Re: [PATCH][PR65802] Mark ifn_va_arg with ECF_NOTHROW

[Richard Biener <rguenther@suse.de>]

On Tue, 21 Apr 2015, Tom de Vries wrote:

> Hi,
> 
> this patch fixes PR65802.
> 
> The problem described in PR65802 is that when compiling the test-case
> (included in the patch below) at -O0, the compiler runs into a gcc_assert ICE
> in redirect_eh_edge_1 during pass_cleanup_eh:
> ...
> 	  gcc_assert (lookup_stmt_eh_lp (throw_stmt) == old_lp_nr);
> ...
> 
> 
> In more detail, during compilation the ifn_va_arg is marked at as a throwing
> function. That causes exception handling code to be generated, with exception
> handling edges:
> ...
> ;; basic block 2, loop depth 0, count 0, freq 0, maybe hot
> ;;  prev block 0, next block 3, flags: (NEW, REACHABLE)
> ;;  pred:       ENTRY (FALLTHRU)
> [LP 1] # .MEM_5 = VDEF <.MEM_4(D)>
> # USE = anything
> # CLB = anything
> _6 = VA_ARG (&cD.2333, 0B);
> ;;  succ:       7 (EH)
> ;;              3 (FALLTHRU)
> ...
> 
> After pass_lower_vaarg, the expansion of ifn_va_arg is spread over several
> basic blocks:
> ...
> ;;   basic block 2, loop depth 0, count 0, freq 0, maybe hot
> ;;    prev block 0, next block 11, flags: (NEW, REACHABLE)
> ;;    pred:       ENTRY (FALLTHRU)
> ;;    succ:       11 [100.0%]  (FALLTHRU)
> 
> ;;   basic block 11, loop depth 0, count 0, freq 0, maybe hot
> ;;    prev block 2, next block 12, flags: (NEW)
> ;;    pred:       2 [100.0%]  (FALLTHRU)
>   # VUSE <.MEM_4(D)>
>   _22 = cD.2333.gp_offsetD.5;
>   if (_22 >= 48)
>     goto <bb 13> (<L6>);
>   else
>     goto <bb 12> (<L5>);
> ;;    succ:       13 (TRUE_VALUE)
> ;;                12 (FALSE_VALUE)
> 
> ;;   basic block 12, loop depth 0, count 0, freq 0, maybe hot
> ;;    prev block 11, next block 13, flags: (NEW)
> ;;    pred:       11 (FALSE_VALUE)
> <L5>:
>   # VUSE <.MEM_4(D)>
>   _23 = cD.2333.reg_save_areaD.8;
>   # VUSE <.MEM_4(D)>
>   _24 = cD.2333.gp_offsetD.5;
>   _25 = (sizetype) _24;
>   addr.1_26 = _23 + _25;
>   # VUSE <.MEM_4(D)>
>   _27 = cD.2333.gp_offsetD.5;
>   _28 = _27 + 8;
>   # .MEM_29 = VDEF <.MEM_4(D)>
>   cD.2333.gp_offsetD.5 = _28;
>   goto <bb 14> (<L7>);
> ;;    succ:       14 (FALLTHRU)
> 
> ;;   basic block 13, loop depth 0, count 0, freq 0, maybe hot
> ;;    prev block 12, next block 14, flags: (NEW)
> ;;    pred:       11 (TRUE_VALUE)
> <L6>:
>   # VUSE <.MEM_4(D)>
>   _30 = cD.2333.overflow_arg_areaD.7;
>   addr.1_31 = _30;
>   _32 = _30 + 8;
>   # .MEM_33 = VDEF <.MEM_4(D)>
>   cD.2333.overflow_arg_areaD.7 = _32;
> ;;    succ:       14 (FALLTHRU)
> 
> ;;   basic block 14, loop depth 0, count 0, freq 0, maybe hot
> ;;    prev block 13, next block 15, flags: (NEW)
> ;;    pred:       12 (FALLTHRU)
> ;;                13 (FALLTHRU)
>   # .MEM_20 = PHI <.MEM_29(12), .MEM_33(13)>
>   # addr.1_21 = PHI <addr.1_26(12), addr.1_31(13)>
> <L7>:
>   # VUSE <.MEM_20>
>   _6 = MEM[(intD.9 * * {ref-all})addr.1_21];
> ;;    succ:       15 (FALLTHRU)
> 
> ;;   basic block 15, loop depth 0, count 0, freq 0, maybe hot
> ;;    prev block 14, next block 3, flags: (NEW)
> ;;    pred:       14 (FALLTHRU)
> ;;    succ:       7 (EH)
> ;;                3 (FALLTHRU)
> ...
> 
> And an ICE is triggered in redirect_eh_edge_1, because the code expects the
> last statement in a BB with an outgoing EH edge to be a throwing statement.
> 
> That's obviously not the case, since bb15 is empty. But also all the other
> statements in the expansion are non-throwing.
> 
> 
> Looking at the representation before the ifn_va_arg, VA_ARG_EXPR is
> non-throwing (even with -fnon-call-exceptions).
> 
> And looking at the situation before the introduction of ifn_va_arg, the
> expansion of VA_ARG_EXPR also didn't contain any throwing statements.
> 
> 
> This patch fixes the ICE by marking ifn_va_arg with ECF_NOTHROW.
> 
> Bootstrapped and reg-tested on x86_64.
> 
> OK for trunk?

Ok.

Thanks,
Richard.

> Thanks,
> - Tom
> 

-- 
Richard Biener <rguenther@suse.de>
SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard, Jennifer Guild,
Dilip Upmanyu, Graham Norton HRB 21284 (AG Nuernberg)