[Git][ghc/ghc][wip/simplifier-tweaks] 85 commits: EPA: Introduce HasAnnotation class
Simon Peyton Jones (@simonpj)
gitlab at gitlab.haskell.org
Fri Oct 20 16:00:44 UTC 2023
Simon Peyton Jones pushed to branch wip/simplifier-tweaks at Glasgow Haskell Compiler / GHC
Commits:
94066d58 by Alan Zimmerman at 2023-10-09T21:35:53-04:00
EPA: Introduce HasAnnotation class
The class is defined as
class HasAnnotation e where
noAnnSrcSpan :: SrcSpan -> e
This generalises noAnnSrcSpan, and allows
noLocA :: (HasAnnotation e) => a -> GenLocated e a
noLocA = L (noAnnSrcSpan noSrcSpan)
- - - - -
8792a1bc by Ben Gamari at 2023-10-09T21:36:29-04:00
Bump unix submodule to v2.8.3.0
- - - - -
e96c51cb by Andreas Klebinger at 2023-10-10T16:44:27+01:00
Add a flag -fkeep-auto-rules to optionally keep auto-generated rules around.
The motivation for the flag is given in #21917.
- - - - -
3ed58cef by Matthew Pickering at 2023-10-10T19:01:22-04:00
hadrian: Add ghcToolchain to tool args list
This allows you to load ghc-toolchain and ghc-toolchain-bin into HLS.
- - - - -
476c02d4 by Matthew Pickering at 2023-10-10T19:01:22-04:00
ghc-toolchain: Normalise triple via config.sub
We were not normalising the target triple anymore like we did with the
old make build system.
Fixes #23856
- - - - -
303dd237 by Matthew Pickering at 2023-10-10T19:01:22-04:00
ghc-toolchain: Add missing vendor normalisation
This is copied from m4/ghc_convert_vendor.m4
Towards #23868
- - - - -
838026c9 by Matthew Pickering at 2023-10-10T19:01:22-04:00
ghc-toolchain: Add loongarch64 to parseArch
Towards #23868
- - - - -
1a5bc0b5 by Matthew Pickering at 2023-10-10T19:01:22-04:00
Add same LD hack to ghc-toolchain
In the ./configure script, if you pass the `LD` variable then this has
the effect of stopping use searching for a linker and hence passing
`-fuse-ld=...`.
We want to emulate this logic in ghc-toolchain, if a use explicilty
specifies `LD` variable then don't add `-fuse-ld=..` with the goal of
making ./configure and ghc-toolchain agree on which flags to use when
using the C compiler as a linker.
This is quite unsavoury as we don't bake the choice of LD into the
configuration anywhere but what's important for now is making
ghc-toolchain and ./configure agree as much as possible.
See #23857 for more discussion
- - - - -
42d50b5a by Ben Gamari at 2023-10-10T19:01:22-04:00
ghc-toolchain: Check for C99 support with -std=c99
Previously we failed to try enabling C99 support with `-std=c99`, as
`autoconf` attempts. This broke on older compilers (e.g. CentOS 7) which
don't enable C99 by default.
Fixes #23879.
- - - - -
da2961af by Matthew Pickering at 2023-10-10T19:01:22-04:00
ghc-toolchain: Add endianess check using __BYTE_ORDER__ macro
In very old toolchains the BYTE_ORDER macro is not set but thankfully
the __BYTE_ORDER__ macro can be used instead.
- - - - -
d8da73cd by Matthew Pickering at 2023-10-10T19:01:22-04:00
configure: AC_PATH_TARGET_TOOL for LD
We want to make sure that LD is set to an absolute path in order to be
consistent with the `LD=$(command -v ld)` call. The AC_PATH_TARGET_TOOL
macro uses the absolute path rather than AC_CHECK_TARGET_TOOL which
might use a relative path.
- - - - -
171f93cc by Matthew Pickering at 2023-10-10T19:01:22-04:00
ghc-toolchain: Check whether we need -std=gnu99 for CPP as well
In ./configure the C99 flag is passed to the C compiler when used as a C
preprocessor. So we also check the same thing in ghc-toolchain.
- - - - -
89a0918d by Matthew Pickering at 2023-10-10T19:01:22-04:00
Check for --target linker flag separately to C compiler
There are situations where the C compiler doesn't accept `--target` but
when used as a linker it does (but doesn't do anything most likely)
In particular with old gcc toolchains, the C compiler doesn't support
--target but when used as a linker it does.
- - - - -
37218329 by Matthew Pickering at 2023-10-10T19:01:22-04:00
Use Cc to compile test file in nopie check
We were attempting to use the C compiler, as a linker, to compile a file
in the nopie check, but that won't work in general as the flags we pass
to the linker might not be compatible with the ones we pass when using
the C compiler.
- - - - -
9b2dfd21 by Matthew Pickering at 2023-10-10T19:01:22-04:00
configure: Error when ghc-toolchain fails to compile
This is a small QOL change as if you are working on ghc-toolchain and it
fails to compile then configure will continue and can give you outdated
results.
- - - - -
1f0de49a by Matthew Pickering at 2023-10-10T19:01:22-04:00
configure: Check whether -no-pie works when the C compiler is used as a linker
`-no-pie` is a flag we pass when using the C compiler as a linker (see
pieCCLDOpts in GHC.Driver.Session) so we should test whether the C
compiler used as a linker supports the flag, rather than just the C
compiler.
- - - - -
62cd2579 by Matthew Pickering at 2023-10-10T19:01:22-04:00
ghc-toolchain: Remove javascript special case for --target detection
emcc when used as a linker seems to ignore the --target flag, and for
consistency with configure which now tests for --target, we remove this
special case.
- - - - -
0720fde7 by Ben Gamari at 2023-10-10T19:01:22-04:00
toolchain: Don't pass --target to emscripten toolchain
As noted in `Note [Don't pass --target to emscripten toolchain]`,
emscripten's `emcc` is rather inconsistent with respect to its treatment
of the `--target` flag. Avoid this by special-casing this toolchain
in the `configure` script and `ghc-toolchain`.
Fixes on aspect of #23744.
- - - - -
6354e1da by Matthew Pickering at 2023-10-10T19:01:22-04:00
hadrian: Don't pass `--gcc-options` as a --configure-arg to cabal configure
Stop passing -gcc-options which mixed together linker flags and
non-linker flags. There's no guarantee the C compiler will accept both
of these in each mode.
- - - - -
c00a4bd6 by Ben Gamari at 2023-10-10T19:01:22-04:00
configure: Probe stage0 link flags
For consistency with later stages and CC.
- - - - -
1f11e7c4 by Sebastian Graf at 2023-10-10T19:01:58-04:00
Stricter Binary.get in GHC.Types.Unit (#23964)
I noticed some thunking while looking at Core.
This change has very modest, but throughout positive ghc/alloc effect:
```
hard_hole_fits(normal) ghc/alloc 283,057,664 281,620,872 -0.5%
geo. mean -0.1%
minimum -0.5%
maximum +0.0%
```
Fixes #23964.
- - - - -
a4f1a181 by Bryan Richter at 2023-10-10T19:02:37-04:00
rel_eng/upload.sh cleanups
- - - - -
80705335 by doyougnu at 2023-10-10T19:03:18-04:00
ci: add javascript label rule
This adds a rule which triggers the javascript job when the "javascript"
label is assigned to an MR.
- - - - -
a2c0fff6 by Matthew Craven at 2023-10-10T19:03:54-04:00
Make 'wWarningFlagsDeps' include every WarningFlag
Fixes #24071.
- - - - -
d055f099 by Jan Hrček at 2023-10-10T19:04:33-04:00
Fix pretty printing of overlap pragmas in TH splices (fixes #24074)
- - - - -
0746b868 by Andreas Klebinger at 2023-10-10T19:05:09-04:00
Aarch64 NCG: Use encoded immediates for literals.
Try to generate
instr x2, <imm>
instead of
mov x1, lit
instr x2, x1
When possible. This get's rid if quite a few redundant
mov instructions.
I believe this causes a metric decrease for LargeRecords as
we reduce register pressure.
-------------------------
Metric Decrease:
LargeRecord
-------------------------
- - - - -
739f4e6f by Andreas Klebinger at 2023-10-10T19:05:09-04:00
AArch NCG: Refactor getRegister'
Remove some special cases which can be handled just as well by the
generic case.
This increases code re-use while also fixing #23749. Since some of the
special case wasn't upholding Note [Signed arithmetic on AArch64].
- - - - -
1b213d33 by Andreas Klebinger at 2023-10-10T19:05:09-04:00
Aarch ncg: Optimize immediate use for address calculations
When the offset doesn't fit into the immediate we now just reuse the
general getRegister' code path which is well optimized to compute the
offset into a register instead of a special case for CmmRegOff.
This means we generate a lot less code under certain conditions which is
why performance metrics for these improve.
-------------------------
Metric Decrease:
T4801
T5321FD
T5321Fun
-------------------------
- - - - -
b7df0732 by John Ericson at 2023-10-11T16:02:11-04:00
RTS configure: Move over mem management checks
These are for heap allocation, a strictly RTS concern.
All of this should boil down to `AC_DEFINE` not `AC_SUBST`, so it
belongs in the RTS configure and should be safe to move without
modification.
The RTS configure one has a new
```
AC_CHECK_SIZEOF([void *])
```
that the top-level configure version didn't have, so that
`ac_cv_sizeof_void_p` is defined. Once more code is moved over in latter
commits, that can go away.
Progress towards #17191
- - - - -
41130a65 by John Ericson at 2023-10-11T16:02:11-04:00
RTS configure: Move over `__thread` check
This used by (@bgamari thinks) the `GCThread` abstraction in the RTS.
All of this should boil down to `AC_DEFINE` not `AC_SUBST`, so it
belongs in the RTS configure and should be safe to move without
modification.
Progress towards #17191
- - - - -
cc5ec2bd by John Ericson at 2023-10-11T16:02:11-04:00
RTS configure: Move over misc function checks
These are for general use in the RTS.
All of this should boil down to `AC_DEFINE` not `AC_SUBST`, so it
belongs in the RTS configure and should be safe to move without
modification.
Progress towards #17191
- - - - -
809e7c2d by John Ericson at 2023-10-11T16:02:11-04:00
RTS configure: Move over `eventfd` check
This check is for the RTS part of the event manager and has a
corresponding part in `base`.
All of this should boil down to `AC_DEFINE` not `AC_SUBST`, so it
belongs in the RTS configure and should be safe to move without
modification.
Progress towards #17191
- - - - -
58f3babf by John Ericson at 2023-10-11T16:02:48-04:00
Split `FP_CHECK_PTHREADS` and move part to RTS configure
`NEED_PTHREAD_LIB` is unused since
3609340743c1b25fdfd0e18b1670dac54c8d8623 (part of the make build
system), and so is no longer defined.
Progress towards #17191
- - - - -
e99cf237 by Moritz Angermann at 2023-10-11T16:03:24-04:00
nativeGen: section flags for .text$foo only
Commit 3ece9856d157c85511d59f9f862ab351bbd9b38b, was supposed to fix
#22834 in !9810.
It does however add "xr" indiscriminatly to .text sections
even if splitSections is disabled. This leads to the assembler saying:
ghc_1.s:7849:0: error:
Warning: Ignoring changed section attributes for .text
|
7849 | .section .text,"xr"
| ^
- - - - -
f383a242 by Sylvain Henry at 2023-10-11T16:04:04-04:00
Modularity: pass TempDir instead of DynFlags (#17957)
- - - - -
34fc28b0 by John Ericson at 2023-10-12T06:48:28-04:00
Test that functions from `mingwex` are available
Ryan wrote these two minimizations, but they never got added to the test
suite.
See #23309, #23378
Co-Authored-By: Ben Gamari <bgamari.foss at gmail.com>
Co-Authored-By: Ryan Scott <ryan.gl.scott at gmail.com>
- - - - -
bdb54a0e by John Ericson at 2023-10-12T06:48:28-04:00
Do not check for the `mingwex` library in `/configure`
See the recent discussion in !10360 --- Cabal will itself check for the
library for the packages that need it, and while the autoconf check
additionally does some other things like define a `HAS_LIBMINGWEX` C
Preprocessor macro, those other things are also unused and unneeded.
Progress towards #17191, which aims to get rid of `/configure` entirely.
- - - - -
43e814e1 by Ben Gamari at 2023-10-12T06:49:40-04:00
base: Introduce move modules into src
The only non-move changes here are whitespace changes to pass the
`whitespace` test and a few testsuite adaptations.
- - - - -
df81536f by Moritz Angermann at 2023-10-12T06:50:16-04:00
[PEi386 linker] Bounds check and null-deref guard
We should resonably be able to expect that we won't exceed the number of
sections if we assume to be dealing with legal object files. We can however
not guarantee that we get some negative values, and while we try to
special case most, we should exclude negative indexing into the sections
array.
We also need to ensure that we do not try to derefences targetSection,
if it is NULL, due to the switch statement.
- - - - -
c74c4f00 by John Ericson at 2023-10-12T10:31:13-04:00
Move apple compat check to RTS configure
- - - - -
c80778ea by John Ericson at 2023-10-12T10:31:13-04:00
Move clock/timer fun checks to RTS configure
Actual library check (which will set the Cabal flag) is left in the
top-level configure for now.
Progress towards #17191
- - - - -
7f9f2686 by John Ericson at 2023-10-12T10:31:13-04:00
Move visibility and "musttail" annotation checks to the RTS configure
All of this should boil down to `AC_DEFINE` not `AC_SUBST`, so it
belongs in the RTS configure and should be safe to move without
modification.
Progress towards #17191
- - - - -
ffb3efe6 by John Ericson at 2023-10-12T10:31:13-04:00
Move leading underscore checks to RTS configure
`CabalLeadingUnderscore` is done via Hadrian already, so we can stop
`AC_SUBST`ing it completely.
- - - - -
25fa4b02 by John Ericson at 2023-10-12T10:31:13-04:00
Move alloca, fork, const, and big endian checks to RTS configure
All of this should boil down to `AC_DEFINE` not `AC_SUBST`, so it
belongs in the RTS configure and should be safe to move without
modification.
- - - - -
5170f42a by John Ericson at 2023-10-12T10:31:13-04:00
Move libdl check to RTS configure
- - - - -
ea7a1447 by John Ericson at 2023-10-12T10:31:13-04:00
Adjust `FP_FIND_LIBFFI`
Just set vars, and `AC_SUBST` in top-level configure.
Don't define `HAVE_SYSTEM_LIBFFI` because nothing is using it. It hasn't
be in used since 3609340743c1b25fdfd0e18b1670dac54c8d8623 (part of the
make build system).
- - - - -
f399812c by John Ericson at 2023-10-12T10:31:13-04:00
Split BFD support to RTS configure
The flag is still in the top-level configure, but the other checks
(which define various macros --- important) are in the RTS configure.
- - - - -
f64f44e9 by John Ericson at 2023-10-12T10:31:13-04:00
Split libm check between top level and RTS
- - - - -
dafc4709 by Moritz Angermann at 2023-10-12T10:31:49-04:00
CgUtils.fixStgRegStmt respect register width
This change ensure that the reg + offset computation is always of the
same size. Before this we could end up with a 64bit register, and then
add a 32bit offset (on 32bit platforms). This not only would fail type
sanity checking, but also incorrectly truncate 64bit values into 32bit
values silently on 32bit architectures.
- - - - -
9e6ef7ba by Matthew Pickering at 2023-10-12T20:35:00-04:00
hadrian: Decrease verbosity of cabal commands
In Normal, most tools do not produce output to stdout unless there are
error conditions.
Reverts 7ed65f5a1bc8e040e318ccff395f53a9bbfd8217
- - - - -
08fc27af by John Ericson at 2023-10-12T20:35:36-04:00
Do not substitute `@...@` for stage-specific values in cabal files
`rts` and `ghc-prim` now no longer have a `*.cabal.in` to set Cabal flag
defaults; instead manual choices are passed to configure in the usual
way.
The old way was fundamentally broken, because it meant we were baking
these Cabal files for a specific stage. Now we only do stage-agnostic
@...@ substitution in cabal files (the GHC version), and so all
stage-specific configuration is properly confined to `_build` and the
right stage dir.
Also `include-ghc-prim` is a flag that no longer exists for `ghc-prim`
(it was removed in 835d8ddbbfb11796ea8a03d1806b7cee38ba17a6) so I got
rid of it.
Co-Authored-By: Matthew Pickering <matthewtpickering at gmail.com>
- - - - -
a0ac8785 by Sebastian Graf at 2023-10-14T19:17:12-04:00
Fix restarts in .ghcid
Using the whole of `hadrian/` restarted in a loop for me.
- - - - -
fea9ecdb by Sebastian Graf at 2023-10-14T19:17:12-04:00
CorePrep: Refactor FloatingBind (#23442)
A drastically improved architecture for local floating in CorePrep
that decouples the decision of whether a float is going to be let- or case-bound
from how far it can float (out of strict contexts, out of lazy contexts, to
top-level).
There are a couple of new Notes describing the effort:
* `Note [Floating in CorePrep]` for the overview
* `Note [BindInfo and FloatInfo]` for the new classification of floats
* `Note [Floats and FloatDecision]` for how FloatInfo is used to inform
floating decisions
This is necessary ground work for proper treatment of Strict fields and
unlifted values at top-level.
Fixes #23442.
NoFib results (omitted = 0.0%):
```
--------------------------------------------------------------------------------
Program Allocs Instrs
--------------------------------------------------------------------------------
pretty 0.0% -1.6%
scc 0.0% -1.7%
--------------------------------------------------------------------------------
Min 0.0% -1.7%
Max 0.0% -0.0%
Geometric Mean -0.0% -0.0%
```
- - - - -
32523713 by Matthew Pickering at 2023-10-14T19:17:49-04:00
hadrian: Move ghcBinDeps into ghcLibDeps
This completes a5227080b57cb51ac34d4c9de1accdf6360b818b, the
`ghc-usage.txt` and `ghci-usage.txt` file are also used by the `ghc`
library so need to make sure they are present in the libdir even if we
are not going to build `ghc-bin`.
This also fixes things for cross compilers because the stage2
cross-compiler requires the ghc-usage.txt file, but we are using
the stage2 lib folder but not building stage3:exe:ghc-bin so
ghc-usage.txt was not being generated.
- - - - -
ec3c4488 by sheaf at 2023-10-14T19:18:29-04:00
Combine GREs when combining in mkImportOccEnv
In `GHC.Rename.Names.mkImportOccEnv`, we sometimes discard one import
item in favour of another, as explained in Note [Dealing with imports]
in `GHC.Rename.Names`. However, this can cause us to lose track of
important parent information.
Consider for example #24084:
module M1 where { class C a where { type T a } }
module M2 ( module M1 ) where { import M1 }
module M3 where { import M2 ( C, T ); instance C () where T () = () }
When processing the import list of `M3`, we start off (for reasons that
are not relevant right now) with two `Avail`s attached to `T`, namely
`C(C, T)` and `T(T)`. We combine them in the `combine` function of
`mkImportOccEnv`; as described in Note [Dealing with imports] we discard
`C(C, T)` in favour of `T(T)`. However, in doing so, we **must not**
discard the information want that `C` is the parent of `T`. Indeed,
losing track of this information can cause errors when importing,
as we could get an error of the form
‘T’ is not a (visible) associated type of class ‘C’
We fix this by combining the two GREs for `T` using `plusGRE`.
Fixes #24084
- - - - -
257c2807 by Ilias Tsitsimpis at 2023-10-14T19:19:07-04:00
hadrian: Pass -DNOSMP to C compiler when needed
Hadrian passes the -DNOSMP flag to GHC when the target doesn't support
SMP, but doesn't pass it to CC as well, leading to the following
compilation error on mips64el:
| Run Cc (FindCDependencies CDep) Stage1: rts/sm/NonMovingScav.c => _build/stage1/rts/build/c/sm/NonMovingScav.o.d
Command line: /usr/bin/mips64el-linux-gnuabi64-gcc -E -MM -MG -MF _build/stage1/rts/build/c/hooks/FlagDefaults.thr_debug_p_o.d -MT _build/stage1/rts/build/c/hooks/FlagDefaults.o -Irts/include -I_build/stage1/rts/build -I_build/stage1/rts/build/include -Irts/include -x c rts/hooks/FlagDefaults.c -Wall -Wextra -Wstrict-prototypes -Wmissing-prototypes -Wmissing-declarations -Winline -Wpointer-arith -Wmissing-noreturn -Wnested-externs -Wredundant-decls -Wundef -fno-strict-aliasing -DTHREADED_RTS -DDEBUG -fomit-frame-pointer -O2 -g -Irts -I_build/stage1/rts/build -DDEBUG -fno-omit-frame-pointer -g3 -O0
===> Command failed with error code: 1
In file included from rts/include/Stg.h:348,
from rts/include/Rts.h:38,
from rts/hooks/FlagDefaults.c:8:
rts/include/stg/SMP.h:416:2: error: #error memory barriers unimplemented on this architecture
416 | #error memory barriers unimplemented on this architecture
| ^~~~~
rts/include/stg/SMP.h:440:2: error: #error memory barriers unimplemented on this architecture
440 | #error memory barriers unimplemented on this architecture
| ^~~~~
rts/include/stg/SMP.h:464:2: error: #error memory barriers unimplemented on this architecture
464 | #error memory barriers unimplemented on this architecture
| ^~~~~
The old make system correctly passed this flag to both GHC and CC [1].
Fix this error by passing -DNOSMP to CC as well.
[1] https://gitlab.haskell.org/ghc/ghc/-/blob/00920f176b0235d5bb52a8e054d89a664f8938fe/rts/ghc.mk#L407
Closes #24082
- - - - -
13d3c613 by John Ericson at 2023-10-14T19:19:42-04:00
Users Guide: Drop dead code for Haddock refs to `parallel`
I noticed while working on !11451 that `@LIBRARY_parallel_UNIT_ID@` was
not substituted. It is dead code -- there is no `parallel-ref` usages
and it doesn't look like there ever was (going back to
3e5d0f188d6c8633e55e9ba6c8941c07e459fa4b), so let's delete it.
- - - - -
fe067577 by Sylvain Henry at 2023-10-18T19:40:25-04:00
Avoid out-of-bound array access in bigNatIsPowerOf2 (fix #24066)
bigNatIndex# in the `where` clause wasn't guarded by "bigNatIsZero a".
- - - - -
cc1625b1 by Sylvain Henry at 2023-10-18T19:40:25-04:00
Bignum: fix right shift of negative BigNat with native backend
- - - - -
cbe4400d by Sylvain Henry at 2023-10-18T19:40:25-04:00
Rts: expose rtsOutOfBoundsAccess symbol
- - - - -
72c7380c by Sylvain Henry at 2023-10-18T19:40:25-04:00
Hadrian: enable `-fcheck-prim-bounds` in validate flavour
This allows T24066 to fail when the bug is present.
Otherwise the out-of-bound access isn't detected as it happens in
ghc-bignum which wasn't compiled with the bounds check.
- - - - -
f9436990 by John Ericson at 2023-10-18T19:41:01-04:00
Make Hadrian solely responsible for substituting `docs/users_guide/ghc_config.py.in`
Fixes #24091
Progress on #23966
Issue #24091 reports that `@ProjectVersion@` is no longer being
substituted in the GHC user's guide. I assume this is a recent issue,
but I am not sure how it's worked since
c1a3ecde720b3bddc2c8616daaa06ee324e602ab; it looks like both Hadrian and
configure are trying to substitute the same `.in` file!
Now only Hadrian does. That is better anyways; already something that
issue #23966 requested.
It seems like we were missing some dependencies in Hadrian. (I really,
really hate that this is possible!) Hopefully it is fixed now.
- - - - -
b12df0bb by John Ericson at 2023-10-18T19:41:37-04:00
`ghcversion.h`: No need to cope with undefined `ProjectPatchLevel*`
Since 4e6c80197f1cc46dfdef0300de46847c7cfbdcb0, these are guaranteed to
be defined. (Guaranteed including a test in the testsuite.)
- - - - -
0295375a by John Ericson at 2023-10-18T19:41:37-04:00
Generate `ghcversion.h` from a `.in` file
Now that there are no conditional sections (see the previous commit), we
can just a do simple substitution rather than pasting it together line
by line.
Progress on #23966
- - - - -
740a1b85 by Krzysztof Gogolewski at 2023-10-19T11:37:20-04:00
Add a regression test for #24064
- - - - -
921fbf2f by Hécate Moonlight at 2023-10-19T11:37:59-04:00
CLC Proposal #182: Export List from Data.List
Proposal link: https://github.com/haskell/core-libraries-committee/issues/182
- - - - -
4f02d3c1 by Sylvain Henry at 2023-10-20T04:01:32-04:00
rts: fix small argument passing on big-endian arch (fix #23387)
- - - - -
b86243b4 by Sylvain Henry at 2023-10-20T04:02:13-04:00
Interpreter: fix literal alignment on big-endian architectures (fix #19261)
Literals weren't correctly aligned on big-endian, despite what the
comment said.
- - - - -
a4b2ec47 by Sylvain Henry at 2023-10-20T04:02:54-04:00
Testsuite: recomp011 and recomp015 are fixed on powerpc
These tests have been fixed but not tested and re-enabled on big-endian
powerpc (see comments in #11260 and #11323)
- - - - -
fded7dd4 by Sebastian Graf at 2023-10-20T04:03:30-04:00
CorePrep: Allow floating dictionary applications in -O0 into a Rec (#24102)
- - - - -
d1b0cc6f by Simon Peyton Jones at 2023-10-20T16:01:11+01:00
Several improvements to the handling of coercions
* Make `mkSymCo` and `mkInstCo` smarter
Fixes #23642
* Fix return role of `SelCo` in the coercion optimiser.
Fixes #23617
* Make the coercion optimiser `opt_trans_rule` work better for newtypes
Fixes #23619
- - - - -
62bc9035 by Simon Peyton Jones at 2023-10-20T16:01:11+01:00
Simplifier improvements
This MR started as: allow the simplifer to do more in one pass,
arising from places I could see the simplifier taking two iterations
where one would do. But it turned into a larger project, because
these changes unexpectedly made inlining blow up, especially join
points in deeply-nested cases.
The net result is good: a 2% improvement in compile time. The table
below shows changes over 1%.
The main changes are:
* The SimplEnv now has a seInlineDepth field, which says how deep
in unfoldings we are. See Note [Inline depth] in Simplify.Env
* Avoid repeatedly simplifying coercions.
see Note [Avoid re-simplifying coercions] in Simplify.Iteration
As you'll see from the Note, this makes use of the seInlineDepth.
* Allow Simplify.Utils.postInlineUnconditionally to inline variables
that are used exactly once. See Note [Post-inline for single-use things].
* Allow Simplify.Iteration.simplAuxBind to inline used-once things.
This is another part of Note [Post-inline for single-use things], and
is really good for reducing simplifier iterations in situations like
case K e of { K x -> blah }
wher x is used once in blah.
* Make GHC.Core.SimpleOpt.exprIsConApp_maybe do some simple case
elimination. Note [Case elim in exprIsConApp_maybe]
* When making join points, don't do so if the join point is so small
it will immediately be inlined. See Note [Duplicating alternatives]
* Do not add an unfolding to a join point at birth. This is a tricky one
and has a long Note [Do not add unfoldings to join points at birth]
It shows up in two places
- In `mkDupableAlt` do not add an inlining
- (trickier) In `simplLetUnfolding` don't add an unfolding for a
fresh join point
I am not fully satisifed with this, but it works and is well documented.
* Many new or rewritten Notes. E.g. Note [Avoiding simplifying repeatedly]
I discovered that GHC.HsToCore.Pmc.Solver.Types.trvVarInfo was very
delicately balanced. It's a small, heavily used, overloaded function
and it's important that it inlines. By a fluke it was before, but at
various times in my journey it stopped doing so. So I added an INLINE
pragma to it.
Metrics: compile_time/bytes allocated
------------------------------------------------
CoOpt_Singletons(normal) -4.3% GOOD
LargeRecord(normal) -23.3% GOOD
PmSeriesS(normal) -2.4%
T11195(normal) -1.7%
T12227(normal) -20.0% GOOD
T12545(normal) -5.4%
T13253-spj(normal) -50.7% GOOD
T13386(normal) -5.1% GOOD
T14766(normal) -2.4% GOOD
T15164(normal) -1.7%
T15304(normal) +1.0%
T15630(normal) -7.7%
T15630a(normal) NEW
T15703(normal) -7.5% GOOD
T16577(normal) -5.1% GOOD
T17516(normal) -3.6%
T18223(normal) -16.8% GOOD
T18282(normal) -1.5%
T18304(normal) +1.9%
T21839c(normal) -3.5% GOOD
T3064(normal) -1.5%
T5030(normal) -16.2% GOOD
T5321Fun(normal) -1.6%
T6048(optasm) -2.1% GOOD
T8095(normal) -6.1% GOOD
T9630(normal) -5.1% GOOD
WWRec(normal) -1.6%
geo. mean -2.1%
minimum -50.7%
maximum +1.9%
Metric Decrease:
CoOpt_Singletons
LargeRecord
T12227
T13253-spj
T13386
T14766
T15703
T16577
T18223
T21839c
T5030
T6048
T8095
T9630
- - - - -
2e5b78c6 by Simon Peyton Jones at 2023-10-20T16:01:11+01:00
Improve postInlineUnconditionally
This commit adds two things to postInlineUnconditionally:
1. Do not postInlineUnconditionally join point, ever.
Doing so does not reduce allocation, which is the main point,
and with join points that are used a lot it can bloat code.
See point (1) of Note [Duplicating join points] in
GHC.Core.Opt.Simplify.Iteration.
2. Do not postInlineUnconditionally a strict (demanded) binding.
It will not allocate a thunk (it'll turn into a case instead)
so again the main point of inlining it doesn't hold. Better
to check per-call-site.
- - - - -
42950140 by Simon Peyton Jones at 2023-10-20T16:01:11+01:00
Update testsuite output
- - - - -
c237da54 by Simon Peyton Jones at 2023-10-20T16:01:11+01:00
Try effect of
* making multi-branch cases not work free (fixes #22423)
* use plan A for dataToTag and tagToEnum
- - - - -
83306277 by Simon Peyton Jones at 2023-10-20T16:01:11+01:00
More changes
* No floating at all for join points
* Never inline j x = I x
Example: integerSignum !j = IS (integerSignum# j)
We want this to inline and then cancel with an enclosing case.
But it won't if we have changed it to
integerSignum x = case x of
IN a -> IS (...)
IS b -> IS (...)
IP c -> IS (...)
This involved changing
- UnfoldingGuidance to not say always-inline for j x = Ix
- callSiteInline to inline join points only if there is a real
benefit
- ok_to_dup_alt in Simplify.Iteration
* Row back (for now) on changes to GHC.Core.Utils.ExprIsCheap
- - - - -
fe109d1f by Simon Peyton Jones at 2023-10-20T16:01:11+01:00
Wibble
- - - - -
5cfc57c5 by Simon Peyton Jones at 2023-10-20T16:01:11+01:00
Wibble
- - - - -
9e512441 by Simon Peyton Jones at 2023-10-20T16:01:11+01:00
Further wibbles
- - - - -
6bcff167 by Simon Peyton Jones at 2023-10-20T16:01:11+01:00
One more wibble
Don't float an unlifted join point
- - - - -
c0e517c8 by Simon Peyton Jones at 2023-10-20T16:01:12+01:00
Small wibbles
The most significant change is to mkSelCo.
- - - - -
efc98bde by Simon Peyton Jones at 2023-10-20T16:01:12+01:00
Temporarily add -DDEBUG to Coercion.hs
- - - - -
49960854 by Simon Peyton Jones at 2023-10-20T16:01:12+01:00
Fix a missing zonk in before mkSelCo
- - - - -
cf6e970f by Simon Peyton Jones at 2023-10-20T16:01:12+01:00
Wibbles
- - - - -
c559190d by Simon Peyton Jones at 2023-10-20T16:01:12+01:00
Tickish comment
- - - - -
26 changed files:
- .ghcid
- .gitignore
- .gitlab/generate-ci/gen_ci.hs
- .gitlab/jobs.yaml
- .gitlab/rel_eng/upload.sh
- compiler/GHC/ByteCode/Asm.hs
- compiler/GHC/CmmToAsm/AArch64/CodeGen.hs
- compiler/GHC/CmmToAsm/AArch64/Instr.hs
- compiler/GHC/CmmToAsm/AArch64/Ppr.hs
- compiler/GHC/CmmToAsm/AArch64/Regs.hs
- compiler/GHC/CmmToAsm/Ppr.hs
- compiler/GHC/Core/Coercion.hs
- compiler/GHC/Core/Coercion/Opt.hs
- compiler/GHC/Core/Opt/Arity.hs
- compiler/GHC/Core/Opt/OccurAnal.hs
- compiler/GHC/Core/Opt/SetLevels.hs
- compiler/GHC/Core/Opt/Simplify/Env.hs
- compiler/GHC/Core/Opt/Simplify/Inline.hs
- compiler/GHC/Core/Opt/Simplify/Iteration.hs
- compiler/GHC/Core/Opt/Simplify/Monad.hs
- compiler/GHC/Core/Opt/Simplify/Utils.hs
- compiler/GHC/Core/Opt/Stats.hs
- compiler/GHC/Core/SimpleOpt.hs
- compiler/GHC/Core/TyCo/Rep.hs
- compiler/GHC/Core/Unfold.hs
- compiler/GHC/Core/Unfold/Make.hs
The diff was not included because it is too large.
View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/compare/c9bcb2f7a195805443c1a5493b2e5698c1acfe44...c559190dcf35edd3379e8b00b0e8bb913db2f339
--
View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/compare/c9bcb2f7a195805443c1a5493b2e5698c1acfe44...c559190dcf35edd3379e8b00b0e8bb913db2f339
You're receiving this email because of your account on gitlab.haskell.org.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.haskell.org/pipermail/ghc-commits/attachments/20231020/abcb01ad/attachment-0001.html>
More information about the ghc-commits
mailing list