eh_table alignment

eh_table alignment

[Bruno Haible]

The calls to ASM_OUTPUT_ALIGN in dwarf2out.c in testgcc-971128 cause a problem
on i386.
Take for example, the following source

================= foo.cc =======================
void foo () { throw 5; }
================================================

and its output (on i486-linux)

================== cc1plus foo.cc -dA -o foo.s =========================
	.file	"foo.cc"
	.version	"01.01"
/ GNU C++ version testgcc-2.8.0 971128 experimental (i486-linux) compiled by GNU C version 2.7.2.
/ options passed: 
/ options enabled:  -fpeephole -ffunction-cse -fkeep-static-consts
/ -fpcc-struct-return -fexceptions -fcommon -fverbose-asm -fgnu-linker
/ -m80387 -mhard-float -mno-soft-float -mieee-fp -mfp-ret-in-387
/ -mschedule-prologue -mcpu=i486 -march=pentium

gcc2_compiled.:
.globl __throw
.text
	.align 16
.globl foo__Fv
	.type	 foo__Fv,@function
foo__Fv:
.LFB1:
	pushl %ebp
.LCFI0:
	movl %esp,%ebp
.LCFI1:
	pushl %ebx
.LCFI2:
	pushl $0
.LCFI3:
	call __tfi
	movl %eax,%eax
	pushl %eax
.LCFI4:
	pushl $4
.LCFI5:
	call __builtin_new
	addl $4,%esp
.LCFI6:
	movl %eax,%eax
	movl %eax,%ebx
	movl $5,(%ebx)
	pushl %ebx
.LCFI7:
	call __cp_push_exception
	addl $12,%esp
.LCFI8:
.L2:
	movl $.L2,__eh_pc
	call __throw
	.align 16
	jmp .L3
	.align 16
	jmp .L1
	.align 16
.L3:
.L1:
	movl -4(%ebp),%ebx
	movl %ebp,%esp
	popl %ebp
	ret
.LFE1:
.Lfe1:
	.size	 foo__Fv,.Lfe1-foo__Fv

#APP
.section	.eh_frame,"aw",@progbits
__FRAME_BEGIN__:
	.4byte	.LLCIE1	/ Length of Common Information Entry
.LSCIE1:
	.4byte	0x0	/ CIE Identifier Tag
	.byte	0x1	/ CIE Version
	.byte	0x0	/ CIE Augmentation (none)
	.byte	0x1	/ ULEB128 0x1 (CIE Code Alignment Factor)
	.byte	0x7c	/ SLEB128 -4 (CIE Data Alignment Factor)
	.byte	0x8	/ CIE RA Column
	.byte	0xc	/ DW_CFA_def_cfa
	.byte	0x4	/ ULEB128 0x4
	.byte	0x4	/ ULEB128 0x4
	.byte	0x88	/ DW_CFA_offset, column 0x8
	.byte	0x1	/ ULEB128 0x1
	.align 4
.LECIE1:
	.set	.LLCIE1,.LECIE1-.LSCIE1
	.4byte	.LLFDE1	/ FDE Length
.LSFDE1:
	.4byte	.LSFDE1-__FRAME_BEGIN__	/ FDE CIE offset
	.4byte	.LFB1	/ FDE initial location
	.4byte	.LFE1-.LFB1	/ FDE address range
	.byte	0x4	/ DW_CFA_advance_loc4
	.4byte	.LCFI0-.LFB1
	.byte	0xe	/ DW_CFA_def_cfa_offset
	.byte	0x8	/ ULEB128 0x8
	.byte	0x85	/ DW_CFA_offset, column 0x5
	.byte	0x2	/ ULEB128 0x2
	.byte	0x4	/ DW_CFA_advance_loc4
	.4byte	.LCFI1-.LCFI0
	.byte	0xd	/ DW_CFA_def_cfa_register
	.byte	0x5	/ ULEB128 0x5
	.byte	0x4	/ DW_CFA_advance_loc4
	.4byte	.LCFI2-.LCFI1
	.byte	0x83	/ DW_CFA_offset, column 0x3
	.byte	0x3	/ ULEB128 0x3
	.byte	0x4	/ DW_CFA_advance_loc4
	.4byte	.LCFI3-.LCFI2
	.byte	0x2e	/ DW_CFA_GNU_args_size
	.byte	0x4	/ ULEB128 0x4
	.byte	0x4	/ DW_CFA_advance_loc4
	.4byte	.LCFI4-.LCFI3
	.byte	0x2e	/ DW_CFA_GNU_args_size
	.byte	0x8	/ ULEB128 0x8
	.byte	0x4	/ DW_CFA_advance_loc4
	.4byte	.LCFI5-.LCFI4
	.byte	0x2e	/ DW_CFA_GNU_args_size
	.byte	0xc	/ ULEB128 0xc
	.byte	0x4	/ DW_CFA_advance_loc4
	.4byte	.LCFI6-.LCFI5
	.byte	0x2e	/ DW_CFA_GNU_args_size
	.byte	0x8	/ ULEB128 0x8
	.byte	0x4	/ DW_CFA_advance_loc4
	.4byte	.LCFI7-.LCFI6
	.byte	0x2e	/ DW_CFA_GNU_args_size
	.byte	0xc	/ ULEB128 0xc
	.byte	0x4	/ DW_CFA_advance_loc4
	.4byte	.LCFI8-.LCFI7
	.byte	0x2e	/ DW_CFA_GNU_args_size
	.byte	0x0	/ ULEB128 0x0
	.align 4
.LEFDE1:
	.set	.LLFDE1,.LEFDE1-.LSFDE1
#NO_APP
	.ident	"GCC: (GNU) testgcc-2.8.0 971128 experimental"
==================================================================

Note that the call to __throw is *after* the label .LCFI8 (which is
the last label covered by an advance_loc4 instruction by the FDE).
This occurs only without -O.

At runtime, the following happens:
- __throw() is called.
- __frame_state_for sees that the FDE at .LSFDE1 is responsible,
  because the call to __throw is between .LFB1 and .LFE1.
- __frame_state_for executes all instructions of the FDE, up to .LEFDE1,
  because that's what it is programmed to do if no DW_CFA_advance_loc4
  matches.
- Among these, ".align 4" is one or more bytes with value 0x90, which is
  interpreted as a DW_CFA_offset instruction for register 16. This sets
  the `saved' bit for register 16.
- __throw then calls copy_reg for this register. Since register 16 is
  not a call-saved register on i386 (it is not a valid register at all),
  abort() is called.

What is the right way to fix this?

  a. Make sure that .LCFI8 is emitted *after* the call to __throw
     (as is done when compiling with "-O") ?

  b. Instead of ".align 4", emit ".align 4,0x00" where the 0x00 is
     DW_CFA_nop. Do all assemblers support this ?

  c. Remove the alignment altogether, and in frame.c change the two
     occurrences of __attribute__ ((packed, aligned (__alignof__ (void *))))
     to __attribute__ ((packed, aligned (1))) . Does gcc know to emit
     unaligned accesses for all CPUs ?

Re: eh_table alignment

[Andreas Schwab]

Bruno Haible <haible@ilog.fr> writes:

|> - Among these, ".align 4" is one or more bytes with value 0x90

IIRC this is fixed in the latest binutils.

-- 
Andreas Schwab                                      "And now for something
schwab@issan.informatik.uni-dortmund.de              completely different"
schwab@gnu.org

Re: eh_table alignment

[Joseph H. Buehler]

Andreas Schwab <schwab@issan.informatik.uni-dortmund.de> writes:

> |> - Among these, ".align 4" is one or more bytes with value 0x90
> 
> IIRC this is fixed in the latest binutils.

Where are we supposed to get the binutils for egcs?  This is another
good thing to put in the install instructions.

Joe Buehler

Re: eh_table alignment

[Jeffrey A Law]

  In message < m2d8jgd696.fsf@altera.gaithersburg.md.us >you write:
  > Andreas Schwab <schwab@issan.informatik.uni-dortmund.de> writes:
  > 
  > > |> - Among these, ".align 4" is one or more bytes with value 0x90
  > > 
  > > IIRC this is fixed in the latest binutils.
  > 
  > Where are we supposed to get the binutils for egcs?  This is another
  > good thing to put in the install instructions.
Actually, we aren't recommend you pick up any binutils snapshots;
just releases -- either binutils-2.8 or the Linux specific releases.
The install instructions and FAQ have pointers to the Linux specific
binutils releases.
jeff