[Bug c/94522] New: Enhancement request: asm goto with outputs

[Bug c/94522] New: Enhancement request: asm goto with outputs

[bp at alien8 dot de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94522

            Bug ID: 94522
           Summary: Enhancement request: asm goto with outputs
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
          Assignee: unassigned at gcc dot gnu.org
          Reporter: bp at alien8 dot de
  Target Milestone: ---

Hi all,

this came up recently on the kernel ML (see
https://lkml.kernel.org/r/20200402134051.GC9352@zn.tnic) and we thought it
might be a good idea to support it. In the same "fallthrough" fashion like
clang:

https://reviews.llvm.org/rG50cac248773

Micha (CCed) says that it should be relatively easy to do that in gcc too so in
case someone feels bored in the future... :)

Some other useful scenarios, courtesy of Linus, would be exception handling, to
avoid register pressure and in the alternatives patching we do in the kernel.
And in general, an output value from the asm goto("") in the fallthrough case
might come in handy.

Thx.

[Bug inline-asm/94522] Enhancement request: asm goto with outputs

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94522

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Severity|normal                      |enhancement
   Last reconfirmed|                            |2020-04-07
     Ever confirmed|0                           |1
             Status|UNCONFIRMED                 |WAITING
          Component|c                           |inline-asm

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Take:
asm goto ("" : "=r"(*m) :....);

Where does the store to *m happen?  Do you replicate it on to the label side
too?

What are the semantics for the above case on clang?  My bet is it is not well
defined and really broken.

[Bug inline-asm/94522] Enhancement request: asm goto with outputs

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94522

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Also reload (and IRA) does not handle output on jumps because of the same
reason why I mention.

[Bug inline-asm/94522] Enhancement request: asm goto with outputs

[matz at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94522

--- Comment #3 from Michael Matz <matz at gcc dot gnu.org> ---
See the llvm link of the respective patch.  They specify that the outputs are
reliable only on the fallthrough (i.e. no goto taken) path (in particular the
outputs might or might not have been changed on the jump paths).  That avoids
all
the problems reload and it's replacements traditionally have with outputs on
jump insns, so it'd be relatively easy for us to support that as well.

This restriction, if documented, isn't even unreasonable and still allows some
interesting uses of goto asms with outputs.

[Bug inline-asm/94522] Enhancement request: asm goto with outputs

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94522

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|WAITING                     |NEW

[Bug inline-asm/94522] Enhancement request: asm goto with outputs

[torvalds@linux-foundation.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94522

Linus Torvalds <torvalds@linux-foundation.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |torvalds@linux-foundation.o
                   |                            |rg

--- Comment #4 from Linus Torvalds <torvalds@linux-foundation.org> ---
(In reply to Andrew Pinski from comment #1)
> Take:
> asm goto ("" : "=r"(*m) :....);
> 
> Where does the store to *m happen?  Do you replicate it on to the label side
> too?
> 
> What are the semantics for the above case on clang?  My bet is it is not
> well defined and really broken.

The outputs are defined to have values only in the fallthrough case.

On the label side, the outputs (whether memory or register) are explicitly
undefined. 

The outputs may obviously not even be in scope except in the fallthrough.

Now, with memory ops, it's obviously quite possible that the programmer then
knows that they wrote an inline asm that did the write to memory before it did
the goto. 

But that's no different from a _non_output_ "asm goto" that you pass a memory
pointer to, so I don't think that's something that is new to this situation.

As to whether this is simple or not to do in gcc - the clang implementation has
been buggy so far, altough it's good enough for testing ;)

[Bug inline-asm/94522] Enhancement request: asm goto with outputs

[torvalds@linux-foundation.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94522

--- Comment #5 from Linus Torvalds <torvalds@linux-foundation.org> ---
Btw, Nick (who is doing this on the clang side, tells me that the tcmalloc
people are looking at using the asm goto with outputs too, so it's not just the
kernel.

If somebody wants to play with it, I do have a patch to use it in the kernel as
a test-case.

HOWEVER - I'm working on cleaning up some of the infrastructure around it, but
at least for now, that patch is a "all or nothing" thing: it unconditionally
requires asm goto w/ inputs support, so there's no fallback to "older compiler
without asm goto with inputs" codepath.

With the cleanups, I hope to have a patch that can swing both ways, but right
now it's very much for testing only, and I don't want to make that patch too
public because it will just break for anybody with a non-experimental compiler.

[Bug inline-asm/94522] Enhancement request: asm goto with outputs

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94522

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Resolution|---                         |FIXED
   Target Milestone|---                         |11.0
             Status|NEW                         |RESOLVED

--- Comment #6 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Fixed in GCC 11 by r11-5002.