In https://reviews.llvm.org/D71059 LLVM 11, the time trace profiler was
extended to support multiple threads.
`timeTraceProfilerInitialize` creates a thread local profiler instance.
When a thread finishes `timeTraceProfilerFinishThread` moves a thread
local instance into a global collection of instances. Finally when all
codegen work is complete `timeTraceProfilerWrite` writes data from the
current thread local instance and the instances in global collection
of instances.
Previously, the profiler was intialized on a single thread only. Since
this thread performs no code generation on its own, the resulting
profile was empty.
Update LLVM codegen to initialize & finish time trace profiler on each
code generation thread.
Add -Z no-unique-section-names to reduce ELF header bloat.
This change adds a new compiler flag that can help reduce the size of ELF binaries that contain many functions.
By default, when enabling function sections (which is the default for most targets), the LLVM backend will generate different section names for each function. For example, a function `func` would generate a section called `.text.func`. Normally this is fine because the linker will merge all those sections into a single one in the binary. However, starting with [LLVM 12](https://github.com/llvm/llvm-project/commit/ee5d1a04), the backend will also generate unique section names for exception handling, resulting in thousands of `.gcc_except_table.*` sections ending up in the final binary because some linkers like LLD don't currently merge or strip these EH sections (see discussion [here](https://reviews.llvm.org/D83655)). This can bloat the ELF headers and string table significantly in binaries that contain many functions.
The new option is analogous to Clang's `-fno-unique-section-names`, and instructs LLVM to generate the same `.text` and `.gcc_except_table` section for each function, resulting in a smaller final binary.
The motivation to add this new option was because we have a binary that ended up with so many ELF sections (over 65,000) that it broke some existing ELF tools, which couldn't handle so many sections.
Here's our old binary:
```
$ readelf --sections old.elf | head -1
There are 71746 section headers, starting at offset 0x2a246508:
$ readelf --sections old.elf | grep shstrtab
[71742] .shstrtab STRTAB 0000000000000000 2977204c ad44bb 00 0 0 1
```
That's an 11MB+ string table. Here's the new binary using this option:
```
$ readelf --sections new.elf | head -1
There are 43 section headers, starting at offset 0x29143ca8:
$ readelf --sections new.elf | grep shstrtab
[40] .shstrtab STRTAB 0000000000000000 29143acc 0001db 00 0 0 1
```
The whole binary size went down by over 20MB, which is quite significant.
This change adds a new compiler flag that can help reduce the size of
ELF binaries that contain many functions.
By default, when enabling function sections (which is the default for most
targets), the LLVM backend will generate different section names for each
function. For example, a function "func" would generate a section called
".text.func". Normally this is fine because the linker will merge all those
sections into a single one in the binary. However, starting with LLVM 12
(llvm/llvm-project@ee5d1a0), the backend will
also generate unique section names for exception handling, resulting in
thousands of ".gcc_except_table.*" sections ending up in the final binary
because some linkers don't currently merge or strip these EH sections.
This can bloat the ELF headers and string table significantly in
binaries that contain many functions.
The new option is analogous to Clang's -fno-unique-section-names, and
instructs LLVM to generate the same ".text" and ".gcc_except_table"
section for each function, resulting in smaller object files and
potentially a smaller final binary.
Implement `#[link_ordinal(n)]`
Allows the use of `#[link_ordinal(n)]` with `#[link(kind = "raw-dylib")]`, allowing Rust to link against DLLs that export symbols by ordinal rather than by name. As long as the ordinal matches, the name of the function in Rust is not required to match the name of the corresponding function in the exporting DLL.
Part of #58713.
Enable AutoFDO.
This largely involves implementing the options debug-info-for-profiling
and profile-sample-use and forwarding them on to LLVM.
AutoFDO can be used on x86-64 Linux like this:
rustc -O -Clink-arg='Wl,--no-rosegment' -Cdebug-info-for-profiling main.rs -o main
perf record -b ./main
create_llvm_prof --binary=main --out=code.prof
rustc -O -Cprofile-sample-use=code.prof main.rs -o main2
Now `main2` will have feedback directed optimization applied to it.
The create_llvm_prof tool can be obtained from this github repository:
https://github.com/google/autofdo
The option -Clink-arg='Wl,--no-rosegment' is necessary to avoid lld
putting an extra RO segment before the executable code, which would make
the binary silently incompatible with create_llvm_prof.
This largely involves implementing the options debug-info-for-profiling
and profile-sample-use and forwarding them on to LLVM.
AutoFDO can be used on x86-64 Linux like this:
rustc -O -Cdebug-info-for-profiling main.rs -o main
perf record -b ./main
create_llvm_prof --binary=main --out=code.prof
rustc -O -Cprofile-sample-use=code.prof main.rs -o main2
Now `main2` will have feedback directed optimization applied to it.
The create_llvm_prof tool can be obtained from this github repository:
https://github.com/google/autofdoFixes#64892.
No functional changes intended.
The LLVM commit
e463b69736
changed an argument of fatal_error_handler_t from std::string to char*.
This adapts RustWrapper accordingly.
thinLTOResolvePrevailingInModule became thinLTOFinalizeInModule and
gained the ability to propagate noRecurse and noUnwind function
attributes. I ran codegen tests with it both on and off, as the upstream
patch uses it in both modes, and the tests pass both ways. Given that,
it seemed reasonable to go ahead and let the propagation be enabled in
rustc, and see what happens. See https://reviews.llvm.org/D36850 for
more examples of how the new version of the function gets used.
Change ab41eef9aca3 in LLVM split MemorySanitizerPass into
MemorySanitizerPass for functions and ModuleMemorySanitizerPass for
modules. There's a related change for ThreadSanitizerPass, and in here
since we're using a ModulePassManager I only add the module flavor of
the pass on LLVM 14.
r? @nikic cc @nagisa
These were deleted in https://reviews.llvm.org/D108614, and in C++ I
definitely see the argument for their removal. I didn't try and
propagate the changes up into higher layers of rustc in this change
because my initial goal was to get rustc working against LLVM HEAD
promptly, but I'm happy to follow up with some refactoring to make the
API on the Rust side match the LLVM API more directly (though the way
the enum works in Rust makes the API less scary IMO).
r? @nagisa cc @nikic
The above-mentioned commit (part of the LLVM 14 development cycle)
removes a method that rustc uses somewhat extensively. We mostly switch
to lower-level methods that exist in all versions of LLVM we use, so no
new ifdef logic is required in most cases.
PassWrapper: adapt for LLVM 14 changes
These API changes appear to have all taken place in
https://reviews.llvm.org/D105007, which moved HWAddressSanitizerPass and
AddressSanitizerPass to only accept their options type as a ctor
argument instead of the sequence of bools etc. This required a couple of
parameter additions, which I made match the default prior to the
mentioned upstream LLVM change.
This patch restores rustc to building (though not quite passing all
tests, I've mailed other patches for those issues) against LLVM HEAD.
These API changes appear to have all taken place in
https://reviews.llvm.org/D105007, which moved HWAddressSanitizerPass and
AddressSanitizerPass to only accept their options type as a ctor
argument instead of the sequence of bools etc. This required a couple of
parameter additions, which I made match the default prior to the
mentioned upstream LLVM change.
This patch restores rustc to building (though not quite passing all
tests, I've mailed other patches for those issues) against LLVM HEAD.
PassWrapper: handle move of OptimizationLevel class out of PassBuilder
This is the first build break of the LLVM 14 cycle, and was caused by
https://reviews.llvm.org/D107025. Mercifully an easy fix.
Rather than relying on `getPointerElementType()` from LLVM function
pointers, we now pass the function type explicitly when building `call`
or `invoke` instructions.
This does not yet support #[link_name] attributes on functions, the #[link_ordinal]
attribute, #[link(kind = "raw-dylib")] on extern blocks in bin crates, or
stdcall functions on 32-bit x86.
This commit updates how rustc compiler metadata is stored in rlibs.
Previously metadata was stored as a raw file that has the same format as
`--emit metadata`. After this commit, however, the metadata is encoded
into a small object file which has one section which is the contents of
the metadata.
The motivation for this commit is to fix a common case where #83730
arises. The problem is that when rustc crates a `dylib` crate type it
needs to include entire rlib files into the dylib, so it passes
`--whole-archive` (or the equivalent) to the linker. The problem with
this, though, is that the linker will attempt to read all files in the
archive. If the metadata file were left as-is (today) then the linker
would generate an error saying it can't read the file. The previous
solution was to alter the rlib just before linking, creating a new
archive in a temporary directory which has the metadata file removed.
This problem from before this commit is now removed if the metadata file
is stored in an object file that the linker can read. The only caveat we
have to take care of is to ensure that the linker never actually
includes the contents of the object file into the final output. We apply
similar tricks as the `.llvmbc` bytecode sections to do this.
This involved changing the metadata loading code a bit, namely updating
some of the LLVM C APIs used to use non-deprecated ones and fiddling
with the lifetimes a bit to get everything to work out. Otherwise though
this isn't intended to be a functional change really, only that metadata
is stored differently in archives now.
This should end up fixing #83730 because by default dylibs will no
longer have their rlib dependencies "altered" meaning that
split-debuginfo will continue to have valid paths pointing at the
original rlibs. (note that we still "alter" rlibs if LTO is enabled to
remove Rust object files and we also "alter" for the #[link(cfg)]
feature, but that's rarely used).
Closes#83730
PassWrapper: update for LLVM change D102093
In https://reviews.llvm.org/D102093 lots of things stopped taking the
DebugLogging boolean parameter. Mercifully we appear to always set
DebugPassManager to false, so I don't think we're losing anything by not
passing this parameter.
In https://reviews.llvm.org/D102093 lots of things stopped taking the
DebugLogging boolean parameter. Mercifully we appear to always set
DebugPassManager to false, so I don't think we're losing anything by not
passing this parameter.
This lets me build against llvm/main as of March 23rd, 2021. I'm not
entirely sure this is _correct_, but it appears to be functionally
identical to what was done in LLVM: existing callsites of
setInlineAsmDiagnosticHandler were moved to SetDiagnosticHandler() on
the context object, which we already set up in both places that we
called setInlineAsmDiagnosticHandler().
`fast-math` implies things like functions not being able to accept as an
argument or return as a result, say, `inf` which made these functions
confusingly named or behaving incorrectly, depending on how you
interpret it. Since the time when these intrinsics have been implemented
the intrinsics user's (stdsimd) approach has changed significantly and
so now it is required that these intrinsics operate normally rather than
in "whatever" way.
Fixes#84268
This should have no real effect in most cases, as e.g. `hidden`
visibility already implies `dso_local` (or at least LLVM IR does not
preserve the `dso_local` setting if the item is already `hidden`), but
it should fix `-Crelocation-model=static` and improve codegen in
executables.
Note that this PR does not exhaustively port the logic in [clang]. Only
the obviously correct portion and what is necessary to fix a regression
from LLVM 12 that relates to `-Crelocation_model=static`.
Fixes#83335
[clang]: 3001d080c8/clang/lib/CodeGen/CodeGenModule.cpp (L945-L1039)
THis came up in the review of #83425: it's hard to imagine a use of
LLVM_VERSION_LE() or LLVM_VERSION_EQ() that's not asking for trouble
when a point release gets created, so let's just discard them to prevent
the issue.
