[NativeAOT] ARM: Implement the EABI #97604
Conversation
ghost
added
the
community-contribution
|
Tagging subscribers to this area: @agocke, @MichalStrehovsky, @jkotas Issue DetailsThis replaces DWARF exception unwinding on linux-arm with the compact EHABI representation. It also emits the Kept as draft since some code paths are missing. It passes the smoke tests though.
|
filipnavara
requested review from
lewing,
pavelsavara and
MichalStrehovsky
as code owners
filipnavara
force-pushed
the
naot-bringup-005
branch
from
ceca9ce to
ce9fe57
Compare
|
|
||
| // The maximum sequence length of the ARM EHABI unwinding code is 1024 | ||
| // bytes. | ||
| byte[] unwindData = new byte[1024]; |
There was a problem hiding this comment.
| byte[] unwindData = new byte[1024]; | |
| var unwindData = new ArrayBuilder<byte>(); |
I assume that the blob is typically much smaller. Allocating full 1024 bytes every time looks wasteful.
There was a problem hiding this comment.
Good catch. It was originally a stackalloc but I could not use that with the local methods. I rewrote it to use ArrayPool now, to match the existing code in DwarfFde. That said, it's probably something that can be rewritten not to allocate at all. I will likely revisit it at later time.
src/coreclr/tools/aot/ILCompiler.Compiler/Compiler/ObjectWriter/ElfObjectWriter.cs
Outdated
Show resolved
Hide resolved
| SectionWriter extabSectionWriter; | ||
|
|
||
| if (ShouldShareSymbol((ObjectNode)nodeWithCodeInfo) || | ||
| sectionWriter.SectionIndex != _managedCodeSectionIndex) |
There was a problem hiding this comment.
What is this special casing of _managedCodeSectionIndex about?
There was a problem hiding this comment.
The specification allows either one .ARM.exidx/.ARM.extab section per code section, or one for each method. This is enforced to some extent by the format since the .ARM.exidx link field has to point to the corresponding code section.
Since we currently produce the unwinding info only for the managed code section this was merely a shortcut to use a single unwinding section for managed code and fallback to one per function for everything else (realistically only COMDAT handled by the first condition at this point).
There was a problem hiding this comment.
Alternatively I can update the code to properly produce .ARM.exidx/.ARM.extab for each section and get rid of the condition...
There was a problem hiding this comment.
Alternatively I can update the code to properly produce .ARM.exidx/.ARM.extab for each section and get rid of the condition...
Done
|
/azp run runtime-nativeaot-outerloop |
|
Azure Pipelines successfully started running 1 pipeline(s). |
This replaces DWARF exception unwinding on linux-arm with the compact EHABI representation. It also emits the
.ARM.attributessection to aid debuggers and linkers.Majority of methods use single unwind word (4 bytes) to describe the method and two words in the index table (8 bytes). For something like the Exceptions smoke test we see the following file sizes of the
Exceptions.oobject file.Before: 16,769,298 bytes
After: 16,431,639 bytes
So, we saved 337,659 bytes and that doesn't even include the EH frame index produced by the linker for DWARF. It's not entirely fair comparison since the DWARF information emitted on ARM was overly verbose but it's a real size saving.