ICE at -O1 and above in both 32-bit and 64-bit modes on x86_64-linux-gnu

[kugan]

Hi Richard,

Thanks for the review.

On 23/09/16 17:19, Richard Biener wrote:
> On Fri, Sep 23, 2016 at 12:24 AM, kugan
> <kugan.vivekanandarajah@linaro.org> wrote:
>> Hi,
>> As Richard pointed out in PR77677, TREE_OVERFLOW is not cleared in IPA-VRP.
>> There are three places in which we set value_range:
>>
>>
>> 1. When value ranges are obtained from SSA_NAME with get_range_info with
>> wide_int_to_tree. In this case we will not have TREE_OVERFLOW set.
>>
>> 2. When we vrp_meet/vrp_intersect_ranges two ranges. It does int_const_binop
>> but AFAIK this does not set TREE_OVERFLOW.
>>
>> 3. When we create range from constant. This is the problem bit and we need
>> to clear TREE_OVERFLOW here.
>>
>> Attached patch clears the TREE_OVERFLOW in 3rd case. Bootstrap and
>> regression testing are ongoing. Is this OK if there is no regression.
>
> Ok.  Though it would be nice to drop it at the source (that is, the point we
> initialize the IPA-CP lattice and its modifications).

In ipa_compute_jump_function_for_egde, value_range lattice is not set 
for constants as this information is already there in IPA_JF_CONSTANT. 
That is, we initialize only when we get it from get_range_info 
(SSA_NAMES); others are set to unknown. Though we can set it at this 
point, it can be inefficient in terms of streaming in/out this data. 
While propagating we get it from IPA_JF_CONSTANT.

Thanks,
Kugan

> Richard.
>
>> Thanks,
>> Kugan
>>
>>
>> gcc/ChangeLog:
>>
>> 2016-09-23  Kugan Vivekanandarajah  <kuganv@linaro.org>
>>
>>         PR ipa/77677
>>         * ipa-cp.c (propagate_vr_accross_jump_function):Drop TREE_OVERFLOW
>>         from constant while creating value range.
>>
>> gcc/testsuite/ChangeLog:
>>
>> 2016-09-23  Kugan Vivekanandarajah  <kuganv@linaro.org>
>>
>>         PR ipa/77677
>>         * gcc.dg/torture/pr77677.c: New test.