[PATCH/RFA] PR target/13250

[PATCH/RFA] PR target/13250

[Kaz Kojima]

Hi,

It seems that PR target/13250
<URL:http://gcc.gnu.org/bugzilla/show_bug.cgi?id=13250>
is a problem caused by a direct substitution operands[1] for parts[?]
in rotlsi3 expand of SH.  The appended patch fixes it.
It's tested with bootstrap 3.4 on sh4-unknown-linux-gnu and regtested
on mainline for x86 cross to sh4-unknown-linux-gnu, though now mainline
fails in building c++ library for this target without -fno-schedule-insns.

Regards,
	kaz
--
	PR target/13250
	* config/sh/sh.md (rotlsi3): Use emit_move_insn.

diff -uprN ORIG/gcc/gcc/config/sh/sh.md LOCAL/gcc/gcc/config/sh/sh.md
--- ORIG/gcc/gcc/config/sh/sh.md	2004-05-14 10:08:34.000000000 +0900
+++ LOCAL/gcc/gcc/config/sh/sh.md	2004-05-27 08:32:37.000000000 +0900
@@ -2194,7 +2194,7 @@
 	parts[0] = gen_reg_rtx (SImode);
 	parts[1] = gen_reg_rtx (SImode);
 	emit_insn (gen_rotlsi3_16 (parts[2-choice], operands[1]));
-	parts[choice-1] = operands[1];
+	emit_move_insn (parts[choice-1], operands[1]);
 	emit_insn (gen_ashlsi3 (parts[0], parts[0], GEN_INT (8)));
 	emit_insn (gen_lshrsi3 (parts[1], parts[1], GEN_INT (8)));
 	emit_insn (gen_iorsi3 (operands[0], parts[0], parts[1]));

Re: [PATCH/RFA] PR target/13250

[Alexandre Oliva]

On May 28, 2004, Kaz Kojima <kkojima@rr.iij4u.or.jp> wrote:

> 	PR target/13250
> 	* config/sh/sh.md (rotlsi3): Use emit_move_insn.

Ok, thanks.

-- 
Alexandre Oliva             http://www.ic.unicamp.br/~oliva/
Red Hat Compiler Engineer   aoliva@{redhat.com, gcc.gnu.org}
Free Software Evangelist  oliva@{lsd.ic.unicamp.br, gnu.org}

Re: [PATCH/RFA] PR target/13250

[Kaz Kojima]

Alexandre Oliva <aoliva@redhat.com> wrote:
>> 	PR target/13250
>> 	* config/sh/sh.md (rotlsi3): Use emit_move_insn.
> 
> Ok, thanks.

Thanks!  Applied to mainline.

This PR is originally against 3.3/3.4.  Mark, is it ok for 3.4?
Bootstrap and regression test is done already on 3.4-branch.

Gaby, can I queue it for 3.3.5 if the tests with it is done
successfully on 3.3?

Regards,
	kaz

Re: [PATCH/RFA] PR target/13250

[Gabriel Dos Reis]

Kaz Kojima <kkojima@rr.iij4u.or.jp> writes:

| Alexandre Oliva <aoliva@redhat.com> wrote:
| >> 	PR target/13250
| >> 	* config/sh/sh.md (rotlsi3): Use emit_move_insn.
| > 
| > Ok, thanks.
| 
| Thanks!  Applied to mainline.
| 
| This PR is originally against 3.3/3.4.  Mark, is it ok for 3.4?
| Bootstrap and regression test is done already on 3.4-branch.
| 
| Gaby, can I queue it for 3.3.5 if the tests with it is done
| successfully on 3.3?

If it goes into 3.4.x, yes.

-- Gaby

Re: [PATCH/RFA] PR target/13250

[Joern Rennecke]

> It's tested with bootstrap 3.4 on sh4-unknown-linux-gnu and regtested
> on mainline for x86 cross to sh4-unknown-linux-gnu, though now mainline
> fails in building c++ library for this target without -fno-schedule-insns.

Strange, are you saying it worked before?  sh-elf has failed to build
in mailine for the past two weeks.
See http://gcc.gnu.org/ml/gcc-patches/2004-05/msg00857.html

Re: [PATCH/RFA] PR target/13250

[Kaz Kojima]

Joern Rennecke <joern.rennecke@superh.com> wrote:
>> It's tested with bootstrap 3.4 on sh4-unknown-linux-gnu and regtested
>> on mainline for x86 cross to sh4-unknown-linux-gnu, though now mainline
>> fails in building c++ library for this target without -fno-schedule-insns.
> 
> Strange, are you saying it worked before?  sh-elf has failed to build
> in mailine for the past two weeks.
> See http://gcc.gnu.org/ml/gcc-patches/2004-05/msg00857.html

I mean that -fno-schedule-insns is needed to build c++ library on
mainline.  ATM, there are many spill failures in building c++ and
java library for sh4-unknown-linux-gnu.
Your patch for post-call group sheduling fixes all of them except
a failure compiling local-inst.cc for -fPIC.

Regards,
	kaz

Re: [PATCH/RFA] PR target/13250

[Joern Rennecke]

> Your patch for post-call group sheduling fixes all of them except
> a failure compiling local-inst.cc for -fPIC.

Thanks for the testcase (sent off-list).
It turns out that the loop optimizer has inserted a jump between
the call and the instruction that copied the return value.
I.e. before loop2 we have:

(call_insn 2975 2974 3094 88 (parallel [
            (set (reg:SI 0 r0)
                (call (mem:SI (symbol_ref:SI ("_ZNKSt6locale2id5_M_idEv") [flags 0x41] <function_decl 0x408077b4 _M_id>) [0 S4 A32])
                    (const_int 0 [0x0])))
            (use (reg:PSI 151 ))
            (use (reg:SI 12 r12))
            (clobber (reg:SI 146 pr))
            (clobber (scratch:SI))
        ]) 172 {call_value_pcrel} (nil)
    (expr_list:REG_EH_REGION (const_int 1 [0x1])
        (nil))
    (expr_list (use (reg:SI 4 r4))
        (nil)))
;; End of basic block 88, registers live:
 (nil)

;; Start of basic block 89, registers live: (nil)
(note 3094 2975 2976 89 [bb 89] NOTE_INSN_BASIC_BLOCK)

(insn 2976 3094 2978 89 (set (reg:SI 182 [ T.678 ])
        (reg:SI 0 r0)) 123 {movsi_ie} (nil)
    (nil))

and after loop2 we got:

(call_insn 2975 2974 3463 90 (parallel [
            (set (reg:SI 0 r0)
                (call (mem:SI (symbol_ref:SI ("_ZNKSt6locale2id5_M_idEv") [flags 0x41] <function_decl 0x408077b4 _M_id>) [0 S4 A32])
                    (const_int 0 [0x0])))
            (use (reg:PSI 151 ))
            (use (reg:SI 12 r12))
            (clobber (reg:SI 146 pr))
            (clobber (scratch:SI))
        ]) 172 {call_value_pcrel} (nil)
    (expr_list:REG_EH_REGION (const_int 1 [0x1])
        (nil))
    (expr_list (use (reg:SI 4 r4))
        (nil)))
;; End of basic block 90, registers live:
 (nil)

;; Start of basic block 91, registers live: (nil)
(note 3463 2975 3465 91 [bb 91] NOTE_INSN_BASIC_BLOCK)

(jump_insn 3465 3463 3466 91 (set (pc)
        (label_ref 3464)) -1 (nil)
    (nil))
;; End of basic block 91, registers live:
 (nil)

(barrier 3466 3465 3250)

;; Start of basic block 92, registers live: (nil)
(code_label/s 3250 3466 3253 92 352 "" [1 uses])

(note 3253 3250 3251 92 [bb 92] NOTE_INSN_BASIC_BLOCK)

(insn 3251 3253 3252 92 (set (reg:SI 374 [ save_eptr.1036 ])
        (reg:SI 4 r4)) 123 {movsi_ie} (nil)
    (nil))

(insn 3252 3251 3097 92 (set (reg:SI 373 [ save_filt.1037 ])
        (reg:SI 5 r5)) 123 {movsi_ie} (nil)
    (nil))
;; End of basic block 92, registers live:
 (nil)

;; Start of basic block 93, registers live: (nil)
(code_label 3097 3252 3102 93 327 "" [16 uses])

(note 3102 3097 3098 93 [bb 93] NOTE_INSN_BASIC_BLOCK)

(insn 3098 3102 3099 93 (set (reg:SI 147 t)
        (eq:SI (reg:SI 373 [ save_filt.1037 ])
            (const_int -1 [0xffffffff]))) 1 {cmpeqsi_t} (nil)
    (nil))

(jump_insn 3099 3098 3254 93 (set (pc)
        (if_then_else (ne (reg:SI 147 t)
                (const_int 0 [0x0]))
            (label_ref 2987)
            (pc))) 153 {branch_true} (nil)
    (expr_list:REG_BR_PROB (const_int 2901 [0xb55])
        (nil)))
;; End of basic block 93, registers live:
 (nil)

;; Start of basic block 94, registers live: (nil)
(note 3254 3099 3151 94 [bb 94] NOTE_INSN_BASIC_BLOCK)

(insn 3151 3254 3152 94 (set (reg:SI 4 r4 [ save_eptr.1036 ])
        (reg:SI 374 [ save_eptr.1036 ])) 123 {movsi_ie} (nil)
    (nil))

(call_insn 3152 3151 3101 94 (parallel [
            (call (mem:SI (symbol_ref:SI ("_Unwind_Resume") [flags 0x41]) [0 S4 A32])
                (const_int 0 [0x0]))
            (use (reg:PSI 151 ))
            (use (reg:SI 12 r12))
            (clobber (reg:SI 146 pr))
            (clobber (scratch:SI))
        ]) 166 {call_pcrel} (nil)
    (expr_list:REG_NORETURN (const_int 0 [0x0])
        (nil))
    (expr_list (use (reg:SI 4 r4 [ save_eptr.1036 ]))
        (nil)))
;; End of basic block 94, registers live:
 (nil)

(barrier 3101 3152 2987)

;; Start of basic block 95, registers live: (nil)
(code_label/s 2987 3101 3095 95 326 "" [2 uses])

(note 3095 2987 2990 95 [bb 95] NOTE_INSN_BASIC_BLOCK)

(insn 2990 3095 2991 95 (set (reg:SI 4 r4 [ save_eptr.1036 ])
        (reg:SI 374 [ save_eptr.1036 ])) 123 {movsi_ie} (nil)
    (nil))

(call_insn 2991 2990 2992 95 (parallel [
            (call (mem:SI (symbol_ref:SI ("__cxa_call_unexpected") [flags 0x41] <function_decl 0x402ac740 __cxa_call_unexpected>) [0 S4 A32])
                (const_int 0 [0x0]))
            (use (reg:PSI 151 ))
            (use (reg:SI 12 r12))
            (clobber (reg:SI 146 pr))
            (clobber (scratch:SI))
        ]) 166 {call_pcrel} (nil)
    (expr_list:REG_NORETURN (const_int 0 [0x0])
        (nil))
    (expr_list (use (reg:SI 4 r4 [ save_eptr.1036 ]))
        (nil)))
;; End of basic block 95, registers live:
 (nil)

(barrier 2992 2991 3464)

;; Start of basic block 96, registers live: (nil)
(code_label 3464 2992 3094 96 373 "" [1 uses])

(note 3094 3464 2976 96 [bb 96] NOTE_INSN_BASIC_BLOCK)

(insn 2976 3094 2978 96 (set (reg:SI 182 [ T.678 ])
        (reg:SI 0 r0)) 123 {movsi_ie} (nil)
    (nil))

This in itself doesn't cause a reload problem, but it means that
we have to look harder to find out if we got a post-call group.

I must admit that I'm at a bit of a loss what the point of the
reordering is.  Unfortunately, the new cfg handling makes it harder
to track down changes to the code that cause them.

#0  make_jump_insn_raw (pattern=0x401d8340) at ../../srcw/gcc/emit-rtl.c:3439
#1  0x08365a7b in emit_jump_insn (x=0x401d8340)
    at ../../srcw/gcc/emit-rtl.c:4548
#2  0x08464794 in gen_jump (operand0=0x40d0191c) at insn-emit.c:9990
#3  0x08309be0 in force_nonfallthru_and_redirect (e=0x40ca3ef4, 
    target=0x406ffef4) at ../../srcw/gcc/cfgrtl.c:1142
#4  0x08309c63 in force_nonfallthru (e=0x40ca3ef4)
    at ../../srcw/gcc/cfgrtl.c:1163
#5  0x08300637 in fixup_reorder_chain () at ../../srcw/gcc/cfglayout.c:778
#6  0x083012ca in cfg_layout_finalize () at ../../srcw/gcc/cfglayout.c:1180
#7  0x085c6859 in rest_of_handle_loop2 (decl=0x4080fb54, insns=0x4077f168)
    at ../../srcw/gcc/passes.c:1337

At any rate, this is how this can be fixed in sched-deps too:

2004-05-28  J"orn Rennecke <joern.rennecke@superh.com>

	* sched-int.h (in_post_call_group_p): Change type to enum.
	* sched-deps.c (sched_analyze_insn):
	(sched_analyze): When in_post_call_group_p is post_call_initial,
	don't add a dependency, but still set SCHED_GROUP_P and CANT_MOVE,
	and also reset in_post_call_group_p to post_call.
	(sched_analyze): When the previous basic block ended in a CALL_INSN,
	initialize in_post_call_group_p as post_call_initial.
	(init_deps): initialize in_post_call_group_p to not_post_call.
	
Index: sched-deps.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/sched-deps.c,v
retrieving revision 1.72
diff -p -r1.72 sched-deps.c
*** sched-deps.c	25 Mar 2004 00:58:57 -0000	1.72
--- sched-deps.c	28 May 2004 17:45:49 -0000
*************** sched_analyze_insn (struct deps *deps, r
*** 1145,1157 ****
        if (src_regno < FIRST_PSEUDO_REGISTER
  	  || dest_regno < FIRST_PSEUDO_REGISTER)
  	{
! 	  set_sched_group_p (insn);
  	  CANT_MOVE (insn) = 1;
  	}
        else
  	{
  	end_call_group:
! 	  deps->in_post_call_group_p = false;
  	}
      }
  }
--- 1145,1165 ----
        if (src_regno < FIRST_PSEUDO_REGISTER
  	  || dest_regno < FIRST_PSEUDO_REGISTER)
  	{
! 	  /* If we are inside a post-call group right at the start of the
! 	     scheduling region, we must not add a dependency.  */
! 	  if (deps->in_post_call_group_p == post_call_initial)
! 	    {
! 	      SCHED_GROUP_P (insn) = 1;
! 	      deps->in_post_call_group_p = post_call;
! 	    }
! 	  else
! 	    set_sched_group_p (insn);
  	  CANT_MOVE (insn) = 1;
  	}
        else
  	{
  	end_call_group:
! 	  deps->in_post_call_group_p = not_post_call;
  	}
      }
  }
*************** sched_analyze (struct deps *deps, rtx he
*** 1168,1173 ****
--- 1176,1213 ----
    if (current_sched_info->use_cselib)
      cselib_init (true);
  
+   /* Before reload, if the previous block ended in a call, show that
+      we are inside a post-call group, so as to keep the lifetimes of
+      hard registers correct.  */
+   if (! reload_completed)
+     {
+       insn = prev_nonnote_insn (head);
+       if (insn && GET_CODE (insn) != CODE_LABEL)
+ 	{
+ 	  if (insn && GET_CODE (insn) == CALL_INSN)
+ 	    deps->in_post_call_group_p = post_call_initial;
+ 	}
+       else if (insn)
+ 	{
+ 	  /* The new loop optimizer can insert jumps in strange places -
+              between a call and the insn that copies the return value.
+ 	     This is in itself harmless, but makes it harder here to
+ 	     recognize post-call groups.  */
+ 	  edge e;
+ 	  basic_block pred_bb;
+ 
+ 	  for (e = BLOCK_FOR_INSN (head)->pred; e; e = e->pred_next)
+ 	    {
+ 	      pred_bb = e->src;
+ 	      insn = BB_END(pred_bb);
+ 	      if (GET_CODE (insn) != JUMP_INSN)
+ 		continue;
+ 	      insn = prev_nonnote_insn (insn);
+ 	      if (insn && GET_CODE (insn) == CALL_INSN)
+ 		deps->in_post_call_group_p = post_call_initial;
+ 	    }
+ 	}
+     }
    for (insn = head;; insn = NEXT_INSN (insn))
      {
        rtx link, end_seq, r0, set;
*************** sched_analyze (struct deps *deps, rtx he
*** 1259,1265 ****
  	  /* Before reload, begin a post-call group, so as to keep the
  	     lifetimes of hard registers correct.  */
  	  if (! reload_completed)
! 	    deps->in_post_call_group_p = true;
  	}
  
        /* See comments on reemit_notes as to why we do this.
--- 1299,1305 ----
  	  /* Before reload, begin a post-call group, so as to keep the
  	     lifetimes of hard registers correct.  */
  	  if (! reload_completed)
! 	    deps->in_post_call_group_p = post_call;
  	}
  
        /* See comments on reemit_notes as to why we do this.
*************** init_deps (struct deps *deps)
*** 1420,1426 ****
    deps->last_pending_memory_flush = 0;
    deps->last_function_call = 0;
    deps->sched_before_next_call = 0;
!   deps->in_post_call_group_p = false;
    deps->libcall_block_tail_insn = 0;
  }
  
--- 1460,1466 ----
    deps->last_pending_memory_flush = 0;
    deps->last_function_call = 0;
    deps->sched_before_next_call = 0;
!   deps->in_post_call_group_p = not_post_call;
    deps->libcall_block_tail_insn = 0;
  }
  

Re: [PATCH/RFA] PR target/13250

[Kaz Kojima]

Joern Rennecke <joern.rennecke@superh.com> wrote:
> It turns out that the loop optimizer has inserted a jump between
> the call and the instruction that copied the return value.
[snip]
> At any rate, this is how this can be fixed in sched-deps too:

All libraries are successfully built on mainline with your new
patch for sh4-unknown-linux-gnu!
I saw compiler segfaults in patch for some C tests because
BB_END returns a null for those cases.  Changing a line like

+ 	      if (! insn || GET_CODE (insn) != JUMP_INSN)
+ 		continue;

seems work for me and cross building and C regression test are
ok with it.  The result of C++ test is same as one without patch
but with libstdc++.so built with -fno-schedule-insns.
I've tested it with i686-pc-linux-gnu bootstrap and regtest.
Now full bootstrap on sh4-unknown-linux-gnu goes on.

Regards,
	kaz

Re: [PATCH/RFA] PR target/13250

[Joern Rennecke]

> 
> Joern Rennecke <joern.rennecke@superh.com> wrote:
> > It turns out that the loop optimizer has inserted a jump between
> > the call and the instruction that copied the return value.
> [snip]
> > At any rate, this is how this can be fixed in sched-deps too:
> 
> All libraries are successfully built on mainline with your new
> patch for sh4-unknown-linux-gnu!
> I saw compiler segfaults in patch for some C tests because
> BB_END returns a null for those cases.  Changing a line like
> 
> + 	      if (! insn || GET_CODE (insn) != JUMP_INSN)
> + 		continue;

I don't think that should ever happen.  Even when we have a syntax
error code is either emitted or it isn't.  The basic block structures
are created for the instructions that are actually present.
Which testcases demonstrate this problem?

Re: [PATCH/RFA] PR target/13250

[Kaz Kojima]

Joern Rennecke <joern.rennecke@superh.com> wrote:
>> I saw compiler segfaults in patch for some C tests because
>> BB_END returns a null for those cases.  Changing a line like
>> 
>> + 	      if (! insn || GET_CODE (insn) != JUMP_INSN)
>> + 		continue;
> 
> I don't think that should ever happen.  Even when we have a syntax
> error code is either emitted or it isn't.  The basic block structures
> are created for the instructions that are actually present.
> Which testcases demonstrate this problem?

The first several testcases are

gcc.c-torture/compile/20000728-1.c  (-O2)
gcc.c-torture/compile/20011229-2.c  (-Os)
gcc.c-torture/compile/20031113-1.c  (-O3)
gcc.c-torture/compile/920623-1.c    (-O2)
gcc.c-torture/compile/930210-1.c    (-O2)

Regards,
	kaz

Re: [PATCH/RFA] PR target/13250

[Joern Rennecke]

> gcc.c-torture/compile/20000728-1.c  (-O2)

It turned out that this was ENTRY_BLOCK_PTR.  So that's all right then
that there are no insns.  This testcase works when I add:

              if (pred_bb == ENTRY_BLOCK_PTR)
                continue;

Re: [PATCH/RFA] PR target/13250

[Kaz Kojima]

Joern Rennecke <joern.rennecke@superh.com> wrote:
>> gcc.c-torture/compile/20000728-1.c  (-O2)
> 
> It turned out that this was ENTRY_BLOCK_PTR.  So that's all right then
> that there are no insns.  This testcase works when I add:
> 
>               if (pred_bb == ENTRY_BLOCK_PTR)
>                 continue;

It fixes all similar failures in make check-gcc on x86 cross to
sh4-unknown-linux-gnu.  There are no problems in building c++
and java libraries too.  I'm trying to bootstrap on native
i686-linux and sh4-linux with this patch.

Regards,
	kaz
 

building sh-elf / sh-linux (Was: Re: [PATCH/RFA] PR target/13250)

[Joern Rennecke]

I'm not really happy with having the scheduler follow through jumps to
find call / call-value-copy pairs.  I have now implemented
the approach of deleting fixup_fallthru_exit_predecessor and making
mid-function epilogues valid.
There was also a real.c bug that got in the way.  I'm appending the patch
set that I am currently testing; I'll submit the two new patches
individually when they are tested.

2004-06-04  J"orn Rennecke <joern.rennecke@superh.com>

	* basic-block.h (could_fall_through): Declare.
	* cfganal.c (can_fallthru): Suceed if the target is EXIT_BLOCK_PTR.
	Fail if the source already has a fallthrough edge to the exit
	block pointer.
	(could_fall_through): New function.
	(make_edges): Check if we already have a fallthrough edge to the
	exit block pointer.
	cfglayout.c (fixup_fallthru_exit_predecessor): Delete.
	(cfg_layout_finalize): Don't call it.
	(fixup_reorder_chain): A fall through to the exit block does not
	require the block to come last.  Add sanity checks.
	* cfgrtl.c (rtl_split_edge): Add special handling of fall through
	edges to the exit block.

2004-06-03  J"orn Rennecke <joern.rennecke@superh.com>

	* real.c (do_add): Initialize r->signalling and r->canonical.

2004-05-14  J"orn Rennecke <joern.rennecke@superh.com>

	* sched-int.h (in_post_call_group_p): Change type to enum.
	* sched-deps.c (sched_analyze_insn):
	(sched_analyze): When in_post_call_group_p is post_call_initial,
	don't add a dependency, but still set SCHED_GROUP_P and CANT_MOVE,
	and also reset in_post_call_group_p to post_call.
	(sched_analyze): When the previous basic block ended in a CALL_INSN,
	initialize in_post_call_group_p as post_call_initial.
	(init_deps): initialize in_post_call_group_p to not_post_call.

Index: basic-block.h
===================================================================
RCS file: /cvs/gcc/gcc/gcc/basic-block.h,v
retrieving revision 1.196
diff -p -r1.196 basic-block.h
*** basic-block.h	15 May 2004 09:39:28 -0000	1.196
--- basic-block.h	4 Jun 2004 14:13:14 -0000
*************** extern void find_sub_basic_blocks (basic
*** 628,633 ****
--- 628,634 ----
  extern void find_many_sub_basic_blocks (sbitmap);
  extern void rtl_make_eh_edge (sbitmap *, basic_block, rtx);
  extern bool can_fallthru (basic_block, basic_block);
+ extern bool could_fall_through (basic_block, basic_block);
  extern void flow_nodes_print (const char *, const sbitmap, FILE *);
  extern void flow_edge_list_print (const char *, const edge *, int, FILE *);
  extern void alloc_aux_for_block (basic_block, int);
Index: cfganal.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/cfganal.c,v
retrieving revision 1.42
diff -p -r1.42 cfganal.c
*** cfganal.c	13 May 2004 06:39:32 -0000	1.42
--- cfganal.c	4 Jun 2004 14:13:14 -0000
*************** bool
*** 103,119 ****
  can_fallthru (basic_block src, basic_block target)
  {
    rtx insn = BB_END (src);
!   rtx insn2 = target == EXIT_BLOCK_PTR ? NULL : BB_HEAD (target);
  
    if (src->next_bb != target)
      return 0;
  
    if (insn2 && !active_insn_p (insn2))
      insn2 = next_active_insn (insn2);
  
    /* ??? Later we may add code to move jump tables offline.  */
    return next_active_insn (insn) == insn2;
  }
  \f
  /* Mark the back edges in DFS traversal.
     Return nonzero if a loop (natural or otherwise) is present.
--- 103,144 ----
  can_fallthru (basic_block src, basic_block target)
  {
    rtx insn = BB_END (src);
!   rtx insn2;
!   edge e;
  
+   if (target == EXIT_BLOCK_PTR)
+     return true;
    if (src->next_bb != target)
      return 0;
+   for (e = src->succ; e; e = e->succ_next)
+     if (e->dest == EXIT_BLOCK_PTR
+ 	&& e->flags & EDGE_FALLTHRU)
+     return 0;
  
+   insn2 = BB_HEAD (target);
    if (insn2 && !active_insn_p (insn2))
      insn2 = next_active_insn (insn2);
  
    /* ??? Later we may add code to move jump tables offline.  */
    return next_active_insn (insn) == insn2;
  }
+ 
+ /* Return nonzero if we could reach target from src by falling through,
+    if the target was made adjacent.  If we already have a fall-through
+    edge to the exit block, we can't do that.  */
+ bool
+ could_fall_through (basic_block src, basic_block target)
+ {
+   edge e;
+ 
+   if (target == EXIT_BLOCK_PTR)
+     return true;
+   for (e = src->succ; e; e = e->succ_next)
+     if (e->dest == EXIT_BLOCK_PTR
+ 	&& e->flags & EDGE_FALLTHRU)
+     return 0;
+   return true;
+ }
  \f
  /* Mark the back edges in DFS traversal.
     Return nonzero if a loop (natural or otherwise) is present.
Index: cfgbuild.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/cfgbuild.c,v
retrieving revision 1.46
diff -p -r1.46 cfgbuild.c
*** cfgbuild.c	27 May 2004 12:56:30 -0000	1.46
--- cfgbuild.c	4 Jun 2004 14:13:14 -0000
*************** make_edges (rtx label_value_list, basic_
*** 313,318 ****
--- 313,319 ----
        rtx insn, x;
        enum rtx_code code;
        int force_fallthru = 0;
+       edge e;
  
        if (GET_CODE (BB_HEAD (bb)) == CODE_LABEL
  	  && LABEL_ALT_ENTRY_P (BB_HEAD (bb)))
*************** make_edges (rtx label_value_list, basic_
*** 435,440 ****
--- 436,447 ----
  
        /* Find out if we can drop through to the next block.  */
        insn = NEXT_INSN (insn);
+       for (e = bb->succ; e; e = e->succ_next)
+ 	if (e->dest == EXIT_BLOCK_PTR && e->flags & EDGE_FALLTHRU)
+ 	  {
+ 	    insn = 0;
+ 	    break;
+ 	  }
        while (insn
  	     && GET_CODE (insn) == NOTE
  	     && NOTE_LINE_NUMBER (insn) != NOTE_INSN_BASIC_BLOCK)
Index: cfglayout.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/cfglayout.c,v
retrieving revision 1.59
diff -p -r1.59 cfglayout.c
*** cfglayout.c	25 May 2004 12:54:53 -0000	1.59
--- cfglayout.c	4 Jun 2004 14:13:14 -0000
*************** static void set_block_levels (tree, int)
*** 53,59 ****
  static void change_scope (rtx, tree, tree);
  
  void verify_insn_chain (void);
- static void fixup_fallthru_exit_predecessor (void);
  static tree insn_scope (rtx);
  static void update_unlikely_executed_notes (basic_block);
  \f
--- 53,58 ----
*************** fixup_reorder_chain (void)
*** 714,719 ****
--- 713,722 ----
  		      && invert_jump (bb_end_insn,
  				      label_for_bb (e_fall->dest), 0))
  		    {
+ #ifdef ENABLE_CHECKING
+ 		      if (!could_fall_through (e_taken->src, e_taken->dest))
+ 			abort ();
+ #endif
  		      e_fall->flags &= ~EDGE_FALLTHRU;
  		      e_taken->flags |= EDGE_FALLTHRU;
  		      update_br_prob_note (bb);
*************** fixup_reorder_chain (void)
*** 731,736 ****
--- 734,743 ----
  	      else if (invert_jump (bb_end_insn,
  				    label_for_bb (e_fall->dest), 0))
  		{
+ #ifdef ENABLE_CHECKING
+ 		  if (!could_fall_through (e_taken->src, e_taken->dest))
+ 		    abort ();
+ #endif
  		  e_fall->flags &= ~EDGE_FALLTHRU;
  		  e_taken->flags |= EDGE_FALLTHRU;
  		  update_br_prob_note (bb);
*************** fixup_reorder_chain (void)
*** 770,776 ****
  	    continue;
  
  	  /* A fallthru to exit block.  */
! 	  if (!bb->rbi->next && e_fall->dest == EXIT_BLOCK_PTR)
  	    continue;
  	}
  
--- 777,783 ----
  	    continue;
  
  	  /* A fallthru to exit block.  */
! 	  if (e_fall->dest == EXIT_BLOCK_PTR)
  	    continue;
  	}
  
*************** verify_insn_chain (void)
*** 910,943 ****
      abort ();
  }
  \f
- /* The block falling through to exit must be the last one in the
-    reordered chain.  Ensure that this condition is met.  */
- static void
- fixup_fallthru_exit_predecessor (void)
- {
-   edge e;
-   basic_block bb = NULL;
- 
-   for (e = EXIT_BLOCK_PTR->pred; e; e = e->pred_next)
-     if (e->flags & EDGE_FALLTHRU)
-       bb = e->src;
- 
-   if (bb && bb->rbi->next)
-     {
-       basic_block c = ENTRY_BLOCK_PTR->next_bb;
- 
-       while (c->rbi->next != bb)
- 	c = c->rbi->next;
- 
-       c->rbi->next = bb->rbi->next;
-       while (c->rbi->next)
- 	c = c->rbi->next;
- 
-       c->rbi->next = bb;
-       bb->rbi->next = NULL;
-     }
- }
- \f
  /* Return true in case it is possible to duplicate the basic block BB.  */
  
  /* We do not want to declare the function in a header file, since it should
--- 917,922 ----
*************** cfg_layout_finalize (void)
*** 1176,1182 ****
    verify_flow_info ();
  #endif
    rtl_register_cfg_hooks ();
-   fixup_fallthru_exit_predecessor ();
    fixup_reorder_chain ();
  
  #ifdef ENABLE_CHECKING
--- 1155,1160 ----
Index: cfgrtl.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/cfgrtl.c,v
retrieving revision 1.118
diff -p -r1.118 cfgrtl.c
*** cfgrtl.c	13 May 2004 06:39:32 -0000	1.118
--- cfgrtl.c	4 Jun 2004 14:13:14 -0000
*************** rtl_split_edge (edge edge_in)
*** 1337,1343 ****
    else
      before = NULL_RTX;
  
!   bb = create_basic_block (before, NULL, edge_in->dest->prev_bb);
  
    /* ??? This info is likely going to be out of date very soon.  */
    if (edge_in->dest->global_live_at_start)
--- 1337,1355 ----
    else
      before = NULL_RTX;
  
!   /* If this is a fall through edge to the exit block, the blocks might be
!      not adjacent, and the right place is the after the source.  */
!   if (edge_in->flags & EDGE_FALLTHRU && edge_in->dest == EXIT_BLOCK_PTR)
!     {
!       before = NEXT_INSN (BB_END (edge_in->src));
!       if (before
! 	  && GET_CODE (before) == NOTE
! 	  && NOTE_LINE_NUMBER (before) == NOTE_INSN_LOOP_END)
! 	before = NEXT_INSN (before);
!       bb = create_basic_block (before, NULL, edge_in->src);
!     }
!   else
!     bb = create_basic_block (before, NULL, edge_in->dest->prev_bb);
  
    /* ??? This info is likely going to be out of date very soon.  */
    if (edge_in->dest->global_live_at_start)
Index: real.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/real.c,v
retrieving revision 1.141
diff -p -r1.141 real.c
*** real.c	22 Mar 2004 00:40:44 -0000	1.141
--- real.c	4 Jun 2004 14:13:15 -0000
*************** do_add (REAL_VALUE_TYPE *r, const REAL_V
*** 580,585 ****
--- 580,591 ----
        abort ();
      }
  
+   /* Make sure all fields in the result are initialized.
+      This is needed so that tree hashing produces consistent results.
+      We can't use get_zero here because A or B might be equal to R.  */
+   r->signalling = 0;
+   r->canonical = 0;
+ 
    /* Swap the arguments such that A has the larger exponent.  */
    dexp = REAL_EXP (a) - REAL_EXP (b);
    if (dexp < 0)
Index: sched-deps.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/sched-deps.c,v
retrieving revision 1.72
diff -p -r1.72 sched-deps.c
*** sched-deps.c	25 Mar 2004 00:58:57 -0000	1.72
--- sched-deps.c	4 Jun 2004 14:13:15 -0000
*************** sched_analyze_insn (struct deps *deps, r
*** 1145,1157 ****
        if (src_regno < FIRST_PSEUDO_REGISTER
  	  || dest_regno < FIRST_PSEUDO_REGISTER)
  	{
! 	  set_sched_group_p (insn);
  	  CANT_MOVE (insn) = 1;
  	}
        else
  	{
  	end_call_group:
! 	  deps->in_post_call_group_p = false;
  	}
      }
  }
--- 1145,1165 ----
        if (src_regno < FIRST_PSEUDO_REGISTER
  	  || dest_regno < FIRST_PSEUDO_REGISTER)
  	{
! 	  /* If we are inside a post-call group right at the start of the
! 	     scheduling region, we must not add a dependency.  */
! 	  if (deps->in_post_call_group_p == post_call_initial)
! 	    {
! 	      SCHED_GROUP_P (insn) = 1;
! 	      deps->in_post_call_group_p = post_call;
! 	    }
! 	  else
! 	    set_sched_group_p (insn);
  	  CANT_MOVE (insn) = 1;
  	}
        else
  	{
  	end_call_group:
! 	  deps->in_post_call_group_p = not_post_call;
  	}
      }
  }
*************** sched_analyze (struct deps *deps, rtx he
*** 1168,1173 ****
--- 1176,1190 ----
    if (current_sched_info->use_cselib)
      cselib_init (true);
  
+   /* Before reload, if the previous block ended in a call, show that
+      we are inside a post-call group, so as to keep the lifetimes of
+      hard registers correct.  */
+   if (! reload_completed && GET_CODE (head) != CODE_LABEL)
+     {
+       insn = prev_nonnote_insn (head);
+       if (insn && GET_CODE (insn) == CALL_INSN)
+ 	deps->in_post_call_group_p = post_call_initial;
+     }
    for (insn = head;; insn = NEXT_INSN (insn))
      {
        rtx link, end_seq, r0, set;
*************** sched_analyze (struct deps *deps, rtx he
*** 1259,1265 ****
  	  /* Before reload, begin a post-call group, so as to keep the
  	     lifetimes of hard registers correct.  */
  	  if (! reload_completed)
! 	    deps->in_post_call_group_p = true;
  	}
  
        /* See comments on reemit_notes as to why we do this.
--- 1276,1282 ----
  	  /* Before reload, begin a post-call group, so as to keep the
  	     lifetimes of hard registers correct.  */
  	  if (! reload_completed)
! 	    deps->in_post_call_group_p = post_call;
  	}
  
        /* See comments on reemit_notes as to why we do this.
*************** init_deps (struct deps *deps)
*** 1420,1426 ****
    deps->last_pending_memory_flush = 0;
    deps->last_function_call = 0;
    deps->sched_before_next_call = 0;
!   deps->in_post_call_group_p = false;
    deps->libcall_block_tail_insn = 0;
  }
  
--- 1437,1443 ----
    deps->last_pending_memory_flush = 0;
    deps->last_function_call = 0;
    deps->sched_before_next_call = 0;
!   deps->in_post_call_group_p = not_post_call;
    deps->libcall_block_tail_insn = 0;
  }
  
Index: sched-int.h
===================================================================
RCS file: /cvs/gcc/gcc/gcc/sched-int.h,v
retrieving revision 1.32
diff -p -r1.32 sched-int.h
*** sched-int.h	23 Feb 2004 17:02:50 -0000	1.32
--- sched-int.h	4 Jun 2004 14:13:15 -0000
*************** struct deps
*** 84,90 ****
  
    /* Used to keep post-call pseudo/hard reg movements together with
       the call.  */
!   bool in_post_call_group_p;
  
    /* Set to the tail insn of the outermost libcall block.
  
--- 84,90 ----
  
    /* Used to keep post-call pseudo/hard reg movements together with
       the call.  */
!   enum { not_post_call, post_call, post_call_initial } in_post_call_group_p;
  
    /* Set to the tail insn of the outermost libcall block.
  

Re: building sh-elf / sh-linux (Was: Re: [PATCH/RFA] PR target/13250)

[Joern Rennecke]

> 2004-06-04  J"orn Rennecke <joern.rennecke@superh.com>
> 
> 	* basic-block.h (could_fall_through): Declare.
> 	* cfganal.c (can_fallthru): Suceed if the target is EXIT_BLOCK_PTR.
> 	Fail if the source already has a fallthrough edge to the exit
> 	block pointer.
> 	(could_fall_through): New function.
> 	(make_edges): Check if we already have a fallthrough edge to the
> 	exit block pointer.
> 	cfglayout.c (fixup_fallthru_exit_predecessor): Delete.
> 	(cfg_layout_finalize): Don't call it.
> 	(fixup_reorder_chain): A fall through to the exit block does not
> 	require the block to come last.  Add sanity checks.
> 	* cfgrtl.c (rtl_split_edge): Add special handling of fall through
> 	edges to the exit block.
> 
> 2004-06-03  J"orn Rennecke <joern.rennecke@superh.com>
> 
> 	* real.c (do_add): Initialize r->signalling and r->canonical.
> 
> 2004-05-14  J"orn Rennecke <joern.rennecke@superh.com>
> 
> 	* sched-int.h (in_post_call_group_p): Change type to enum.
> 	* sched-deps.c (sched_analyze_insn):
> 	(sched_analyze): When in_post_call_group_p is post_call_initial,
> 	don't add a dependency, but still set SCHED_GROUP_P and CANT_MOVE,
> 	and also reset in_post_call_group_p to post_call.
> 	(sched_analyze): When the previous basic block ended in a CALL_INSN,
> 	initialize in_post_call_group_p as post_call_initial.
> 	(init_deps): initialize in_post_call_group_p to not_post_call.

Sorry about the noise I created when trying to bootstrap with a Red Hat
gcc 2.96 compiler.

Using a gcc 3.0.2 bootstrap compiler, I could bootstrap both original and
patched sources, and the set of regressions is also identical for both.

Re: building sh-elf / sh-linux (Was: Re: [PATCH/RFA] PR target/13250)

[Joern Rennecke]

> 
> I'm not really happy with having the scheduler follow through jumps to
> find call / call-value-copy pairs.  I have now implemented
> the approach of deleting fixup_fallthru_exit_predecessor and making
> mid-function epilogues valid.
> There was also a real.c bug that got in the way.  I'm appending the patch
> set that I am currently testing; I'll submit the two new patches
> individually when they are tested.
> 
> 2004-06-04  J"orn Rennecke <joern.rennecke@superh.com>
> 
> 	* basic-block.h (could_fall_through): Declare.
> 	* cfganal.c (can_fallthru): Suceed if the target is EXIT_BLOCK_PTR.
> 	Fail if the source already has a fallthrough edge to the exit
> 	block pointer.
> 	(could_fall_through): New function.
> 	(make_edges): Check if we already have a fallthrough edge to the
> 	exit block pointer.
> 	cfglayout.c (fixup_fallthru_exit_predecessor): Delete.
> 	(cfg_layout_finalize): Don't call it.
> 	(fixup_reorder_chain): A fall through to the exit block does not
> 	require the block to come last.  Add sanity checks.
> 	* cfgrtl.c (rtl_split_edge): Add special handling of fall through
> 	edges to the exit block.
> 
> 2004-06-03  J"orn Rennecke <joern.rennecke@superh.com>
> 
> 	* real.c (do_add): Initialize r->signalling and r->canonical.
> 
> 2004-05-14  J"orn Rennecke <joern.rennecke@superh.com>
> 
> 	* sched-int.h (in_post_call_group_p): Change type to enum.
> 	* sched-deps.c (sched_analyze_insn):
> 	(sched_analyze): When in_post_call_group_p is post_call_initial,
> 	don't add a dependency, but still set SCHED_GROUP_P and CANT_MOVE,
> 	and also reset in_post_call_group_p to post_call.
> 	(sched_analyze): When the previous basic block ended in a CALL_INSN,
I've chosen a i686-pc-linux-gnu host for a bootstrap test.
For the unmodified control, I got a failure during stage1 building
crtbegin.o - cc1 throws a floating point exception. The patched
compiler has moved on to stage2 in the meantime.

Likewise, sh-elf (using sources from the 26th May as baseline) has suceeded
building and the sh1 big endian results are:

                === gcc Summary for sh-hms-sim ===

# of expected passes            25703
# of unexpected failures        29
# of unexpected successes       1
# of expected failures          68
# of unresolved testcases       5
# of untested testcases         28
# of unsupported tests          368

Considering that the controls don't build at all, I suppose I can say the
regression test passed.

Re: building sh-elf / sh-linux (Was: Re: [PATCH/RFA] PR target/13250)

[Joern Rennecke]

> I've chosen a i686-pc-linux-gnu host for a bootstrap test.
> For the unmodified control, I got a failure during stage1 building
> crtbegin.o - cc1 throws a floating point exception. The patched
> compiler has moved on to stage2 in the meantime.

Hmm, it's actually a stage2 failure, and it is the same with
and without patches.

Program received signal SIGFPE, Arithmetic exception.
0x08421ace in narrow_str_to_charconst (pfile=0x8590518, str=
      {len = 2, text = 0x85bcc78 "\n"}, pchars_seen=0xbfffe650, 
    unsignedp=0xbfffe654) at ../../srcw/libcpp/charset.c:1214
1214      size_t max_chars = CPP_OPTION (pfile, int_precision) / width;
(gdb) p width
$1 = 0
(gdb) l
1209    static cppchar_t
1210    narrow_str_to_charconst (cpp_reader *pfile, cpp_string str,
1211                             unsigned int *pchars_seen, int *unsignedp)
1212    {
1213      size_t width = CPP_OPTION (pfile, char_precision);
1214      size_t max_chars = CPP_OPTION (pfile, int_precision) / width;