>The attached patch prevents the creation of reg-moves for definitions
>with MODE_CC and thus solves this ICE.
>
>Currently testing and bootstrap on ppc64-redhat-linux, enabling SMS on
>loops with SC 1.
>
>OK for 4.7 once testing completes?
Yes, thanks for catching this. Shouldn't we prevent creating such
regmoves for (the other case of) intra-loop anti-deps as well?
>> While true, I doubt either PPC or MIPS really benefit from moving
>> around registers of CCmode.
IIRC, PPC has eight 4-bit CR's and supports efficient copying of *one
bit* from one CR to another (via cror), which would require special
regmove handling. Hence I share your doubt.
Ayal.
On Thu, Dec 22, 2011 at 10:53 AM, Richard Sandiford
<richard.sandif...@linaro.org> wrote:
> Richard Henderson <r...@redhat.com> writes:
>> On 12/20/2011 09:47 AM, Richard Sandiford wrote:
>>> Revital Eres <revital.e...@linaro.org> writes:
>>>> +/* Return true if one of the definitions in INSN has MODE_CC. Otherwise
>>>> + return false. */
>>>> +static bool
>>>> +def_has_ccmode_p (rtx insn)
>>>> +{
>>>> + df_ref *def;
>>>> +
>>>> + for (def = DF_INSN_DEFS (insn); *def; def++)
>>>> + {
>>>> + enum machine_mode mode = GET_MODE (DF_REF_REG (*def));
>>>> +
>>>> + if (GET_MODE_CLASS (mode) == MODE_CC)
>>>> + return true;
>>>> + }
>>>> +
>>>> + return false;
>>>> +}
>>>
>>> FWIW, an alternative might be to test have_regs_of_mode[(int) mode].
>>> That says whether there are any allocatable (non-fixed) registers
>>> of the given mode.
>>
>> While true, I doubt either PPC or MIPS really benefit from moving
>> around registers of CCmode. Certainly MIPS has no way of easily
>> moving CCmode registers around. It's a rather complicated reload,
>> that.
>>
>> I'd be very tempted to simply go with the original patch.
>
> OK, sorry for the noise.
>
> Richard
