Width: W0).regify(R: Reg0), FromN: 8), Outputs); |
846 | case A2_sath: |
847 | return rr0(eSXT(A1: RegisterCell::self(Reg: 0, Width: W0).regify(R: Reg0), FromN: 16), Outputs); |
848 | case A2_satub: |
849 | return rr0(eZXT(A1: RegisterCell::self(Reg: 0, Width: W0).regify(R: Reg0), FromN: 8), Outputs); |
850 | case A2_satuh: |
851 | return rr0(eZXT(A1: RegisterCell::self(Reg: 0, Width: W0).regify(R: Reg0), FromN: 16), Outputs); |
852 | |
853 | // Bit count: |
854 | |
855 | case S2_cl0: |
856 | case S2_cl0p: |
857 | // Always produce a 32-bit result. |
858 | return rr0(eCLB(rc(1), B: false/*bit*/, W: 32), Outputs); |
859 | case S2_cl1: |
860 | case S2_cl1p: |
861 | return rr0(eCLB(rc(1), B: true/*bit*/, W: 32), Outputs); |
862 | case S2_clb: |
863 | case S2_clbp: { |
864 | uint16_t W1 = getRegBitWidth(RR: Reg[1]); |
865 | RegisterCell R1 = rc(1); |
866 | BT::BitValue TV = R1[W1-1]; |
867 | if (TV.is(T: 0) || TV.is(T: 1)) |
868 | return rr0(eCLB(A1: R1, B: TV, W: 32), Outputs); |
869 | break; |
870 | } |
871 | case S2_ct0: |
872 | case S2_ct0p: |
873 | return rr0(eCTB(rc(1), B: false/*bit*/, W: 32), Outputs); |
874 | case S2_ct1: |
875 | case S2_ct1p: |
876 | return rr0(eCTB(rc(1), B: true/*bit*/, W: 32), Outputs); |
877 | case S5_popcountp: |
878 | // TODO |
879 | break; |
880 | |
881 | case C2_all8: { |
882 | RegisterCell P1 = rc(1); |
883 | bool Has0 = false, All1 = true; |
884 | for (uint16_t i = 0; i < 8/*XXX*/; ++i) { |
885 | if (!P1[i].is(T: 1)) |
886 | All1 = false; |
887 | if (!P1[i].is(T: 0)) |
888 | continue; |
889 | Has0 = true; |
890 | break; |
891 | } |
892 | if (!Has0 && !All1) |
893 | break; |
894 | RegisterCell RC(W0); |
895 | RC.fill(B: 0, E: W0, V: (All1 ? BT::BitValue::One : BT::BitValue::Zero)); |
896 | return rr0(RC, Outputs); |
897 | } |
898 | case C2_any8: { |
899 | RegisterCell P1 = rc(1); |
900 | bool Has1 = false, All0 = true; |
901 | for (uint16_t i = 0; i < 8/*XXX*/; ++i) { |
902 | if (!P1[i].is(T: 0)) |
903 | All0 = false; |
904 | if (!P1[i].is(T: 1)) |
905 | continue; |
906 | Has1 = true; |
907 | break; |
908 | } |
909 | if (!Has1 && !All0) |
910 | break; |
911 | RegisterCell RC(W0); |
912 | RC.fill(B: 0, E: W0, V: (Has1 ? BT::BitValue::One : BT::BitValue::Zero)); |
913 | return rr0(RC, Outputs); |
914 | } |
915 | case C2_and: |
916 | return rr0(eAND(rc(1), rc(2)), Outputs); |
917 | case C2_andn: |
918 | return rr0(eAND(rc(1), A2: eNOT(rc(2))), Outputs); |
919 | case C2_not: |
920 | return rr0(eNOT(rc(1)), Outputs); |
921 | case C2_or: |
922 | return rr0(eORL(rc(1), rc(2)), Outputs); |
923 | case C2_orn: |
924 | return rr0(eORL(rc(1), A2: eNOT(rc(2))), Outputs); |
925 | case C2_xor: |
926 | return rr0(eXOR(rc(1), rc(2)), Outputs); |
927 | case C4_and_and: |
928 | return rr0(eAND(rc(1), A2: eAND(rc(2), rc(3))), Outputs); |
929 | case C4_and_andn: |
930 | return rr0(eAND(rc(1), A2: eAND(rc(2), A2: eNOT(rc(3)))), Outputs); |
931 | case C4_and_or: |
932 | return rr0(eAND(rc(1), A2: eORL(rc(2), rc(3))), Outputs); |
933 | case C4_and_orn: |
934 | return rr0(eAND(rc(1), A2: eORL(rc(2), A2: eNOT(rc(3)))), Outputs); |
935 | case C4_or_and: |
936 | return rr0(eORL(rc(1), A2: eAND(rc(2), rc(3))), Outputs); |
937 | case C4_or_andn: |
938 | return rr0(eORL(rc(1), A2: eAND(rc(2), A2: eNOT(rc(3)))), Outputs); |
939 | case C4_or_or: |
940 | return rr0(eORL(rc(1), A2: eORL(rc(2), rc(3))), Outputs); |
941 | case C4_or_orn: |
942 | return rr0(eORL(rc(1), A2: eORL(rc(2), A2: eNOT(rc(3)))), Outputs); |
943 | case C2_bitsclr: |
944 | case C2_bitsclri: |
945 | case C2_bitsset: |
946 | case C4_nbitsclr: |
947 | case C4_nbitsclri: |
948 | case C4_nbitsset: |
949 | // TODO |
950 | break; |
951 | case S2_tstbit_i: |
952 | case S4_ntstbit_i: { |
953 | BT::BitValue V = rc(1)[im(2)]; |
954 | if (V.is(T: 0) || V.is(T: 1)) { |
955 | // If instruction is S2_tstbit_i, test for 1, otherwise test for 0. |
956 | bool TV = (Opc == S2_tstbit_i); |
957 | BT::BitValue F = V.is(T: TV) ? BT::BitValue::One : BT::BitValue::Zero; |
958 | return rr0(RegisterCell(W0).fill(B: 0, E: W0, V: F), Outputs); |
959 | } |
960 | break; |
961 | } |
962 | |
963 | default: |
964 | // For instructions that define a single predicate registers, store |
965 | // the low 8 bits of the register only. |
966 | if (unsigned DefR = getUniqueDefVReg(MI)) { |
967 | if (MRI.getRegClass(Reg: DefR) == &Hexagon::PredRegsRegClass) { |
968 | BT::RegisterRef PD(DefR, 0); |
969 | uint16_t RW = getRegBitWidth(RR: PD); |
970 | uint16_t PW = 8; // XXX Pred size: getRegBitWidth(Reg[1]); |
971 | RegisterCell RC = RegisterCell::self(Reg: DefR, Width: RW); |
972 | RC.fill(B: PW, E: RW, V: BT::BitValue::Zero); |
973 | putCell(RR: PD, RC, M&: Outputs); |
974 | return true; |
975 | } |
976 | } |
977 | return MachineEvaluator::evaluate(MI, Inputs, Outputs); |
978 | } |
979 | #undef im |
980 | #undef rc |
981 | #undef op |
982 | return false; |
983 | } |
984 | |
985 | bool HexagonEvaluator::evaluate(const MachineInstr &BI, |
986 | const CellMapType &Inputs, |
987 | BranchTargetList &Targets, |
988 | bool &FallsThru) const { |
989 | // We need to evaluate one branch at a time. TII::analyzeBranch checks |
990 | // all the branches in a basic block at once, so we cannot use it. |
991 | unsigned Opc = BI.getOpcode(); |
992 | bool SimpleBranch = false; |
993 | bool Negated = false; |
994 | switch (Opc) { |
995 | case Hexagon::J2_jumpf: |
996 | case Hexagon::J2_jumpfpt: |
997 | case Hexagon::J2_jumpfnew: |
998 | case Hexagon::J2_jumpfnewpt: |
999 | Negated = true; |
1000 | [[fallthrough]]; |
1001 | case Hexagon::J2_jumpt: |
1002 | case Hexagon::J2_jumptpt: |
1003 | case Hexagon::J2_jumptnew: |
1004 | case Hexagon::J2_jumptnewpt: |
1005 | // Simple branch: if([!]Pn) jump ... |
1006 | // i.e. Op0 = predicate, Op1 = branch target. |
1007 | SimpleBranch = true; |
1008 | break; |
1009 | case Hexagon::J2_jump: |
1010 | Targets.insert(X: BI.getOperand(i: 0).getMBB()); |
1011 | FallsThru = false; |
1012 | return true; |
1013 | default: |
1014 | // If the branch is of unknown type, assume that all successors are |
1015 | // executable. |
1016 | return false; |
1017 | } |
1018 | |
1019 | if (!SimpleBranch) |
1020 | return false; |
1021 | |
1022 | // BI is a conditional branch if we got here. |
1023 | RegisterRef PR = BI.getOperand(i: 0); |
1024 | RegisterCell PC = getCell(RR: PR, M: Inputs); |
1025 | const BT::BitValue &Test = PC[0]; |
1026 | |
1027 | // If the condition is neither true nor false, then it's unknown. |
1028 | if (!Test.is(T: 0) && !Test.is(T: 1)) |
1029 | return false; |
1030 | |
1031 | // "Test.is(!Negated)" means "branch condition is true". |
1032 | if (!Test.is(T: !Negated)) { |
1033 | // Condition known to be false. |
1034 | FallsThru = true; |
1035 | return true; |
1036 | } |
1037 | |
1038 | Targets.insert(X: BI.getOperand(i: 1).getMBB()); |
1039 | FallsThru = false; |
1040 | return true; |
1041 | } |
1042 | |
1043 | unsigned HexagonEvaluator::getUniqueDefVReg(const MachineInstr &MI) const { |
1044 | unsigned DefReg = 0; |
1045 | for (const MachineOperand &Op : MI.operands()) { |
1046 | if (!Op.isReg() || !Op.isDef()) |
1047 | continue; |
1048 | Register R = Op.getReg(); |
1049 | if (!R.isVirtual()) |
1050 | continue; |
1051 | if (DefReg != 0) |
1052 | return 0; |
1053 | DefReg = R; |
1054 | } |
1055 | return DefReg; |
1056 | } |
1057 | |
1058 | bool HexagonEvaluator::evaluateLoad(const MachineInstr &MI, |
1059 | const CellMapType &Inputs, |
1060 | CellMapType &Outputs) const { |
1061 | using namespace Hexagon; |
1062 | |
1063 | if (TII.isPredicated(MI)) |
1064 | return false; |
1065 | assert(MI.mayLoad() && "A load that mayn't?" ); |
1066 | unsigned Opc = MI.getOpcode(); |
1067 | |
1068 | uint16_t BitNum; |
1069 | bool SignEx; |
1070 | |
1071 | switch (Opc) { |
1072 | default: |
1073 | return false; |
1074 | |
1075 | #if 0 |
1076 | // memb_fifo |
1077 | case L2_loadalignb_pbr: |
1078 | case L2_loadalignb_pcr: |
1079 | case L2_loadalignb_pi: |
1080 | // memh_fifo |
1081 | case L2_loadalignh_pbr: |
1082 | case L2_loadalignh_pcr: |
1083 | case L2_loadalignh_pi: |
1084 | // membh |
1085 | case L2_loadbsw2_pbr: |
1086 | case L2_loadbsw2_pci: |
1087 | case L2_loadbsw2_pcr: |
1088 | case L2_loadbsw2_pi: |
1089 | case L2_loadbsw4_pbr: |
1090 | case L2_loadbsw4_pci: |
1091 | case L2_loadbsw4_pcr: |
1092 | case L2_loadbsw4_pi: |
1093 | // memubh |
1094 | case L2_loadbzw2_pbr: |
1095 | case L2_loadbzw2_pci: |
1096 | case L2_loadbzw2_pcr: |
1097 | case L2_loadbzw2_pi: |
1098 | case L2_loadbzw4_pbr: |
1099 | case L2_loadbzw4_pci: |
1100 | case L2_loadbzw4_pcr: |
1101 | case L2_loadbzw4_pi: |
1102 | #endif |
1103 | |
1104 | case L2_loadrbgp: |
1105 | case L2_loadrb_io: |
1106 | case L2_loadrb_pbr: |
1107 | case L2_loadrb_pci: |
1108 | case L2_loadrb_pcr: |
1109 | case L2_loadrb_pi: |
1110 | case PS_loadrbabs: |
1111 | case L4_loadrb_ap: |
1112 | case L4_loadrb_rr: |
1113 | case L4_loadrb_ur: |
1114 | BitNum = 8; |
1115 | SignEx = true; |
1116 | break; |
1117 | |
1118 | case L2_loadrubgp: |
1119 | case L2_loadrub_io: |
1120 | case L2_loadrub_pbr: |
1121 | case L2_loadrub_pci: |
1122 | case L2_loadrub_pcr: |
1123 | case L2_loadrub_pi: |
1124 | case PS_loadrubabs: |
1125 | case L4_loadrub_ap: |
1126 | case L4_loadrub_rr: |
1127 | case L4_loadrub_ur: |
1128 | BitNum = 8; |
1129 | SignEx = false; |
1130 | break; |
1131 | |
1132 | case L2_loadrhgp: |
1133 | case L2_loadrh_io: |
1134 | case L2_loadrh_pbr: |
1135 | case L2_loadrh_pci: |
1136 | case L2_loadrh_pcr: |
1137 | case L2_loadrh_pi: |
1138 | case PS_loadrhabs: |
1139 | case L4_loadrh_ap: |
1140 | case L4_loadrh_rr: |
1141 | case L4_loadrh_ur: |
1142 | BitNum = 16; |
1143 | SignEx = true; |
1144 | break; |
1145 | |
1146 | case L2_loadruhgp: |
1147 | case L2_loadruh_io: |
1148 | case L2_loadruh_pbr: |
1149 | case L2_loadruh_pci: |
1150 | case L2_loadruh_pcr: |
1151 | case L2_loadruh_pi: |
1152 | case L4_loadruh_rr: |
1153 | case PS_loadruhabs: |
1154 | case L4_loadruh_ap: |
1155 | case L4_loadruh_ur: |
1156 | BitNum = 16; |
1157 | SignEx = false; |
1158 | break; |
1159 | |
1160 | case L2_loadrigp: |
1161 | case L2_loadri_io: |
1162 | case L2_loadri_pbr: |
1163 | case L2_loadri_pci: |
1164 | case L2_loadri_pcr: |
1165 | case L2_loadri_pi: |
1166 | case L2_loadw_locked: |
1167 | case PS_loadriabs: |
1168 | case L4_loadri_ap: |
1169 | case L4_loadri_rr: |
1170 | case L4_loadri_ur: |
1171 | case LDriw_pred: |
1172 | BitNum = 32; |
1173 | SignEx = true; |
1174 | break; |
1175 | |
1176 | case L2_loadrdgp: |
1177 | case L2_loadrd_io: |
1178 | case L2_loadrd_pbr: |
1179 | case L2_loadrd_pci: |
1180 | case L2_loadrd_pcr: |
1181 | case L2_loadrd_pi: |
1182 | case L4_loadd_locked: |
1183 | case PS_loadrdabs: |
1184 | case L4_loadrd_ap: |
1185 | case L4_loadrd_rr: |
1186 | case L4_loadrd_ur: |
1187 | BitNum = 64; |
1188 | SignEx = true; |
1189 | break; |
1190 | } |
1191 | |
1192 | const MachineOperand &MD = MI.getOperand(i: 0); |
1193 | assert(MD.isReg() && MD.isDef()); |
1194 | RegisterRef RD = MD; |
1195 | |
1196 | uint16_t W = getRegBitWidth(RR: RD); |
1197 | assert(W >= BitNum && BitNum > 0); |
1198 | RegisterCell Res(W); |
1199 | |
1200 | for (uint16_t i = 0; i < BitNum; ++i) |
1201 | Res[i] = BT::BitValue::self(Self: BT::BitRef(RD.Reg, i)); |
1202 | |
1203 | if (SignEx) { |
1204 | const BT::BitValue &Sign = Res[BitNum-1]; |
1205 | for (uint16_t i = BitNum; i < W; ++i) |
1206 | Res[i] = BT::BitValue::ref(V: Sign); |
1207 | } else { |
1208 | for (uint16_t i = BitNum; i < W; ++i) |
1209 | Res[i] = BT::BitValue::Zero; |
1210 | } |
1211 | |
1212 | putCell(RR: RD, RC: Res, M&: Outputs); |
1213 | return true; |
1214 | } |
1215 | |
1216 | bool HexagonEvaluator::evaluateFormalCopy(const MachineInstr &MI, |
1217 | const CellMapType &Inputs, |
1218 | CellMapType &Outputs) const { |
1219 | // If MI defines a formal parameter, but is not a copy (loads are handled |
1220 | // in evaluateLoad), then it's not clear what to do. |
1221 | assert(MI.isCopy()); |
1222 | |
1223 | RegisterRef RD = MI.getOperand(i: 0); |
1224 | RegisterRef RS = MI.getOperand(i: 1); |
1225 | assert(RD.Sub == 0); |
1226 | if (!RS.Reg.isPhysical()) |
1227 | return false; |
1228 | RegExtMap::const_iterator F = VRX.find(Val: RD.Reg); |
1229 | if (F == VRX.end()) |
1230 | return false; |
1231 | |
1232 | uint16_t EW = F->second.Width; |
1233 | // Store RD's cell into the map. This will associate the cell with a virtual |
1234 | // register, and make zero-/sign-extends possible (otherwise we would be ex- |
1235 | // tending "self" bit values, which will have no effect, since "self" values |
1236 | // cannot be references to anything). |
1237 | putCell(RR: RD, RC: getCell(RR: RS, M: Inputs), M&: Outputs); |
1238 | |
1239 | RegisterCell Res; |
1240 | // Read RD's cell from the outputs instead of RS's cell from the inputs: |
1241 | if (F->second.Type == ExtType::SExt) |
1242 | Res = eSXT(A1: getCell(RR: RD, M: Outputs), FromN: EW); |
1243 | else if (F->second.Type == ExtType::ZExt) |
1244 | Res = eZXT(A1: getCell(RR: RD, M: Outputs), FromN: EW); |
1245 | |
1246 | putCell(RR: RD, RC: Res, M&: Outputs); |
1247 | return true; |
1248 | } |
1249 | |
1250 | unsigned HexagonEvaluator::getNextPhysReg(unsigned PReg, unsigned Width) const { |
1251 | using namespace Hexagon; |
1252 | |
1253 | bool Is64 = DoubleRegsRegClass.contains(Reg: PReg); |
1254 | assert(PReg == 0 || Is64 || IntRegsRegClass.contains(PReg)); |
1255 | |
1256 | static const unsigned Phys32[] = { R0, R1, R2, R3, R4, R5 }; |
1257 | static const unsigned Phys64[] = { D0, D1, D2 }; |
1258 | const unsigned Num32 = sizeof(Phys32)/sizeof(unsigned); |
1259 | const unsigned Num64 = sizeof(Phys64)/sizeof(unsigned); |
1260 | |
1261 | // Return the first parameter register of the required width. |
1262 | if (PReg == 0) |
1263 | return (Width <= 32) ? Phys32[0] : Phys64[0]; |
1264 | |
1265 | // Set Idx32, Idx64 in such a way that Idx+1 would give the index of the |
1266 | // next register. |
1267 | unsigned Idx32 = 0, Idx64 = 0; |
1268 | if (!Is64) { |
1269 | while (Idx32 < Num32) { |
1270 | if (Phys32[Idx32] == PReg) |
1271 | break; |
1272 | Idx32++; |
1273 | } |
1274 | Idx64 = Idx32/2; |
1275 | } else { |
1276 | while (Idx64 < Num64) { |
1277 | if (Phys64[Idx64] == PReg) |
1278 | break; |
1279 | Idx64++; |
1280 | } |
1281 | Idx32 = Idx64*2+1; |
1282 | } |
1283 | |
1284 | if (Width <= 32) |
1285 | return (Idx32+1 < Num32) ? Phys32[Idx32+1] : 0; |
1286 | return (Idx64+1 < Num64) ? Phys64[Idx64+1] : 0; |
1287 | } |
1288 | |
1289 | unsigned HexagonEvaluator::getVirtRegFor(unsigned PReg) const { |
1290 | for (std::pair<unsigned,unsigned> P : MRI.liveins()) |
1291 | if (P.first == PReg) |
1292 | return P.second; |
1293 | return 0; |
1294 | } |
1295 | |