1 //===- SSEDomainFix.cpp - Use proper int/float domain for SSE ---*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains the SSEDomainFix pass.
12 // Some SSE instructions like mov, and, or, xor are available in different
13 // variants for different operand types. These variant instructions are
14 // equivalent, but on Nehalem and newer cpus there is extra latency
15 // transferring data between integer and floating point domains.
17 // This pass changes the variant instructions to minimize domain crossings.
19 //===----------------------------------------------------------------------===//
21 #define DEBUG_TYPE "sse-domain-fix"
22 #include "X86InstrInfo.h"
23 #include "llvm/CodeGen/MachineFunctionPass.h"
24 #include "llvm/CodeGen/MachineRegisterInfo.h"
25 #include "llvm/ADT/DepthFirstIterator.h"
26 #include "llvm/Support/Allocator.h"
27 #include "llvm/Support/Debug.h"
28 #include "llvm/Support/raw_ostream.h"
31 /// A DomainValue is a bit like LiveIntervals' ValNo, but it also keeps track
32 /// of execution domains.
34 /// An open DomainValue represents a set of instructions that can still switch
35 /// execution domain. Multiple registers may refer to the same open
36 /// DomainValue - they will eventually be collapsed to the same execution
39 /// A collapsed DomainValue represents a single register that has been forced
40 /// into one of more execution domains. There is a separate collapsed
41 /// DomainValue for each register, but it may contain multiple execution
42 /// domains. A register value is initially created in a single execution
43 /// domain, but if we were forced to pay the penalty of a domain crossing, we
44 /// keep track of the fact the the register is now available in multiple
48 // Basic reference counting.
51 // Bitmask of available domains. For an open DomainValue, it is the still
52 // possible domains for collapsing. For a collapsed DomainValue it is the
53 // domains where the register is available for free.
54 unsigned AvailableDomains;
56 // Position of the last defining instruction.
59 // Twiddleable instructions using or defining these registers.
60 SmallVector<MachineInstr*, 8> Instrs;
62 // A collapsed DomainValue has no instructions to twiddle - it simply keeps
63 // track of the domains where the registers are already available.
64 bool isCollapsed() const { return Instrs.empty(); }
66 // Is domain available?
67 bool hasDomain(unsigned domain) const {
68 return AvailableDomains & (1u << domain);
71 // Mark domain as available.
72 void addDomain(unsigned domain) {
73 AvailableDomains |= 1u << domain;
76 // Restrict to a single domain available.
77 void setSingleDomain(unsigned domain) {
78 AvailableDomains = 1u << domain;
81 // Return bitmask of domains that are available and in mask.
82 unsigned getCommonDomains(unsigned mask) const {
83 return AvailableDomains & mask;
86 // First domain available.
87 unsigned getFirstDomain() const {
88 return CountTrailingZeros_32(AvailableDomains);
91 DomainValue() { clear(); }
94 Refs = AvailableDomains = Dist = 0;
100 static const unsigned NumRegs = 16;
103 class SSEDomainFixPass : public MachineFunctionPass {
105 SpecificBumpPtrAllocator<DomainValue> Allocator;
106 SmallVector<DomainValue*,16> Avail;
109 const X86InstrInfo *TII;
110 const TargetRegisterInfo *TRI;
111 MachineBasicBlock *MBB;
112 DomainValue **LiveRegs;
113 typedef DenseMap<MachineBasicBlock*,DomainValue**> LiveOutMap;
118 SSEDomainFixPass() : MachineFunctionPass(&ID) {}
120 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
121 AU.setPreservesAll();
122 MachineFunctionPass::getAnalysisUsage(AU);
125 virtual bool runOnMachineFunction(MachineFunction &MF);
127 virtual const char *getPassName() const {
128 return "SSE execution domain fixup";
133 int RegIndex(unsigned Reg);
135 // DomainValue allocation.
136 DomainValue *Alloc(int domain = -1);
137 void Recycle(DomainValue*);
139 // LiveRegs manipulations.
140 void SetLiveReg(int rx, DomainValue *DV);
142 void Force(int rx, unsigned domain);
143 void Collapse(DomainValue *dv, unsigned domain);
144 bool Merge(DomainValue *A, DomainValue *B);
146 void enterBasicBlock();
147 void visitGenericInstr(MachineInstr*);
148 void visitSoftInstr(MachineInstr*, unsigned mask);
149 void visitHardInstr(MachineInstr*, unsigned domain);
153 char SSEDomainFixPass::ID = 0;
155 /// Translate TRI register number to an index into our smaller tables of
156 /// interesting registers. Return -1 for boring registers.
157 int SSEDomainFixPass::RegIndex(unsigned reg) {
158 // Registers are sorted lexicographically.
159 // We just need them to be consecutive, ordering doesn't matter.
160 assert(X86::XMM9 == X86::XMM0+NumRegs-1 && "Unexpected sort");
162 return reg < NumRegs ? (int) reg : -1;
165 DomainValue *SSEDomainFixPass::Alloc(int domain) {
166 DomainValue *dv = Avail.empty() ?
167 new(Allocator.Allocate()) DomainValue :
168 Avail.pop_back_val();
171 dv->addDomain(domain);
175 void SSEDomainFixPass::Recycle(DomainValue *dv) {
176 assert(dv && "Cannot recycle NULL");
181 /// Set LiveRegs[rx] = dv, updating reference counts.
182 void SSEDomainFixPass::SetLiveReg(int rx, DomainValue *dv) {
183 assert(unsigned(rx) < NumRegs && "Invalid index");
185 LiveRegs = new DomainValue*[NumRegs];
186 std::fill(LiveRegs, LiveRegs+NumRegs, (DomainValue*)0);
189 if (LiveRegs[rx] == dv)
192 assert(LiveRegs[rx]->Refs && "Bad refcount");
193 if (--LiveRegs[rx]->Refs == 0) Recycle(LiveRegs[rx]);
199 // Kill register rx, recycle or collapse any DomainValue.
200 void SSEDomainFixPass::Kill(int rx) {
201 assert(unsigned(rx) < NumRegs && "Invalid index");
202 if (!LiveRegs || !LiveRegs[rx]) return;
204 // Before killing the last reference to an open DomainValue, collapse it to
205 // the first available domain.
206 if (LiveRegs[rx]->Refs == 1 && !LiveRegs[rx]->isCollapsed())
207 Collapse(LiveRegs[rx], LiveRegs[rx]->getFirstDomain());
212 /// Force register rx into domain.
213 void SSEDomainFixPass::Force(int rx, unsigned domain) {
214 assert(unsigned(rx) < NumRegs && "Invalid index");
216 if (LiveRegs && (dv = LiveRegs[rx])) {
217 if (dv->isCollapsed())
218 dv->addDomain(domain);
219 else if (dv->hasDomain(domain))
220 Collapse(dv, domain);
222 // This is an incompatible open DomainValue. Collapse it to whatever and force
223 // the new value into domain. This costs a domain crossing.
224 Collapse(dv, dv->getFirstDomain());
225 assert(LiveRegs[rx] && "Not live after collapse?");
226 LiveRegs[rx]->addDomain(domain);
229 // Set up basic collapsed DomainValue.
230 SetLiveReg(rx, Alloc(domain));
234 /// Collapse open DomainValue into given domain. If there are multiple
235 /// registers using dv, they each get a unique collapsed DomainValue.
236 void SSEDomainFixPass::Collapse(DomainValue *dv, unsigned domain) {
237 assert(dv->hasDomain(domain) && "Cannot collapse");
239 // Collapse all the instructions.
240 while (!dv->Instrs.empty())
241 TII->SetSSEDomain(dv->Instrs.pop_back_val(), domain);
242 dv->setSingleDomain(domain);
244 // If there are multiple users, give them new, unique DomainValues.
245 if (LiveRegs && dv->Refs > 1)
246 for (unsigned rx = 0; rx != NumRegs; ++rx)
247 if (LiveRegs[rx] == dv)
248 SetLiveReg(rx, Alloc(domain));
251 /// Merge - All instructions and registers in B are moved to A, and B is
253 bool SSEDomainFixPass::Merge(DomainValue *A, DomainValue *B) {
254 assert(!A->isCollapsed() && "Cannot merge into collapsed");
255 assert(!B->isCollapsed() && "Cannot merge from collapsed");
258 // Restrict to the domains that A and B have in common.
259 unsigned common = A->getCommonDomains(B->AvailableDomains);
262 A->AvailableDomains = common;
263 A->Dist = std::max(A->Dist, B->Dist);
264 A->Instrs.append(B->Instrs.begin(), B->Instrs.end());
265 for (unsigned rx = 0; rx != NumRegs; ++rx)
266 if (LiveRegs[rx] == B)
271 void SSEDomainFixPass::enterBasicBlock() {
272 // Try to coalesce live-out registers from predecessors.
273 for (MachineBasicBlock::livein_iterator i = MBB->livein_begin(),
274 e = MBB->livein_end(); i != e; ++i) {
275 int rx = RegIndex(*i);
276 if (rx < 0) continue;
277 for (MachineBasicBlock::const_pred_iterator pi = MBB->pred_begin(),
278 pe = MBB->pred_end(); pi != pe; ++pi) {
279 LiveOutMap::const_iterator fi = LiveOuts.find(*pi);
280 if (fi == LiveOuts.end()) continue;
281 DomainValue *pdv = fi->second[rx];
283 if (!LiveRegs || !LiveRegs[rx]) {
288 // We have a live DomainValue from more than one predecessor.
289 if (LiveRegs[rx]->isCollapsed()) {
290 // We are already collapsed, but predecessor is not. Force him.
291 unsigned domain = LiveRegs[rx]->getFirstDomain();
292 if (!pdv->isCollapsed() && pdv->hasDomain(domain))
293 Collapse(pdv, domain);
297 // Currently open, merge in predecessor.
298 if (!pdv->isCollapsed())
299 Merge(LiveRegs[rx], pdv);
301 Force(rx, pdv->getFirstDomain());
306 // A hard instruction only works in one domain. All input registers will be
307 // forced into that domain.
308 void SSEDomainFixPass::visitHardInstr(MachineInstr *mi, unsigned domain) {
309 // Collapse all uses.
310 for (unsigned i = mi->getDesc().getNumDefs(),
311 e = mi->getDesc().getNumOperands(); i != e; ++i) {
312 MachineOperand &mo = mi->getOperand(i);
313 if (!mo.isReg()) continue;
314 int rx = RegIndex(mo.getReg());
315 if (rx < 0) continue;
319 // Kill all defs and force them.
320 for (unsigned i = 0, e = mi->getDesc().getNumDefs(); i != e; ++i) {
321 MachineOperand &mo = mi->getOperand(i);
322 if (!mo.isReg()) continue;
323 int rx = RegIndex(mo.getReg());
324 if (rx < 0) continue;
330 // A soft instruction can be changed to work in other domains given by mask.
331 void SSEDomainFixPass::visitSoftInstr(MachineInstr *mi, unsigned mask) {
332 // Bitmask of available domains for this instruction after taking collapsed
333 // operands into account.
334 unsigned available = mask;
336 // Scan the explicit use operands for incoming domains.
337 SmallVector<int, 4> used;
339 for (unsigned i = mi->getDesc().getNumDefs(),
340 e = mi->getDesc().getNumOperands(); i != e; ++i) {
341 MachineOperand &mo = mi->getOperand(i);
342 if (!mo.isReg()) continue;
343 int rx = RegIndex(mo.getReg());
344 if (rx < 0) continue;
345 if (DomainValue *dv = LiveRegs[rx]) {
346 // Bitmask of domains that dv and available have in common.
347 unsigned common = dv->getCommonDomains(available);
348 // Is it possible to use this collapsed register for free?
349 if (dv->isCollapsed()) {
350 // Restrict available domains to the ones in common with the operand.
351 // If there are no common domains, we must pay the cross-domain
352 // penalty for this operand.
353 if (common) available = common;
355 // Open DomainValue is compatible, save it for merging.
358 // Open DomainValue is not compatible with instruction. It is useless
364 // If the collapsed operands force a single domain, propagate the collapse.
365 if (isPowerOf2_32(available)) {
366 unsigned domain = CountTrailingZeros_32(available);
367 TII->SetSSEDomain(mi, domain);
368 visitHardInstr(mi, domain);
372 // Kill off any remaining uses that don't match available, and build a list of
373 // incoming DomainValues that we want to merge.
374 SmallVector<DomainValue*,4> doms;
375 for (SmallVector<int, 4>::iterator i=used.begin(), e=used.end(); i!=e; ++i) {
377 DomainValue *dv = LiveRegs[rx];
378 // This useless DomainValue could have been missed above.
379 if (!dv->getCommonDomains(available)) {
383 // sorted, uniqued insert.
384 bool inserted = false;
385 for (SmallVector<DomainValue*,4>::iterator i = doms.begin(), e = doms.end();
386 i != e && !inserted; ++i) {
389 else if (dv->Dist < (*i)->Dist) {
398 // doms are now sorted in order of appearance. Try to merge them all, giving
399 // priority to the latest ones.
401 while (!doms.empty()) {
403 dv = doms.pop_back_val();
407 DomainValue *latest = doms.pop_back_val();
408 if (Merge(dv, latest)) continue;
410 // If latest didn't merge, it is useless now. Kill all registers using it.
411 for (SmallVector<int,4>::iterator i=used.begin(), e=used.end(); i != e; ++i)
412 if (LiveRegs[*i] == latest)
416 // dv is the DomainValue we are going to use for this instruction.
420 dv->AvailableDomains = available;
421 dv->Instrs.push_back(mi);
423 // Finally set all defs and non-collapsed uses to dv.
424 for (unsigned i = 0, e = mi->getDesc().getNumOperands(); i != e; ++i) {
425 MachineOperand &mo = mi->getOperand(i);
426 if (!mo.isReg()) continue;
427 int rx = RegIndex(mo.getReg());
428 if (rx < 0) continue;
429 if (!LiveRegs || !LiveRegs[rx] || (mo.isDef() && LiveRegs[rx]!=dv)) {
436 void SSEDomainFixPass::visitGenericInstr(MachineInstr *mi) {
437 // Process explicit defs, kill any XMM registers redefined.
438 for (unsigned i = 0, e = mi->getDesc().getNumDefs(); i != e; ++i) {
439 MachineOperand &mo = mi->getOperand(i);
440 if (!mo.isReg()) continue;
441 int rx = RegIndex(mo.getReg());
442 if (rx < 0) continue;
447 bool SSEDomainFixPass::runOnMachineFunction(MachineFunction &mf) {
449 TII = static_cast<const X86InstrInfo*>(MF->getTarget().getInstrInfo());
450 TRI = MF->getTarget().getRegisterInfo();
454 assert(NumRegs == X86::VR128RegClass.getNumRegs() && "Bad regclass");
456 // If no XMM registers are used in the function, we can skip it completely.
457 bool anyregs = false;
458 for (TargetRegisterClass::const_iterator I = X86::VR128RegClass.begin(),
459 E = X86::VR128RegClass.end(); I != E; ++I)
460 if (MF->getRegInfo().isPhysRegUsed(*I)) {
464 if (!anyregs) return false;
466 MachineBasicBlock *Entry = MF->begin();
467 SmallPtrSet<MachineBasicBlock*, 16> Visited;
468 for (df_ext_iterator<MachineBasicBlock*, SmallPtrSet<MachineBasicBlock*, 16> >
469 DFI = df_ext_begin(Entry, Visited), DFE = df_ext_end(Entry, Visited);
473 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); I != E;
475 MachineInstr *mi = I;
476 if (mi->isDebugValue()) continue;
478 std::pair<uint16_t, uint16_t> domp = TII->GetSSEDomain(mi);
481 visitSoftInstr(mi, domp.second);
483 visitHardInstr(mi, domp.first);
485 visitGenericInstr(mi);
488 // Save live registers at end of MBB - used by enterBasicBlock().
490 LiveOuts.insert(std::make_pair(MBB, LiveRegs));
494 // Clear the LiveOuts vectors. Should we also collapse any remaining
496 for (LiveOutMap::const_iterator i = LiveOuts.begin(), e = LiveOuts.end();
501 Allocator.DestroyAll();
506 FunctionPass *llvm::createSSEDomainFixPass() {
507 return new SSEDomainFixPass();