<html>
<head>
<meta http-equiv="Content-Type" content="text/html; charset=iso-8859-1">
<style type="text/css" style="display:none;"> P {margin-top:0;margin-bottom:0;} </style>
</head>
<body dir="ltr">
<div style="font-family: Calibri, Arial, Helvetica, sans-serif; font-size: 12pt; color: rgb(0, 0, 0); background-color: rgb(255, 255, 255);">
Replies inlined.</div>
<div id="appendonsend"></div>
<div style="font-family: Calibri, Arial, Helvetica, sans-serif; font-size: 12pt; color: rgb(0, 0, 0);">
<br>
</div>
<hr tabindex="-1" style="display:inline-block; width:98%">
<div id="divRplyFwdMsg" dir="ltr"><font face="Calibri, sans-serif" color="#000000" style="font-size: 11pt;"><b>From:</b> maskray@google.com <maskray@google.com><br>
<b>Sent:</b> Thursday, May 27, 2021 5:23 PM<br>
<b>To:</b> Alexander Yermolovich <ayermolo@fb.com><br>
<b>Cc:</b> llvm-dev@lists.llvm.org <llvm-dev@lists.llvm.org>; Wenlei He <wenlei@fb.com>; Modi Mo <modimo@fb.com>; dblaikie@gmail.com <dblaikie@gmail.com><br>
<b>Subject:</b> Re: [RFC] Changing .llvm.call-graph-profile to use relocations</font>
<div> </div>
</div>
<div class="BodyFragment"><font size="2"><span style="font-size:11pt">
<div class="PlainText">On 2021-05-27, Alexander Yermolovich wrote:<br>
>I was thinking keeping both structures for backward compatibility in case object files with old representation are fed in to new llvm-objdump, and even lld. For example if someone will use older clang release with lld/llvm-objdump after this change.<br>
>For just changing the value and keeping the name. Looks like it will leave a gap in this sequential sequence. If a new flag is added later down the road and in this case latest clang used with older lld/llvm-objdump this will also present a problem as older
tools will interpret new flag as SHT_LLVM_CALL_GRAPH_PROFILE.<br>
>Not sure if these are valid concerns, what do you think?<br>
>If we go for clean state, then maybe leave 0x6fff4c02 entry and change it to SHT_LLVM_CALL_GRAPH_PROFILE_DEPRICATED, and emit warning if objects with it are passed in.<br>
> SHT_LLVM_ODRTAB = 0x6fff4c00, // LLVM ODR table.<br>
> SHT_LLVM_LINKER_OPTIONS = 0x6fff4c01, // LLVM Linker Options.<br>
> SHT_LLVM_CALL_GRAPH_PROFILE = 0x6fff4c02, // LLVM Call Graph Profile.<br>
> SHT_LLVM_ADDRSIG = 0x6fff4c03, // List of address-significant symbols<br>
> // for safe ICF.<br>
> SHT_LLVM_DEPENDENT_LIBRARIES =<br>
> 0x6fff4c04, // LLVM Dependent Library Specifiers.<br>
> SHT_LLVM_SYMPART = 0x6fff4c05, // Symbol partition specification.<br>
> SHT_LLVM_PART_EHDR = 0x6fff4c06, // ELF header for loadable partition.<br>
> SHT_LLVM_PART_PHDR = 0x6fff4c07, // Phdrs for loadable partition.<br>
> SHT_LLVM_BB_ADDR_MAP = 0x6fff4c08, // LLVM Basic Block Address Map.<br>
<br>
Do you have measurement how well SHT_LLVM_CALL_GRAPH_PROFILE optimizes?<br>
My understanding is that with ThinLTO+PGO it is has very tiny benefit.<br>
[Alex] I do not. Wenlei mentioned he can dig up some numbers.<br>
<br>
The value reading old format is low. I don't expect users want to mix<br>
newer object files with old object files and want to have optimization<br>
from the old object files. SHT_LLVM_CALL_GRAPH_PROFILE_DEPRICATED is<br>
just adding maintenance cost.<br>
<br>
Changing the value but retaining the name is sufficient for the various<br>
compatibility issues.<br>
<br>
Reid's size concern is valid. Have you measured the size overhead?</div>
<div class="PlainText">[Alex] Good point. I'll measure once I implement new approach with SHT_REL + <span style="color: rgb(0, 0, 0); background-color: rgb(255, 255, 255); display: inline !important;">R_X86_64_NONE.</span></div>
<div class="PlainText"><br>
We can use the SHT_REL format for SHT_LLVM_CALL_GRAPH_PROFILE<br>
relocations, which brings down the per entry cost from 24 bytes to 16<br>
bytes for ELFCLASS64. (It is a bit unfortunate that the Linux kernel<br>
does not support ELFCLASS32 executables on a 64-bit architecture.<br>
For most usage we are using the small code model and don't benefit<br>
from 64-bit addresses/sizes.)<br>
<br>
>________________________________<br>
>From: maskray@google.com <maskray@google.com><br>
>Sent: Thursday, May 27, 2021 11:27 AM<br>
>To: Alexander Yermolovich <ayermolo@fb.com><br>
>Cc: llvm-dev@lists.llvm.org <llvm-dev@lists.llvm.org>; Wenlei He <wenlei@fb.com>; Modi Mo <modimo@fb.com>; dblaikie@gmail.com <dblaikie@gmail.com><br>
>Subject: Re: [RFC] Changing .llvm.call-graph-profile to use relocations<br>
><br>
>On 2021-05-27, Alexander Yermolovich wrote:<br>
>>Thank you for feedback, let me try to use R_X86_64_NONE, and introduce another type. . I thought R_X86_64_32 would be less impactful as structure of is preserved, but it is space wasteful.<br>
>>For type name: ELF::SHT_LLVM_CALL_GRAPH_PROFILE_RELA ?<br>
>><br>
>>Alex<br>
><br>
>Preversing the structure is not needed because the symbol representation<br>
>is changed anyway.<br>
><br>
>You just need to change the value in<br>
>llvm/llvm/include/llvm/BinaryFormat/ELF.h<br>
>The name doesn't need to change.<br>
><br>
>>From: maskray@google.com <maskray@google.com><br>
>>Sent: Wednesday, May 26, 2021 5:03 PM<br>
>>To: Alexander Yermolovich <ayermolo@fb.com><br>
>>Cc: llvm-dev@lists.llvm.org <llvm-dev@lists.llvm.org>; Wenlei He <wenlei@fb.com>; Modi Mo <modimo@fb.com>; dblaikie@gmail.com <dblaikie@gmail.com><br>
>>Subject: Re: [RFC] Changing .llvm.call-graph-profile to use relocations<br>
>><br>
>>On 2021-05-26, Alexander Yermolovich wrote:<br>
>>>Currently when .llvm.call-graph-profile is created by llvm it explicitly encodes the symbol indices. This section is basically a black box for post processing tools. For example, if we run strip -s on the object files the symbol table changes, but indices
in that section do not.<br>
>>><br>
>>>We propose to change this section to use relocations. The Frequency will still be in the .llvm.call-graph-profile, but symbol information will be in relocation section. In LLD information from both sections is used to reconstruct call graph profile. Relocations
themselves will never be applied.<br>
>><br>
>>Yes, a relocation based approach will be more robust and can fix the<br>
>>.llvm_addrsig issue <a href="https://sourceware.org/bugzilla/show_bug.cgi?id=23817">https://sourceware.org/bugzilla/show_bug.cgi?id=23817</a>
<br>
>><br>
>>The relocation type doesn't matter. Your implementation uses R_X86_64_32 and from/to/value/from/to/value/from/to/value<br>
>>An alternative design is R_X86_64_NONE + value/value/value, i.e. from/to do not occupy space in the content.<br>
>>We will get a 3x space saving.<br>
>><br>
>>We need to change the section type since changing representation is incompatible<br>
>>and the sections from old object files should be ignored.<br>
>>The new section type will be ignored by old LLVM tools as well.<br>
>><br>
>>>With this approach post processing tools that handle relocations correctly work for this section also.<br>
>>><br>
>>>One thing is section is marked with SHF_EXCLUDE.<br>
>>>From spec<br>
>>>"This section is excluded from input to the link-edit of an executable or shared object. This flag is ignored if the SHF_ALLOC flag is also set, or if relocations exist against the section."<br>
>>><br>
>>>So technically speaking it needs to be kept, and presumably relocations applied, but LLD follows gold and ld and discards sections marked as SHF_EXCLUDE even with relocations. So, I think this approach should be fine.
<a href="https://reviews.llvm.org/D24966">https://reviews.llvm.org/D24966</a>
<br>
>><br>
>>This is Solaris's spec (since 1996), not standard ELF's. It can be advisory but<br>
>>our behavior (mostly GNU ABI+Linux ABI+LLVM extensions) does not necessarily<br>
>>follow it. I cannot find a definition in the x86-64 psABI or a GNU ABI<br>
>>document.<br>
>><br>
>>I think the behavior as implemented in gold and LLD is more useful.<br>
>><br>
>>>Finally, this bug seems similar to <a href="https://sourceware.org/bugzilla/show_bug.cgi?id=23817">https://sourceware.org/bugzilla/show_bug.cgi?id=23817</a> . Proposed solution for that was also to use relocations.<br>
>>><br>
>>>Implementation: <a href="https://reviews.llvm.org/D103212">https://reviews.llvm.org/D103212</a>
<br>
</div>
</span></font></div>
</body>
</html>