Re: [PATCH] PR80101: Fix ICE in store_data_bypass_p

[Segher Boessenkool <segher@kernel.crashing.org>]

On Mon, Apr 10, 2017 at 06:38:07PM +0100, Richard Sandiford wrote:
> Segher Boessenkool <segher@kernel.crashing.org> writes:
> > On Fri, Apr 07, 2017 at 10:39:03AM +0200, Eric Botcazou wrote:
> >> > Or we could just change "blockage" and wait for the next bug report.
> >> 
> >> That's my suggestion, yes.
> >> 
> >> > Alternatively, we can arrange for the bypass functions to not ICE.  We
> >> > can do that specific to these rs6000 pipeline descriptions, by having
> >> > our own version of store_data_bypass_p; or we can make that function
> >> > work for all insns (its definition works fine for insn pairs where
> >> > not both the producer and consumer are SETs).  That's what Kelvin's
> >> > patch does.  What is the value in ICEing here?
> >> 
> >> Telling the back-end writer that something may be wrong somewhere instead of 
> >> silently accepting nonsense?
> >
> > Why is it nonsense?  The predicate gives the answer to the question
> > "given these insns A and B, does A feed data that B stores in memory".
> > That is a perfectly valid question to ask of any two insns.
> 
> Agreed FWIW, but for:
> 
> @@ -3701,7 +3704,8 @@ store_data_bypass_p (rtx_insn *out_insn, rtx_insn
>              if (GET_CODE (out_exp) == CLOBBER)
>                continue;
> 
> -            gcc_assert (GET_CODE (out_exp) == SET);
> +	    if (GET_CODE (out_exp) != SET)
> +	      return false;
> 
>              if (reg_mentioned_p (SET_DEST (out_exp), SET_DEST (in_set)))
>                return false;
> 
> how about instead changing the CLOBBER check so that we continue
> when it isn't a SET?  That would allow things like UNSPECs and
> USEs as well.

Yeah that sounds good.  Kelvin, could you try that please?


Segher