llvm dev - May 2016 - Do LLVM passes have different behavior depending on the target triple and/or the optimization level selected?

Hi again,


Thanks for answering my previous question :)

There is something else I do not understand about the passes in LLVM.

Does the behavior of some LLVM passes change depending on the 'target 
triple' passed (either inside the LLVM IR or passed as command line 
parameter) to the LLVM Optimizer (LLVM 'opt' tool)?
Is the behavior of the passes when called by LLVM 'opt' target specific 
(architecture/subarchitecture, vendor, operating system, etc...)?


If that is the case, then when I'm cross-compiling for a given target 
(e.g., target SPARC-ELF on a X86-LINUX machine) I must always pass the 
target triple.
If that is not the case, they I only need to pass the target triple to 
the 'llc' static compiler when gerating the assembly code for the target
platform.


Additionally, someone pointed me to this commit, suggesting that the 
behavior of some passes depends on the selected optimization level.
Is this correct?

If so, is it possible to execute only some of the passes in -O2 (or in a 
different order), but have them individually behaving as if -O2 was 
being used?

Is there any direct mechanism to control compiler passes individually?


Regards,
Ricardo

Hi Ricardo,

Yes, the optimizer's behavior can, and does, change depending on the target
triple, so you definitely need to pass the target triple to the optimizer
when cross-compiling.
Furthermore, even your frontend will often care about the triple (for a
simple example, consider sizeof in C). So you really ought to use the same
triple through the entire compilation process.

Michael

On Tue, May 10, 2016 at 10:21 AM, Ricardo Nobre via llvm-dev <
llvm-dev at lists.llvm.org> wrote:
> Hi again,
>
>
> Thanks for answering my previous question :)
>
> There is something else I do not understand about the passes in LLVM.
>
> Does the behavior of some LLVM passes change depending on the 'target
> triple' passed (either inside the LLVM IR or passed as command line
> parameter) to the LLVM Optimizer (LLVM 'opt' tool)?
> Is the behavior of the passes when called by LLVM 'opt' target
specific
> (architecture/subarchitecture, vendor, operating system, etc...)?
>
>
> If that is the case, then when I'm cross-compiling for a given target
> (e.g., target SPARC-ELF on a X86-LINUX machine) I must always pass the
> target triple.
> If that is not the case, they I only need to pass the target triple to the
> 'llc' static compiler when gerating the assembly code for the
target
> platform.
>
>
> Additionally, someone pointed me to this commit, suggesting that the
> behavior of some passes depends on the selected optimization level.
> Is this correct?
>
> If so, is it possible to execute only some of the passes in -O2 (or in a
> different order), but have them individually behaving as if -O2 was being
> used?
>
> Is there any direct mechanism to control compiler passes individually?
>
>
> Regards,
> Ricardo
> _______________________________________________
> 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/20160510/32f314f1/attachment.html>