1 //===--------- SCEVAffinator.cpp  - Create Scops from LLVM IR -------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // Create a polyhedral description for a SCEV value.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "polly/Support/SCEVAffinator.h"
15 #include "polly/Options.h"
16 #include "polly/ScopInfo.h"
17 #include "polly/Support/GICHelper.h"
18 #include "polly/Support/SCEVValidator.h"
19 #include "polly/Support/ScopHelper.h"
20 #include "isl/aff.h"
21 #include "isl/local_space.h"
22 #include "isl/set.h"
23 #include "isl/val.h"
24 
25 using namespace llvm;
26 using namespace polly;
27 
28 static cl::opt<bool> IgnoreIntegerWrapping(
29     "polly-ignore-integer-wrapping",
30     cl::desc("Do not build run-time checks to proof absence of integer "
31              "wrapping"),
32     cl::Hidden, cl::ZeroOrMore, cl::init(false), cl::cat(PollyCategory));
33 
34 // The maximal number of basic sets we allow during the construction of a
35 // piecewise affine function. More complex ones will result in very high
36 // compile time.
37 static int const MaxDisjunctionsInPwAff = 100;
38 
39 // The maximal number of bits for which a zero-extend is modeled precisely.
40 static unsigned const MaxZextSmallBitWidth = 7;
41 
42 // The maximal number of bits for which a truncate is modeled precisely.
43 static unsigned const MaxTruncateSmallBitWidth = 31;
44 
45 /// Return true if a zero-extend from @p Width bits is precisely modeled.
46 static bool isPreciseZeroExtend(unsigned Width) {
47   return Width <= MaxZextSmallBitWidth;
48 }
49 
50 /// Return true if a truncate from @p Width bits is precisely modeled.
51 static bool isPreciseTruncate(unsigned Width) {
52   return Width <= MaxTruncateSmallBitWidth;
53 }
54 
55 /// Add the number of basic sets in @p Domain to @p User
56 static isl_stat addNumBasicSets(__isl_take isl_set *Domain,
57                                 __isl_take isl_aff *Aff, void *User) {
58   auto *NumBasicSets = static_cast<unsigned *>(User);
59   *NumBasicSets += isl_set_n_basic_set(Domain);
60   isl_set_free(Domain);
61   isl_aff_free(Aff);
62   return isl_stat_ok;
63 }
64 
65 /// Helper to free a PWACtx object.
66 static void freePWACtx(__isl_take PWACtx &PWAC) {
67   isl_pw_aff_free(PWAC.first);
68   isl_set_free(PWAC.second);
69 }
70 
71 /// Helper to copy a PWACtx object.
72 static __isl_give PWACtx copyPWACtx(const __isl_keep PWACtx &PWAC) {
73   return std::make_pair(isl_pw_aff_copy(PWAC.first), isl_set_copy(PWAC.second));
74 }
75 
76 /// Determine if @p PWAC is too complex to continue.
77 ///
78 /// Note that @p PWAC will be "free" (deallocated) if this function returns
79 /// true, but not if this function returns false.
80 static bool isTooComplex(PWACtx &PWAC) {
81   unsigned NumBasicSets = 0;
82   isl_pw_aff_foreach_piece(PWAC.first, addNumBasicSets, &NumBasicSets);
83   if (NumBasicSets <= MaxDisjunctionsInPwAff)
84     return false;
85   freePWACtx(PWAC);
86   return true;
87 }
88 
89 /// Return the flag describing the possible wrapping of @p Expr.
90 static SCEV::NoWrapFlags getNoWrapFlags(const SCEV *Expr) {
91   if (auto *NAry = dyn_cast<SCEVNAryExpr>(Expr))
92     return NAry->getNoWrapFlags();
93   return SCEV::NoWrapMask;
94 }
95 
96 static void combine(__isl_keep PWACtx &PWAC0, const __isl_take PWACtx &PWAC1,
97                     isl_pw_aff *(Fn)(isl_pw_aff *, isl_pw_aff *)) {
98   PWAC0.first = Fn(PWAC0.first, PWAC1.first);
99   PWAC0.second = isl_set_union(PWAC0.second, PWAC1.second);
100 }
101 
102 /// Set the possible wrapping of @p Expr to @p Flags.
103 static const SCEV *setNoWrapFlags(ScalarEvolution &SE, const SCEV *Expr,
104                                   SCEV::NoWrapFlags Flags) {
105   auto *NAry = dyn_cast<SCEVNAryExpr>(Expr);
106   if (!NAry)
107     return Expr;
108 
109   SmallVector<const SCEV *, 8> Ops(NAry->op_begin(), NAry->op_end());
110   switch (Expr->getSCEVType()) {
111   case scAddExpr:
112     return SE.getAddExpr(Ops, Flags);
113   case scMulExpr:
114     return SE.getMulExpr(Ops, Flags);
115   case scAddRecExpr:
116     return SE.getAddRecExpr(Ops, cast<SCEVAddRecExpr>(Expr)->getLoop(), Flags);
117   default:
118     return Expr;
119   }
120 }
121 
122 static __isl_give isl_pw_aff *getWidthExpValOnDomain(unsigned Width,
123                                                      __isl_take isl_set *Dom) {
124   auto *Ctx = isl_set_get_ctx(Dom);
125   auto *WidthVal = isl_val_int_from_ui(Ctx, Width);
126   auto *ExpVal = isl_val_2exp(WidthVal);
127   return isl_pw_aff_val_on_domain(Dom, ExpVal);
128 }
129 
130 SCEVAffinator::SCEVAffinator(Scop *S, LoopInfo &LI)
131     : S(S), Ctx(S->getIslCtx()), SE(*S->getSE()), LI(LI),
132       TD(S->getFunction().getParent()->getDataLayout()) {}
133 
134 SCEVAffinator::~SCEVAffinator() {
135   for (auto &CachedPair : CachedExpressions)
136     freePWACtx(CachedPair.second);
137 }
138 
139 Loop *SCEVAffinator::getScope() { return BB ? LI.getLoopFor(BB) : nullptr; }
140 
141 void SCEVAffinator::interpretAsUnsigned(__isl_keep PWACtx &PWAC,
142                                         unsigned Width) {
143   auto *PWA = PWAC.first;
144   auto *NonNegDom = isl_pw_aff_nonneg_set(isl_pw_aff_copy(PWA));
145   auto *NonNegPWA = isl_pw_aff_intersect_domain(isl_pw_aff_copy(PWA),
146                                                 isl_set_copy(NonNegDom));
147   auto *ExpPWA = getWidthExpValOnDomain(Width, isl_set_complement(NonNegDom));
148   PWAC.first = isl_pw_aff_union_add(NonNegPWA, isl_pw_aff_add(PWA, ExpPWA));
149 }
150 
151 void SCEVAffinator::takeNonNegativeAssumption(PWACtx &PWAC) {
152   auto *NegPWA = isl_pw_aff_neg(isl_pw_aff_copy(PWAC.first));
153   auto *NegDom = isl_pw_aff_pos_set(NegPWA);
154   PWAC.second = isl_set_union(PWAC.second, isl_set_copy(NegDom));
155   auto *Restriction = BB ? NegDom : isl_set_params(NegDom);
156   auto DL = BB ? BB->getTerminator()->getDebugLoc() : DebugLoc();
157   S->recordAssumption(UNSIGNED, Restriction, DL, AS_RESTRICTION, BB);
158 }
159 
160 __isl_give PWACtx SCEVAffinator::getPWACtxFromPWA(__isl_take isl_pw_aff *PWA) {
161   return std::make_pair(
162       PWA, isl_set_empty(isl_space_set_alloc(Ctx, 0, NumIterators)));
163 }
164 
165 __isl_give PWACtx SCEVAffinator::getPwAff(const SCEV *Expr, BasicBlock *BB) {
166   this->BB = BB;
167 
168   if (BB) {
169     auto *DC = S->getDomainConditions(BB);
170     NumIterators = isl_set_n_dim(DC);
171     isl_set_free(DC);
172   } else
173     NumIterators = 0;
174 
175   auto *Scope = getScope();
176   S->addParams(getParamsInAffineExpr(&S->getRegion(), Scope, Expr, SE));
177 
178   return visit(Expr);
179 }
180 
181 __isl_give PWACtx SCEVAffinator::checkForWrapping(const SCEV *Expr,
182                                                   PWACtx PWAC) const {
183   // If the SCEV flags do contain NSW (no signed wrap) then PWA already
184   // represents Expr in modulo semantic (it is not allowed to overflow), thus we
185   // are done. Otherwise, we will compute:
186   //   PWA = ((PWA + 2^(n-1)) mod (2 ^ n)) - 2^(n-1)
187   // whereas n is the number of bits of the Expr, hence:
188   //   n = bitwidth(ExprType)
189 
190   if (IgnoreIntegerWrapping || (getNoWrapFlags(Expr) & SCEV::FlagNSW))
191     return PWAC;
192 
193   auto *PWA = PWAC.first;
194   auto *PWAMod = addModuloSemantic(isl_pw_aff_copy(PWA), Expr->getType());
195   auto *NotEqualSet = isl_pw_aff_ne_set(isl_pw_aff_copy(PWA), PWAMod);
196   PWAC.second = isl_set_union(PWAC.second, isl_set_copy(NotEqualSet));
197   PWAC.second = isl_set_coalesce(PWAC.second);
198 
199   const DebugLoc &Loc = BB ? BB->getTerminator()->getDebugLoc() : DebugLoc();
200   NotEqualSet = BB ? NotEqualSet : isl_set_params(NotEqualSet);
201   NotEqualSet = isl_set_coalesce(NotEqualSet);
202 
203   if (isl_set_is_empty(NotEqualSet))
204     isl_set_free(NotEqualSet);
205   else
206     S->recordAssumption(WRAPPING, NotEqualSet, Loc, AS_RESTRICTION, BB);
207 
208   return PWAC;
209 }
210 
211 __isl_give isl_pw_aff *
212 SCEVAffinator::addModuloSemantic(__isl_take isl_pw_aff *PWA,
213                                  Type *ExprType) const {
214   unsigned Width = TD.getTypeSizeInBits(ExprType);
215   isl_ctx *Ctx = isl_pw_aff_get_ctx(PWA);
216 
217   isl_val *ModVal = isl_val_int_from_ui(Ctx, Width);
218   ModVal = isl_val_2exp(ModVal);
219 
220   isl_set *Domain = isl_pw_aff_domain(isl_pw_aff_copy(PWA));
221   isl_pw_aff *AddPW = getWidthExpValOnDomain(Width - 1, Domain);
222 
223   PWA = isl_pw_aff_add(PWA, isl_pw_aff_copy(AddPW));
224   PWA = isl_pw_aff_mod_val(PWA, ModVal);
225   PWA = isl_pw_aff_sub(PWA, AddPW);
226 
227   return PWA;
228 }
229 
230 bool SCEVAffinator::hasNSWAddRecForLoop(Loop *L) const {
231   for (const auto &CachedPair : CachedExpressions) {
232     auto *AddRec = dyn_cast<SCEVAddRecExpr>(CachedPair.first.first);
233     if (!AddRec)
234       continue;
235     if (AddRec->getLoop() != L)
236       continue;
237     if (AddRec->getNoWrapFlags() & SCEV::FlagNSW)
238       return true;
239   }
240 
241   return false;
242 }
243 
244 __isl_give PWACtx SCEVAffinator::visit(const SCEV *Expr) {
245 
246   auto Key = std::make_pair(Expr, BB);
247   PWACtx PWAC = CachedExpressions[Key];
248   if (PWAC.first)
249     return copyPWACtx(PWAC);
250 
251   auto ConstantAndLeftOverPair = extractConstantFactor(Expr, SE);
252   auto *Factor = ConstantAndLeftOverPair.first;
253   Expr = ConstantAndLeftOverPair.second;
254 
255   // In case the scev is a valid parameter, we do not further analyze this
256   // expression, but create a new parameter in the isl_pw_aff. This allows us
257   // to treat subexpressions that we cannot translate into an piecewise affine
258   // expression, as constant parameters of the piecewise affine expression.
259   if (isl_id *Id = S->getIdForParam(Expr)) {
260     isl_space *Space = isl_space_set_alloc(Ctx, 1, NumIterators);
261     Space = isl_space_set_dim_id(Space, isl_dim_param, 0, Id);
262 
263     isl_set *Domain = isl_set_universe(isl_space_copy(Space));
264     isl_aff *Affine = isl_aff_zero_on_domain(isl_local_space_from_space(Space));
265     Affine = isl_aff_add_coefficient_si(Affine, isl_dim_param, 0, 1);
266 
267     PWAC = getPWACtxFromPWA(isl_pw_aff_alloc(Domain, Affine));
268   } else {
269     PWAC = SCEVVisitor<SCEVAffinator, PWACtx>::visit(Expr);
270     PWAC = checkForWrapping(Expr, PWAC);
271   }
272 
273   if (!Factor->getType()->isIntegerTy(1))
274     combine(PWAC, visitConstant(Factor), isl_pw_aff_mul);
275 
276   // For compile time reasons we need to simplify the PWAC before we cache and
277   // return it.
278   PWAC.first = isl_pw_aff_coalesce(PWAC.first);
279   PWAC = checkForWrapping(Key.first, PWAC);
280 
281   CachedExpressions[Key] = copyPWACtx(PWAC);
282   return PWAC;
283 }
284 
285 __isl_give PWACtx SCEVAffinator::visitConstant(const SCEVConstant *Expr) {
286   ConstantInt *Value = Expr->getValue();
287   isl_val *v;
288 
289   // LLVM does not define if an integer value is interpreted as a signed or
290   // unsigned value. Hence, without further information, it is unknown how
291   // this value needs to be converted to GMP. At the moment, we only support
292   // signed operations. So we just interpret it as signed. Later, there are
293   // two options:
294   //
295   // 1. We always interpret any value as signed and convert the values on
296   //    demand.
297   // 2. We pass down the signedness of the calculation and use it to interpret
298   //    this constant correctly.
299   v = isl_valFromAPInt(Ctx, Value->getValue(), /* isSigned */ true);
300 
301   isl_space *Space = isl_space_set_alloc(Ctx, 0, NumIterators);
302   isl_local_space *ls = isl_local_space_from_space(Space);
303   return getPWACtxFromPWA(isl_pw_aff_from_aff(isl_aff_val_on_domain(ls, v)));
304 }
305 
306 __isl_give PWACtx
307 SCEVAffinator::visitTruncateExpr(const SCEVTruncateExpr *Expr) {
308   // Truncate operations are basically modulo operations, thus we can
309   // model them that way. However, for large types we assume the operand
310   // to fit in the new type size instead of introducing a modulo with a very
311   // large constant.
312 
313   auto *Op = Expr->getOperand();
314   auto OpPWAC = visit(Op);
315 
316   unsigned Width = TD.getTypeSizeInBits(Expr->getType());
317   bool Precise = isPreciseTruncate(Width);
318 
319   if (Precise) {
320     OpPWAC.first = addModuloSemantic(OpPWAC.first, Expr->getType());
321     return OpPWAC;
322   }
323 
324   auto *Dom = isl_pw_aff_domain(isl_pw_aff_copy(OpPWAC.first));
325   auto *ExpPWA = getWidthExpValOnDomain(Width - 1, Dom);
326   auto *GreaterDom =
327       isl_pw_aff_ge_set(isl_pw_aff_copy(OpPWAC.first), isl_pw_aff_copy(ExpPWA));
328   auto *SmallerDom =
329       isl_pw_aff_lt_set(isl_pw_aff_copy(OpPWAC.first), isl_pw_aff_neg(ExpPWA));
330   auto *OutOfBoundsDom = isl_set_union(SmallerDom, GreaterDom);
331   OpPWAC.second = isl_set_union(OpPWAC.second, isl_set_copy(OutOfBoundsDom));
332   S->recordAssumption(UNSIGNED, OutOfBoundsDom, DebugLoc(), AS_RESTRICTION, BB);
333 
334   return OpPWAC;
335 }
336 
337 __isl_give PWACtx
338 SCEVAffinator::visitZeroExtendExpr(const SCEVZeroExtendExpr *Expr) {
339   // A zero-extended value can be interpreted as a piecewise defined signed
340   // value. If the value was non-negative it stays the same, otherwise it
341   // is the sum of the original value and 2^n where n is the bit-width of
342   // the original (or operand) type. Examples:
343   //   zext i8 127 to i32 -> { [127] }
344   //   zext i8  -1 to i32 -> { [256 + (-1)] } = { [255] }
345   //   zext i8  %v to i32 -> [v] -> { [v] | v >= 0; [256 + v] | v < 0 }
346   //
347   // However, LLVM/Scalar Evolution uses zero-extend (potentially lead by a
348   // truncate) to represent some forms of modulo computation. The left-hand side
349   // of the condition in the code below would result in the SCEV
350   // "zext i1 <false, +, true>for.body" which is just another description
351   // of the C expression "i & 1 != 0" or, equivalently, "i % 2 != 0".
352   //
353   //   for (i = 0; i < N; i++)
354   //     if (i & 1 != 0 /* == i % 2 */)
355   //       /* do something */
356   //
357   // If we do not make the modulo explicit but only use the mechanism described
358   // above we will get the very restrictive assumption "N < 3", because for all
359   // values of N >= 3 the SCEVAddRecExpr operand of the zero-extend would wrap.
360   // Alternatively, we can make the modulo in the operand explicit in the
361   // resulting piecewise function and thereby avoid the assumption on N. For the
362   // example this would result in the following piecewise affine function:
363   // { [i0] -> [(1)] : 2*floor((-1 + i0)/2) = -1 + i0;
364   //   [i0] -> [(0)] : 2*floor((i0)/2) = i0 }
365   // To this end we can first determine if the (immediate) operand of the
366   // zero-extend can wrap and, in case it might, we will use explicit modulo
367   // semantic to compute the result instead of emitting non-wrapping
368   // assumptions.
369   //
370   // Note that operands with large bit-widths are less likely to be negative
371   // because it would result in a very large access offset or loop bound after
372   // the zero-extend. To this end one can optimistically assume the operand to
373   // be positive and avoid the piecewise definition if the bit-width is bigger
374   // than some threshold (here MaxZextSmallBitWidth).
375   //
376   // We choose to go with a hybrid solution of all modeling techniques described
377   // above. For small bit-widths (up to MaxZextSmallBitWidth) we will model the
378   // wrapping explicitly and use a piecewise defined function. However, if the
379   // bit-width is bigger than MaxZextSmallBitWidth we will employ overflow
380   // assumptions and assume the "former negative" piece will not exist.
381 
382   auto *Op = Expr->getOperand();
383   unsigned Width = TD.getTypeSizeInBits(Op->getType());
384 
385   bool Precise = isPreciseZeroExtend(Width);
386 
387   auto Flags = getNoWrapFlags(Op);
388   auto NoWrapFlags = ScalarEvolution::setFlags(Flags, SCEV::FlagNSW);
389   bool OpCanWrap = Precise && !(Flags & SCEV::FlagNSW);
390   if (OpCanWrap)
391     Op = setNoWrapFlags(SE, Op, NoWrapFlags);
392 
393   auto OpPWAC = visit(Op);
394   if (OpCanWrap)
395     OpPWAC.first = addModuloSemantic(OpPWAC.first, Op->getType());
396 
397   // If the width is to big we assume the negative part does not occur.
398   if (!Precise) {
399     takeNonNegativeAssumption(OpPWAC);
400     return OpPWAC;
401   }
402 
403   // If the width is small build the piece for the non-negative part and
404   // the one for the negative part and unify them.
405   interpretAsUnsigned(OpPWAC, Width);
406   return OpPWAC;
407 }
408 
409 __isl_give PWACtx
410 SCEVAffinator::visitSignExtendExpr(const SCEVSignExtendExpr *Expr) {
411   // As all values are represented as signed, a sign extension is a noop.
412   return visit(Expr->getOperand());
413 }
414 
415 __isl_give PWACtx SCEVAffinator::visitAddExpr(const SCEVAddExpr *Expr) {
416   PWACtx Sum = visit(Expr->getOperand(0));
417 
418   for (int i = 1, e = Expr->getNumOperands(); i < e; ++i) {
419     combine(Sum, visit(Expr->getOperand(i)), isl_pw_aff_add);
420     if (isTooComplex(Sum))
421       return std::make_pair(nullptr, nullptr);
422   }
423 
424   return Sum;
425 }
426 
427 __isl_give PWACtx SCEVAffinator::visitMulExpr(const SCEVMulExpr *Expr) {
428   PWACtx Prod = visit(Expr->getOperand(0));
429 
430   for (int i = 1, e = Expr->getNumOperands(); i < e; ++i) {
431     combine(Prod, visit(Expr->getOperand(i)), isl_pw_aff_mul);
432     if (isTooComplex(Prod))
433       return std::make_pair(nullptr, nullptr);
434   }
435 
436   return Prod;
437 }
438 
439 __isl_give PWACtx SCEVAffinator::visitAddRecExpr(const SCEVAddRecExpr *Expr) {
440   assert(Expr->isAffine() && "Only affine AddRecurrences allowed");
441 
442   auto Flags = Expr->getNoWrapFlags();
443 
444   // Directly generate isl_pw_aff for Expr if 'start' is zero.
445   if (Expr->getStart()->isZero()) {
446     assert(S->contains(Expr->getLoop()) &&
447            "Scop does not contain the loop referenced in this AddRec");
448 
449     PWACtx Step = visit(Expr->getOperand(1));
450     isl_space *Space = isl_space_set_alloc(Ctx, 0, NumIterators);
451     isl_local_space *LocalSpace = isl_local_space_from_space(Space);
452 
453     unsigned loopDimension = S->getRelativeLoopDepth(Expr->getLoop());
454 
455     isl_aff *LAff = isl_aff_set_coefficient_si(
456         isl_aff_zero_on_domain(LocalSpace), isl_dim_in, loopDimension, 1);
457     isl_pw_aff *LPwAff = isl_pw_aff_from_aff(LAff);
458 
459     Step.first = isl_pw_aff_mul(Step.first, LPwAff);
460     return Step;
461   }
462 
463   // Translate AddRecExpr from '{start, +, inc}' into 'start + {0, +, inc}'
464   // if 'start' is not zero.
465   // TODO: Using the original SCEV no-wrap flags is not always safe, however
466   //       as our code generation is reordering the expression anyway it doesn't
467   //       really matter.
468   const SCEV *ZeroStartExpr =
469       SE.getAddRecExpr(SE.getConstant(Expr->getStart()->getType(), 0),
470                        Expr->getStepRecurrence(SE), Expr->getLoop(), Flags);
471 
472   PWACtx Result = visit(ZeroStartExpr);
473   PWACtx Start = visit(Expr->getStart());
474   combine(Result, Start, isl_pw_aff_add);
475   return Result;
476 }
477 
478 __isl_give PWACtx SCEVAffinator::visitSMaxExpr(const SCEVSMaxExpr *Expr) {
479   PWACtx Max = visit(Expr->getOperand(0));
480 
481   for (int i = 1, e = Expr->getNumOperands(); i < e; ++i) {
482     combine(Max, visit(Expr->getOperand(i)), isl_pw_aff_max);
483     if (isTooComplex(Max))
484       return std::make_pair(nullptr, nullptr);
485   }
486 
487   return Max;
488 }
489 
490 __isl_give PWACtx SCEVAffinator::visitUMaxExpr(const SCEVUMaxExpr *Expr) {
491   llvm_unreachable("SCEVUMaxExpr not yet supported");
492 }
493 
494 __isl_give PWACtx SCEVAffinator::visitUDivExpr(const SCEVUDivExpr *Expr) {
495   // The handling of unsigned division is basically the same as for signed
496   // division, except the interpretation of the operands. As the divisor
497   // has to be constant in both cases we can simply interpret it as an
498   // unsigned value without additional complexity in the representation.
499   // For the dividend we could choose from the different representation
500   // schemes introduced for zero-extend operations but for now we will
501   // simply use an assumption.
502   auto *Dividend = Expr->getLHS();
503   auto *Divisor = Expr->getRHS();
504   assert(isa<SCEVConstant>(Divisor) &&
505          "UDiv is no parameter but has a non-constant RHS.");
506 
507   auto DividendPWAC = visit(Dividend);
508   auto DivisorPWAC = visit(Divisor);
509 
510   if (SE.isKnownNegative(Divisor)) {
511     // Interpret negative divisors unsigned. This is a special case of the
512     // piece-wise defined value described for zero-extends as we already know
513     // the actual value of the constant divisor.
514     unsigned Width = TD.getTypeSizeInBits(Expr->getType());
515     auto *DivisorDom = isl_pw_aff_domain(isl_pw_aff_copy(DivisorPWAC.first));
516     auto *WidthExpPWA = getWidthExpValOnDomain(Width, DivisorDom);
517     DivisorPWAC.first = isl_pw_aff_add(DivisorPWAC.first, WidthExpPWA);
518   }
519 
520   // TODO: One can represent the dividend as piece-wise function to be more
521   //       precise but therefor a heuristic is needed.
522 
523   // Assume a non-negative dividend.
524   takeNonNegativeAssumption(DividendPWAC);
525 
526   combine(DividendPWAC, DivisorPWAC, isl_pw_aff_div);
527   DividendPWAC.first = isl_pw_aff_floor(DividendPWAC.first);
528 
529   return DividendPWAC;
530 }
531 
532 __isl_give PWACtx SCEVAffinator::visitSDivInstruction(Instruction *SDiv) {
533   assert(SDiv->getOpcode() == Instruction::SDiv && "Assumed SDiv instruction!");
534 
535   auto *Scope = getScope();
536   auto *Divisor = SDiv->getOperand(1);
537   auto *DivisorSCEV = SE.getSCEVAtScope(Divisor, Scope);
538   auto DivisorPWAC = visit(DivisorSCEV);
539   assert(isa<SCEVConstant>(DivisorSCEV) &&
540          "SDiv is no parameter but has a non-constant RHS.");
541 
542   auto *Dividend = SDiv->getOperand(0);
543   auto *DividendSCEV = SE.getSCEVAtScope(Dividend, Scope);
544   auto DividendPWAC = visit(DividendSCEV);
545   combine(DividendPWAC, DivisorPWAC, isl_pw_aff_tdiv_q);
546   return DividendPWAC;
547 }
548 
549 __isl_give PWACtx SCEVAffinator::visitSRemInstruction(Instruction *SRem) {
550   assert(SRem->getOpcode() == Instruction::SRem && "Assumed SRem instruction!");
551 
552   auto *Scope = getScope();
553   auto *Divisor = SRem->getOperand(1);
554   auto *DivisorSCEV = SE.getSCEVAtScope(Divisor, Scope);
555   auto DivisorPWAC = visit(DivisorSCEV);
556   assert(isa<ConstantInt>(Divisor) &&
557          "SRem is no parameter but has a non-constant RHS.");
558 
559   auto *Dividend = SRem->getOperand(0);
560   auto *DividendSCEV = SE.getSCEVAtScope(Dividend, Scope);
561   auto DividendPWAC = visit(DividendSCEV);
562   combine(DividendPWAC, DivisorPWAC, isl_pw_aff_tdiv_r);
563   return DividendPWAC;
564 }
565 
566 __isl_give PWACtx SCEVAffinator::visitUnknown(const SCEVUnknown *Expr) {
567   if (Instruction *I = dyn_cast<Instruction>(Expr->getValue())) {
568     switch (I->getOpcode()) {
569     case Instruction::IntToPtr:
570       return visit(SE.getSCEVAtScope(I->getOperand(0), getScope()));
571     case Instruction::PtrToInt:
572       return visit(SE.getSCEVAtScope(I->getOperand(0), getScope()));
573     case Instruction::SDiv:
574       return visitSDivInstruction(I);
575     case Instruction::SRem:
576       return visitSRemInstruction(I);
577     default:
578       break; // Fall through.
579     }
580   }
581 
582   llvm_unreachable(
583       "Unknowns SCEV was neither parameter nor a valid instruction.");
584 }
585