[llvm-dev] Status of Intel JCC Mitigations and Next Steps

[Philip Reames via llvm-dev]

TLDR - We have a choice to make about assembler support, and a 
disagreement about how to move forward.  Community input needed.


Background

Intel has a hardware bug in Skylake and later whose mitigation requires 
padding of branches to avoid performance degradation. Background here: 
https://www.intel.com/content/dam/support/us/en/documents/processors/mitigations-jump-conditional-code-erratum.pdf

We now have in tree support for alignment of such branches via nop 
padding, and limited support for padding existing instructions with 
either prefixes or larger immediate values.  This has survived several 
days of dedicated testing and appears to be reasonably robust.  The 
padding support applies both to branch alignment for the mitigation, but 
also normal align directives.

The original patches proposed a somewhat different approach than we've 
ended up taking - primarily because of memory overhead concerns.  
However, there was also a deeper disagreement on the original review 
threads (D70157 and others) which was never settled, and we seem to be 
at a point where this needs attention. In short, the question is how 
assembler support should be handled.


The Choice

The problematic use case comes when assembling user provided .s files.  
(Instead of the more restricted output of the compiler.) Our basic 
choice is do we want to force a new directive syntax (and thus a source 
code change to use the new feature), or attempt to automatically infer 
where it's safe to insert padding?

The options as I see them:

  * Assembler directives w/explicit opt in - In this model, assembler
    input is assumed to only enable padding in regions where it is safe
    to do so.
  * Automagic assembler - In this model, the assembler is responsible
    for inferring where it is legal to pad without breaking user
    expectations.

(I'll stop and disclaim that I'm strongly in favor of the former.  I've 
tried to describe the pros/cons of each, but my perspective is 
definitely biased.)

The difference between the two is a huge amount of complexity, and a 
very fundamental correctness risk.  The basic problem is that assemblers 
have to handle unconstrained inputs, and IMO, the semantics of assembler 
as used in practice is so under specified that it's really hard to infer 
semantics in any useful way.  As a couple of examples, is the fault 
behavior of an instruction well defined?  Is the label near an 
instruction used by the signal handler?  Is that data byte just before 
an instruction actually decoded as part of the instruction?

The benefit of the later option is that existing assembly files can be 
used without modification.  This is a huge advantage in terms of ease of 
mitigation for existing code bases.  It's also the approach the original 
patch sets for GCC took.

In the original review thread(s), I had taken the position that we 
should reject the automagic assembler based on the correctness concerns 
mentioned.  I had thought the consensus in the review was clearly in 
that direction as well, but this has recently come up again.  Given 
that, I wanted to open it to a wider audience.


Why am I pushing for a decision now?

There are two major reasons.  First, there have recently been a couple 
of patches posted and landed (D76176, and D76052) building towards the 
automagic assembler variant.  And second, I've started getting review 
comments (https://reviews.llvm.org/D76398#1930383) which block forward 
progress on generalized padding support assuming the automagic 
interpretation.  Implementing the automatic assembler variant for prefix 
and immediate padding adds substantial complexity and I would very much 
like not to bother with if I don't have to.


Current implementation details

We have support in the integrated assembler only for autopadding 
suppression.  This allows a LLVM based compiler to effectively apply 
padding selectively.  In particular, we've instrumented lowering from MI 
to MC (X86MCInstLowering.cpp) to selectively disable padding around 
constructs which are thought to be problematic.  We do not have an 
agreed upon syntax for this in assembler; the code that got checked in 
is modeled closely around the last seriously discussed variant (see 
below).  This support is able to use all of the padding variants: nop, 
prefix, and immediate.

We also have limited support in the assembler for not inserting nops 
between fragments where doing so would break known idioms. The list of 
such idioms is, IMO, ad hoc.  This assembler support does not include 
prefix or immediate padding.


Philip

p.s. For those interested, here's roughly what the last round of 
assembler syntax I remember being discussed looked like.

.autopadding
.noautopadding

These two directives would respectively enable and disable automatic 
padding of instructions within the region defined.  It's presumed to be 
legal to insert nops between instructions, modify encodings, or 
otherwise adjust offsets of instruction boundaries within the region to 
achieve target specific desired alignments. Similarly, it's presumed not 
to be legal to change relative offsets outside an explicitly enabled 
region.  (Except for existing cases - e.g. relaxation of branches, etc...)

The assembler would provide a command line flag which conceptually 
wrapped the whole file in a pair of enable/disable directives.

We'd previously discussed a variant with push/pop semantics and more 
fine grained control over alignment requests, but I believe we decided 
that was overkill in the end.  (I walked away with that impression based 
on the integrated assembler work at least.)


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20200324/3999fdc8/attachment.html>