Rodrigo Mesquita pushed to branch wip/romes/25636 at Glasgow Haskell Compiler / GHC
Commits:
55eab80d by Sylvain Henry at 2025-11-20T17:33:13-05:00
Build external interpreter program on demand (#24731)
This patch teaches GHC how to build the external interpreter program
when it is missing. As long as we have the `ghci` library, doing this is
trivial so most of this patch is refactoring for doing it sanely.
- - - - -
08bbc028 by Rodrigo Mesquita at 2025-11-20T17:33:54-05:00
Add tests for #23973 and #26565
These were fixed by 4af4f0f070f83f948e49ad5d7835fd91b8d3f0e6 in !10417
- - - - -
6b42232c by sheaf at 2025-11-20T17:34:35-05:00
Mark T26410_ffi as fragile on Windows
As seen in #26595, this test intermittently fails on Windows.
This commit marks it as fragile, until we get around to fixing it.
- - - - -
b7b7c049 by Andrew Lelechenko at 2025-11-21T21:04:01+00:00
Add nubOrd / nubOrdBy to Data.List and Data.List.NonEmpty
As per https://github.com/haskell/core-libraries-committee/issues/336
- - - - -
352d5462 by Marc Scholten at 2025-11-22T10:33:03-05:00
Fix haddock test runner to handle UTF-8 output
xhtml 3000.4.0.0 now produces UTF-8 output instead of escaping non-ASCII characters.
When using --test-accept it previously wrote files in the wrong encoding
because they have not been decoded properly when reading the files.
- - - - -
48a3ed57 by Simon Peyton Jones at 2025-11-25T15:33:54+00:00
Add a fast-path for args=[] to occAnalApp
In the common case of having not arguments, occAnalApp
was doing redundant work.
- - - - -
951e5ed9 by Simon Peyton Jones at 2025-11-25T15:33:54+00:00
Fix a performance hole in the occurrence analyser
As #26425 showed, the clever stuff in
Note [Occurrence analysis for join points]
does a lot of duplication of usage details. This patch
improved matters with a little fancy footwork. It is
described in the new (W4) of the same Note.
Compile-time allocations go down slightly. Here are the changes
of +/- 0.5% or more:
T13253(normal) 329,369,244 326,395,544 -0.9%
T13253-spj(normal) 66,410,496 66,095,864 -0.5%
T15630(normal) 129,797,200 128,663,136 -0.9%
T15630a(normal) 129,212,408 128,027,560 -0.9%
T16577(normal) 6,756,706,896 6,723,028,512 -0.5%
T18282(normal) 128,462,070 125,808,584 -2.1% GOOD
T18698a(normal) 208,418,305 202,037,336 -3.1% GOOD
T18730(optasm) 136,981,756 136,208,136 -0.6%
T18923(normal) 58,103,088 57,745,840 -0.6%
T19695(normal) 1,386,306,272 1,365,609,416 -1.5%
T26425(normal) 3,344,402,957 2,457,811,664 -26.5% GOOD
T6048(optasm) 79,763,816 79,212,760 -0.7%
T9020(optasm) 225,278,408 223,682,440 -0.7%
T9961(normal) 303,810,717 300,729,168 -1.0% GOOD
geo. mean -0.5%
minimum -26.5%
maximum +0.4%
Metric Decrease:
T18282
T18698a
T26425
T9961
- - - - -
f1959dfc by Simon Peyton Jones at 2025-11-26T11:58:07+00:00
Remove a quadratic-cost assertion check in mkCoreApp
See the new Note [Assertion checking in mkCoreApp]
- - - - -
98fa0d36 by Simon Hengel at 2025-11-27T17:54:57-05:00
Fix typo in docs/users_guide/exts/type_families.rst
- - - - -
5b97e5ce by Simon Hengel at 2025-11-27T17:55:37-05:00
Fix broken RankNTypes example in user's guide
- - - - -
fa2aaa00 by Simon Peyton Jones at 2025-11-27T17:56:18-05:00
Switch off specialisation in ExactPrint
In !15057 (where we re-introduced -fpolymoprhic-specialisation) we found
that ExactPrint's compile time blew up by a factor of 5. It turned out
to be caused by bazillions of specialisations of `markAnnotated`.
Since ExactPrint isn't perf-critical, it does not seem worth taking
the performance hit, so this patch switches off specialisation in
this one module.
- - - - -
1fd25987 by Simon Peyton Jones at 2025-11-27T17:56:18-05:00
Switch -fpolymorphic-specialisation on by default
This patch addresses #23559.
Now that !10479 has landed and #26329 is fixed, we can switch on
polymorphic specialisation by default, addressing a bunch of other
tickets listed in #23559.
Metric changes:
* CoOpt_Singleton: +4% compiler allocations: we just get more
specialisations
* info_table_map_perf: -20% decrease in compiler allocations.
This is caused by using -fno-specialise in ExactPrint.hs
Without that change we get a 4x blow-up in compile time;
see !15058 for details
Metric Decrease:
info_table_map_perf
Metric Increase:
CoOpt_Singletons
- - - - -
b7fe7445 by Matthew Pickering at 2025-11-27T17:56:59-05:00
rts: Fix a deadlock with eventlog flush interval and RTS shutdown
The ghc_ticker thread attempts to flush at the eventlog tick interval, this requires
waiting to take all capabilities.
At the same time, the main thread is shutting down, the schedule is
stopped and then we wait for the ticker thread to finish.
Therefore we are deadlocked.
The solution is to use `newBoundTask/exitMyTask`, so that flushing can
cooperate with the scheduler shutdown.
Fixes #26573
- - - - -
1d4a1229 by sheaf at 2025-11-27T17:58:02-05:00
SimpleOpt: don't subst in pushCoercionIntoLambda
It was noticed in #26589 that the change in 15b311be was incorrect:
the simple optimiser carries two different substitution-like pieces of
information: 'soe_subst' (from InVar to OutExpr) and 'soe_inl'
(from InId to InExpr). It is thus incorrect to have 'pushCoercionIntoLambda'
apply the substitution from 'soe_subst' while discarding 'soe_inl'
entirely, which is what was done in 15b311be.
Instead, we change back pushCoercionIntoLambda to take an InScopeSet,
and optimise the lambda before calling 'pushCoercionIntoLambda' to avoid
mixing InExpr with OutExpr, or mixing two InExpr with different
environments. We can then call 'soeZapSubst' without problems.
Fixes #26588 #26589
- - - - -
84a087d5 by Sylvain Henry at 2025-11-28T17:35:28-05:00
Fix PIC jump tables on Windows (#24016)
Avoid overflows in jump tables by using a base label closer to the jump
targets. See added Note [Jump tables]
- - - - -
82db7042 by Zubin Duggal at 2025-11-28T17:36:10-05:00
rts/linker/PEi386: Copy strings before they are inserted into LoadedDllCache. The original strings are temporary and might be freed at an arbitrary point.
Fixes #26613
- - - - -
ff3f0d09 by Ben Gamari at 2025-11-29T18:34:28-05:00
gitlab-ci: Run ghcup-metadata jobs on OpenCape runners
This significantly reduces our egress traffic
and makes the jobs significantly faster.
- - - - -
ef0dc33b by Matthew Pickering at 2025-11-29T18:35:10-05:00
Use 'OsPath' in getModificationTimeIfExists
This part of the compiler is quite hot during recompilation checking in
particular since the filepaths will be translated to a string. It is
better to use the 'OsPath' native function, which turns out to be easy
to do.
- - - - -
fa3bd0a6 by Georgios Karachalias at 2025-11-29T18:36:05-05:00
Use OsPath in PkgDbRef and UnitDatabase, not FilePath
- - - - -
0d7c05ec by Ben Gamari at 2025-12-01T03:13:46-05:00
hadrian: Place user options after package arguments
This makes it easier for the user to override the default package
arguments with `UserSettings.hs`.
Fixes #25821.
-------------------------
Metric Decrease:
T14697
-------------------------
- - - - -
3b2c4598 by Vladislav Zavialov at 2025-12-01T03:14:29-05:00
Namespace-specified wildcards in import/export lists (#25901)
This change adds support for top-level namespace-specified wildcards
`type ..` and `data ..` to import and export lists.
Examples:
import M (type ..) -- imports all type and class constructors from M
import M (data ..) -- imports all data constructors and terms from M
module M (type .., f) where
-- exports all type and class constructors defined in M,
-- plus the function 'f'
The primary intended usage of this feature is in combination with module
aliases, allowing namespace disambiguation:
import Data.Proxy as T (type ..) -- T.Proxy is unambiguously the type constructor
import Data.Proxy as D (data ..) -- D.Proxy is unambiguously the data constructor
The patch accounts for the interactions of wildcards with:
* Imports with `hiding` clauses
* Import warnings -Wunused-imports, -Wdodgy-imports
* Export warnings -Wduplicate-exports, -Wdodgy-exports
Summary of the changes:
1. Move the NamespaceSpecifier type from GHC.Hs.Binds to GHC.Hs.Basic,
making it possible to use it in more places in the AST.
2. Extend the AST (type: IE) with a representation of `..`, `type ..`,
and `data ..` (constructor: IEWholeNamespace). Per the proposal, the
plain `..` is always rejected with a dedicated error message.
3. Extend the grammar in Parser.y with productions for `..`, `type ..`,
and `data ..` in both import and export lists.
4. Implement wildcard imports by updating the `filterImports` function
in GHC.Rename.Names; the logic for IEWholeNamespace is roughly
modeled after the Nothing (no explicit import list) case.
5. Implement wildcard exports by updating the `exports_from_avail`
function in GHC.Tc.Gen.Export; the logic for IEWholeNamespace is
closely modeled after the IEModuleContents case.
6. Refactor and extend diagnostics to report the new warnings and
errors. See PsErrPlainWildcardImport, DodgyImportsWildcard,
PsErrPlainWildcardExport, DodgyExportsWildcard,
TcRnDupeWildcardExport.
Note that this patch is specifically about top-level import/export
items. Subordinate import/export items are left unchanged.
- - - - -
c71faa76 by Luite Stegeman at 2025-12-01T03:16:05-05:00
rts: Handle overflow of ELF section header string table
If the section header string table is stored in a section greater
than or equal to SHN_LORESERVE (0xff00), the 16-bit field e_shstrndx
in the ELF header does not contain the section number, but rather
an overflow value SHN_XINDEX (0xffff) indicating that we need to look
elsewhere.
This fixes the linker by not using e_shstrndx directly but calling
elf_shstrndx, which correctly handles the SHN_XINDEX value.
Fixes #26603
- - - - -
ab20eb54 by Mike Pilgrem at 2025-12-01T22:46:55+00:00
Re CLC issue 292 Warn GHC.Internal.List.{init,last} are partial
Also corrects the warning for `tail` to refer to `Data.List.uncons` (like the existing warning for `head`).
In module `Settings.Warnings`, applies `-Wno-x-partial` to the `filepath`, and `parsec` packages (outside GHC's repository).
Also bumps submodules.
- - - - -
fc1d7f79 by Jade Lovelace at 2025-12-02T11:04:09-05:00
docs: fix StandaloneKindSignatures in DataKinds docs
These should be `type` as otherwise GHC reports a duplicate definition
error.
- - - - -
beae879b by Rodrigo Mesquita at 2025-12-03T15:42:37+01:00
task: Substitute some datatypes for newtypes
* Substitutes some data type declarations for newtype declarations
* Adds comment to `LlvmConfigCache`, which must decidedly not be a
newtype.
Fixes #23555
- - - - -
3bd7dd44 by mangoiv at 2025-12-04T04:36:45-05:00
Renamer: reinstate the template haskell level check in notFound
Out-of-scope names might be caused by a staging error, as is explained by
Note [Out of scope might be a staging error] in GHC.Tc.Utils.Env.hs.
This logic was assumed to be dead code after 217caad1 and has thus been
removed. This commit reintroduces it and thus fixes issue #26099.
- - - - -
0318010b by Zubin Duggal at 2025-12-04T04:37:27-05:00
testlib: Optionally include the way name in the expected output file
This allows us to have different outputs for different ways.
- - - - -
6d945fdd by Zubin Duggal at 2025-12-04T04:37:27-05:00
testsuite: Accept output of tests failing in ext-interp way due to differing compilation requirements
Fixes #26552
- - - - -
0ffc5243 by Cheng Shao at 2025-12-04T04:38:09-05:00
devx: minor fixes for compile_flags.txt
This patch includes minor fixes for compile_flags.txt to improve
developer experience when using clangd as language server to hack on
RTS C sources:
- Ensure `-fPIC` is passed and `__PIC__` is defined, to be coherent
with `-DDYNAMIC` and ensure the `__PIC__` guarded code paths are
indexed
- Add the missing `-DRtsWay` definition, otherwise a few source files
like `RtsUtils.c` and `Trace.c` would produce clangd errors
- - - - -
e36a5fcb by Matthew Pickering at 2025-12-05T16:25:57-05:00
Add support for building bytecode libraries
A bytecode library is a collection of bytecode files (.gbc) and a
library which combines together additional object files.
A bytecode library is created by invoking GHC with the `-bytecodelib`
flag.
A library can be created from in-memory `ModuleByteCode` linkables or
by passing `.gbc` files as arguments on the command line.
Fixes #26298
- - - - -
8f9ae339 by Matthew Pickering at 2025-12-05T16:25:57-05:00
Load bytecode libraries to satisfy package dependencies
This commit allows you to use a bytecode library to satisfy a package
dependency when using the interpreter.
If a user enables `-fprefer-byte-code`, then if a package provides a
bytecode library, that will be loaded and used to satisfy the
dependency.
The main change is to separate the relevant parts of the `LoaderState`
into external and home package byte code. Bytecode is loaded into either
the home package or external part (similar to HPT/EPS split), HPT
bytecode can be unloaded. External bytecode is never unloaded.
The unload function has also only been called with an empty list of
"stable linkables" for a long time. It has been modified to directly
implement a complete unloading of the home package bytecode linkables.
At the moment, the bytecode libraries are found in the "library-dirs"
field from the package description. In the future when `Cabal`
implements support for "bytecode-library-dirs" field, we can read the
bytecode libraries from there. No changes to the Cabal submodule are
necessary at the moment.
Four new tests are added in testsuite/tests/cabal, which generate fake
package descriptions and test loading the libraries into GHCi.
Fixes #26298
- - - - -
54458ce4 by mangoiv at 2025-12-05T16:26:50-05:00
ExplicitLevelImports: improve documentation of the code
- more explicit names for variable names like `flg` or `topLevel`
- don't pass the same value twice to functions
- some explanations of interesting but undocumented code paths
- adjust comment to not mention non-existent error message
- - - - -
c7061392 by mangoiv at 2025-12-05T16:27:42-05:00
driver: don't expect nodes to exist when checking paths between them
In `mgQueryZero`, previously node lookups were expected to never fail,
i.e. it was expected that when calculating the path between two nodes in
a zero level import graph, both nodes would always exist. This is not
the case, e.g. in some situations involving exact names (see the
test-case). The fix is to first check whether the node is present in the
graph at all, instead of panicking, just to report that there is no
path.
Closes #26568
- - - - -
d6cf8463 by Peng Fan at 2025-12-06T11:06:28-05:00
NCG/LA64: Simplify genCCall into two parts
genCCall is too long, so it's been simplified into two parts:
genPrim and genLibCCall.
Suggested by Andreas Klebinger
- - - - -
9d371d23 by Matthew Pickering at 2025-12-06T11:07:09-05:00
hadrian: Use a response file to invoke GHC for dep gathering.
In some cases we construct an argument list too long for GHC to
handle directly on windows. This happens when we generate
the dependency file because the command line will contain
references to a large number of .hs files.
To avoid this we now invoke GHC using a response file when
generating dependencies to sidestep length limitations.
Note that we only pass the actual file names in the dependency
file. Why? Because this side-steps #26560
- - - - -
0043bfb0 by Marc Scholten at 2025-12-06T11:08:03-05:00
update xhtml to 3000.4.0.0
haddock-api: bump xhtml bounds
haddock-api: use lazy text instead of string to support xhtml 3000.4.0.0
Bumping submodule xhtml to 3000.4.0.0
add xhtml to stage0Packages
remove unused import of writeUtf8File
Remove redundant import
Update haddock golden files for xhtml 3000.4.0.0
Metric Decrease:
haddock.Cabal
haddock.base
- - - - -
fc958fc9 by Julian Ospald at 2025-12-06T11:08:53-05:00
rts: Fix object file format detection in loadArchive
Commit 76d1041dfa4b96108cfdd22b07f2b3feb424dcbe seems to
have introduced this bug, ultimately leading to failure of
test T11788. I can only theorize that this test isn't run
in upstream's CI, because they don't build a static GHC.
The culprit is that we go through the thin archive, trying
to follow the members on the filesystem, but don't
re-identify the new object format of the member. This pins
`object_fmt` to `NotObject` from the thin archive.
Thanks to @angerman for spotting this.
- - - - -
0f297f6e by mangoiv at 2025-12-06T11:09:44-05:00
users' guide: don't use f strings in the python script to ensure compatibility with python 3.5
- - - - -
3bfe7aa2 by Matthew Pickering at 2025-12-07T12:18:57-05:00
ci: Try using multi repl in ghc-in-ghci test
This should be quite a bit faster than the ./hadrian/ghci command as it
doesn't properly build all the dependencies.
- - - - -
2ef1601a by Rodrigo Mesquita at 2025-12-07T12:19:38-05:00
Stack.Decode: Don't error on bitmap size 0
A RET_BCO may have a bitmap with no payload.
In that case, the bitmap = 0.
One can observe this by using -ddump-bcos and interpreting
```
main = pure ()
```
Observe, for instance, that the BCO for this main function has size 0:
```
ProtoBCO Main.main#0:
\u []
breakmain:Main,0() GHC.Internal.Base.pure
GHC.Internal.Base.$fApplicativeIO GHC.Internal.Tuple.()
bitmap: 0 []
BRK_FUN <breakarray> main:Main 0 <cc>
PACK () 0
PUSH_G GHC.Internal.Base.$fApplicativeIO
PUSH_APPLY_PP
PUSH_G GHC.Internal.Base.pure
ENTER
```
Perhaps we never tried to decode a stack in which a BCO like this was
present. However, for the debugger, we want to decode stacks of threads
stopped at breakpoints, and these kind of BCOs do get on a stack under
e.g. `stg_apply_interp_info` frames.
See the accompanying test in the next commit for an example to trigger
the bug this commit fixes.
Fixes #26640
- - - - -
747153d2 by Rodrigo Mesquita at 2025-12-07T12:19:38-05:00
Add test for #26640
- - - - -
d4b1e353 by Simon Hengel at 2025-12-10T00:00:02-05:00
Fix syntax error in gadt_syntax.rst
- - - - -
91cc8be6 by Cheng Shao at 2025-12-10T00:00:43-05:00
ci: fix "ci.sh clean" to address frequent out of space error on windows runners
This patch fixes the `ci.sh clean` logic to address frequent out of
space error on windows runners; previously it didn't clean up the
inplace mingw blobs, which is the largest source of space leak on
windows runners. See added comment for detailed explanation.
- - - - -
fe2b79f4 by Recursion Ninja at 2025-12-10T08:34:18-05:00
Narrow before optimising MUL/DIV/REM into shifts
The MUL/DIV/REM operations can be optimised into shifts when one of the
operands is a constant power of 2. However, as literals in Cmm are
stored as 'Integer', for this to be correct we first need to narrow the
literal to the appropriate width before checking whether the literal is
a power of 2.
Fixes #25664
- - - - -
06c2349c by Recursion Ninja at 2025-12-10T08:34:58-05:00
Decouple 'Language.Haskell.Syntax.Type' from 'GHC.Utils.Panic'
- Remove the *original* defintion of 'hsQTvExplicit' defined within 'Language.Haskell.Syntax.Type'
- Redefine 'hsQTvExplicit' as 'hsq_explicit' specialized to 'GhcPass' exported by 'GHC.Utils.Panic'
- Define 'hsQTvExplicitBinders' as 'hsq_explicit' specialized to 'DocNameI' exported by 'Haddock.GhcUtils'.
- Replace all call sites of the original 'hsQTvExplicit' definition with either:
1. 'hsQTvExplicit' updated definition
2. 'hsQTvExplicitBinders'
All call sites never entered the 'XLHsQTyVars' constructor branch, but a call to 'panic' existed on this code path because the type system was not strong enought to guarantee that the 'XLHsQTyVars' construction was impossible.
These two specialized functions provide the type system with enough information to make that guarantee, and hence the dependancy on 'panic' can be removed.
- - - - -
ac0815d5 by sheaf at 2025-12-10T23:39:57-05:00
Quantify arg before mult in function arrows
As noted in #23764, we expect quantification order to be left-to-right,
so that in a type such as
a %m -> b
the inferred quantification order should be [a, m, b] and not [m, a, b].
This was addressed in commit d31fbf6c, but that commit failed to update
some other functions such as GHC.Core.TyCo.FVs.tyCoFVsOfType.
This affects Haddock, as whether we print an explicit forall or not
depends on whether the inferred quantification order matches the actual
quantification order.
- - - - -
2caf796e by sheaf at 2025-12-10T23:39:57-05:00
Haddock: improvements to ty-var quantification
This commit makes several improvements to how Haddock deals with the
quantification of type variables:
1. In pattern synonyms, Haddock used to jumble up universal and
existential quantification. That is now fixed, fixing #26252.
Tested in the 'PatternSyns2' haddock-html test.
2. The logic for computing whether to use an explicit kind annotation
for a type variable quantified in a forall was not even wrong.
This commit improves the heuristic, but it will always remain an
imperfect heuristic (lest we actually run kind inference again).
In the future (#26271), we hope to avoid reliance on this heuristic.
- - - - -
b14bdd59 by Teo Camarasu at 2025-12-10T23:40:38-05:00
Add explicit export list to GHC.Num
Let's make clear what this module exports to allow us to easily deprecate and remove some of these in the future. Resolves https://gitlab.haskell.org/ghc/ghc/-/issues/26625
- - - - -
d99f8326 by Cheng Shao at 2025-12-11T19:14:18-05:00
compiler: remove unused CPP code in foreign stub
This patch removes unused CPP code in the generated foreign stub:
- `#define IN_STG_CODE 0` is not needed, since `Rts.h` already
includes this definition
- The `if defined(__cplusplus)` code paths are not needed in the `.c`
file, since we don't generate C++ stubs and don't include C++
headers in our stubs. But it still needs to be present in the `.h`
header since it might be later included into C++ source files.
- - - - -
46c9746f by Cheng Shao at 2025-12-11T19:14:57-05:00
configure: bump LlvmMaxVersion to 22
This commit bumps LlvmMaxVersion to 22; 21.x releases have been
available since Aug 26th, 2025 and there's no regressions with 21.x so
far. This bump is also required for updating fedora image to 43.
- - - - -
96fce8d0 by Cheng Shao at 2025-12-12T01:17:51+01:00
hadrian: add support for building with UndefinedBehaviorSanitizer
This patch adds a +ubsan flavour transformer to hadrian to build all
stage1+ C/C++ code with UndefinedBehaviorSanitizer. This is
particularly useful to catch potential undefined behavior in the RTS
codebase.
- - - - -
f7a06d8c by Cheng Shao at 2025-12-12T01:17:51+01:00
ci: update alpine/fedora & add ubsan job
This patch updates alpine image to 3.23, fedora image to 43, and adds
a `x86_64-linux-fedora43-validate+debug_info+ubsan` job that's run in
validate/nightly pipelines to catch undefined behavior in the RTS
codebase.
- - - - -
2ccd11ca by Cheng Shao at 2025-12-12T01:17:51+01:00
rts: fix zero-length VLA undefined behavior in interpretBCO
This commit fixes a zero-length VLA undefined behavior in interpretBCO, caught by UBSan:
```
+rts/Interpreter.c:3133:19: runtime variable length array bound evaluates to non-positive value 0
```
- - - - -
4156ed19 by Cheng Shao at 2025-12-12T01:17:51+01:00
rts: fix unaligned ReadSpB in interpretBCO
This commit fixes unaligned ReadSpB in interpretBCO, caught by UBSan:
```
+rts/Interpreter.c:2174:64: runtime load of misaligned address 0x004202059dd1 for type 'StgWord', which requires 8 byte alignment
```
To perform proper unaligned read, we define StgUnalignedWord as a type
alias of StgWord with aligned(1) attribute, and load StgUnalignedWord
instead of StgWord in ReadSpB, so the C compiler is aware that we're
not loading with natural alignment.
- - - - -
fef89fb9 by Cheng Shao at 2025-12-12T01:17:51+01:00
rts: fix signed integer overflow in subword arithmetic in interpretBCO
This commit fixes signed integer overflow in subword arithmetic in
interpretBCO, see added note for detailed explanation.
- - - - -
3c001377 by Cheng Shao at 2025-12-13T05:03:15-05:00
ci: use treeless fetch for perf notes
This patch improves the ci logic for fetching perf notes by using
treeless fetch
(https://github.blog/open-source/git/get-up-to-speed-with-partial-clone-and-s...),
to avoid downloading all blobs of the perf notes repo at once, and
only fetch the actually required blobs on-demand when needed. This
makes the initial `test-metrics.sh pull` operation much faster, and
also more robust, since we are seeing an increasing rate of 504 errors
in CI when fetching all perf notes at once, which is a major source of
CI flakiness at this point.
Co-authored-by: Codex