GitLab

Summary:
This hugs heritage gone away with commit f1dffa02
in 2001
Signed-off-by: Sergei Trofimovich <slyfox@gentoo.org>

Test Plan: tried to find HAVE_BIN_SH in internet, no success

Reviewers: simonmar, austin

Reviewed By: austin

Subscribers: phaskell, simonmar, relrod, ezyang, carter

Differential Revision: https://phabricator.haskell.org/D117

Update Haddock submodule

The change here is to make INCOHERENT slightly more permissive:

  if the selected candidate is incoherent
  then ignore all unifying candidates

This allows us to move the {-# INCOHERENT #-} pragma from
  from   instance Typeable (f a)
  to     Typeable (n:Nat) and Typable (s:Symbol)
where it belongs, and where Trac #9242 said it should be.

I don't think this will affect anyone.

I've updated the user manual.

Please adjust the perf number on your platform if/when it fails. It
should improve slightly. Updates submodule.

This  +4% increase (from -1% before) was caused by 1fc60ea1. But that
commit did not cause any other regressions, so I’m not investigating
further.

Summary:
New function signal_exit_code hides differences between target platforms
handling of fatal error signals and the applications' exit codes. E.g.
on Linux the application exit code which receives fatal error signal
is encoded as 128 + signal value. On the other hand on Solaris
the application exit code is signal value alone.

Test Plan: validated on Linux and tested on Solaris

Reviewers: austin, simonmar

Reviewed By: simonmar

Subscribers: phaskell, simonmar, relrod, ezyang, carter

Differential Revision: https://phabricator.haskell.org/D122

Summary:
Caught by './validate --slow' in 'ghci/scripts/ghci024'.

Commit 20787529 changed indentation a bit:

    --- ./ghci/scripts/ghci024.stdout       2014-07-31 12:05:34.000000000 +0300
    +++ ./ghci/scripts/ghci024.run.stdout   2014-08-07 17:19:23.000000000 +0300
    @@ -33,4 +33,4 @@
     active package flags: none
     ~~~~~~~~~~ Testing :show packages, including the ghc package
     active package flags:
    -  -package ghc
    +-package ghc

Patch restores indentation.
Signed-off-by: Sergei Trofimovich <slyfox@gentoo.org>

Test Plan: passed validate, tested manually

Reviewers: austin, ezyang

Reviewed By: ezyang

Subscribers: phaskell, simonmar, relrod, ezyang, carter

Differential Revision: https://phabricator.haskell.org/D126

Signed-off-by: Edward Z. Yang <ezyang@cs.stanford.edu>

I bisected the performance difference in Haddock and found it was due to
d6aec63c009c4e57181900eb03847d7dc0fc3c7c, which I accidentally picked up
when updating Haddock 00b8f8c5.  The
performance regression is justified by the fact that we are now actually
processing URLs in Haddock comments that we were not previously, so
there would be more allocation.  Time use was not affected.

The TODOs simply reflect the fact that we need updated numbers for
32-bit Linux and Windows.  Please add them when you get a chance.
Signed-off-by: Edward Z. Yang <ezyang@cs.stanford.edu>

Summary:
Previously, we allocated uniques for strings starting at zero, which
means the tag bits in the unique are zero, which means that printing a
Unique for a string will start with a null byte.  This is bad.  So
instead, start our numbering with the tag byte as '$' (as in $tring).
This is hard coded so we don't have to worry about the optimizer
reducing the expression.
Signed-off-by: Edward Z. Yang <ezyang@cs.stanford.edu>

Test Plan: validate

Reviewers: hvr, simonmar, austin

Subscribers: simonmar, relrod, ezyang, carter

Differential Revision: https://phabricator.haskell.org/D123

GHC Trac Issues: #9413

According to the definition of has_side_effets in PrimOp,
raise# clearly has side effects!  In practice it makes little
difference becuase the fact that it returns bottom is more
important... but still it's better to say it right.

Mainly in Simplify.rebuildCase.  The old code wasn't wrong, but I kept
mis-understanding it.  This patch cuts splits out "pure seq" from "strict
let", which makes it much easier to grok.

In fixing Trac #9390 I discovered that I *still* didn't really understand
what the can_fail and has_side_effects properties of a PrimOp mean, precisely.

The big new things I learned are

* has_side_effects needs to be true only of *write* effects,
  Reads (which are, strictly speaking, effects) don't matter here.

* has_side_effects must be true of primops that can throw a synchronous
  Haskell exception (eg raiseIO#)

* can_fail is true only of primops that can cause an *unchecked* (not
  Haskell) system exception, like divide by zero, or accessing memory
  out of range through an array read or write.

I've documented all this now.  The changes in this patch are only
in comments.

It's not obvious why the simplifier generates code that correctly satisfies
the let/app invariant.   This patch does some minor refactoring, but the main
point is to document pre-conditions to key functions, namely that the rhs
passed in satisfies the let/app invariant.

There shouldn't be any change in behaviour.

Vectorise.Generic.PAMethods.buildToArrPReprs was building an expression like
   pvoids# (lengthSels2# sels)
which does not satisfy the let/app invariant.  It should be more like
  case lengthSels2# sels of l -> pvoids# l

This was caught by Core Lint (once it was taught to check for the invariant)

This is actually the bug that triggered Trac #9390.  We had
an unboxed tuple (# writeArray# ..., () #), and that writeArray#
argument isn't ok-for-speculation, so disobeys the invariant.

The desugaring of unboxed tuples was to blame; the fix is easy.

We were inadvertently destroying the let/app invariant,
by floating into an unlifted function argument.

If we have an invariant, Lint should jolly well check it.

(And indeed, adding this test throws up Lint errors that
are fixed in separate patches.)

The issue is presented in Makefile logic where it attempts to start
linker_unload and pass it HSinteger-gmp library for unload,
but the library name is prefixed with two directories names. The first
is of ghc's integer-gmp/build itself and another is the directory name
passed to --with-gmp-libraries= configure parameter. The testcase then
fails on unloading integer-gmp/build directory thinking that this is
a library to unload. The issue is solved by cuting (head -1) the first
library name from the list and using this for unloading the HSinteger-gmp
library. I use head -1 instead of cut -d ' ' here since ghc may be
installed into the directory with space(s) in its name like in the case
when running validate.

Thanks to slyfox for noticing this reregression.

This reverts commit 65e5dbcd.

Signed-off-by: Edward Z. Yang <ezyang@cs.stanford.edu>

Summary:
Previously, if you ran trace "foo\0bar", the output was truncated so
that everything after the null byte was omitted.  This was terrible.
Now we filter out null bytes, and emit an extra trace saying that
null bytes were filtered out.

NB: we CANNOT fix debugBelch, because all printf variants *always*
respect null bytes, even if you're using string precision such as
%.*s.  The alternative would have been to introduce a new function
debugRawBelch which did not use format strings and took an explicit
string length, but I decided we generally should avoid putting null
bytes in our trace messages, and warn the user.
Signed-off-by: Edward Z. Yang <ezyang@cs.stanford.edu>

Test Plan: validate

Reviewers: hvr, austin

Subscribers: simonmar, relrod, ezyang, carter

Differential Revision: https://phabricator.haskell.org/D121

GHC Trac Issues: #9395

Signed-off-by: Edward Z. Yang <ezyang@cs.stanford.edu>

Summary:
Another macro borrowed from hugs, gone aways in
commit 528a7d2cSigned-off-by: Sergei Trofimovich <slyfox@gentoo.org>

Test Plan: build-tested

Reviewers: simonmar, austin, ezyang

Reviewed By: austin, ezyang

Subscribers: phaskell, simonmar, relrod, ezyang, carter

Differential Revision: https://phabricator.haskell.org/D118

Signed-off-by: Edward Z. Yang <ezyang@cs.stanford.edu>

Summary:
This patch set adds support for extra syntax on -package and related
arguments which allow you to thin and rename modules from a package.
For example, this argument:

    -package "base (Data.Bool as Bam, Data.List)"

adds two more modules into scope, Bam and Data.List, without adding
any of base's other modules to scope.

These flags are additive: so, for example, saying:

    -hide-all-packages -package base -package "base (Data.Bool as Bam)"

will provide both the normal bindings for modules in base, as well as
the module Bam.

There is also a new debug flag -ddump-mod-map which prints the state
of the module mapping database.  H = hidden, E = exposed (so for
example EH says the module in question is exported, but in a hidden
package.)

Module suggestions have been minorly overhauled to work better with reexports:
if you have -package "base (Data.Bool as Bam)" and mispell Bam, GHC
will suggest "Did you mean Bam (defined via package flags to be
base:Data.Bool)"; and generally you will get more accurate information.
Also, fix a bug where we suggest the -package flag when we really need
the -package-key flag.

NB: The renaming afforded here does *not* affect what wired in
symbols GHC generates.  (But it does affect implicit prelude!)

ToDo: add 'hiding' functionality, to make it easier to support the alternative
prelude use-case.

ToDo: Cabal support
Signed-off-by: Edward Z. Yang <ezyang@cs.stanford.edu>

Test Plan: new tests and validate

Reviewers: simonpj, simonmar, hvr, austin

Subscribers: simonmar, relrod, ezyang, carter

Differential Revision: https://phabricator.haskell.org/D113

GHC Trac Issues: #9375

You can parametrize over the different selection by using a
different PackageArg.  This helps reduce code duplication.
Signed-off-by: Edward Z. Yang <ezyang@cs.stanford.edu>

Instead of building a multiply indirected data structure and querying
it on every import, we now have two data structures moduleToPkgConf
and moduleToPkgConfAll.  moduleToPkgConf is a single-level UniqFM that
is intended to be used for most valid imports; however, it does not
contain any information useful for error reporting.  If an error is
occurred, we then query moduleToPkgConfAll, which contains a more
comprehensive view of the package database.  This field is lazily
initialized (so this means we're retaining the package database list,
but this should be fine because we're already maintaining the entries
of the list.)  Additionally, the full view doesn't keep track of a boolean
toggle for visibility/exposure anymore, but instead tracks the *provenance*
of how the module binding came to be (the ModuleOrigin data type).

Additionally, we move the logic for determining if a module is exposed
or not from Finder.lhs and put it in Packages.lhs; this information is
communicated via the LookupResult data type.  Unfortunately, we can't
directly return a FindResult, because this data type is defined in
HscTypes which depends on Packages.  This is going to change some more
in the near future when I add thinning/renaming to package flags; the
error messages will need to be more flexible.

I've also slightly changed the semantics of error messages for package
qualified imports.  Previously, if we didn't find any package qualified
imports, but there were hidden modules in a *different* package, the error
message would prefer mentioning those as opposed to providing suggestions.
Now, if a module is hidden but in the wrong package, we won't mention it;
instead, it will get mentioned with the other module suggestions.  I
was too lazy to write a test, but I can add one if people would like.

The module reexport bug was, package q reexported p:P as Conflict,
and package r reexported p:P2 as Conflict, this was *not* reported as
a conflict, because the old logic incorrectly decided that P and P2 were
the same module on account of being from the same package.  The logic here
has been corrected.

Contains haddock submodule update.
Signed-off-by: Edward Z. Yang <ezyang@cs.stanford.edu>