[llvm-dev] Finding Size of X86 instruction in MachineFunctionPass

[Craig Topper via llvm-dev]

Relaxation causes instructions to get larger. We optimistically assume that
we can encode a relative jump using a 1 byte immediate until we prove that
we can't. We will then switch to a 4 byte immediate.

~Craig


On Tue, Jul 10, 2018 at 2:06 PM John Criswell via llvm-dev <
llvm-dev at lists.llvm.org> wrote:

> On 7/10/18 1:48 PM, Cranmer, Joshua via llvm-dev wrote:
> > The actual size of even a MC instruction will change during
> relaxation--we don't choose between 8-bit jumps and wider jumps until
> relaxation. You can get the actual encoding of an MCInst via MCCodeEmitter,
> but if you expect relaxation to occur, or particularly if you're testing
> before register allocation, the size could only be an estimate and
> shouldn't be trusted for exact size.
>
> Thanks.  I should be doing this as the last pass before emission to the
> object file, so it is post register allocation.
>
> Will relaxation cause an instruction to become larger or only smaller?
> If the instruction ends up being smaller, that is fine. If it can get
> larger, that could be a problem.
>
> I'll take a look at MCCodeEmitter.
>
> Thanks,
>
> John Criswell
>
> >
> > -----Original Message-----
> > From: llvm-dev [mailto:llvm-dev-bounces at lists.llvm.org] On Behalf Of
> Matthias Braun via llvm-dev
> > Sent: Tuesday, July 10, 2018 13:21
> > To: John Criswell <jtcriswel at gmail.com>
> > Cc: llvm-dev at lists.llvm.org
> > Subject: Re: [llvm-dev] Finding Size of X86 instruction in
> MachineFunctionPass
> >
> > I don't think this API exists for machine function passes. It also
> wouldn't be too useful because the size of the instruction depends on
> details like which register is assigned  which could be changed by later
> machine passes. Of course during emission the MC layer should have an idea
> of how big instructions become, so your best bet is probably to modify the
> compiler at this point (I cannot give tips to that though as I'm not that
> familiar with these parts).
> >
> > - Matthias
> >
> >> On Jul 10, 2018, at 7:29 AM, John Criswell via llvm-dev <
> llvm-dev at lists.llvm.org> wrote:
> >>
> >> Dear All,
> >>
> >> Is there a way in newer versions of LLVM (4.0 and higher) to find the
> size of an x86 instruction within a MachineFunctionPass?  If not, is there
> something I can do in the MC layer to do this?
> >>
> >> I'm trying to modify LLVM to implement something like Native Client in
> which groups of instructions are aligned at a specific alignment and have
> the same length.  My current approach is to split each MachineBasicBlock so
> that it is less than or equal to the desired size and then align each
> MachineBasicBlock at the proper alignment.  However, I've been poking
> around doxygen and can't find anything that will give me the size of an x86
> instruction.
> >>
> >> Thanks in advance,
> >>
> >> John Criswell
> >>
> >> --
> >> John Criswell
> >> Assistant Professor
> >> Department of Computer Science, University of Rochester
> >> http://www.cs.rochester.edu/u/criswell
> >>
> >> _______________________________________________
> >> LLVM Developers mailing list
> >> llvm-dev at lists.llvm.org
> >> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
> > _______________________________________________
> > LLVM Developers mailing list
> > llvm-dev at lists.llvm.org
> > http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
> > _______________________________________________
> > LLVM Developers mailing list
> > llvm-dev at lists.llvm.org
> > http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>
>
> --
> John Criswell
> Assistant Professor
> Department of Computer Science, University of Rochester
> http://www.cs.rochester.edu/u/criswell
>
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20180710/96066af7/attachment.html>