1@node poly_int
2@chapter Sizes and offsets as runtime invariants
3@cindex polynomial integers
4@findex poly_int
5
6GCC allows the size of a hardware register to be a runtime invariant
7rather than a compile-time constant.  This in turn means that various
8sizes and offsets must also be runtime invariants rather than
9compile-time constants, such as:
10
11@itemize @bullet
12@item
13the size of a general @code{machine_mode} (@pxref{Machine Modes});
14
15@item
16the size of a spill slot;
17
18@item
19the offset of something within a stack frame;
20
21@item
22the number of elements in a vector;
23
24@item
25the size and offset of a @code{mem} rtx (@pxref{Regs and Memory}); and
26
27@item
28the byte offset in a @code{subreg} rtx (@pxref{Regs and Memory}).
29@end itemize
30
31The motivating example is the Arm SVE ISA, whose vector registers can be
32any multiple of 128 bits between 128 and 2048 inclusive.  The compiler
33normally produces code that works for all SVE register sizes, with the
34actual size only being known at runtime.
35
36GCC's main representation of such runtime invariants is the
37@code{poly_int} class.  This chapter describes what @code{poly_int}
38does, lists the available operations, and gives some general
39usage guidelines.
40
41@menu
42* Overview of @code{poly_int}::
43* Consequences of using @code{poly_int}::
44* Comparisons involving @code{poly_int}::
45* Arithmetic on @code{poly_int}s::
46* Alignment of @code{poly_int}s::
47* Computing bounds on @code{poly_int}s::
48* Converting @code{poly_int}s::
49* Miscellaneous @code{poly_int} routines::
50* Guidelines for using @code{poly_int}::
51@end menu
52
53@node Overview of @code{poly_int}
54@section Overview of @code{poly_int}
55
56@cindex @code{poly_int}, runtime value
57We define indeterminates @var{x1}, @dots{}, @var{xn} whose values are
58only known at runtime and use polynomials of the form:
59
60@smallexample
61@var{c0} + @var{c1} * @var{x1} + @dots{} + @var{cn} * @var{xn}
62@end smallexample
63
64to represent a size or offset whose value might depend on some
65of these indeterminates.  The coefficients @var{c0}, @dots{}, @var{cn}
66are always known at compile time, with the @var{c0} term being the
67``constant'' part that does not depend on any runtime value.
68
69GCC uses the @code{poly_int} class to represent these coefficients.
70The class has two template parameters: the first specifies the number of
71coefficients (@var{n} + 1) and the second specifies the type of the
72coefficients.  For example, @samp{poly_int<2, unsigned short>} represents
73a polynomial with two coefficients (and thus one indeterminate), with each
74coefficient having type @code{unsigned short}.  When @var{n} is 0,
75the class degenerates to a single compile-time constant @var{c0}.
76
77@cindex @code{poly_int}, template parameters
78@findex NUM_POLY_INT_COEFFS
79The number of coefficients needed for compilation is a fixed
80property of each target and is specified by the configuration macro
81@code{NUM_POLY_INT_COEFFS}.  The default value is 1, since most targets
82do not have such runtime invariants.  Targets that need a different
83value should @code{#define} the macro in their @file{@var{cpu}-modes.def}
84file.  @xref{Back End}.
85
86@cindex @code{poly_int}, invariant range
87@code{poly_int} makes the simplifying requirement that each indeterminate
88must be a nonnegative integer.  An indeterminate value of 0 should usually
89represent the minimum possible runtime value, with @var{c0} specifying
90the value in that case.
91
92For example, when targetting the Arm SVE ISA, the single indeterminate
93represents the number of 128-bit blocks in a vector @emph{beyond the minimum
94length of 128 bits}.  Thus the number of 64-bit doublewords in a vector
95is 2 + 2 * @var{x1}.  If an aggregate has a single SVE vector and 16
96additional bytes, its total size is 32 + 16 * @var{x1} bytes.
97
98The header file @file{poly-int-types.h} provides typedefs for the
99most common forms of @code{poly_int}, all having
100@code{NUM_POLY_INT_COEFFS} coefficients:
101
102@cindex @code{poly_int}, main typedefs
103@table @code
104@item poly_uint16
105a @samp{poly_int} with @code{unsigned short} coefficients.
106
107@item poly_int64
108a @samp{poly_int} with @code{HOST_WIDE_INT} coefficients.
109
110@item poly_uint64
111a @samp{poly_int} with @code{unsigned HOST_WIDE_INT} coefficients.
112
113@item poly_offset_int
114a @samp{poly_int} with @code{offset_int} coefficients.
115
116@item poly_wide_int
117a @samp{poly_int} with @code{wide_int} coefficients.
118
119@item poly_widest_int
120a @samp{poly_int} with @code{widest_int} coefficients.
121@end table
122
123Since the main purpose of @code{poly_int} is to represent sizes and
124offsets, the last two typedefs are only rarely used.
125
126@node Consequences of using @code{poly_int}
127@section Consequences of using @code{poly_int}
128
129The two main consequences of using polynomial sizes and offsets are that:
130
131@itemize
132@item
133there is no total ordering between the values at compile time, and
134
135@item
136some operations might yield results that cannot be expressed as a
137@code{poly_int}.
138@end itemize
139
140For example, if @var{x} is a runtime invariant, we cannot tell at
141compile time whether:
142
143@smallexample
1443 + 4@var{x} <= 1 + 5@var{x}
145@end smallexample
146
147since the condition is false when @var{x} <= 1 and true when @var{x} >= 2.
148
149Similarly, @code{poly_int} cannot represent the result of:
150
151@smallexample
152(3 + 4@var{x}) * (1 + 5@var{x})
153@end smallexample
154
155since it cannot (and in practice does not need to) store powers greater
156than one.  It also cannot represent the result of:
157
158@smallexample
159(3 + 4@var{x}) / (1 + 5@var{x})
160@end smallexample
161
162The following sections describe how we deal with these restrictions.
163
164@cindex @code{poly_int}, use in target-independent code
165As described earlier, a @code{poly_int<1, @var{T}>} has no indeterminates
166and so degenerates to a compile-time constant of type @var{T}.  It would
167be possible in that case to do all normal arithmetic on the @var{T},
168and to compare the @var{T} using the normal C++ operators.  We deliberately
169prevent target-independent code from doing this, since the compiler needs
170to support other @code{poly_int<@var{n}, @var{T}>} as well, regardless of
171the current target's @code{NUM_POLY_INT_COEFFS}.
172
173@cindex @code{poly_int}, use in target-specific code
174However, it would be very artificial to force target-specific code
175to follow these restrictions if the target has no runtime indeterminates.
176There is therefore an implicit conversion from @code{poly_int<1, @var{T}>}
177to @var{T} when compiling target-specific translation units.
178
179@node Comparisons involving @code{poly_int}
180@section Comparisons involving @code{poly_int}
181
182In general we need to compare sizes and offsets in two situations:
183those in which the values need to be ordered, and those in which
184the values can be unordered.  More loosely, the distinction is often
185between values that have a definite link (usually because they refer to the
186same underlying register or memory location) and values that have
187no definite link.  An example of the former is the relationship between
188the inner and outer sizes of a subreg, where we must know at compile time
189whether the subreg is paradoxical, partial, or complete.  An example of
190the latter is alias analysis: we might want to check whether two
191arbitrary memory references overlap.
192
193Referring back to the examples in the previous section, it makes sense
194to ask whether a memory reference of size @samp{3 + 4@var{x}} overlaps
195one of size @samp{1 + 5@var{x}}, but it does not make sense to have a
196subreg in which the outer mode has @samp{3 + 4@var{x}} bytes and the
197inner mode has @samp{1 + 5@var{x}} bytes (or vice versa).  Such subregs
198are always invalid and should trigger an internal compiler error
199if formed.
200
201The underlying operators are the same in both cases, but the distinction
202affects how they are used.
203
204@menu
205* Comparison functions for @code{poly_int}::
206* Properties of the @code{poly_int} comparisons::
207* Comparing potentially-unordered @code{poly_int}s::
208* Comparing ordered @code{poly_int}s::
209* Checking for a @code{poly_int} marker value::
210* Range checks on @code{poly_int}s::
211* Sorting @code{poly_int}s::
212@end menu
213
214@node Comparison functions for @code{poly_int}
215@subsection Comparison functions for @code{poly_int}
216
217@code{poly_int} provides the following routines for checking whether
218a particular condition ``may be'' (might be) true:
219
220@example
221maybe_lt maybe_le maybe_eq maybe_ge maybe_gt
222                  maybe_ne
223@end example
224
225The functions have their natural meaning:
226
227@table @samp
228@item maybe_lt(@var{a}, @var{b})
229Return true if @var{a} might be less than @var{b}.
230
231@item maybe_le(@var{a}, @var{b})
232Return true if @var{a} might be less than or equal to @var{b}.
233
234@item maybe_eq(@var{a}, @var{b})
235Return true if @var{a} might be equal to @var{b}.
236
237@item maybe_ne(@var{a}, @var{b})
238Return true if @var{a} might not be equal to @var{b}.
239
240@item maybe_ge(@var{a}, @var{b})
241Return true if @var{a} might be greater than or equal to @var{b}.
242
243@item maybe_gt(@var{a}, @var{b})
244Return true if @var{a} might be greater than @var{b}.
245@end table
246
247For readability, @code{poly_int} also provides ``known'' inverses of these
248functions:
249
250@example
251known_lt (@var{a}, @var{b}) == !maybe_ge (@var{a}, @var{b})
252known_le (@var{a}, @var{b}) == !maybe_gt (@var{a}, @var{b})
253known_eq (@var{a}, @var{b}) == !maybe_ne (@var{a}, @var{b})
254known_ge (@var{a}, @var{b}) == !maybe_lt (@var{a}, @var{b})
255known_gt (@var{a}, @var{b}) == !maybe_le (@var{a}, @var{b})
256known_ne (@var{a}, @var{b}) == !maybe_eq (@var{a}, @var{b})
257@end example
258
259@node Properties of the @code{poly_int} comparisons
260@subsection Properties of the @code{poly_int} comparisons
261
262All ``maybe'' relations except @code{maybe_ne} are transitive, so for example:
263
264@smallexample
265maybe_lt (@var{a}, @var{b}) && maybe_lt (@var{b}, @var{c}) implies maybe_lt (@var{a}, @var{c})
266@end smallexample
267
268for all @var{a}, @var{b} and @var{c}.  @code{maybe_lt}, @code{maybe_gt}
269and @code{maybe_ne} are irreflexive, so for example:
270
271@smallexample
272!maybe_lt (@var{a}, @var{a})
273@end smallexample
274
275is true for all @var{a}.  @code{maybe_le}, @code{maybe_eq} and @code{maybe_ge}
276are reflexive, so for example:
277
278@smallexample
279maybe_le (@var{a}, @var{a})
280@end smallexample
281
282is true for all @var{a}.  @code{maybe_eq} and @code{maybe_ne} are symmetric, so:
283
284@smallexample
285maybe_eq (@var{a}, @var{b}) == maybe_eq (@var{b}, @var{a})
286maybe_ne (@var{a}, @var{b}) == maybe_ne (@var{b}, @var{a})
287@end smallexample
288
289for all @var{a} and @var{b}.  In addition:
290
291@smallexample
292maybe_le (@var{a}, @var{b}) == maybe_lt (@var{a}, @var{b}) || maybe_eq (@var{a}, @var{b})
293maybe_ge (@var{a}, @var{b}) == maybe_gt (@var{a}, @var{b}) || maybe_eq (@var{a}, @var{b})
294maybe_lt (@var{a}, @var{b}) == maybe_gt (@var{b}, @var{a})
295maybe_le (@var{a}, @var{b}) == maybe_ge (@var{b}, @var{a})
296@end smallexample
297
298However:
299
300@smallexample
301maybe_le (@var{a}, @var{b}) && maybe_le (@var{b}, @var{a}) does not imply !maybe_ne (@var{a}, @var{b}) [== known_eq (@var{a}, @var{b})]
302maybe_ge (@var{a}, @var{b}) && maybe_ge (@var{b}, @var{a}) does not imply !maybe_ne (@var{a}, @var{b}) [== known_eq (@var{a}, @var{b})]
303@end smallexample
304
305One example is again @samp{@var{a} == 3 + 4@var{x}}
306and @samp{@var{b} == 1 + 5@var{x}}, where @samp{maybe_le (@var{a}, @var{b})},
307@samp{maybe_ge (@var{a}, @var{b})} and @samp{maybe_ne (@var{a}, @var{b})}
308all hold.  @code{maybe_le} and @code{maybe_ge} are therefore not antisymetric
309and do not form a partial order.
310
311From the above, it follows that:
312
313@itemize @bullet
314@item
315All ``known'' relations except @code{known_ne} are transitive.
316
317@item
318@code{known_lt}, @code{known_ne} and @code{known_gt} are irreflexive.
319
320@item
321@code{known_le}, @code{known_eq} and @code{known_ge} are reflexive.
322@end itemize
323
324Also:
325
326@smallexample
327known_lt (@var{a}, @var{b}) == known_gt (@var{b}, @var{a})
328known_le (@var{a}, @var{b}) == known_ge (@var{b}, @var{a})
329known_lt (@var{a}, @var{b}) implies !known_lt (@var{b}, @var{a})  [asymmetry]
330known_gt (@var{a}, @var{b}) implies !known_gt (@var{b}, @var{a})
331known_le (@var{a}, @var{b}) && known_le (@var{b}, @var{a}) == known_eq (@var{a}, @var{b}) [== !maybe_ne (@var{a}, @var{b})]
332known_ge (@var{a}, @var{b}) && known_ge (@var{b}, @var{a}) == known_eq (@var{a}, @var{b}) [== !maybe_ne (@var{a}, @var{b})]
333@end smallexample
334
335@code{known_le} and @code{known_ge} are therefore antisymmetric and are
336partial orders.  However:
337
338@smallexample
339known_le (@var{a}, @var{b}) does not imply known_lt (@var{a}, @var{b}) || known_eq (@var{a}, @var{b})
340known_ge (@var{a}, @var{b}) does not imply known_gt (@var{a}, @var{b}) || known_eq (@var{a}, @var{b})
341@end smallexample
342
343For example, @samp{known_le (4, 4 + 4@var{x})} holds because the runtime
344indeterminate @var{x} is a nonnegative integer, but neither
345@code{known_lt (4, 4 + 4@var{x})} nor @code{known_eq (4, 4 + 4@var{x})} hold.
346
347@node Comparing potentially-unordered @code{poly_int}s
348@subsection Comparing potentially-unordered @code{poly_int}s
349
350In cases where there is no definite link between two @code{poly_int}s,
351we can usually make a conservatively-correct assumption.  For example,
352the conservative assumption for alias analysis is that two references
353@emph{might} alias.
354
355One way of checking whether [@var{begin1}, @var{end1}) might overlap
356[@var{begin2}, @var{end2}) using the @code{poly_int} comparisons is:
357
358@smallexample
359maybe_gt (@var{end1}, @var{begin2}) && maybe_gt (@var{end2}, @var{begin1})
360@end smallexample
361
362and another (equivalent) way is:
363
364@smallexample
365!(known_le (@var{end1}, @var{begin2}) || known_le (@var{end2}, @var{begin1}))
366@end smallexample
367
368However, in this particular example, it is better to use the range helper
369functions instead.  @xref{Range checks on @code{poly_int}s}.
370
371@node Comparing ordered @code{poly_int}s
372@subsection Comparing ordered @code{poly_int}s
373
374In cases where there is a definite link between two @code{poly_int}s,
375such as the outer and inner sizes of subregs, we usually require the sizes
376to be ordered by the @code{known_le} partial order.  @code{poly_int} provides
377the following utility functions for ordered values:
378
379@table @samp
380@item ordered_p (@var{a}, @var{b})
381Return true if @var{a} and @var{b} are ordered by the @code{known_le}
382partial order.
383
384@item ordered_min (@var{a}, @var{b})
385Assert that @var{a} and @var{b} are ordered by @code{known_le} and return the
386minimum of the two.  When using this function, please add a comment explaining
387why the values are known to be ordered.
388
389@item ordered_max (@var{a}, @var{b})
390Assert that @var{a} and @var{b} are ordered by @code{known_le} and return the
391maximum of the two.  When using this function, please add a comment explaining
392why the values are known to be ordered.
393@end table
394
395For example, if a subreg has an outer mode of size @var{outer} and an
396inner mode of size @var{inner}:
397
398@itemize @bullet
399@item
400the subreg is complete if known_eq (@var{inner}, @var{outer})
401
402@item
403otherwise, the subreg is paradoxical if known_le (@var{inner}, @var{outer})
404
405@item
406otherwise, the subreg is partial if known_le (@var{outer}, @var{inner})
407
408@item
409otherwise, the subreg is ill-formed
410@end itemize
411
412Thus the subreg is only valid if
413@samp{ordered_p (@var{outer}, @var{inner})} is true.  If this condition
414is already known to be true then:
415
416@itemize @bullet
417@item
418the subreg is complete if known_eq (@var{inner}, @var{outer})
419
420@item
421the subreg is paradoxical if maybe_lt (@var{inner}, @var{outer})
422
423@item
424the subreg is partial if maybe_lt (@var{outer}, @var{inner})
425@end itemize
426
427with the three conditions being mutually exclusive.
428
429Code that checks whether a subreg is valid would therefore generally
430check whether @code{ordered_p} holds (in addition to whatever other
431checks are required for subreg validity).  Code that is dealing
432with existing subregs can assert that @code{ordered_p} holds
433and use either of the classifications above.
434
435@node Checking for a @code{poly_int} marker value
436@subsection Checking for a @code{poly_int} marker value
437
438It is sometimes useful to have a special ``marker value'' that is not
439meant to be taken literally.  For example, some code uses a size
440of -1 to represent an unknown size, rather than having to carry around
441a separate boolean to say whether the size is known.
442
443The best way of checking whether something is a marker value is
444@code{known_eq}.  Conversely the best way of checking whether something
445is @emph{not} a marker value is @code{maybe_ne}.
446
447Thus in the size example just mentioned, @samp{known_eq (size, -1)} would
448check for an unknown size and @samp{maybe_ne (size, -1)} would check for a
449known size.
450
451@node Range checks on @code{poly_int}s
452@subsection Range checks on @code{poly_int}s
453
454As well as the core comparisons
455(@pxref{Comparison functions for @code{poly_int}}), @code{poly_int} provides
456utilities for various kinds of range check.  In each case the range
457is represented by a start position and a size rather than a start
458position and an end position; this is because the former is used
459much more often than the latter in GCC@.  Also, the sizes can be
460-1 (or all ones for unsigned sizes) to indicate a range with a known
461start position but an unknown size.  All other sizes must be nonnegative.
462A range of size 0 does not contain anything or overlap anything.
463
464@table @samp
465@item known_size_p (@var{size})
466Return true if @var{size} represents a known range size, false if it
467is -1 or all ones (for signed and unsigned types respectively).
468
469@item ranges_maybe_overlap_p (@var{pos1}, @var{size1}, @var{pos2}, @var{size2})
470Return true if the range described by @var{pos1} and @var{size1} @emph{might}
471overlap the range described by @var{pos2} and @var{size2} (in other words,
472return true if we cannot prove that the ranges are disjoint).
473
474@item ranges_known_overlap_p (@var{pos1}, @var{size1}, @var{pos2}, @var{size2})
475Return true if the range described by @var{pos1} and @var{size1} is known to
476overlap the range described by @var{pos2} and @var{size2}.
477
478@item known_subrange_p (@var{pos1}, @var{size1}, @var{pos2}, @var{size2})
479Return true if the range described by @var{pos1} and @var{size1} is known to
480be contained in the range described by @var{pos2} and @var{size2}.
481
482@item maybe_in_range_p (@var{value}, @var{pos}, @var{size})
483Return true if @var{value} @emph{might} be in the range described by
484@var{pos} and @var{size} (in other words, return true if we cannot
485prove that @var{value} is outside that range).
486
487@item known_in_range_p (@var{value}, @var{pos}, @var{size})
488Return true if @var{value} is known to be in the range described
489by @var{pos} and @var{size}.
490
491@item endpoint_representable_p (@var{pos}, @var{size})
492Return true if the range described by @var{pos} and @var{size} is
493open-ended or if the endpoint (@var{pos} + @var{size}) is representable
494in the same type as @var{pos} and @var{size}.  The function returns false
495if adding @var{size} to @var{pos} makes conceptual sense but could overflow.
496@end table
497
498There is also a @code{poly_int} version of the @code{IN_RANGE_P} macro:
499
500@table @samp
501@item coeffs_in_range_p (@var{x}, @var{lower}, @var{upper})
502Return true if every coefficient of @var{x} is in the inclusive range
503[@var{lower}, @var{upper}].  This function can be useful when testing
504whether an operation would cause the values of coefficients to
505overflow.
506
507Note that the function does not indicate whether @var{x} itself is in the
508given range.  @var{x} can be either a constant or a @code{poly_int}.
509@end table
510
511@node Sorting @code{poly_int}s
512@subsection Sorting @code{poly_int}s
513
514@code{poly_int} provides the following routine for sorting:
515
516@table @samp
517@item compare_sizes_for_sort (@var{a}, @var{b})
518Compare @var{a} and @var{b} in reverse lexicographical order (that is,
519compare the highest-indexed coefficients first).  This can be useful when
520sorting data structures, since it has the effect of separating constant
521and non-constant values.  If all values are nonnegative, the constant
522values come first.
523
524Note that the values do not necessarily end up in numerical order.
525For example, @samp{1 + 1@var{x}} would come after @samp{100} in the sort order,
526but may well be less than @samp{100} at run time.
527@end table
528
529@node Arithmetic on @code{poly_int}s
530@section Arithmetic on @code{poly_int}s
531
532Addition, subtraction, negation and bit inversion all work normally for
533@code{poly_int}s.  Multiplication by a constant multiplier and left
534shifting by a constant shift amount also work normally.  General
535multiplication of two @code{poly_int}s is not supported and is not
536useful in practice.
537
538Other operations are only conditionally supported: the operation
539might succeed or might fail, depending on the inputs.
540
541This section describes both types of operation.
542
543@menu
544* Using @code{poly_int} with C++ arithmetic operators::
545* @code{wi} arithmetic on @code{poly_int}s::
546* Division of @code{poly_int}s::
547* Other @code{poly_int} arithmetic::
548@end menu
549
550@node Using @code{poly_int} with C++ arithmetic operators
551@subsection Using @code{poly_int} with C++ arithmetic operators
552
553The following C++ expressions are supported, where @var{p1} and @var{p2}
554are @code{poly_int}s and where @var{c1} and @var{c2} are scalars:
555
556@smallexample
557-@var{p1}
558~@var{p1}
559
560@var{p1} + @var{p2}
561@var{p1} + @var{c2}
562@var{c1} + @var{p2}
563
564@var{p1} - @var{p2}
565@var{p1} - @var{c2}
566@var{c1} - @var{p2}
567
568@var{c1} * @var{p2}
569@var{p1} * @var{c2}
570
571@var{p1} << @var{c2}
572
573@var{p1} += @var{p2}
574@var{p1} += @var{c2}
575
576@var{p1} -= @var{p2}
577@var{p1} -= @var{c2}
578
579@var{p1} *= @var{c2}
580@var{p1} <<= @var{c2}
581@end smallexample
582
583These arithmetic operations handle integer ranks in a similar way
584to C++.  The main difference is that every coefficient narrower than
585@code{HOST_WIDE_INT} promotes to @code{HOST_WIDE_INT}, whereas in
586C++ everything narrower than @code{int} promotes to @code{int}.
587For example:
588
589@smallexample
590poly_uint16     + int          -> poly_int64
591unsigned int    + poly_uint16  -> poly_int64
592poly_int64      + int          -> poly_int64
593poly_int32      + poly_uint64  -> poly_uint64
594uint64          + poly_int64   -> poly_uint64
595poly_offset_int + int32        -> poly_offset_int
596offset_int      + poly_uint16  -> poly_offset_int
597@end smallexample
598
599In the first two examples, both coefficients are narrower than
600@code{HOST_WIDE_INT}, so the result has coefficients of type
601@code{HOST_WIDE_INT}.  In the other examples, the coefficient
602with the highest rank ``wins''.
603
604If one of the operands is @code{wide_int} or @code{poly_wide_int},
605the rules are the same as for @code{wide_int} arithmetic.
606
607@node @code{wi} arithmetic on @code{poly_int}s
608@subsection @code{wi} arithmetic on @code{poly_int}s
609
610As well as the C++ operators, @code{poly_int} supports the following
611@code{wi} routines:
612
613@smallexample
614wi::neg (@var{p1}, &@var{overflow})
615
616wi::add (@var{p1}, @var{p2})
617wi::add (@var{p1}, @var{c2})
618wi::add (@var{c1}, @var{p1})
619wi::add (@var{p1}, @var{p2}, @var{sign}, &@var{overflow})
620
621wi::sub (@var{p1}, @var{p2})
622wi::sub (@var{p1}, @var{c2})
623wi::sub (@var{c1}, @var{p1})
624wi::sub (@var{p1}, @var{p2}, @var{sign}, &@var{overflow})
625
626wi::mul (@var{p1}, @var{c2})
627wi::mul (@var{c1}, @var{p1})
628wi::mul (@var{p1}, @var{c2}, @var{sign}, &@var{overflow})
629
630wi::lshift (@var{p1}, @var{c2})
631@end smallexample
632
633These routines just check whether overflow occurs on any individual
634coefficient; it is not possible to know at compile time whether the
635final runtime value would overflow.
636
637@node Division of @code{poly_int}s
638@subsection Division of @code{poly_int}s
639
640Division of @code{poly_int}s is possible for certain inputs.  The functions
641for division return true if the operation is possible and in most cases
642return the results by pointer.  The routines are:
643
644@table @samp
645@item multiple_p (@var{a}, @var{b})
646@itemx multiple_p (@var{a}, @var{b}, &@var{quotient})
647Return true if @var{a} is an exact multiple of @var{b}, storing the result
648in @var{quotient} if so.  There are overloads for various combinations
649of polynomial and constant @var{a}, @var{b} and @var{quotient}.
650
651@item constant_multiple_p (@var{a}, @var{b})
652@itemx constant_multiple_p (@var{a}, @var{b}, &@var{quotient})
653Like @code{multiple_p}, but also test whether the multiple is a
654compile-time constant.
655
656@item can_div_trunc_p (@var{a}, @var{b}, &@var{quotient})
657@itemx can_div_trunc_p (@var{a}, @var{b}, &@var{quotient}, &@var{remainder})
658Return true if we can calculate @samp{trunc (@var{a} / @var{b})} at compile
659time, storing the result in @var{quotient} and @var{remainder} if so.
660
661@item can_div_away_from_zero_p (@var{a}, @var{b}, &@var{quotient})
662Return true if we can calculate @samp{@var{a} / @var{b}} at compile time,
663rounding away from zero.  Store the result in @var{quotient} if so.
664
665Note that this is true if and only if @code{can_div_trunc_p} is true.
666The only difference is in the rounding of the result.
667@end table
668
669There is also an asserting form of division:
670
671@table @samp
672@item exact_div (@var{a}, @var{b})
673Assert that @var{a} is a multiple of @var{b} and return
674@samp{@var{a} / @var{b}}.  The result is a @code{poly_int} if @var{a}
675is a @code{poly_int}.
676@end table
677
678@node Other @code{poly_int} arithmetic
679@subsection Other @code{poly_int} arithmetic
680
681There are tentative routines for other operations besides division:
682
683@table @samp
684@item can_ior_p (@var{a}, @var{b}, &@var{result})
685Return true if we can calculate @samp{@var{a} | @var{b}} at compile time,
686storing the result in @var{result} if so.
687@end table
688
689Also, ANDs with a value @samp{(1 << @var{y}) - 1} or its inverse can be
690treated as alignment operations.  @xref{Alignment of @code{poly_int}s}.
691
692In addition, the following miscellaneous routines are available:
693
694@table @samp
695@item coeff_gcd (@var{a})
696Return the greatest common divisor of all nonzero coefficients in
697@var{a}, or zero if @var{a} is known to be zero.
698
699@item common_multiple (@var{a}, @var{b})
700Return a value that is a multiple of both @var{a} and @var{b}, where
701one value is a @code{poly_int} and the other is a scalar.  The result
702will be the least common multiple for some indeterminate values but
703not necessarily for all.
704
705@item force_common_multiple (@var{a}, @var{b})
706Return a value that is a multiple of both @code{poly_int} @var{a} and
707@code{poly_int} @var{b}, asserting that such a value exists.  The
708result will be the least common multiple for some indeterminate values
709but not necessarily for all.
710
711When using this routine, please add a comment explaining why the
712assertion is known to hold.
713@end table
714
715Please add any other operations that you find to be useful.
716
717@node Alignment of @code{poly_int}s
718@section Alignment of @code{poly_int}s
719
720@code{poly_int} provides various routines for aligning values and for querying
721misalignments.  In each case the alignment must be a power of 2.
722
723@table @samp
724@item can_align_p (@var{value}, @var{align})
725Return true if we can align @var{value} up or down to the nearest multiple
726of @var{align} at compile time.  The answer is the same for both directions.
727
728@item can_align_down (@var{value}, @var{align}, &@var{aligned})
729Return true if @code{can_align_p}; if so, set @var{aligned} to the greatest
730aligned value that is less than or equal to @var{value}.
731
732@item can_align_up (@var{value}, @var{align}, &@var{aligned})
733Return true if @code{can_align_p}; if so, set @var{aligned} to the lowest
734aligned value that is greater than or equal to @var{value}.
735
736@item known_equal_after_align_down (@var{a}, @var{b}, @var{align})
737Return true if we can align @var{a} and @var{b} down to the nearest
738@var{align} boundary at compile time and if the two results are equal.
739
740@item known_equal_after_align_up (@var{a}, @var{b}, @var{align})
741Return true if we can align @var{a} and @var{b} up to the nearest
742@var{align} boundary at compile time and if the two results are equal.
743
744@item aligned_lower_bound (@var{value}, @var{align})
745Return a result that is no greater than @var{value} and that is aligned
746to @var{align}.  The result will the closest aligned value for some
747indeterminate values but not necessarily for all.
748
749For example, suppose we are allocating an object of @var{size} bytes
750in a downward-growing stack whose current limit is given by @var{limit}.
751If the object requires @var{align} bytes of alignment, the new stack
752limit is given by:
753
754@smallexample
755aligned_lower_bound (@var{limit} - @var{size}, @var{align})
756@end smallexample
757
758@item aligned_upper_bound (@var{value}, @var{align})
759Likewise return a result that is no less than @var{value} and that is
760aligned to @var{align}.  This is the routine that would be used for
761upward-growing stacks in the scenario just described.
762
763@item known_misalignment (@var{value}, @var{align}, &@var{misalign})
764Return true if we can calculate the misalignment of @var{value}
765with respect to @var{align} at compile time, storing the result in
766@var{misalign} if so.
767
768@item known_alignment (@var{value})
769Return the minimum alignment that @var{value} is known to have
770(in other words, the largest alignment that can be guaranteed
771whatever the values of the indeterminates turn out to be).
772Return 0 if @var{value} is known to be 0.
773
774@item force_align_down (@var{value}, @var{align})
775Assert that @var{value} can be aligned down to @var{align} at compile
776time and return the result.  When using this routine, please add a
777comment explaining why the assertion is known to hold.
778
779@item force_align_up (@var{value}, @var{align})
780Likewise, but aligning up.
781
782@item force_align_down_and_div (@var{value}, @var{align})
783Divide the result of @code{force_align_down} by @var{align}.  Again,
784please add a comment explaining why the assertion in @code{force_align_down}
785is known to hold.
786
787@item force_align_up_and_div (@var{value}, @var{align})
788Likewise for @code{force_align_up}.
789
790@item force_get_misalignment (@var{value}, @var{align})
791Assert that we can calculate the misalignment of @var{value} with
792respect to @var{align} at compile time and return the misalignment.
793When using this function, please add a comment explaining why
794the assertion is known to hold.
795@end table
796
797@node Computing bounds on @code{poly_int}s
798@section Computing bounds on @code{poly_int}s
799
800@code{poly_int} also provides routines for calculating lower and upper bounds:
801
802@table @samp
803@item constant_lower_bound (@var{a})
804Assert that @var{a} is nonnegative and return the smallest value it can have.
805
806@item constant_lower_bound_with_limit (@var{a}, @var{b})
807Return the least value @var{a} can have, given that the context in
808which @var{a} appears guarantees that the answer is no less than @var{b}.
809In other words, the caller is asserting that @var{a} is greater than or
810equal to @var{b} even if @samp{known_ge (@var{a}, @var{b})} doesn't hold.
811
812@item constant_upper_bound_with_limit (@var{a}, @var{b})
813Return the greatest value @var{a} can have, given that the context in
814which @var{a} appears guarantees that the answer is no greater than @var{b}.
815In other words, the caller is asserting that @var{a} is less than or equal
816to @var{b} even if @samp{known_le (@var{a}, @var{b})} doesn't hold.
817
818@item lower_bound (@var{a}, @var{b})
819Return a value that is always less than or equal to both @var{a} and @var{b}.
820It will be the greatest such value for some indeterminate values
821but necessarily for all.
822
823@item upper_bound (@var{a}, @var{b})
824Return a value that is always greater than or equal to both @var{a} and
825@var{b}.  It will be the least such value for some indeterminate values
826but necessarily for all.
827@end table
828
829@node Converting @code{poly_int}s
830@section Converting @code{poly_int}s
831
832A @code{poly_int<@var{n}, @var{T}>} can be constructed from up to
833@var{n} individual @var{T} coefficients, with the remaining coefficients
834being implicitly zero.  In particular, this means that every
835@code{poly_int<@var{n}, @var{T}>} can be constructed from a single
836scalar @var{T}, or something compatible with @var{T}.
837
838Also, a @code{poly_int<@var{n}, @var{T}>} can be constructed from
839a @code{poly_int<@var{n}, @var{U}>} if @var{T} can be constructed
840from @var{U}.
841
842The following functions provide other forms of conversion,
843or test whether such a conversion would succeed.
844
845@table @samp
846@item @var{value}.is_constant ()
847Return true if @code{poly_int} @var{value} is a compile-time constant.
848
849@item @var{value}.is_constant (&@var{c1})
850Return true if @code{poly_int} @var{value} is a compile-time constant,
851storing it in @var{c1} if so.  @var{c1} must be able to hold all
852constant values of @var{value} without loss of precision.
853
854@item @var{value}.to_constant ()
855Assert that @var{value} is a compile-time constant and return its value.
856When using this function, please add a comment explaining why the
857condition is known to hold (for example, because an earlier phase
858of analysis rejected non-constants).
859
860@item @var{value}.to_shwi (&@var{p2})
861Return true if @samp{poly_int<@var{N}, @var{T}>} @var{value} can be
862represented without loss of precision as a
863@samp{poly_int<@var{N}, @code{HOST_WIDE_INT}>}, storing it in that
864form in @var{p2} if so.
865
866@item @var{value}.to_uhwi (&@var{p2})
867Return true if @samp{poly_int<@var{N}, @var{T}>} @var{value} can be
868represented without loss of precision as a
869@samp{poly_int<@var{N}, @code{unsigned HOST_WIDE_INT}>}, storing it in that
870form in @var{p2} if so.
871
872@item @var{value}.force_shwi ()
873Forcibly convert each coefficient of @samp{poly_int<@var{N}, @var{T}>}
874@var{value} to @code{HOST_WIDE_INT}, truncating any that are out of range.
875Return the result as a @samp{poly_int<@var{N}, @code{HOST_WIDE_INT}>}.
876
877@item @var{value}.force_uhwi ()
878Forcibly convert each coefficient of @samp{poly_int<@var{N}, @var{T}>}
879@var{value} to @code{unsigned HOST_WIDE_INT}, truncating any that are
880out of range.  Return the result as a
881@samp{poly_int<@var{N}, @code{unsigned HOST_WIDE_INT}>}.
882
883@item wi::shwi (@var{value}, @var{precision})
884Return a @code{poly_int} with the same value as @var{value}, but with
885the coefficients converted from @code{HOST_WIDE_INT} to @code{wide_int}.
886@var{precision} specifies the precision of the @code{wide_int} cofficients;
887if this is wider than a @code{HOST_WIDE_INT}, the coefficients of
888@var{value} will be sign-extended to fit.
889
890@item wi::uhwi (@var{value}, @var{precision})
891Like @code{wi::shwi}, except that @var{value} has coefficients of
892type @code{unsigned HOST_WIDE_INT}.  If @var{precision} is wider than
893a @code{HOST_WIDE_INT}, the coefficients of @var{value} will be
894zero-extended to fit.
895
896@item wi::sext (@var{value}, @var{precision})
897Return a @code{poly_int} of the same type as @var{value}, sign-extending
898every coefficient from the low @var{precision} bits.  This in effect
899applies @code{wi::sext} to each coefficient individually.
900
901@item wi::zext (@var{value}, @var{precision})
902Like @code{wi::sext}, but for zero extension.
903
904@item poly_wide_int::from (@var{value}, @var{precision}, @var{sign})
905Convert @var{value} to a @code{poly_wide_int} in which each coefficient
906has @var{precision} bits.  Extend the coefficients according to
907@var{sign} if the coefficients have fewer bits.
908
909@item poly_offset_int::from (@var{value}, @var{sign})
910Convert @var{value} to a @code{poly_offset_int}, extending its coefficients
911according to @var{sign} if they have fewer bits than @code{offset_int}.
912
913@item poly_widest_int::from (@var{value}, @var{sign})
914Convert @var{value} to a @code{poly_widest_int}, extending its coefficients
915according to @var{sign} if they have fewer bits than @code{widest_int}.
916@end table
917
918@node Miscellaneous @code{poly_int} routines
919@section Miscellaneous @code{poly_int} routines
920
921@table @samp
922@item print_dec (@var{value}, @var{file}, @var{sign})
923@itemx print_dec (@var{value}, @var{file})
924Print @var{value} to @var{file} as a decimal value, interpreting
925the coefficients according to @var{sign}.  The final argument is
926optional if @var{value} has an inherent sign; for example,
927@code{poly_int64} values print as signed by default and
928@code{poly_uint64} values print as unsigned by default.
929
930This is a simply a @code{poly_int} version of a wide-int routine.
931@end table
932
933@node Guidelines for using @code{poly_int}
934@section Guidelines for using @code{poly_int}
935
936One of the main design goals of @code{poly_int} was to make it easy
937to write target-independent code that handles variable-sized registers
938even when the current target has fixed-sized registers.  There are two
939aspects to this:
940
941@itemize
942@item
943The set of @code{poly_int} operations should be complete enough that
944the question in most cases becomes ``Can we do this operation on these
945particular @code{poly_int} values?  If not, bail out'' rather than
946``Are these @code{poly_int} values constant?  If so, do the operation,
947otherwise bail out''.
948
949@item
950If target-independent code compiles and runs correctly on a target
951with one value of @code{NUM_POLY_INT_COEFFS}, and if the code does not
952use asserting functions like @code{to_constant}, it is reasonable to
953assume that the code also works on targets with other values of
954@code{NUM_POLY_INT_COEFFS}.  There is no need to check this during
955everyday development.
956@end itemize
957
958So the general principle is: if target-independent code is dealing
959with a @code{poly_int} value, it is better to operate on it as a
960@code{poly_int} if at all possible, choosing conservatively-correct
961behavior if a particular operation fails.  For example, the following
962code handles an index @code{pos} into a sequence of vectors that each
963have @code{nunits} elements:
964
965@smallexample
966/* Calculate which vector contains the result, and which lane of
967   that vector we need.  */
968if (!can_div_trunc_p (pos, nunits, &vec_entry, &vec_index))
969  @{
970    if (dump_enabled_p ())
971      dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
972                       "Cannot determine which vector holds the"
973                       " final result.\n");
974    return false;
975  @}
976@end smallexample
977
978However, there are some contexts in which operating on a
979@code{poly_int} is not possible or does not make sense.  One example
980is when handling static initializers, since no current target supports
981the concept of a variable-length static initializer.  In these
982situations, a reasonable fallback is:
983
984@smallexample
985if (@var{poly_value}.is_constant (&@var{const_value}))
986  @{
987    @dots{}
988    /* Operate on @var{const_value}.  */
989    @dots{}
990  @}
991else
992  @{
993    @dots{}
994    /* Conservatively correct fallback.  */
995    @dots{}
996  @}
997@end smallexample
998
999@code{poly_int} also provides some asserting functions like
1000@code{to_constant}.  Please only use these functions if there is a
1001good theoretical reason to believe that the assertion cannot fire.
1002For example, if some work is divided into an analysis phase and an
1003implementation phase, the analysis phase might reject inputs that are
1004not @code{is_constant}, in which case the implementation phase can
1005reasonably use @code{to_constant} on the remaining inputs.  The assertions
1006should not be used to discover whether a condition ever occurs ``in the
1007field''; in other words, they should not be used to restrict code to
1008constants at first, with the intention of only implementing a
1009@code{poly_int} version if a user hits the assertion.
1010
1011If a particular asserting function like @code{to_constant} is needed
1012more than once for the same reason, it is probably worth adding a
1013helper function or macro for that situation, so that the justification
1014only needs to be given once.  For example:
1015
1016@smallexample
1017/* Return the size of an element in a vector of size SIZE, given that
1018   the vector has NELTS elements.  The return value is in the same units
1019   as SIZE (either bits or bytes).
1020
1021   to_constant () is safe in this situation because vector elements are
1022   always constant-sized scalars.  */
1023#define vector_element_size(SIZE, NELTS) \
1024  (exact_div (SIZE, NELTS).to_constant ())
1025@end smallexample
1026
1027Target-specific code in @file{config/@var{cpu}} only needs to handle
1028non-constant @code{poly_int}s if @code{NUM_POLY_INT_COEFFS} is greater
1029than one.  For other targets, @code{poly_int} degenerates to a compile-time
1030constant and is often interchangable with a normal scalar integer.
1031There are two main exceptions:
1032
1033@itemize
1034@item
1035Sometimes an explicit cast to an integer type might be needed, such as to
1036resolve ambiguities in a @code{?:} expression, or when passing values
1037through @code{...} to things like print functions.
1038
1039@item
1040Target macros are included in target-independent code and so do not
1041have access to the implicit conversion to a scalar integer.
1042If this becomes a problem for a particular target macro, the
1043possible solutions, in order of preference, are:
1044
1045@itemize
1046@item
1047Convert the target macro to a target hook (for all targets).
1048
1049@item
1050Put the target's implementation of the target macro in its
1051@file{@var{cpu}.c} file and call it from the target macro in the
1052@file{@var{cpu}.h} file.
1053
1054@item
1055Add @code{to_constant ()} calls where necessary.  The previous option
1056is preferable because it will help with any future conversion of the
1057macro to a hook.
1058@end itemize
1059@end itemize
1060
1061