[PATCH v3 10/15] gas: Skip SFrame FDE if FP without RA on stack

[Jens Remus]

Hello Indu,

Am 12.04.2024 um 16:47 schrieb Jens Remus:
> The SFrame format cannot represent the frame pointer (FP) being saved
> on the stack without the return address (RA) also being saved on the
> stack, if RA tracking is used.

[...]

> diff --git a/gas/gen-sframe.c b/gas/gen-sframe.c
> index a3b6f75cfe85..87be3eb05ad2 100644
> --- a/gas/gen-sframe.c
> +++ b/gas/gen-sframe.c
> @@ -1439,6 +1439,25 @@ sframe_do_fde (struct sframe_xlate_ctx *xlate_ctx,
>   	= get_dw_fde_end_addrS (xlate_ctx->dw_fde);
>       }
>   
> +#ifdef SFRAME_FRE_RA_TRACKING
> +  if (sframe_ra_tracking_p ())
> +    {
> +      struct sframe_row_entry *fre;
> +
> +      /* Iterate over the scratchpad FREs and validate them.  */
> +      for (fre = xlate_ctx->first_fre; fre; fre = fre->next)
> +	{
> +	  /* SFrame format cannot represent FP on stack without RA on stack.  */
> +	  if (fre->ra_loc != SFRAME_FRE_ELEM_LOC_STACK
> +	      && fre->bp_loc == SFRAME_FRE_ELEM_LOC_STACK)
> +	    {
> +	      as_warn (_("skipping SFrame FDE due to FP without RA on stack"));
> +	      return SFRAME_XLATE_ERR_NOTREPRESENTED;
> +	    }
> +	}
> +    }
> +#endif /* SFRAME_FRE_RA_TRACKING  */
> +
>     return SFRAME_XLATE_OK;
>   }

I noticed that above new warning is erroneously emitted when assembling 
the following CFI directive sequence with option "-alh" (to output a 
listing of the assembly; probably any "-a[...]") on a SFrame enabled 
target, that uses FP and RA tracking.

.cfi_offset <fp-regno>, <fp-offset>
.cfi_offset <ra-regno>, <ra-offset>

The reason is that with listings enabled there is an additional DWARF 
DW_CFA_advance_loc CFI instruction (with a zero advance) between both 
DW_CFA_offset instructions, that the DWARF .eh_frame generator is able 
to process correctly, but causes the .sframe generator to choke.

Additionally with this patch reverted "bad" SFrame information is 
generated (see example below), where there are multiple SFrame FREs for 
the same PC start address.
Note that the FP-tracking information erroneously being displayed in the 
RA-tracking column, is why I introduced this new warning message. I will 
send two alternative patches how to potentially resolve that soon.

$ cat test_fpra_min.s
         .cfi_sections .sframe, .eh_frame
         .cfi_startproc
         stmg    %r11,%r15,88(%r15)
         .cfi_rel_offset 11, 88
         .cfi_rel_offset 14, 112
         la      %r11,0
         la      %r14,0
.Lreturn:
         lmg     %r11,%r15,88(%r15)
         .cfi_restore 14
         .cfi_restore 11
         br      %r14
         .cfi_endproc

$ ojbdump --sframe test_fpra_without-alh.o
...
   Function Index :

     func idx [0]: pc = 0x0, size = 22 bytes
     STARTPC         CFA       FP        RA
     0000000000000000  sp+160    u         u
     0000000000000006  sp+160    c-72      c-48
     0000000000000014  sp+160    u         u

$ objdump --sframe test_fpra_with_alh.o
...
   Function Index :

     func idx [0]: pc = 0x0, size = 22 bytes
     STARTPC         CFA       FP        RA
     0000000000000000  sp+160    u         u
     0000000000000006  sp+160    u         c-72
     0000000000000006  sp+160    c-72      c-48
     0000000000000014  sp+160    u         c-72
     0000000000000014  sp+160    u         u

Note that the outputs of "objdump -Wf" and "objdump -WF" are identical 
in both cases (with and without option "-alh").

Debugging of the SFrame processing of the DWARF CFI instructions shows 
that with option "-a" there are additional DW_CFA_advance_loc:

DW_CFA_def_cfa: reg=15 offset=160
DW_CFA_advance_loc: lab1=L0, lab2=L0
DW_CFA_offset: reg=11 offset=-72
DW_CFA_advance_loc: lab1=L0, lab2=L0   <-- only with -a
DW_CFA_offset: reg=14 offset=-48
DW_CFA_advance_loc: lab1=L0, lab2=L0
DW_CFA_restore: reg=14
DW_CFA_advance_loc: lab1=L0, lab2=L0   <-- only with -a
DW_CFA_restore: reg=11

Debugging of the CFI directive processing in gas/dw2gencfi.c shows the 
following:

- With option "-a" cfi_add_advance_loc() is invoked more often in 
dot_cfi() due to the condition (symbol_get_frag 
(frchain_now->frch_cfi_data->last_address) != frag_now) evaluating to true.

- output_cfi_insn() of case DW_CFA_advance_loc enters the condition 
(symbol_get_frag (to) == symbol_get_frag (from)) without option "-a" and 
enters the else condition with option "-a". The else path has an 
interesting comment that suggests that there is logic to relax an 
advance by zero at a later stage:

"... Call frag_grow with the sum of room needed by frag_more and 
frag_var to preallocate space ensuring that the DW_CFA_advance_loc4 is 
in the fixed part of the rs_cfa frag, so that the relax machinery can 
remove the advance_loc should it advance by zero."

I don't have a clue how to resolve this potential issue in the SFrame 
generation. I could not figure out how to detect the advance of zero in 
the SFrame processing of DW_CFA_advance_loc, so that it could be treated 
special.
I can open a ticket in the Sourceware Bugzilla, if you agree that this 
is an issue.

Thanks and regards,
Jens
-- 
Jens Remus
Linux on Z Development (D3303) and z/VSE Support
+49-7031-16-1128 Office
jremus@de.ibm.com

IBM

IBM Deutschland Research & Development GmbH; Vorsitzender des 
Aufsichtsrats: Wolfgang Wendt; Geschäftsführung: David Faller; Sitz der 
Gesellschaft: Böblingen; Registergericht: Amtsgericht Stuttgart, HRB 243294
IBM Data Privacy Statement: https://www.ibm.com/privacy/