gcc failed to bootstrap on Linux/mipsel

gcc failed to bootstrap on Linux/mipsel

[H . J . Lu]

On Linux/mipsel, as of Sun Jan 27 11:10:21 PST 2002, I got

./xgcc -B./ -B/usr/gcc-3.1/mipsel-pc-linux-gnu/bin/ -isystem
/usr/gcc-3.1/mipsel-pc-linux-gnu/include -isystem
/usr/gcc-3.1/mipsel-pc-linux-gnu/sys-include -dumpspecs > tmp-specs
make[4]: *** [specs] Error 139
make[4]: Leaving directory `/export/build/gnu/gcc-mips/build-mipsel-linux/gcc'
make[3]: *** [stage2_build] Error 2
make[3]: Leaving directory `/export/build/gnu/gcc-mips/build-mipsel-linux/gcc'
make[2]: *** [bootstrap] Error 2

The last good one for me is Jan 24 16:56 PST 2002. I am tracking down
which patch breaks Linux/mipsel now. Irix6.5/mips seems ok. It may be
another ABI/ednian bug.


H.J.

Re: gcc failed to bootstrap on Linux/mipsel

[H . J . Lu]

On Sun, Jan 27, 2002 at 05:34:41PM -0800, H . J . Lu wrote:
> On Linux/mipsel, as of Sun Jan 27 11:10:21 PST 2002, I got
> 
> ./xgcc -B./ -B/usr/gcc-3.1/mipsel-pc-linux-gnu/bin/ -isystem
> /usr/gcc-3.1/mipsel-pc-linux-gnu/include -isystem
> /usr/gcc-3.1/mipsel-pc-linux-gnu/sys-include -dumpspecs > tmp-specs
> make[4]: *** [specs] Error 139
> make[4]: Leaving directory `/export/build/gnu/gcc-mips/build-mipsel-linux/gcc'
> make[3]: *** [stage2_build] Error 2
> make[3]: Leaving directory `/export/build/gnu/gcc-mips/build-mipsel-linux/gcc'
> make[2]: *** [bootstrap] Error 2
> 
> The last good one for me is Jan 24 16:56 PST 2002. I am tracking down
> which patch breaks Linux/mipsel now. Irix6.5/mips seems ok. It may be
> another ABI/ednian bug.

I got:

GOOD: Jan 26 20:30 PST 2002
BAD: Jan 27 10:00 PST 2002

The only relevant change is 

http://gcc.gnu.org/ml/gcc-patches/2002-01/msg01929.html

Richard, could you please take a look at it?

Thanks.


H.J.

Re: gcc failed to bootstrap on Linux/mipsel

[Richard Henderson]

On Mon, Jan 28, 2002 at 12:39:00PM -0800, H . J . Lu wrote:
> I got:
> 
> GOOD: Jan 26 20:30 PST 2002
> BAD: Jan 27 10:00 PST 2002
> 
> The only relevant change is 
> 
> http://gcc.gnu.org/ml/gcc-patches/2002-01/msg01929.html
> 
> Richard, could you please take a look at it?

As usual:

Are you give me any more information than this?  Like finding
the miscompiled file or function?  Otherwise I can't help you.


r~

Re: gcc failed to bootstrap on Linux/mipsel

[H . J . Lu]

On Mon, Jan 28, 2002 at 12:52:50PM -0800, Richard Henderson wrote:
> On Mon, Jan 28, 2002 at 12:39:00PM -0800, H . J . Lu wrote:
> > I got:
> > 
> > GOOD: Jan 26 20:30 PST 2002
> > BAD: Jan 27 10:00 PST 2002
> > 
> > The only relevant change is 
> > 
> > http://gcc.gnu.org/ml/gcc-patches/2002-01/msg01929.html
> > 
> > Richard, could you please take a look at it?
> 
> As usual:
> 
> Are you give me any more information than this?  Like finding
> the miscompiled file or function?  Otherwise I can't help you.
 
I only have a very slow mips box. You may know, it only failed when
bootstrapping on Linux/mipsel. It will take some time to track down
which files are miscompiled. But I have to fix other Linux/mips bugs
first :-).

I only have a semi-automatic setup for bootstrapping gcc on Linux/mips.
I don't have the resources to do more in a timely manner. But I want to
inform everyone as soon as I can when a patch breaks Linux/mips.



H.J.

Re: gcc failed to bootstrap on Linux/mipsel

[H . J . Lu]

On Tue, Jan 29, 2002 at 11:09:07PM -0800, H . J . Lu wrote:
> On Mon, Jan 28, 2002 at 12:52:50PM -0800, Richard Henderson wrote:
> > On Mon, Jan 28, 2002 at 12:39:00PM -0800, H . J . Lu wrote:
> > > I got:
> > > 
> > > GOOD: Jan 26 20:30 PST 2002
> > > BAD: Jan 27 10:00 PST 2002
> > > 
> > > The only relevant change is 
> > > 
> > > http://gcc.gnu.org/ml/gcc-patches/2002-01/msg01929.html
> > > 
> > > Richard, could you please take a look at it?
> > 
> > As usual:
> > 
> > Are you give me any more information than this?  Like finding
> > the miscompiled file or function?  Otherwise I can't help you.
>  
> I only have a very slow mips box. You may know, it only failed when
> bootstrapping on Linux/mipsel. It will take some time to track down
> which files are miscompiled. But I have to fix other Linux/mips bugs
> first :-).

Richarrd, almost every single file in gcc is miscompiled by your change.
../../../gcc-mips/build-mipsel-rearden-linux/gcc/asm/cppinit.s is the
good one and asm/cppinit.s is the bad one.

# gdb cc1
(gdb) r -v
Starting program: /export/build/gnu/gcc/build-mipsel-linux/gcc/cc1 -v
GNU CPP version 3.1 20020127 (experimental) (cpplib) (MIPSel GNU/Linux with ELF)

Program received signal SIGSEGV, Segmentation fault.
0x00479f34 in init_standard_includes (pfile=0x100744b0)
    at /home/hjl/work/gnu/src/gcc/gcc/gcc/cppinit.c:793
793       GET_ENV_PATH_LIST (path, "CPATH");
(gdb) bt
#0  0x00479f34 in init_standard_includes (pfile=0x100744b0)
    at /home/hjl/work/gnu/src/gcc/gcc/gcc/cppinit.c:793
#1  0x0047a6d0 in cpp_read_main_file (pfile=0x100744b0, fname=0x78fc5c "", 
    table=0x3) at /home/hjl/work/gnu/src/gcc/gcc/gcc/cppinit.c:919
#2  0x00426dd0 in init_c_lex (filename=0x78fc5c "")
    at /home/hjl/work/gnu/src/gcc/gcc/gcc/c-lex.c:149
#3  0x0045dd74 in c_common_init (filename=0x0)
    at /home/hjl/work/gnu/src/gcc/gcc/gcc/c-common.c:4041
#4  0x00466c3c in c_objc_common_init (filename=0x0)
    at /home/hjl/work/gnu/src/gcc/gcc/gcc/c-objc-common.c:211
#5  0x00424340 in c_init (
    filename=0x7fff7f9a "/export/build/gnu/gcc/build-mipsel-linux/gcc/cc1")
    at /home/hjl/work/gnu/src/gcc/gcc/gcc/c-lang.c:91
#6  0x0066358c in lang_dependent_init (name=0x7fff7a34 ".{ÿ\177_{ÿ\177")
    at /home/hjl/work/gnu/src/gcc/gcc/gcc/toplev.c:5067
#7  0x00663abc in do_compile ()
    at /home/hjl/work/gnu/src/gcc/gcc/gcc/toplev.c:5176
#8  0x00663c68 in toplev_main (argc=2, argv=0x7fff7a34)
    at /home/hjl/work/gnu/src/gcc/gcc/gcc/toplev.c:5209
#9  0x0047ee7c in main (argc=2147450778, argv=0x79ad96)
    at /home/hjl/work/gnu/src/gcc/gcc/gcc/main.c:35
(gdb) p/x $pc
$3 = 0x479f34
(gdb) p/x $s8
$4 = 0x0
0x479f20 <init_standard_includes+88>:   nop
0x479f24 <init_standard_includes+92>:   lw      t9,-22896(gp)
0x479f28 <init_standard_includes+96>:   nop
0x479f2c <init_standard_includes+100>:  jalr    t9
0x479f30 <init_standard_includes+104>:  nop
0x479f34 <init_standard_includes+108>:  lw      gp,16(s8)
0x479f38 <init_standard_includes+112>:  nop
0x479f3c <init_standard_includes+116>:  move    s8,sp

I hope it is enough for you to fix the breakage.

Thanks.


H.J.
----
diff -upr ../../../gcc-mips/build-mipsel-rearden-linux/gcc/asm/cppinit.s asm/cppinit.s
--- ../../../gcc-mips/build-mipsel-rearden-linux/gcc/asm/cppinit.s	Wed Jan 30 00:00:53 2002
+++ asm/cppinit.s	Tue Jan 29 23:59:57 2002
@@ -2,7 +2,7 @@
 	.section .mdebug.abi32
 	.previous
 	.abicalls
-	.stabs	"/export/build/gnu/gcc-mips/build-mipsel-rearden-linux/gcc/",100,0,0,$Ltext0
+	.stabs	"/export/build/gnu/gcc/build-mipsel-rearden-linux/gcc/",100,0,0,$Ltext0
 	.stabs	"/home/hjl/work/gnu/src/gcc/gcc/gcc/cppinit.c",100,0,0,$Ltext0
 	.text
 $Ltext0:
@@ -1382,10 +1382,10 @@ $LM70:
 	.stabn 68,0,291,$LM71-remove_dup_dir
 $LM71:
 $LBE6:
-	lw	$31,36($sp)
 	move	$2,$17
-	lw	$16,24($sp)
+	lw	$31,36($sp)
 	lw	$17,28($sp)
+	lw	$16,24($sp)
 	#nop
 	.set	noreorder
 	.set	nomacro
@@ -2016,8 +2016,8 @@ $LM127:
 	sw	$31,36($sp)
 	sw	$17,28($sp)
 	sw	$16,24($sp)
-	sw	$28,32($sp)
 	move	$17,$4
+	sw	$28,32($sp)
 	.stabn 68,0,483,$LM128-cpp_create_reader
 $LM128:
 $LBB12:
@@ -2194,8 +2194,8 @@ $LM178:
 	.stabn 68,0,536,$LM179-cpp_create_reader
 $LM179:
 $LBE12:
-	lw	$31,36($sp)
 	move	$2,$16
+	lw	$31,36($sp)
 	lw	$17,28($sp)
 	lw	$16,24($sp)
 	#nop
@@ -2417,8 +2417,8 @@ $LM215:
 	.stabn 68,0,596,$LM216-cpp_destroy
 $LM216:
 $LBE13:
-	lw	$31,44($sp)
 	move	$2,$16
+	lw	$31,44($sp)
 	lw	$19,36($sp)
 	lw	$18,32($sp)
 	lw	$17,28($sp)
@@ -3265,16 +3265,16 @@ $LBB21:
 	.stabn 68,0,793,$LM261-init_standard_includes
 $LM261:
 	la	$4,$LC48
-	.stabn 68,0,782,$LM262-init_standard_includes
+	.stabn 68,0,785,$LM262-init_standard_includes
 $LM262:
-	move	$fp,$sp
-	.stabn 68,0,785,$LM263-init_standard_includes
-$LM263:
 	lw	$22,396($19)
-	.stabn 68,0,793,$LM264-init_standard_includes
-$LM264:
+	.stabn 68,0,793,$LM263-init_standard_includes
+$LM263:
 	la	$25,getenv
 	jal	$31,$25
+	.stabn 68,0,782,$LM264-init_standard_includes
+$LM264:
+	move	$fp,$sp
 	.stabn 68,0,794,$LM265-init_standard_includes
 $LM265:
 	.set	noreorder
@@ -3419,37 +3419,41 @@ $LBB23:
 	srl	$2,$2,3
 	sll	$2,$2,3
 	subu	$sp,$sp,$2
-	addu	$21,$sp,16
 	.stabn 68,0,823,$LM275-init_standard_includes
 $LM275:
 	move	$4,$22
+	.stabn 68,0,822,$LM276-init_standard_includes
+$LM276:
+	addu	$21,$sp,16
+	.stabn 68,0,823,$LM277-init_standard_includes
+$LM277:
 	la	$25,strlen
 	jal	$31,$25
-	.stabn 68,0,825,$LM276-init_standard_includes
-$LM276:
+	.stabn 68,0,825,$LM278-init_standard_includes
+$LM278:
 	la	$5,cpp_GCC_INCLUDE_DIR
 	move	$4,$21
 	move	$6,$20
-	.stabn 68,0,823,$LM277-init_standard_includes
-$LM277:
+	.stabn 68,0,823,$LM279-init_standard_includes
+$LM279:
 	move	$23,$2
-	.stabn 68,0,825,$LM278-init_standard_includes
-$LM278:
+	.stabn 68,0,825,$LM280-init_standard_includes
+$LM280:
 	la	$25,memcpy
 	jal	$31,$25
-	.stabn 68,0,826,$LM279-init_standard_includes
-$LM279:
+	.stabn 68,0,826,$LM281-init_standard_includes
+$LM281:
 	addu	$2,$21,$20
 	sb	$0,0($2)
-	.stabn 68,0,828,$LM280-init_standard_includes
-$LM280:
+	.stabn 68,0,828,$LM282-init_standard_includes
+$LM282:
 	la	$18,cpp_include_defaults
 	lw	$2,0($18)
 	#nop
 	beq	$2,$0,$L209
 $L185:
-	.stabn 68,0,831,$LM281-init_standard_includes
-$LM281:
+	.stabn 68,0,831,$LM283-init_standard_includes
+$LM283:
 	lw	$2,8($18)
 	move	$5,$21
 	.set	noreorder
@@ -3466,14 +3470,14 @@ $LM281:
 	#nop
 	bne	$2,$0,$L184
 $L187:
-	.stabn 68,0,836,$LM282-init_standard_includes
-$LM282:
+	.stabn 68,0,836,$LM284-init_standard_includes
+$LM284:
 	lw	$4,0($18)
 	la	$25,memcmp
 	jal	$31,$25
 	beq	$2,$0,$L207
-	.stabn 68,0,828,$LM283-init_standard_includes
-$LM283:
+	.stabn 68,0,828,$LM285-init_standard_includes
+$LM285:
 $L184:
 	addu	$18,$18,16
 $L208:
@@ -3481,44 +3485,44 @@ $L208:
 	#nop
 	bne	$2,$0,$L185
 	j	$L181
-	.stabn 68,0,839,$LM284-init_standard_includes
-$LM284:
+	.stabn 68,0,839,$LM286-init_standard_includes
+$LM286:
 $L207:
 $LBB24:
 	lw	$4,0($18)
 	la	$25,strlen
 	jal	$31,$25
-	.stabn 68,0,840,$LM285-init_standard_includes
-$LM285:
+	.stabn 68,0,840,$LM287-init_standard_includes
+$LM287:
 	addu	$4,$23,$2
 	subu	$4,$4,$20
-	.stabn 68,0,841,$LM286-init_standard_includes
-$LM286:
-	addu	$4,$4,1
-	.stabn 68,0,839,$LM287-init_standard_includes
-$LM287:
-	move	$16,$2
 	.stabn 68,0,841,$LM288-init_standard_includes
 $LM288:
+	addu	$4,$4,1
+	.stabn 68,0,839,$LM289-init_standard_includes
+$LM289:
+	move	$16,$2
+	.stabn 68,0,841,$LM290-init_standard_includes
+$LM290:
 	la	$25,xmalloc
 	jal	$31,$25
-	.stabn 68,0,842,$LM289-init_standard_includes
-$LM289:
+	.stabn 68,0,842,$LM291-init_standard_includes
+$LM291:
 	move	$4,$2
 	move	$5,$22
 	move	$6,$23
-	.stabn 68,0,841,$LM290-init_standard_includes
-$LM290:
+	.stabn 68,0,841,$LM292-init_standard_includes
+$LM292:
 	move	$17,$2
-	.stabn 68,0,843,$LM291-init_standard_includes
-$LM291:
+	.stabn 68,0,843,$LM293-init_standard_includes
+$LM293:
 	subu	$16,$16,$20
-	.stabn 68,0,842,$LM292-init_standard_includes
-$LM292:
+	.stabn 68,0,842,$LM294-init_standard_includes
+$LM294:
 	la	$25,memcpy
 	jal	$31,$25
-	.stabn 68,0,843,$LM293-init_standard_includes
-$LM293:
+	.stabn 68,0,843,$LM295-init_standard_includes
+$LM295:
 	lw	$5,0($18)
 	addu	$16,$16,1
 	addu	$5,$5,$20
@@ -3526,8 +3530,8 @@ $LM293:
 	move	$6,$16
 	la	$25,memcpy
 	jal	$31,$25
-	.stabn 68,0,847,$LM294-init_standard_includes
-$LM294:
+	.stabn 68,0,847,$LM296-init_standard_includes
+$LM296:
 	lw	$7,12($18)
 	move	$4,$19
 	move	$5,$17
@@ -3541,8 +3545,8 @@ $LM294:
 	.set	macro
 	.set	reorder
 
-	.stabn 68,0,813,$LM295-init_standard_includes
-$LM295:
+	.stabn 68,0,813,$LM297-init_standard_includes
+$LM297:
 $L205:
 $LBE24:
 $LBE23:
@@ -3552,8 +3556,8 @@ $LBE23:
 	jal	$31,$25
 	j	$L180
 $L175:
-	.stabn 68,0,809,$LM296-init_standard_includes
-$LM296:
+	.stabn 68,0,809,$LM298-init_standard_includes
+$LM298:
 	la	$4,$LC52
 $L202:
 	la	$25,getenv
@@ -3566,29 +3570,29 @@ $L202:
 	.set	reorder
 
 $L173:
-	.stabn 68,0,806,$LM297-init_standard_includes
-$LM297:
+	.stabn 68,0,806,$LM299-init_standard_includes
+$LM299:
 	la	$4,$LC51
-	.stabn 68,0,807,$LM298-init_standard_includes
-$LM298:
+	.stabn 68,0,807,$LM300-init_standard_includes
+$LM300:
 	j	$L202
 $L204:
 	bne	$4,$0,$L168
-	.stabn 68,0,800,$LM299-init_standard_includes
-$LM299:
+	.stabn 68,0,800,$LM301-init_standard_includes
+$LM301:
 	la	$4,$LC49
-	.stabn 68,0,801,$LM300-init_standard_includes
-$LM300:
+	.stabn 68,0,801,$LM302-init_standard_includes
+$LM302:
 	j	$L202
 $L171:
-	.stabn 68,0,803,$LM301-init_standard_includes
-$LM301:
+	.stabn 68,0,803,$LM303-init_standard_includes
+$LM303:
 	la	$4,$LC50
-	.stabn 68,0,804,$LM302-init_standard_includes
-$LM302:
+	.stabn 68,0,804,$LM304-init_standard_includes
+$LM304:
 	j	$L202
-	.stabn 68,0,795,$LM303-init_standard_includes
-$LM303:
+	.stabn 68,0,795,$LM305-init_standard_includes
+$LM305:
 $L203:
 	move	$5,$16
 	move	$6,$0

Re: gcc failed to bootstrap on Linux/mipsel

[Richard Henderson]

On Wed, Jan 30, 2002 at 12:23:23AM -0800, H . J . Lu wrote:
> Richarrd, almost every single file in gcc is miscompiled by your change.

"Different" != "miscompiled".  There were zero functional changes
in the diff you posted.  The breakage is elsewhere.


r~

Re: gcc failed to bootstrap on Linux/mipsel

[H . J . Lu]

On Wed, Jan 30, 2002 at 12:36:22AM -0800, Richard Henderson wrote:
> On Wed, Jan 30, 2002 at 12:23:23AM -0800, H . J . Lu wrote:
> > Richarrd, almost every single file in gcc is miscompiled by your change.
> 
> "Different" != "miscompiled".  There were zero functional changes

It is miscompiled.

> in the diff you posted.  The breakage is elsewhere.

Please read

http://gcc.gnu.org/ml/gcc-patches/2002-01/msg02117.html

one more time. The good init_standard_includes has

	.frame  $fp,72,$31 
	...
	move    $fp,$sp
	...
	la      $25,getenv
	jal     $31,$25

But with your change, it becomes

	.frame  $fp,72,$31 
	...
	la      $25,getenv
	jal     $31,$25
	...
	move    $fp,$sp

The o32 assmebler has to turn

	la      $25,getenv
	jal     $31,$25

into
	la      $25,getenv
	jal     $31,$25
	lw      gp,16(frame_register)

That is why it failed at

0x479f34 <init_standard_includes+108>:  lw      gp,16(s8)

since you moved the initialization of s8/fp after "jal". That is a very
bad idea. You have to set the frame register before making any external
calls so that the assembler can restore gp after the function return.
Could you please fix your change?

Thanks.


H.J.

Re: gcc failed to bootstrap on Linux/mipsel

[Richard Henderson]

On Wed, Jan 30, 2002 at 09:05:55AM -0800, H . J . Lu wrote:
> The o32 assmebler has to turn
> 
> 	la      $25,getenv
> 	jal     $31,$25
> 
> into
> 	la      $25,getenv
> 	jal     $31,$25
> 	lw      gp,16(frame_register)
> 
> That is why it failed at
> 
> 0x479f34 <init_standard_includes+108>:  lw      gp,16(s8)
> 
> since you moved the initialization of s8/fp after "jal".

If you knew what the problem was, why didn't you say so rather
than make me guess?

Fixed.


r~


	* sched-deps.c (sched_analyze): Make a call read the frame pointer.

Index: sched-deps.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/sched-deps.c,v
retrieving revision 1.32
diff -c -p -d -r1.32 sched-deps.c
*** sched-deps.c	2002/01/29 22:11:08	1.32
--- sched-deps.c	2002/01/30 18:03:14
*************** sched_analyze (deps, head, tail)
*** 1267,1286 ****
  	    }
  	  else
  	    {
- 	      /* A call may read and modify global register variables.
- 		 Other call-clobbered hard regs may be clobbered.  We
- 		 don't know what set of fixed registers might be used
- 		 by the function.  It is certain that the stack pointer
- 		 is among them, but be conservative.  */
  	      for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
  		if (global_regs[i])
  		  {
  		    SET_REGNO_REG_SET (reg_pending_sets, i);
  		    SET_REGNO_REG_SET (reg_pending_uses, i);
  		  }
  		else if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i))
  		  SET_REGNO_REG_SET (reg_pending_clobbers, i);
  		else if (fixed_regs[i])
  		  SET_REGNO_REG_SET (reg_pending_uses, i);
  	    }
  
--- 1267,1295 ----
  	    }
  	  else
  	    {
  	      for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
+ 		/* A call may read and modify global register variables.  */
  		if (global_regs[i])
  		  {
  		    SET_REGNO_REG_SET (reg_pending_sets, i);
  		    SET_REGNO_REG_SET (reg_pending_uses, i);
  		  }
+ 		/* Other call-clobbered hard regs may be clobbered.  */
  		else if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i))
  		  SET_REGNO_REG_SET (reg_pending_clobbers, i);
+ 		/* We don't know what set of fixed registers might be used
+ 		   by the function, but it is certain that the stack pointer
+ 		   is among them, but be conservative.  */
  		else if (fixed_regs[i])
+ 		  SET_REGNO_REG_SET (reg_pending_uses, i);
+ 		/* The frame pointer is normally not used by the function
+ 		   itself, but by the debugger.  */
+ 		/* ??? MIPS o32 is an exception.  It uses the frame pointer
+ 		   in the macro expansion of jal but does not represent this
+ 		   fact in the call_insn rtl.  */
+ 		else if (i == FRAME_POINTER_REGNUM
+ 			 || (i == HARD_FRAME_POINTER_REGNUM
+ 			     && (! reload_completed || frame_pointer_needed)))
  		  SET_REGNO_REG_SET (reg_pending_uses, i);
  	    }
  

Re: gcc failed to bootstrap on Linux/mipsel

[John David Anglin]

> The last good one for me is Jan 24 16:56 PST 2002. I am tracking down
> which patch breaks Linux/mipsel now. Irix6.5/mips seems ok. It may be
> another ABI/ednian bug.

I am seeing increased testsuite failures on hppa2.0w-hp-hpux11.11 as of
Jan. 26 09:10:13 UTC.  My last good build was as of Jan. 25 05:05:05 UTC.
I have looked at one of the new failure (g++.abi/cxa_vec.C).  The failure
appears to be a new bug in the loop pass.  After every call on the PA,
the pic offset table register needs to be restored.  For some reason,
the loop pass pulled the insn doing this restore out of the loop, apparently
ignoring an indirect call which has a USE for the register.  As a result,
the pic offset table register is not valid after the first call in the 
loop and a seg fault occurs.

Dave
-- 
J. David Anglin                                  dave.anglin@nrc.ca
National Research Council of Canada              (613) 990-0752 (FAX: 952-6605)

Re: gcc failed to bootstrap on Linux/mipsel

[H . J . Lu]

On Sun, Jan 27, 2002 at 09:39:30PM -0500, John David Anglin wrote:
> > The last good one for me is Jan 24 16:56 PST 2002. I am tracking down
> > which patch breaks Linux/mipsel now. Irix6.5/mips seems ok. It may be
> > another ABI/ednian bug.
> 
> I am seeing increased testsuite failures on hppa2.0w-hp-hpux11.11 as of
> Jan. 26 09:10:13 UTC.  My last good build was as of Jan. 25 05:05:05 UTC.

It may be a different bug. "Jan 26 14:00:00 UTC 2002" seems ok on
Linux/mipsel.

> I have looked at one of the new failure (g++.abi/cxa_vec.C).  The failure
> appears to be a new bug in the loop pass.  After every call on the PA,
> the pic offset table register needs to be restored.  For some reason,
> the loop pass pulled the insn doing this restore out of the loop, apparently
> ignoring an indirect call which has a USE for the register.  As a result,
> the pic offset table register is not valid after the first call in the 
> loop and a seg fault occurs.
> 


H.J.

Re: gcc failed to bootstrap on Linux/mipsel

[Geoff Keating]

"John David Anglin" <dave@hiauly1.hia.nrc.ca> writes:

> After every call on the PA, the pic offset table register needs to
> be restored.  For some reason, the loop pass pulled the insn doing
> this restore out of the loop, apparently ignoring an indirect call
> which has a USE for the register.  As a result,

You mean CLOBBER, right?  If it's just a USE, of course the restore
can be hoisted.

-- 
- Geoffrey Keating <geoffk@geoffk.org> <geoffk@redhat.com>

Re: gcc failed to bootstrap on Linux/mipsel

[John David Anglin]

> > After every call on the PA, the pic offset table register needs to
> > be restored.  For some reason, the loop pass pulled the insn doing
> > this restore out of the loop, apparently ignoring an indirect call
> > which has a USE for the register.  As a result,
> 
> You mean CLOBBER, right?  If it's just a USE, of course the restore
> can be hoisted.

The reason that the restore shouldn't be hoisted is
PIC_OFFSET_TABLE_REG_CALL_CLOBBERED is defined.  This
only applies when pic code is being generated. This
register is also one of the CALL_USED_REGISTERS, and
one of the FIXED_REGISTERS when pic code is being generated.

I have identified the patch that introduced the problem:

<http://gcc.gnu.org/ml/gcc-patches/2002-01/msg01836.html>.

> While working on a loop optimization issue I noticed that gen_rtx_REG
> does not return the same rtx pointer when called with
> PIC_OFFSET_TABLE_REGNUM like it does if called with
> RETURN_ADDRESS_POINTER_REGNUM or STACK_POINTER_REGNUM.  This comes up in
> loop.c (loop_regs_scan) where we loop:
> 
> for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
>       if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i)
> 	    && rtx_varies_p (gen_rtx_REG (Pmode, i), /*for_alias=*/1))

rtx_varies_p is called with for_alias=1, so the pic_offset_table_rtx
is going to be treated as a constant even though
PIC_OFFSET_TABLE_REG_CALL_CLOBBERED is set.  This apparently enables
the restore to be hoisted.

Dave
-- 
J. David Anglin                                  dave.anglin@nrc.ca
National Research Council of Canada              (613) 990-0752 (FAX: 952-6605)

Re: gcc failed to bootstrap on Linux/mipsel

[John David Anglin]

> > for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
> >       if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i)
> > 	    && rtx_varies_p (gen_rtx_REG (Pmode, i), /*for_alias=*/1))

I see the problem.  The comment for the above is incorrect:

  /* Invalidate all hard registers clobbered by calls.  With one exception:
     a call-clobbered PIC register is still function-invariant for our
     purposes, since we can hoist any PIC calculations out of the loop.
     Thus the call to rtx_varies_p.  */

You can hoist any insns involving the PIC register except those
that set it.

Dave
-- 
J. David Anglin                                  dave.anglin@nrc.ca
National Research Council of Canada              (613) 990-0752 (FAX: 952-6605)

Re: gcc failed to bootstrap on Linux/mipsel

[Richard Henderson]

On Mon, Jan 28, 2002 at 12:04:49PM -0500, John David Anglin wrote:
> > > for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
> > >       if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i)
> > > 	    && rtx_varies_p (gen_rtx_REG (Pmode, i), /*for_alias=*/1))
> 
> I see the problem.  The comment for the above is incorrect:
> 
>   /* Invalidate all hard registers clobbered by calls.  With one exception:
>      a call-clobbered PIC register is still function-invariant for our
>      purposes, since we can hoist any PIC calculations out of the loop.
>      Thus the call to rtx_varies_p.  */
> 
> You can hoist any insns involving the PIC register except those
> that set it.

Bugger all.  Well, I guess we'll just have to do for_alias=0
for now.  I don't think we can fix this properly for 3.1.


r~

Re: gcc failed to bootstrap on Linux/mipsel

[John David Anglin]

> > You can hoist any insns involving the PIC register except those
> > that set it.
> 
> Bugger all.  Well, I guess we'll just have to do for_alias=0
> for now.  I don't think we can fix this properly for 3.1.

Is this obviously wrong?  I'm trying it now.

Dave
-- 
J. David Anglin                                  dave.anglin@nrc.ca
National Research Council of Canada              (613) 990-0752 (FAX: 952-6605)

2002-01-28  John David Anglin  <dave@hiauly1.hia.nrc.ca>

	* loop.c (scan_loop): Don't hoist insns that set the pic offset table
	register if PIC_OFFSET_TABLE_REG_CALL_CLOBBERED is defined and we are
	generating PIC code.

--- loop.c.orig	Fri Jan 25 00:04:25 2002
+++ loop.c	Mon Jan 28 13:49:58 2002
@@ -762,6 +762,9 @@
       if (GET_CODE (p) == INSN
 	  && (set = single_set (p))
 	  && GET_CODE (SET_DEST (set)) == REG
+#ifdef PIC_OFFSET_TABLE_REG_CALL_CLOBBERED
+	  && (! flag_pic || SET_DEST (set) != pic_offset_table_rtx)
+#endif
 	  && ! regs->array[REGNO (SET_DEST (set))].may_not_optimize)
 	{
 	  int tem1 = 0;

Re: gcc failed to bootstrap on Linux/mipsel

[Richard Henderson]

On Mon, Jan 28, 2002 at 02:12:40PM -0500, John David Anglin wrote:
> +#ifdef PIC_OFFSET_TABLE_REG_CALL_CLOBBERED
> +	  && (! flag_pic || SET_DEST (set) != pic_offset_table_rtx)
> +#endif

Well flag_pic definitely shouldn't be here.  ia64 for instance
uses pic_offset_table_rtx all the time.


r~

Re: gcc failed to bootstrap on Linux/mipsel

[John David Anglin]

> On Mon, Jan 28, 2002 at 02:12:40PM -0500, John David Anglin wrote:
> > +#ifdef PIC_OFFSET_TABLE_REG_CALL_CLOBBERED
> > +	  && (! flag_pic || SET_DEST (set) != pic_offset_table_rtx)
> > +#endif
> 
> Well flag_pic definitely shouldn't be here.  ia64 for instance
> uses pic_offset_table_rtx all the time.

What about testing to see if the PIC_OFFSET_TABLE_REGNUM isn't a fixed
register?

Dave
-- 
J. David Anglin                                  dave.anglin@nrc.ca
National Research Council of Canada              (613) 990-0752 (FAX: 952-6605)

Re: gcc failed to bootstrap on Linux/mipsel

[Richard Henderson]

On Mon, Jan 28, 2002 at 02:31:51PM -0500, John David Anglin wrote:
> What about testing to see if the PIC_OFFSET_TABLE_REGNUM isn't a fixed
> register?

I guess that would be ok.  Actualy, should change gen_reg_RTX
to check that and have loop test vs pic_offset_table_rtx.


r~

Re: gcc failed to bootstrap on Linux/mipsel

[John David Anglin]

> > What about testing to see if the PIC_OFFSET_TABLE_REGNUM isn't a fixed
> > register?
> 
> I guess that would be ok.  Actualy, should change gen_reg_RTX
> to check that and have loop test vs pic_offset_table_rtx.

This fixes cxa_vec.C and several hundred objc failures.  Tested on
hppa2.0w-hp-hpux11.11.

OK?

Dave
-- 
J. David Anglin                                  dave.anglin@nrc.ca
National Research Council of Canada              (613) 990-0752 (FAX: 952-6605)

2002-01-28  John David Anglin  <dave@hiauly1.hia.nrc.ca>

	* emit-rtl.c (gen_rtx_REG): Check that the PIC_OFFSET_TABLE_REGNUM
	is a fixed register before returning pic_offset_table_rtx.
	* loop.c (scan_loop): Don't hoist insns that set pic_offset_table_rtx
	when PIC_OFFSET_TABLE_REG_CALL_CLOBBERED is defined.

--- emit-rtl.c.orig	Mon Jan 28 12:11:37 2002
+++ emit-rtl.c	Mon Jan 28 15:06:47 2002
@@ -411,7 +411,8 @@
       if (regno == RETURN_ADDRESS_POINTER_REGNUM)
 	return return_address_pointer_rtx;
 #endif
-      if (regno == PIC_OFFSET_TABLE_REGNUM)
+      if (regno == PIC_OFFSET_TABLE_REGNUM
+	  && fixed_regs[PIC_OFFSET_TABLE_REGNUM])
         return pic_offset_table_rtx;
       if (regno == STACK_POINTER_REGNUM)
 	return stack_pointer_rtx;
--- loop.c.orig	Fri Jan 25 00:04:25 2002
+++ loop.c	Mon Jan 28 15:00:03 2002
@@ -762,6 +762,9 @@
       if (GET_CODE (p) == INSN
 	  && (set = single_set (p))
 	  && GET_CODE (SET_DEST (set)) == REG
+#ifdef PIC_OFFSET_TABLE_REG_CALL_CLOBBERED
+	  && SET_DEST (set) != pic_offset_table_rtx
+#endif
 	  && ! regs->array[REGNO (SET_DEST (set))].may_not_optimize)
 	{
 	  int tem1 = 0;

Re: gcc failed to bootstrap on Linux/mipsel

[Richard Henderson]

On Mon, Jan 28, 2002 at 06:56:56PM -0500, John David Anglin wrote:
> 	* emit-rtl.c (gen_rtx_REG): Check that the PIC_OFFSET_TABLE_REGNUM
> 	is a fixed register before returning pic_offset_table_rtx.
> 	* loop.c (scan_loop): Don't hoist insns that set pic_offset_table_rtx
> 	when PIC_OFFSET_TABLE_REG_CALL_CLOBBERED is defined.

Ok.


r~