Re: Delay RTL initialization until it is really needed

[Richard Sandiford <rdsandiford at googlemail dot com>]

Jan Hubicka <hubicka@ucw.cz> writes:
> Hi,

> IRA initialization shows high in profiles even when building lto
> objects.  This patch simply delays RTL backend initialization until we
> really decide to output a function.  In some cases this avoids the
> initialization completely (like in the case of LTO but also user
> target attributes) and there is some hope for better cache locality.
>
> Basic idea is to have two flags saying whether lang and target
> dependent bits needs initialization and check it when starting
> function codegen.
>
> Bootstrapped/regtested x86_64-linux, testing also at AIX. Ok if it passes?
>
> Honza
>
> 	* toplev.c (backend_init_target): Move init_emit_regs and init_regs to...
> 	(backend_init) ... here; skip ira_init_once and backend_init_target.
> 	(target_reinit) ... and here; clear this_target_rtl->lang_dependent_initialized.
> 	(lang_dependent_init_target): Clear this_target_rtl->lang_dependent_initialized;
> 	break out rtl initialization to ...
> 	(initialize_rtl): ... here; call also backend_init_target and ira_init_once.
> 	* toplev.h (initialize_rtl): New function.
> 	* function.c: Include toplev.h
> 	(init_function_start): Call initialize_rtl.
> 	* rtl.h (target_rtl): Add target_specific_initialized,
> 	lang_dependent_initialized.
> Index: toplev.c
> ===================================================================
> --- toplev.c	(revision 211837)
> +++ toplev.c	(working copy)
> @@ -1583,14 +1583,6 @@ backend_init_target (void)
>    /* Initialize alignment variables.  */
>    init_alignments ();
>  
> -  /* This reinitializes hard_frame_pointer, and calls init_reg_modes_target()
> -     to initialize reg_raw_mode[].  */
> -  init_emit_regs ();
> -
> -  /* This invokes target hooks to set fixed_reg[] etc, which is
> -     mode-dependent.  */
> -  init_regs ();
> -
>    /* This depends on stack_pointer_rtx.  */
>    init_fake_stack_mems ();
>  
> @@ -1632,9 +1624,13 @@ backend_init (void)
>    init_varasm_once ();
>    save_register_info ();
>  
> -  /* Initialize the target-specific back end pieces.  */
> -  ira_init_once ();
> -  backend_init_target ();
> +  /* Middle end needs this initialization for default mem attributes
> +     used by early calls to make_decl_rtl.  */
> +  init_emit_regs ();
> +
> +  /* Middle end needs this initialization for mode tables used to assign
> +     modes to vector variables.  */
> +  init_regs ();

This causes a segfault on gcc.target/mips/umips-store16-1.c.  The register
asm:

register unsigned int global asm ("$16");

causes us to globalise $16 and call reinit_regs.  reinit_regs in turn
calls ira_init, but IRA hasn't been initialised at this point and
prerequisites like init_fake_stack_mems haven't yet been called.

Does the patch below look OK?

> @@ -1686,6 +1682,31 @@ lang_dependent_init_target (void)
>       front end is initialized.  It also depends on the HAVE_xxx macros
>       generated from the target machine description.  */
>    init_optabs ();
> +  this_target_rtl->lang_dependent_initialized = false;
> +}
> +
> +/* Perform initializations that are lang-dependent or target-dependent.
> +   but matters only for late optimizations and RTL generation.  */
> +
> +void
> +initialize_rtl (void)
> +{
> +  static int initialized_once;
> +
> +  /* Initialization done just once per compilation, but delayed
> +     till code generation.  */
> +  if (!initialized_once)
> +    ira_init_once ();
> +  initialized_once = true;
> +
> +  /* Target specific RTL backend initialization.  */
> +  if (!this_target_rtl->target_specific_initialized)
> +    backend_init_target ();
> +  this_target_rtl->target_specific_initialized = true;
> +
> +  if (this_target_rtl->lang_dependent_initialized)
> +    return;
> +  this_target_rtl->lang_dependent_initialized = true;
>  
>    /* The following initialization functions need to generate rtl, so
>       provide a dummy function context for them.  */

Why do you need both these flags?  We only call this function once
the language has been initialised, so we should always be initialising
both sets of information (backend_init_target and the stuff after
the comment above, from the old lang_dependent_init_target).

How about the second patch below, still under testing?  The new assert
is OK for target_reinit because it has:

  this_target_rtl->target_specific_initialized = false;

  /* This initializes hard_frame_pointer, and calls init_reg_modes_target()
     to initialize reg_raw_mode[].  */
  init_emit_regs ();

  /* This invokes target hooks to set fixed_reg[] etc, which is
     mode-dependent.  */
  init_regs ();

  /* Reinitialize lang-dependent parts.  */
  lang_dependent_init_target ();

i.e. it sets the flag to say that the RTL stuff hasn't been initialised
and then goes on to initialise everything that needs to be deferred.

Thanks,
Richard


gcc/
	PR rtl-optimization/61629
	* reginfo.c (reinit_regs): Only call ira_init and recog_init if
	they have already been initialized.

Index: gcc/reginfo.c
===================================================================
--- gcc/reginfo.c	2014-07-16 07:59:00.039669668 +0100
+++ gcc/reginfo.c	2014-07-16 07:59:00.397672987 +0100
@@ -533,8 +533,11 @@ reinit_regs (void)
   init_regs ();
   /* caller_save needs to be re-initialized.  */
   caller_save_initialized_p = false;
-  ira_init ();
-  recog_init ();
+  if (this_target_rtl->target_specific_initialized)
+    {
+      ira_init ();
+      recog_init ();
+    }
 }
 
 /* Initialize some fake stack-frame MEM references for use in


gcc/
	* rtl.h (target_rtl): Remove lang_dependent_initialized.
	* toplev.c (initialize_rtl): Don't use it.  Move previously
	"language-dependent" calls to...
	(backend_init): ...here.
	(lang_dependent_init_target): Don't set lang_dependent_initialized.
	Assert that RTL initialization hasn't happend yet.

Index: gcc/rtl.h
===================================================================
--- gcc/rtl.h	2014-07-11 11:55:14.265158363 +0100
+++ gcc/rtl.h	2014-07-16 08:16:23.245370141 +0100
@@ -2517,7 +2517,6 @@ struct GTY(()) target_rtl {
 
   /* Track if RTL has been initialized.  */
   bool target_specific_initialized;
-  bool lang_dependent_initialized;
 };
 
 extern GTY(()) struct target_rtl default_target_rtl;
Index: gcc/toplev.c
===================================================================
--- gcc/toplev.c	2014-07-11 11:54:41.604838961 +0100
+++ gcc/toplev.c	2014-07-16 08:22:36.226034738 +0100
@@ -1604,6 +1604,10 @@ backend_init_target (void)
      on a mode change.  */
   init_expmed ();
   init_lower_subreg ();
+  init_set_costs ();
+
+  init_expr_target ();
+  ira_init ();
 
   /* We may need to recompute regno_save_code[] and regno_restore_code[]
      after a mode change as well.  */
@@ -1682,7 +1686,8 @@ lang_dependent_init_target (void)
      front end is initialized.  It also depends on the HAVE_xxx macros
      generated from the target machine description.  */
   init_optabs ();
-  this_target_rtl->lang_dependent_initialized = false;
+
+  gcc_assert (!this_target_rtl->target_specific_initialized);
 }
 
 /* Perform initializations that are lang-dependent or target-dependent.
@@ -1701,26 +1706,10 @@ initialize_rtl (void)
 
   /* Target specific RTL backend initialization.  */
   if (!this_target_rtl->target_specific_initialized)
-    backend_init_target ();
-  this_target_rtl->target_specific_initialized = true;
-
-  if (this_target_rtl->lang_dependent_initialized)
-    return;
-  this_target_rtl->lang_dependent_initialized = true;
-
-  /* The following initialization functions need to generate rtl, so
-     provide a dummy function context for them.  */
-  init_dummy_function_start ();
-
-  /* Do the target-specific parts of expr initialization.  */
-  init_expr_target ();
-
-  /* Although the actions of these functions are language-independent,
-     they use optabs, so we cannot call them from backend_init.  */
-  init_set_costs ();
-  ira_init ();
-
-  expand_dummy_function_end ();
+    {
+      backend_init_target ();
+      this_target_rtl->target_specific_initialized = true;
+    }
 }
 
 /* Language-dependent initialization.  Returns nonzero on success.  */