[3/3] lto: respect jobserver in parallel WPA streaming

Patch can bootstrap on x86_64-linux-gnu and survives regression tests.

Ready to be installed?
Thanks,
Martin

	PR lto/106328

gcc/ChangeLog:

	* jobserver.h (struct jobserver_info): Add pipefd.
	(jobserver_info::connect): New.
	(jobserver_info::disconnect): Likewise.
	(jobserver_info::get_token): Likewise.
	(jobserver_info::return_token): Likewise.

gcc/lto/ChangeLog:

	* lto.cc (wait_for_child): Decrement nruns once a process
	finishes.
	(stream_out_partitions): Use job server if active.
	(do_whole_program_analysis): Likewise.
---
 gcc/jobserver.h | 54 ++++++++++++++++++++++++++++++++++++++++++++++++
 gcc/lto/lto.cc  | 55 +++++++++++++++++++++++++++++++++++++------------
 2 files changed, 96 insertions(+), 13 deletions(-)

On Tue, Aug 9, 2022 at 2:02 PM Martin Liška <mliska@suse.cz> wrote:
>
> Patch can bootstrap on x86_64-linux-gnu and survives regression tests.
>
> Ready to be installed?
> Thanks,
> Martin
>
>         PR lto/106328
>
> gcc/ChangeLog:
>
>         * jobserver.h (struct jobserver_info): Add pipefd.
>         (jobserver_info::connect): New.
>         (jobserver_info::disconnect): Likewise.
>         (jobserver_info::get_token): Likewise.
>         (jobserver_info::return_token): Likewise.
>
> gcc/lto/ChangeLog:
>
>         * lto.cc (wait_for_child): Decrement nruns once a process
>         finishes.
>         (stream_out_partitions): Use job server if active.
>         (do_whole_program_analysis): Likewise.
> ---
>  gcc/jobserver.h | 54 ++++++++++++++++++++++++++++++++++++++++++++++++
>  gcc/lto/lto.cc  | 55 +++++++++++++++++++++++++++++++++++++------------
>  2 files changed, 96 insertions(+), 13 deletions(-)
>
> diff --git a/gcc/jobserver.h b/gcc/jobserver.h
> index 856e326ddfc..2a7dc9f4113 100644
> --- a/gcc/jobserver.h
> +++ b/gcc/jobserver.h
> @@ -31,6 +31,18 @@ struct jobserver_info
>    /* Default constructor.  */
>    jobserver_info ();
>
> +  /* Connect to the server.  */
> +  void connect ();
> +
> +  /* Disconnect from the server.  */
> +  void disconnect ();
> +
> +  /* Get token from the server.  */
> +  bool get_token ();
> +
> +  /* Return token to the server.  */
> +  void return_token ();
> +
>    /* Error message if there is a problem.  */
>    string error_msg = "";
>    /* Skipped MAKEFLAGS where --jobserver-auth is skipped.  */
> @@ -41,6 +53,8 @@ struct jobserver_info
>    int wfd = -1;
>    /* Named pipe path.  */
>    string pipe_path = "";
> +  /* Pipe file descriptor.  */
> +  int pipefd = -1;
>    /* Return true if jobserver is active.  */
>    bool is_active = false;
>  };
> @@ -97,4 +111,44 @@ jobserver_info::jobserver_info ()
>      error_msg = "jobserver is not available: " + error_msg;
>  }
>
> +void
> +jobserver_info::connect ()
> +{
> +  if (!pipe_path.empty ())
> +    pipefd = open (pipe_path.c_str (), O_RDWR);
> +}
> +
> +void
> +jobserver_info::disconnect ()

If it's not a template and the methods not inline won't we get multiple
definition errors when this include file is included in multiple places
from an executable?  opts-common.cc might be a good place to
stuff it, and opts.h to declare (if it were not for std::string, so possibly
jobserver.h is good enough, or opts-jobserver.h?)

> +{
> +  if (!pipe_path.empty ())
> +    {
> +      gcc_assert (close (pipefd) == 0);
> +      pipefd = -1;
> +    }
> +}
> +
> +bool
> +jobserver_info::get_token ()
> +{
> +  int fd = pipe_path.empty () ? rfd : pipefd;
> +  char c;
> +  unsigned n = read (fd, &c, 1);
> +  if (n != 1)
> +    {
> +      gcc_assert (errno == EAGAIN);
> +      return false;
> +    }
> +  else
> +    return true;
> +}
> +
> +void
> +jobserver_info::return_token ()
> +{
> +  int fd = pipe_path.empty () ? wfd : pipefd;
> +  char c = 'G';
> +  gcc_assert (write (fd, &c, 1) == 1);
> +}
> +
>  #endif /* GCC_JOBSERVER_H */
> diff --git a/gcc/lto/lto.cc b/gcc/lto/lto.cc
> index 31b0c1862f7..56266195ead 100644
> --- a/gcc/lto/lto.cc
> +++ b/gcc/lto/lto.cc
> @@ -54,11 +54,17 @@ along with GCC; see the file COPYING3.  If not see
>  #include "attribs.h"
>  #include "builtins.h"
>  #include "lto-common.h"
> -
> +#include "jobserver.h"
>
>  /* Number of parallel tasks to run, -1 if we want to use GNU Make jobserver.  */
>  static int lto_parallelism;
>
> +/* Number of active WPA streaming processes.  */
> +static int nruns = 0;
> +
> +/* GNU make's jobserver info.  */
> +static jobserver_info *jinfo = NULL;
> +
>  /* Return true when NODE has a clone that is analyzed (i.e. we need
>     to load its body even if the node itself is not needed).  */
>
> @@ -205,6 +211,12 @@ wait_for_child ()
>                      "streaming subprocess was killed by signal");
>      }
>    while (!WIFEXITED (status) && !WIFSIGNALED (status));
> +
> +    --nruns;
> +
> +    /* Return token to the jobserver if active.  */
> +    if (jinfo != NULL && jinfo->is_active)
> +      jinfo->return_token ();
>  }
>  #endif
>
> @@ -228,25 +240,35 @@ stream_out_partitions (char *temp_filename, int blen, int min, int max,
>                        bool ARG_UNUSED (last))
>  {
>  #ifdef HAVE_WORKING_FORK
> -  static int nruns;
> -
>    if (lto_parallelism <= 1)
>      {
>        stream_out_partitions_1 (temp_filename, blen, min, max);
>        return;
>      }
>
> -  /* Do not run more than LTO_PARALLELISM streamings
> -     FIXME: we ignore limits on jobserver.  */
>    if (lto_parallelism > 0 && nruns >= lto_parallelism)
> -    {
> -      wait_for_child ();
> -      nruns --;
> -    }
> +    wait_for_child ();
> +
>    /* If this is not the last parallel partition, execute new
>       streaming process.  */
>    if (!last)
>      {
> +      if (jinfo != NULL && jinfo->is_active)
> +       while (true)
> +         {
> +           if (jinfo->get_token ())
> +             break;
> +           if (nruns > 0)
> +             wait_for_child ();
> +           else
> +             {
> +               /* There are no free tokens, lets do the job outselves.  */
> +               stream_out_partitions_1 (temp_filename, blen, min, max);
> +               asm_nodes_output = true;
> +               return;
> +             }
> +         }
> +
>        pid_t cpid = fork ();
>
>        if (!cpid)
> @@ -264,10 +286,12 @@ stream_out_partitions (char *temp_filename, int blen, int min, int max,
>    /* Last partition; stream it and wait for all children to die.  */
>    else
>      {
> -      int i;
>        stream_out_partitions_1 (temp_filename, blen, min, max);
> -      for (i = 0; i < nruns; i++)
> +      while (nruns > 0)
>         wait_for_child ();
> +
> +      if (jinfo != NULL && jinfo->is_active)
> +       jinfo->disconnect ();
>      }
>    asm_nodes_output = true;
>  #else
> @@ -460,9 +484,14 @@ do_whole_program_analysis (void)
>
>    lto_parallelism = 1;
>
> -  /* TODO: jobserver communication is not supported, yet.  */
>    if (!strcmp (flag_wpa, "jobserver"))
> -    lto_parallelism = param_max_lto_streaming_parallelism;
> +    {
> +      jinfo = new jobserver_info ();
> +      if (jinfo->is_active)
> +       jinfo->connect ();
> +
> +      lto_parallelism = param_max_lto_streaming_parallelism;
> +    }
>    else
>      {
>        lto_parallelism = atoi (flag_wpa);
> --
> 2.37.1
>

On 8/10/22 12:52, Richard Biener wrote:
> If it's not a template and the methods not inline won't we get multiple
> definition errors when this include file is included in multiple places
> from an executable?  opts-common.cc might be a good place to
> stuff it, and opts.h to declare (if it were not for std::string, so possibly
> jobserver.h is good enough, or opts-jobserver.h?)

Works for me, updated in v2.

Martin