| Age | Commit message (Collapse) | Author |
|
- In the definition of structures in order.v, displays are removed from
parameters of mixins and fields of classes internally and now only appear in
parameters of structures. Consequently, each mixin is now parameterized by a
class rather than a structure, and the corresponding factory parameterized by
a structure is provided to replace the use of the mixin. These factories have
the same names as in the mixins before this change except that `bLatticeMixin`
and `tbLatticeMixin` have been renamed to `bottomMixin` and `topMixin`
respectively.
- Added a factory `distrLatticePOrderMixin` in order.v to build a
`distrLatticeType` from a `porderType`.
|
|
|
|
fix "Nothing to inject" warnings
|
|
|
|
|
|
fix the doc for ubnP in ssrnat
|
|
|
|
|
|
|
|
|
|
Co-authored-by: Cyril Cohen <CohenCyril@users.noreply.github.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
silencing warnings in individual packages
|
|
|
|
|
|
|
|
|
|
- deprecating `fintype.arg_(min|max)P`
- removing dangling comments connecting min max and meet join
- better compatibility module
- removing broken notations with `\min` and `\max` (no neutral available)
- fixing `incompare` and `incomparel` in order.v
- adding missing elimination lemmas (`(comparable_)?(max|min)E[lg][et]`)
- adding missing `(comparable|real)_arg(min|max)P`
- CHANGELOG update
|
|
- mostly gathered the changes from previous commits
- add `minrC`
- minor doc addition to `order.v`
|
|
- min and max can now be used in a partial order (sometimes under preconditions)
- min and max can now be used in a numDomainType (sometimes under preconditions)
|
|
- `Order.max` and `Order.min` are now convertible to `maxn` and `minn`
- Inserting `(fun _ _ => erefl)` in `LeOrderMixin` and `LtOrderMixin`
gives `meet`/`join` which are convertible to `min`/`max`
- `Order.max` and `Order.min` are not convertible to former `Num.max` and `Num.min`
|
|
- max and min are not defined in terms or meet and join anymore
- extremum_inP is a generalization of extremum suitable for a partial order
- arg_min and arg_max are usable in a porderType
- equivalences between min and meet, and max and join are proven for orderType
- leP, ltP, ltgtP, and `[l]?comparable_.*P` lemmas have been generalized to take this into account
- total_display was completely removed
|
|
* Missing mono lemmas
|
|
|
|
This change reduces
- use of numerical occurrence selectors (#436) and
- use of non ssreflect tactics such as `auto`,
and improves use of comparison predicates such as `posnP`, `leqP`, `ltnP`,
`ltngtP`, and `eqVneq`.
|
|
Remove the tuple extensions in order.v that is available in tuple.v
|
|
|
|
This fixes two issues:
- dual_finLatticeType was missing, and
- dual_finDistrLatticeType was just an identity function.
|
|
adding deprecations in ssrnat
|
|
|
|
fixes #469
|
|
switching long suffixes to short suffixes
|
|
+ `odd_add` -> `oddD`
+ `odd_sub` -> `oddB`
+ `take_addn` -> `takeD`
+ `rot_addn` -> `rotD`
+ `nseq_addn` -> `nseqD`
fixes #359
|
|
Documentation typos
|
|
|
|
|
|
the->this
Co-Authored-By: Yves Bertot <yves.bertot@inria.fr>
|
|
|
|
|
|
%AC annotation are for backward compatilibity with coq <= 8.9
|
|
This replaces opA, opC, opAC, opCA, ... and any combinations of them
- Right now the rewrite relies on an rather efficient computation
of perm_eq using a "spaghetti sort" in O(n log n)
- Wrongly formed AC statements send error messages showing the
discrepancy between LHS and RHS patterns.
Usage :
rewrite [pattern](AC operator pattern-shape re-ordering)
rewrite [pattern](ACl operator re-ordering)
- pattern is optional, as usual,
- operator must have a canonical Monoid.com_law structure
(additions, multiplications, conjunction and disjunction do)
- pattern-shape is expressed using the syntax
p := n | p * p'
where "*" is purely formal
and n > 0 is number of left associated symbols
examples of pattern shapes:
+ 4 represents (n * m * p * q)
+ (1*2) represents (n * (m * p))
- re-ordering is expressed using the syntax
s := n | s * s'
where "*" is purely formal and n is the position in the LHS
If the ACl variant is used, the pattern-shape defaults to the
pattern fully associated to the left i.e. n i.e (x * y * ...)
Examples of re-orderings:
- ACl op ((0*1)*2) is the identity (and should fail to rewrite)
- opAC == ACl op ((0*2)*1) == AC op 3 ((0*2)*1)
- opCA == AC op (2*1) (0*1*2)
- rewrite opCA -opA == rewrite (ACl op (0*(2*1))
- opACA == AC (2*2) ((0*2)*(1*3))
|
|
remove deprecated commands whose deprecation was introduced in 1.9.0
|
|
Extend comparison predicates for nat with minn and maxn and reorder arguments of those in order.v
|
|
fixes #418
|