Josh Meredith authored 2 years ago and Cheng Shao committed 2 months ago

(cherry picked from commit d69ecac2)

(cherry picked from commit 1fad5bf1)

(cherry picked from commit 280c49af)

(cherry picked from commit 94673d41)
(cherry picked from commit f13a2942)
(cherry picked from commit 97551107)

(cherry picked from commit 64fba310)
(cherry picked from commit b35707a9)
(cherry picked from commit ead1c3cc)

objc-hi/objcxx-hi should be skipped when cross compiling. The existing
opsys('darwin') predicate only asserts the host system is darwin but
tells us nothing about the target, hence the oversight.

(cherry picked from commit 595c0894)
(cherry picked from commit 5273d3d0)
(cherry picked from commit 6babf56f)

(cherry picked from commit e17f2df9)
(cherry picked from commit 9a2f3313)
(cherry picked from commit 08acf27f)

T17920 was marked as fragile on wasm before; it can be trivially fixed
by avoiding calling variadic printf() in cmm.

(cherry picked from commit 7b4c1998)
(cherry picked from commit c4060ce3)
(cherry picked from commit b64e0f63)

(cherry picked from commit b1e0c313)
(cherry picked from commit 34882192)
(cherry picked from commit 94adf918)

Ever since the removal of the make build system, the in tree lndir
hasn't been actually built, so this patch removes it.

(cherry picked from commit d3a050d2)
(cherry picked from commit 04860c24)

hs_try_putmvar002 includes pthread.h and doesn't work on targets
without this header (e.g. wasm32). It doesn't need to include this
header at all. This was previously unnoticed by wasm CI, though recent
toolchain upgrade brought in upstream changes that completely removes
pthread.h in the single-threaded wasm32-wasi sysroot, therefore we
need to handle that change.

(cherry picked from commit 2ce2a493)
(cherry picked from commit f581edb0)

On posix platforms, when performing read/write on FDs, we check the
nonblocking flag first. For FDs without this flag (e.g. stdout), we
call fdReady() first, which in turn calls poll() to wait for I/O to be
available on that FD. This is problematic for wasm32-wasi: although
select()/poll() is supported via the poll_oneoff() wasi syscall, that
syscall is rather heavyweight and runtime behavior differs in
different wasi implementations. The issue is even worse when targeting
browsers, given there's no satisfactory way to implement async I/O as
a synchronous syscall, so existing JS polyfills for wasi often give up
and simply return ENOSYS.

Before we have a proper I/O manager that avoids poll_oneoff() for
async I/O on wasm, this patch improves the status quo a lot by merely
pretending all FDs are "nonblocking". Read/write on FDs will directly
invoke read()/write(), which are much more reliably handled in
existing wasi implementations, especially those in browsers.

Fixes #23275 and the following test cases: T7773 isEOF001 openFile009
T4808 cgrun025

Approved by CLC proposal #234:
https://github.com/haskell/core-libraries-committee/issues/234

(cherry picked from commit 2eca52b4)
(cherry picked from commit 288692dd)

(cherry picked from commit af771148)
(cherry picked from commit 84c875d1)

The wasm backend didn't properly make use of all Cmm global registers
due to #24347. Now that it is fixed, this patch re-enables full
register mapping for wasm32, and we can now generate smaller & faster
wasm modules that doesn't always spill arguments onto the stack. Fixes #22460 #24152.

(cherry picked from commit 0cda2b8b)
(cherry picked from commit f1f5068b398b1effb837add38ecc5303dc9a381f)
(cherry picked from commit 1e695750)

ghc-api tests for wasm32 are more likely to timeout due to the large
wasm module sizes, especially when testing with wasm native tail
calls, given wasmtime's handling of tail call opcodes are suboptimal
at the moment. It makes sense to increase timeout specifically for
these tests on wasm32. This doesn't affect other targets, and for
wasm32 we don't increase timeout for all tests, so not to risk letting
major performance regressions slip through the testsuite.

(cherry picked from commit 07ab5cc1)
(cherry picked from commit c4d1edb1)

This patch includes all wasm32-specific testsuite fixes.

(cherry picked from commit bd2bfdec)

Sylvain Henry authored 2 years ago and Cheng Shao committed 2 months ago

req_cmm is more informative than js_skip

(cherry picked from commit 239202a2)

This patch adds missing annotations (req_th, req_dynamic_lib_support,
req_rts_linker) to some tests. They were discovered when testing
wasm32, though it's better to be explicit about what features they
require, rather than simply adding when(arch('wasm32'), skip).

(cherry picked from commit b174a110)

This patch adds the req_host_target_ghc predicate to the testsuite to
assert the ghc compiler being tested can compile both host/target
code. When testing cross GHCs this is not supported yet, but it may
change in the future.

(cherry picked from commit cb933665)

This patch adds the req_ghc_with_threaded_rts predicate to the
testsuite to assert the platform has threaded RTS, and mark some tests
as req_ghc_with_threaded_rts. Also makes ghc_with_threaded_rts a
config field instead of a global variable.

(cherry picked from commit aa6afe8a)

Sylvain Henry authored 2 years ago and Cheng Shao committed 2 months ago

See #22630 and !9552

This commit:
 - splits req_smp into req_target_smp and req_ghc_smp
 - changes the testsuite driver to calculate req_ghc_smp
 - changes a handful of tests to use req_target_smp instead of req_smp
 - changes a handful of tests to use req_host_smp when needed

The problem:
 - the problem this solves is the ambiguity surrounding req_smp
 - on master req_smp was used to express the constraint that the program
 being compiled supports smp _and_ that the host RTS (i.e., the RTS used
 to compile the program) supported smp. Normally that is fine, but in
 cross compilation this is not always the case as was discovered in #22630.

The solution:
 - Differentiate the two constraints:
   - use req_target_smp to say the RTS the compiled program is linked
   with (and the platform) supports smp
   - use req_host_smp to say the RTS the host is linked with supports smp

WIP: fix req_smp (target vs ghc)

add flag to separate bootstrapper

split req_smp -> req_target_smp and req_ghc_smp

update tests smp flags

cleanup and add some docstrings

only set ghc_with_smp to bootstrapper on S1 or CC

Only set ghc_with_smp to bootstrapperWithSMP of when testing stage 1
and cross compiling

test the RTS in config/ghc not hadrian

re-add ghc_with_smp

fix and align req names

fix T11760 to use req_host_smp

test the rts directly, avoid python 3.5 limitation

test the compiler in a try block

align out of tree and in tree withSMP flags

mark failing tests as host req smp

testsuite: req_host_smp --> req_ghc_smp

Fix ghc vs host, fix ghc_with_smp leftover

(cherry picked from commit 06036d93)

Andreas Klebinger authored 4 months ago and Luite Stegeman committed 3 months ago

As #14497 showed black holes can appear inside large objects when
we capture a computation and later blackhole it like we do for AP_STACK
closures.

Fixes #24791

(cherry picked from commit 6cee2272)

We found that speculative evaluation can increase the amount of
allocations in some circumstances. This patch adds new flags for
selectively disabling speculative evaluation, allowing us to
test the effect of the optimization.

The new flags are:

  -fspec-eval
     globally enable speculative evaluation

  -fspec-eval-dictfun
     enable speculative evaluation for dictionary functions (no effect
     if speculative evaluation is globally disabled)

The new flags are on by default for all optimisation levels.

See #25284

(cherry picked from commit 23099752)

ARATA Mizuki authored 3 months ago and Luite Stegeman committed 3 months ago

With a recent LLVM, `llc -version` emits the version on the first line
if the vendor is set.  It emits the version on the second line
otherwise.

Therefore, we need to check the both lines to detect the version.

GHC now emits a warning if it fails to detect the LLVM version,
so we can notice if the output of `llc -version` changes in the future.

Also, the warning for using LLVM < 10 on s390x is removed, because
we assume LLVM >= 13 now.

This fixes the definition of __GLASGOW_HASKELL_LLVM__ macro.

Fixes #25606

(cherry picked from commit a928c326)

Jaro Reinders authored 1 year ago and Luite Stegeman committed 3 months ago

In #22010 we established that Int was not always sufficient to store all
the uniques we generate during compilation on 32-bit platforms. This
commit addresses that problem by using Word64 instead of Int for
uniques.

The core of the change is in GHC.Core.Types.Unique and
GHC.Core.Types.Unique.Supply. However, the representation of uniques is
used in many other places, so those needed changes too. Additionally, the RTS
has been extended with an atomic_inc64 operation.

One major change from this commit is the introduction of the Word64Set and
Word64Map data types. These are adapted versions of IntSet and IntMap
from the containers package. These are planned to be upstreamed in the
future.

As a natural consequence of these changes, the compiler will be a bit
slower and take more space on 32-bit platforms. Our CI tests indicate
around a 5% residency increase.

Metric Increase:
    CoOpt_Read
    CoOpt_Singletons
    LargeRecord
    ManyAlternatives
    ManyConstructors
    MultiComponentModules
    MultiComponentModulesRecomp
    MultiLayerModulesTH_OneShot
    RecordUpdPerf
    T10421
    T10547
    T12150
    T12227
    T12234
    T12425
    T12707
    T13035
    T13056
    T13253
    T13253-spj
    T13379
    T13386
    T13719
    T14683
    T14697
    T14766
    T15164
    T15703
    T16577
    T16875
    T17516
    T18140
    T18223
    T18282
    T18304
    T18698a
    T18698b
    T18923
    T1969
    T19695
    T20049
    T21839c
    T3064
    T3294
    T4801
    T5030
    T5321FD
    T5321Fun
    T5631
    T5642
    T5837
    T6048
    T783
    T8095
    T9020
    T9198
    T9233
    T9630
    T9675
    T9872a
    T9872b
    T9872b_defer
    T9872c
    T9872d
    T9961
    TcPlugin_RewritePerf
    UniqLoop
    WWRec
    hard_hole_fits
    MultiLayerModulesTH_Make
    T21839r
    mhu-perf

Metric Decrease:
    MultiLayerModulesTH_Make

(cherry picked from commit 9edcb1fb)

Ben Gamari authored 4 months ago and Luite Stegeman committed 3 months ago

(cherry picked from commit 683115a4)

Sebastian Graf authored 5 months ago and Luite Stegeman committed 3 months ago

This applies the same treatment to `prompt#` as for `catch#`.
See `Note [Strictness for mask/unmask/catch/prompt]`.

Fixes #25439.

(cherry picked from commit 00d58ae1)

Simon Peyton Jones authored 6 months ago and Luite Stegeman committed 3 months ago

This MR fixes #25325

See GHC.Tc.Types.Constraint, Note [Insoluble Wanteds], especially (IW2)

There is a small change in the error message for T14172, but it looks
entirely acceptable to me.

(cherry picked from commit 083703a1)

Sylvain Henry authored 7 months ago and Luite Stegeman committed 3 months ago

A kqueue file descriptor isn't inherited by a child created with fork.
As such we mustn't try to close this file descriptor as we would close a
random one, e.g. the one used by timerfd.

Fix #24672

(cherry picked from commit e7a26d7a)

Matthew Pickering authored 7 months ago and Luite Stegeman committed 3 months ago

-working-dir definitely needs more serious testing, there are some easy
ways to test this.

* Modify Cabal to call ghc using -working-dir rather than changing
  directory.
* Modify the testsuite to run ghc using `-working-dir` rather than
  running GHC with cwd = temporary directory.

However this will have to wait until after 9.12.

Fixes #25150

(cherry picked from commit c20d5186)

Sylvain Henry authored 8 months ago and Luite Stegeman committed 3 months ago

Support for rubbish dynamic literals was added in #24664. This patch
does the same for static literals.

Fix #25177

(cherry picked from commit 5092dbff)

Sylvain Henry authored 8 months ago and Luite Stegeman committed 3 months ago

This patch is part of the patches upstreamed from haskell.nix.
See https://github.com/input-output-hk/haskell.nix/pull/1960 for the
original report/patch.

(cherry picked from commit c749bdfd)

Sylvain Henry authored 8 months ago and Luite Stegeman committed 3 months ago

- beef6135 enabled the use of
  MO_Add/MO_Sub for 64-bit operations in the C and LLVM backends
- 6755d833 did the same for the x86 NCG
  backend

However we store some literal values as `Int` in the compiler. As a
result, some Cmm optimizations transformed target 64-bit literals into
compiler `Int`. If the compiler is 32-bit, this leads to computing with
wrong literals (see #24893 and #24700).

This patch disables these Cmm optimizations for 32-bit compilers. This
is unsatisfying (optimizations shouldn't be compiler-word-size
dependent) but it fixes the bug and it makes the patch easy to backport.
A proper fix would be much more invasive but it shall be implemented in
the future.

Co-authored-by: amesgen <amesgen@amesgen.de>
(cherry picked from commit 7446a09a)

Matthew Pickering authored 8 months ago and Luite Stegeman committed 3 months ago

It was assumed that module names were unique but that isn't true with
multiple units.

The fix is quite simple, maintain a set of `(ModuleName, UnitId)` and
query that to see whether the module has been specified.

Fixes #25122

(cherry picked from commit 951ce3d5)

Simon Peyton Jones authored 9 months ago and Luite Stegeman committed 3 months ago

See Note [Case-of-case and full laziness]
in GHC.Driver.Config.Core.Opt.Simplify

(cherry picked from commit de5d9852)

Andreas Klebinger authored 9 months ago and Luite Stegeman committed 3 months ago

It seems we reserve 8 registers instead of four for global regs
based on the layout in Note [AArch64 Register assignments].

I'm not sure it's neccesary, but for now we just accept this state of
affairs and simple update -fregs-graph to account for this.

(cherry picked from commit 3f89ab92)

Peter Trommler authored 9 months ago and Luite Stegeman committed 3 months ago

Promote 8 bit and 16 bit signed arguments by sign extension.

Fixes #25018

(cherry picked from commit a82121b3)

Zubin authored 11 months ago and Luite Stegeman committed 3 months ago

We need to recompile if this flag is changed because later modules might depend on the
simplified core for this module if -fprefer-bytecode is enabled.

Fixes #24656

(cherry picked from commit dddc9dff)