js/src/jit/shared/MoveEmitter-x86-shared.cpp@b8a032363ba2
js/src/jit/shared/MoveEmitter-x86-shared.cpp
Thu, 22 Jan 2015 13:21:57 +0100
- author
- Michael Schloh von Bennewitz <michael@schloh.com>
- date
- Thu, 22 Jan 2015 13:21:57 +0100
- branch
- TOR_BUG_9701
- changeset 15
- b8a032363ba2
- permissions
- -rw-r--r--
Incorporate requested changes from Mozilla in review:
https://bugzilla.mozilla.org/show_bug.cgi?id=1123480#c6
1 /* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 4 -*-
2 * vim: set ts=8 sts=4 et sw=4 tw=99:
3 * This Source Code Form is subject to the terms of the Mozilla Public
4 * License, v. 2.0. If a copy of the MPL was not distributed with this
5 * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
7 #include "jit/shared/MoveEmitter-x86-shared.h"
9 using namespace js;
10 using namespace js::jit;
12 MoveEmitterX86::MoveEmitterX86(MacroAssemblerSpecific &masm)
13 : inCycle_(false),
14 masm(masm),
15 pushedAtCycle_(-1)
16 {
17 pushedAtStart_ = masm.framePushed();
18 }
20 // Examine the cycle in moves starting at position i. Determine if it's a
21 // simple cycle consisting of all register-to-register moves in a single class,
22 // and whether it can be implemented entirely by swaps.
23 size_t
24 MoveEmitterX86::characterizeCycle(const MoveResolver &moves, size_t i,
25 bool *allGeneralRegs, bool *allFloatRegs)
26 {
27 size_t swapCount = 0;
29 for (size_t j = i; ; j++) {
30 const MoveOp &move = moves.getMove(j);
32 // If it isn't a cycle of registers of the same kind, we won't be able
33 // to optimize it.
34 if (!move.to().isGeneralReg())
35 *allGeneralRegs = false;
36 if (!move.to().isFloatReg())
37 *allFloatRegs = false;
38 if (!*allGeneralRegs && !*allFloatRegs)
39 return -1;
41 // Stop iterating when we see the last one.
42 if (j != i && move.isCycleEnd())
43 break;
45 // Check that this move is actually part of the cycle. This is
46 // over-conservative when there are multiple reads from the same source,
47 // but that's expected to be rare.
48 if (move.from() != moves.getMove(j + 1).to()) {
49 *allGeneralRegs = false;
50 *allFloatRegs = false;
51 return -1;
52 }
54 swapCount++;
55 }
57 // Check that the last move cycles back to the first move.
58 const MoveOp &move = moves.getMove(i + swapCount);
59 if (move.from() != moves.getMove(i).to()) {
60 *allGeneralRegs = false;
61 *allFloatRegs = false;
62 return -1;
63 }
65 return swapCount;
66 }
68 // If we can emit optimized code for the cycle in moves starting at position i,
69 // do so, and return true.
70 bool
71 MoveEmitterX86::maybeEmitOptimizedCycle(const MoveResolver &moves, size_t i,
72 bool allGeneralRegs, bool allFloatRegs, size_t swapCount)
73 {
74 if (allGeneralRegs && swapCount <= 2) {
75 // Use x86's swap-integer-registers instruction if we only have a few
76 // swaps. (x86 also has a swap between registers and memory but it's
77 // slow.)
78 for (size_t k = 0; k < swapCount; k++)
79 masm.xchg(moves.getMove(i + k).to().reg(), moves.getMove(i + k + 1).to().reg());
80 return true;
81 }
83 if (allFloatRegs && swapCount == 1) {
84 // There's no xchg for xmm registers, but if we only need a single swap,
85 // it's cheap to do an XOR swap.
86 FloatRegister a = moves.getMove(i).to().floatReg();
87 FloatRegister b = moves.getMove(i + 1).to().floatReg();
88 masm.xorpd(a, b);
89 masm.xorpd(b, a);
90 masm.xorpd(a, b);
91 return true;
92 }
94 return false;
95 }
97 void
98 MoveEmitterX86::emit(const MoveResolver &moves)
99 {
100 for (size_t i = 0; i < moves.numMoves(); i++) {
101 const MoveOp &move = moves.getMove(i);
102 const MoveOperand &from = move.from();
103 const MoveOperand &to = move.to();
105 if (move.isCycleEnd()) {
106 JS_ASSERT(inCycle_);
107 completeCycle(to, move.type());
108 inCycle_ = false;
109 continue;
110 }
112 if (move.isCycleBegin()) {
113 JS_ASSERT(!inCycle_);
115 // Characterize the cycle.
116 bool allGeneralRegs = true, allFloatRegs = true;
117 size_t swapCount = characterizeCycle(moves, i, &allGeneralRegs, &allFloatRegs);
119 // Attempt to optimize it to avoid using the stack.
120 if (maybeEmitOptimizedCycle(moves, i, allGeneralRegs, allFloatRegs, swapCount)) {
121 i += swapCount;
122 continue;
123 }
125 // Otherwise use the stack.
126 breakCycle(to, move.endCycleType());
127 inCycle_ = true;
128 }
130 // A normal move which is not part of a cycle.
131 switch (move.type()) {
132 case MoveOp::FLOAT32:
133 emitFloat32Move(from, to);
134 break;
135 case MoveOp::DOUBLE:
136 emitDoubleMove(from, to);
137 break;
138 case MoveOp::INT32:
139 emitInt32Move(from, to);
140 break;
141 case MoveOp::GENERAL:
142 emitGeneralMove(from, to);
143 break;
144 default:
145 MOZ_ASSUME_UNREACHABLE("Unexpected move type");
146 }
147 }
148 }
150 MoveEmitterX86::~MoveEmitterX86()
151 {
152 assertDone();
153 }
155 Address
156 MoveEmitterX86::cycleSlot()
157 {
158 if (pushedAtCycle_ == -1) {
159 // Reserve stack for cycle resolution
160 masm.reserveStack(sizeof(double));
161 pushedAtCycle_ = masm.framePushed();
162 }
164 return Address(StackPointer, masm.framePushed() - pushedAtCycle_);
165 }
167 Address
168 MoveEmitterX86::toAddress(const MoveOperand &operand) const
169 {
170 if (operand.base() != StackPointer)
171 return Address(operand.base(), operand.disp());
173 JS_ASSERT(operand.disp() >= 0);
175 // Otherwise, the stack offset may need to be adjusted.
176 return Address(StackPointer, operand.disp() + (masm.framePushed() - pushedAtStart_));
177 }
179 // Warning, do not use the resulting operand with pop instructions, since they
180 // compute the effective destination address after altering the stack pointer.
181 // Use toPopOperand if an Operand is needed for a pop.
182 Operand
183 MoveEmitterX86::toOperand(const MoveOperand &operand) const
184 {
185 if (operand.isMemoryOrEffectiveAddress())
186 return Operand(toAddress(operand));
187 if (operand.isGeneralReg())
188 return Operand(operand.reg());
190 JS_ASSERT(operand.isFloatReg());
191 return Operand(operand.floatReg());
192 }
194 // This is the same as toOperand except that it computes an Operand suitable for
195 // use in a pop.
196 Operand
197 MoveEmitterX86::toPopOperand(const MoveOperand &operand) const
198 {
199 if (operand.isMemory()) {
200 if (operand.base() != StackPointer)
201 return Operand(operand.base(), operand.disp());
203 JS_ASSERT(operand.disp() >= 0);
205 // Otherwise, the stack offset may need to be adjusted.
206 // Note the adjustment by the stack slot here, to offset for the fact that pop
207 // computes its effective address after incrementing the stack pointer.
208 return Operand(StackPointer,
209 operand.disp() + (masm.framePushed() - sizeof(void *) - pushedAtStart_));
210 }
211 if (operand.isGeneralReg())
212 return Operand(operand.reg());
214 JS_ASSERT(operand.isFloatReg());
215 return Operand(operand.floatReg());
216 }
218 void
219 MoveEmitterX86::breakCycle(const MoveOperand &to, MoveOp::Type type)
220 {
221 // There is some pattern:
222 // (A -> B)
223 // (B -> A)
224 //
225 // This case handles (A -> B), which we reach first. We save B, then allow
226 // the original move to continue.
227 switch (type) {
228 case MoveOp::FLOAT32:
229 if (to.isMemory()) {
230 masm.loadFloat32(toAddress(to), ScratchFloatReg);
231 masm.storeFloat32(ScratchFloatReg, cycleSlot());
232 } else {
233 masm.storeFloat32(to.floatReg(), cycleSlot());
234 }
235 break;
236 case MoveOp::DOUBLE:
237 if (to.isMemory()) {
238 masm.loadDouble(toAddress(to), ScratchFloatReg);
239 masm.storeDouble(ScratchFloatReg, cycleSlot());
240 } else {
241 masm.storeDouble(to.floatReg(), cycleSlot());
242 }
243 break;
244 #ifdef JS_CODEGEN_X64
245 case MoveOp::INT32:
246 // x64 can't pop to a 32-bit destination, so don't push.
247 if (to.isMemory()) {
248 masm.load32(toAddress(to), ScratchReg);
249 masm.store32(ScratchReg, cycleSlot());
250 } else {
251 masm.store32(to.reg(), cycleSlot());
252 }
253 break;
254 #endif
255 #ifndef JS_CODEGEN_X64
256 case MoveOp::INT32:
257 #endif
258 case MoveOp::GENERAL:
259 masm.Push(toOperand(to));
260 break;
261 default:
262 MOZ_ASSUME_UNREACHABLE("Unexpected move type");
263 }
264 }
266 void
267 MoveEmitterX86::completeCycle(const MoveOperand &to, MoveOp::Type type)
268 {
269 // There is some pattern:
270 // (A -> B)
271 // (B -> A)
272 //
273 // This case handles (B -> A), which we reach last. We emit a move from the
274 // saved value of B, to A.
275 switch (type) {
276 case MoveOp::FLOAT32:
277 JS_ASSERT(pushedAtCycle_ != -1);
278 JS_ASSERT(pushedAtCycle_ - pushedAtStart_ >= sizeof(float));
279 if (to.isMemory()) {
280 masm.loadFloat32(cycleSlot(), ScratchFloatReg);
281 masm.storeFloat32(ScratchFloatReg, toAddress(to));
282 } else {
283 masm.loadFloat32(cycleSlot(), to.floatReg());
284 }
285 break;
286 case MoveOp::DOUBLE:
287 JS_ASSERT(pushedAtCycle_ != -1);
288 JS_ASSERT(pushedAtCycle_ - pushedAtStart_ >= sizeof(double));
289 if (to.isMemory()) {
290 masm.loadDouble(cycleSlot(), ScratchFloatReg);
291 masm.storeDouble(ScratchFloatReg, toAddress(to));
292 } else {
293 masm.loadDouble(cycleSlot(), to.floatReg());
294 }
295 break;
296 #ifdef JS_CODEGEN_X64
297 case MoveOp::INT32:
298 JS_ASSERT(pushedAtCycle_ != -1);
299 JS_ASSERT(pushedAtCycle_ - pushedAtStart_ >= sizeof(int32_t));
300 // x64 can't pop to a 32-bit destination.
301 if (to.isMemory()) {
302 masm.load32(cycleSlot(), ScratchReg);
303 masm.store32(ScratchReg, toAddress(to));
304 } else {
305 masm.load32(cycleSlot(), to.reg());
306 }
307 break;
308 #endif
309 #ifndef JS_CODEGEN_X64
310 case MoveOp::INT32:
311 #endif
312 case MoveOp::GENERAL:
313 JS_ASSERT(masm.framePushed() - pushedAtStart_ >= sizeof(intptr_t));
314 masm.Pop(toPopOperand(to));
315 break;
316 default:
317 MOZ_ASSUME_UNREACHABLE("Unexpected move type");
318 }
319 }
321 void
322 MoveEmitterX86::emitInt32Move(const MoveOperand &from, const MoveOperand &to)
323 {
324 if (from.isGeneralReg()) {
325 masm.move32(from.reg(), toOperand(to));
326 } else if (to.isGeneralReg()) {
327 JS_ASSERT(from.isMemory());
328 masm.load32(toAddress(from), to.reg());
329 } else {
330 // Memory to memory gpr move.
331 JS_ASSERT(from.isMemory());
332 #ifdef JS_CODEGEN_X64
333 // x64 has a ScratchReg. Use it.
334 masm.load32(toAddress(from), ScratchReg);
335 masm.move32(ScratchReg, toOperand(to));
336 #else
337 // No ScratchReg; bounce it off the stack.
338 masm.Push(toOperand(from));
339 masm.Pop(toPopOperand(to));
340 #endif
341 }
342 }
344 void
345 MoveEmitterX86::emitGeneralMove(const MoveOperand &from, const MoveOperand &to)
346 {
347 if (from.isGeneralReg()) {
348 masm.mov(from.reg(), toOperand(to));
349 } else if (to.isGeneralReg()) {
350 JS_ASSERT(from.isMemoryOrEffectiveAddress());
351 if (from.isMemory())
352 masm.loadPtr(toAddress(from), to.reg());
353 else
354 masm.lea(toOperand(from), to.reg());
355 } else if (from.isMemory()) {
356 // Memory to memory gpr move.
357 #ifdef JS_CODEGEN_X64
358 // x64 has a ScratchReg. Use it.
359 masm.loadPtr(toAddress(from), ScratchReg);
360 masm.mov(ScratchReg, toOperand(to));
361 #else
362 // No ScratchReg; bounce it off the stack.
363 masm.Push(toOperand(from));
364 masm.Pop(toPopOperand(to));
365 #endif
366 } else {
367 // Effective address to memory move.
368 JS_ASSERT(from.isEffectiveAddress());
369 #ifdef JS_CODEGEN_X64
370 // x64 has a ScratchReg. Use it.
371 masm.lea(toOperand(from), ScratchReg);
372 masm.mov(ScratchReg, toOperand(to));
373 #else
374 // This is tricky without a ScratchReg. We can't do an lea. Bounce the
375 // base register off the stack, then add the offset in place. Note that
376 // this clobbers FLAGS!
377 masm.Push(from.base());
378 masm.Pop(toPopOperand(to));
379 masm.addPtr(Imm32(from.disp()), toOperand(to));
380 #endif
381 }
382 }
384 void
385 MoveEmitterX86::emitFloat32Move(const MoveOperand &from, const MoveOperand &to)
386 {
387 if (from.isFloatReg()) {
388 if (to.isFloatReg())
389 masm.moveFloat32(from.floatReg(), to.floatReg());
390 else
391 masm.storeFloat32(from.floatReg(), toAddress(to));
392 } else if (to.isFloatReg()) {
393 masm.loadFloat32(toAddress(from), to.floatReg());
394 } else {
395 // Memory to memory move.
396 JS_ASSERT(from.isMemory());
397 masm.loadFloat32(toAddress(from), ScratchFloatReg);
398 masm.storeFloat32(ScratchFloatReg, toAddress(to));
399 }
400 }
402 void
403 MoveEmitterX86::emitDoubleMove(const MoveOperand &from, const MoveOperand &to)
404 {
405 if (from.isFloatReg()) {
406 if (to.isFloatReg())
407 masm.moveDouble(from.floatReg(), to.floatReg());
408 else
409 masm.storeDouble(from.floatReg(), toAddress(to));
410 } else if (to.isFloatReg()) {
411 masm.loadDouble(toAddress(from), to.floatReg());
412 } else {
413 // Memory to memory move.
414 JS_ASSERT(from.isMemory());
415 masm.loadDouble(toAddress(from), ScratchFloatReg);
416 masm.storeDouble(ScratchFloatReg, toAddress(to));
417 }
418 }
420 void
421 MoveEmitterX86::assertDone()
422 {
423 JS_ASSERT(!inCycle_);
424 }
426 void
427 MoveEmitterX86::finish()
428 {
429 assertDone();
431 masm.freeStack(masm.framePushed() - pushedAtStart_);
432 }
