| Age | Commit message (Collapse) | Author |
|---|
|
The `emitcodes` string type was used in both a functional and an
imperative way, so we have to handle it with care in order to preserve
the previous optimizations and semantics.
|
|
Maxime points out that -notop cannot be used as the kernel requires
all constants to belong into a module. Indeed:
```
$ rlwrap ./bin/coqtop -notop
Coq < Definition foo := True.
Toplevel input, characters 0-23:
> Definition foo := True.
> ^^^^^^^^^^^^^^^^^^^^^^^
Error: No session module started (use -top dir)
Coq < Module M. Definition foo := True. End M.
Module M is defined
Coq < Locate foo.
Constant If you see this, it's a bug.M.foo
(shorter name to refer to it in current context is M.foo)
```
My rationale for the removal is that this kind of incomplete features
are often confusing to newcomers ─ it has happened to me many times ─
as it can be seen for example in #397 .
|
|
No functional change.
|
|
No functional change, we create the new string beforehand and
`id_of_string` will become a noop with `-safe-string`.
|
|
It was always set to `greedy:true`.
|
|
This performance bug was introduced 9 years ago in a8b0345, where the
responsibility of normalizing the term went from ccnv to eqappr in Reduction.
As a result, all recursive calls to eqappr that were preemptively reducing the
term ended up calling whd_stack twice, once by themselves, and once in the
subsequent call to eqappr.
This caused an important slowdown for conversion-intensive proofs, as the
whd_stack calls CClosure.zip to perfom in-place term sharing, leading to
useless huge re-allocations and repetitive write barriers.
Now that eqappr always head-normalizes the term beforehand, we simply don't
call whd_stack anymore when jumping to eqappr.
|
|
|
|
We unroll the map operation by hand in two performance-critical cases so as
not to call the generic array allocation function in OCaml, and allocate
directly on the minor heap instead. The generic array function is slow because
it needs to discriminate between float and non-float arrays. The unrolling
replaces this by a simple increment to the minor heap pointer and moves from
the stack.
The quantity of unrolling was determined by experimental measures on various Coq
developments. It looks like most of the maps are for small arrays of size lesser
or equal to 4, so this is what is implemented here. We could probably extend it
to an even bigger number, but that would result in ugly code. From what I've seen,
virtually all maps are of size less than 16, so that we could probably be almost
optimal by going up to 16 unrollings, but the code tradeoffs are not obvious. Maybe
when we have PPX?
|
|
|
|
This allows the decoupling of the notions of context containing kernel
terms and context containing tactic-level terms.
|
|
|
|
This allows to factorize code and prevents the unnecessary use of back and
forth conversions between the various types of terms.
Note that functions from typing may now raise errors as PretypeError rather
than TypeError, because they call the proper wrapper. I think that they were
wrongly calling the kernel because of an overlook of open modules.
|
|
|
|
|
|
|
|
|
|
The native compiler doesn't support `Require` inside `Module` sections
in some cases, we improve the error message. See:
https://coq.inria.fr/bugs/show_bug.cgi?id=4335
This patch improves the error message and gives the user some
feedback.
|
|
This was observable in long proofs, because side effects kept being
duplicated, leading to an additional cost linear in the size of the proof.
This commit touches kernel files, but the corresponding API is only used
in tactic-facing code so that the side_effects type remains opaque. Thus
it does not affect the kernel safety.
|
|
|
|
This was deactivated by 27c9346 and made an optimization moot in eauto. This
optimization was physically checking for equality, but as lists where rebuilt
by the mapping, this was never true. Some contribs were thus quite slower,
including persistent-union-find which was twice slower.
This 2-line patch fixes it by trying to preserve sharing as much as possible.
Note that we should still do something about eauto, because it does redo
useless work a lot whenever the environment only changes a bit, while we could
cache this computation.
|
|
cofixpoint (bug #5286).
|
|
cofixpoint (bug #5286).
|
|
This also fixes decide equality, discriminate, ... (see e.g. #5279).
|
|
This is really [mv fast_typeops.ml{,i} typeops.ml{,i}] plus trivial
changes in the other files, the real changes are in the parent commit.
|
|
This brings the fix in cad44fc for #2996 to the copy of
Fast_typeops.check_hyps_inclusion.
Fast_typeops.constant_type checks the universe constraints instead of
outputting them. Since everyone who used Typeops.constant_type just
discarded the constraints they've been switched to constant_type_in
which should be the same in Fast_typeops and Typeops.
There are some small differences in the interfaces:
- Typeops.type_of_projection <->
Fast_typeops.type_of_projection_constant to avoid collision with the
internally used type_of_projection (which gives the type of [Proj(p,c)]).
- check_hyps_inclusion takes [('a -> constr)] and ['a] instead of
[constr] for reporting errors.
|
|
|
|
This was yet another bug in the VM long multiplication, that
I unfortunately introduced in ebc509ed2. It was impacting only 32-bit
architectures.
In the future, I'll try to make sure that
1) we provide unit tests for integer arithmetic (my int63 branch ships
with such tests)
2) our continuous testing infrastructure runs the test suite on a 32-bit
architecture. I tried to set up such an instance, but failed. Waiting
for support reply.
|
|
Was PR#366: Univs: fix bug 5208
|
|
|
|
Was PR#377: Univs: fix bug #5180
|
|
We make this warning configurable and disabled by default.
|
|
|
|
In the kernel's generic conversion, backtrack on UniverseInconsistency
for the unfolding heuristic (single backtracking point in reduction).
This exception can be raised in the univ_compare structure to produce
better error messages when the generic conversion function is called
from higher level code in reductionops.ml, which itself is called during
unification in evarconv.ml.
Inside the kernel, the infer and check variants of conversion never
raise UniverseInconsistency though, so this does not change the behavior
of the kernel.
|
|
Universes are kept in normal form w.r.t. equality but not the <=
relation, so the previous check worked almost always but was actually
too strict! In cases like (max(Set,u) = u) when u is declared >= Set it
was failing to find an equality. Applying the KISS principle:
u = v <-> u <= v /\ v <= u.
Fix invariant breakage that triggered the discovery of the check_eq bug as well. No algebraic universes should appear in a term position (on the left of
a colon in a typing judgment), this was not the case when an algebraic
universe instantiated an evar that appeared in the term. We force their
universe variable status to change in refresh_universes to avoid this.
Fix ind sort inference: Use syntactic universe equality for inductive
sort inference instead of check_leq (which now correctly takes
constraints into account) and simplify code
|
|
|
|
If the result had its 30th bit set, then all the high part of the result
on a 64-bit architecture would end up being set, thus breaking subsequent
computations.
This patch also fixes the incorrectly parenthesized definition of
uint32_of_value, which by luck was never misused.
|
|
|
|
Strangely enough, the checker seems to rely on an outdated decompose_app
function which is not the same as the kernel, as the latter is sensitive
to casts. Cast-manipulating functions from the kernel are only used on
upper layers, and thus was moved there.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
The bytecode interpreter ensures that the stack space available at some
points is above a static threshold. However, arbitrary large stack space
can be needed between two check points, leading to segmentation faults
in some cases.
We track the use of stack space at compilation time and add
an instruction to ensure greater stack capacity when required. This is
inspired from OCaml's PR#339 and PR#7168.
Patch written with Benjamin Grégoire.
|
|
|
|
|
|
|
|
|
|
|
