| Age | Commit message (Collapse) | Author |
|---|
|
|
|
* hasse/dialyzer/memory_opt/OTP-14127:
dialyzer: Attempt to reduce memory consumption
dialyzer: Do not spawn all workers at once
dialyzer: Write PLT in subprocess
dialyzer: Remove code for non-started workers
dialyzer: Sort graphs topologically
dialyzer: Compress some PLT tables
dialyzer: Check return value of digrahp:add_edge()
dialyzer: Use less memory for the PLT when analyzing
typer: Fix a bug regarding the -T option
dialyzer: Do not keep the code server's type info on the heap
|
|
* maint:
dialyzer: Fix a Dialyzer warning
|
|
|
|
The 'deps' field is reset when no longer used, which could possibly
save some memory.
|
|
Spawning all worker processes at once has the potential to increase
peak memory consumption. Therefore the implementation is now slightly
modified: `20 * dialyzer_utils:parallelism()' processes are running in
parallel. 20 i quite a big factor, but seems necessary to keep all
schedulers busy according to the Observer application's Load Charts,
with a 100 ms update interval.
|
|
At common case, which will otherwise leave a big heap.
|
|
Since SCCs and modules are now topologically sorted (relative the
active graph), we can safely assume that workers have been started.
|
|
Although some variable names indicate that lists of SCCs and modules
are sorted, that was not always the case. The effect on the execution
time of sorting them doesn't seem to be significant.
dialyzer_coordinator:sccs_to_pids() should now always return an empty
second list (not yet started workers).
Since the condensation of graphs often needs a lot of heap memory, it
is run in a separate process.
|
|
Compressing the signatures or the contract records would cost time.
The contract records are still held in two places (code server and
PLT) in many cases.
|
|
|
|
The two tables for (record) types and exported types are no longer
kept in memory (by the PLT) when analyzing, which reduces memory
consumption.
A first step towards removing (or replacing) the #plt{} record.
|
|
The -T option has not been working for a long time.
|
|
The table used during translation of forms to types used to be a
dictionary of the type info (-type(), -opaque(), and -record()) for
all modules. Now it is reduced to the modules that are needed to
handle references to types in other modules.
The PLT:s type info is still kept on the heap.
|
|
* maint:
dialyzer: Increase time limit for tests
dialyzer: Optimize typesig
dialyzer: Optimize evaluation of complex code
dialyzer: Optimize collection of variables
Conflicts:
lib/dialyzer/src/dialyzer_typesig.erl
|
|
dialyzer_typesig:traverse_scc() now resets the context for each
element of the SCC. Since the results of traversing the elements are
saved in the 'cmap' table, there is no need to create an SCC
conjunction. For huge SCCs this saves quite some time: the
lift_lists() function added one element at a time, which made the
calculation of deps very slow.
|
|
Improve the evaluation of long lists of constraints.
|
|
* maint:
dialyzer: Compact 'file' annotations in Core code
dialyzer: Try to reduce memory usage
dialyzer: Use less memory when translating contracts
dialyzer: Use maps instaed of dict
dialyzer: Use maps instead of dict for module contracts map
dialyzer: Compress a few more ETS tables
dialyzer: Optimize memory consumption
dialyzer: Reduce memory consumption during 'remote' phase
dialyzer: Update code for finding parallelism
compiler: Do not spawn process when dialyzing
dialyzer: Reduce ETS usage during the typesig phase
dialyzer: Optimize graph condensation
dialyzer: Do not send full PLTs as messages
|
|
File annotations {file, File} in Core Erlang code is replaced by
{file, FakeFile}, where FakeFile is "0", "1", ..., in order to save
memory. When a warning message is created, FakeFile is translated to
the original File (a bit awkward but easy to do).
The Core code stored in ETS tables needs less space, but more
important is that a less heap memory is used in the dataflow,
warnings, and typesig phases, allowing more parallelism.
|
|
The translation from forms to types is done in a separate process
in an attempt to reduce peak memory usage.
Expect further optimizations as it is probably not feasible
in the long run to keep all type information on the heap.
|
|
|
|
|
|
|
|
|
|
Some SCC info is no longer cached. It reduces the peak memory
consumption of workers during the typesig phase, and seems to cost
very little time.
|
|
The cache used for speeding up the translation from forms to types is
no longer global, but reset per module. The peak memory consumption
is reduced, and the added time seems to be small.
|
|
Using erlang:system_info(schedulers_online) for determining
parallelism is recommended by OTP experts. It takes care of "+S 4:1"
and simultaneous loading of HiPE code.
|
|
The condensed graph of SCCs occupies less ETS memory. A table
translating to and from SCC to a unique integer is introduced.
|
|
By not using ETS when calculating the condensation of graphs, peak
heap memory consumption is reduced.
|
|
The mini PLT is extended to hold all data of the full PLT, and the
full PLT is restored when needed (for storing the PLT on file).
|
|
|
|
|
|
|
|
* maint:
dialyzer: Substitute 'opacity' for 'opaqueness'
dialyzer: Improve a warning message
dialyzer: Improve a warning message
dialyzer: Correct a warnings message
|
|
Suggested by Kostis.
|
|
Messages regarding guards with orelse/andalso could look
like "Clause guard cannot succeed. The variable A was matched
against the type any()". Now they look like as if or/and is
used: "Guard test is_integer(A::atom()) can never succeed".
|
|
Messages like "Invalid type specification for function
para3:exp_adt/0. The success typing is () -> 3" now look like
"The specification for para3:exp_adt/0 has an opaque subtype
para3_adt:exp1(para3_adt:exp2()) which is violated by the success
typing () -> 3".
The old message did not give any clue as to what invalidated the
contract, namely the opaque subtype.
|
|
Messages like "The attempt to match a term of type rec_api:f()
against the variable _ breaks the opaqueness of rec_adt:f()" now
look like "The attempt to match a term of type rec_adt:f()
against the record field 'f' with type rec_api:f() breaks the
opaqueness of the term".
|
|
* maint:
dialyzer: Fix error handling of bad -dialyzer() attributes
|
|
If a dialyzer module attribute references an undefined (local)
function an error tuple is thrown but no caught. This leads to a
'nocatch' error and not the intended "clean" error.
|
|
The previous variable names can be generated by
projects like LFE and Elixir, leading to possible
conflicts. Our first to choice to solve such conflicts
was to use $ but that's not a valid variable name in core.
Therefore we picked @ which is currently supported and
still reduces the chance of conflicts.
|
|
|
|
Let it handle `dial_warning()` input as well, to match what `format_warning/2` expects.
|
|
Warnings about clauses that cannot match and are also compiler generated
are suppressed unless none of the clauses return.
This feature is useful for non-Erlang BEAM languages (such as Elixir)
that compile to Erlang and expand certain language constructs into case
statements. In that case, as long as the language construct can succeed,
these warnings are undesired and appear spurious to users that do not
check the Erlang code that their program expands into.
Thanks to José Valim for the test (slightly modified).
|
|
|
|
|
|
|
|
|
|
* stavros/dialyzer/fix_call_site_analysis/OTP-13655/PR-1092:
Fix a bug in Dialyzer related to call-site analysis
|
|
Dialyzer's "dataflow" module is using information from the callgraph to
determine which functions may be called at a particular call-site. Unfortunately
this information can include functions that are certainly not among the possible
choices. We don't want to emit warnings in such cases, so a "reasonable"
compromise is to stay silent if there are many possible funs and at least one of
them can succeed.
Bug reported by Dan Gudmundsson, test shrunk down by Magnus Lång.
|
