[cfe-dev] Crash on conditional operation with address_space pointer

[John McCall via cfe-dev]

> On Aug 3, 2018, at 7:56 PM, Leonard Chan <leonardchan at google.com> wrote:
> Oh. Should an error be printed here instead, or it's fine to let it
> crash I suppose since the code shouldn't be generate in the first
> place?
> 
> It seems that at least 3 tests expect this conditional operator to
> print a "pointer type mismatch" warning.

The compiler should definitely never crash.  That sounds like the right diagnostic
we should be emitting, except I don't think it can be just a warning for an
address-space mismatch.

John.

> On Fri, Aug 3, 2018 at 4:50 PM John McCall <rjmccall at apple.com> wrote:
>> 
>>> On Aug 3, 2018, at 7:42 PM, Leonard Chan <leonardchan at google.com> wrote:
>>> Does this mean the check during the BitCast codegen shouldn't be there
>>> in the first place?
>> 
>> We should not be generating code for this program, correct.
>> 
>> John.
>> 
>>> On Fri, Aug 3, 2018 at 4:28 PM John McCall <rjmccall at apple.com> wrote:
>>>> 
>>>>> On Aug 3, 2018, at 7:03 PM, Leonard Chan via cfe-dev <cfe-dev at lists.llvm.org> wrote:
>>>>> Compiling the following causes clang to crash
>>>>> 
>>>>> char *cmp(__attribute__((address_space(1))) char *x,
>>>>> __attribute__((address_space(2))) char *y) {
>>>>> return x < y ? x : y;
>>>>> }
>>>>> 
>>>>> with the message: "wrong cast for pointers in different address
>>>>> spaces(must be an address space cast)!"
>>>>> 
>>>>> The problem seems to be that BitCast is used when
>>>>> AddressSpaceConversion should be used instead when casting the `y` to
>>>>> a void *. The error is thrown during code gen where a BitCast expects
>>>>> the source and dest address space to be the same.
>>>>> 
>>>>> I submitted a patch (https://reviews.llvm.org/D50278) for a fix where
>>>>> we check for differing address spaces in the operands and perform an
>>>>> AddressSpaceConversion cast if anyone wants to review it. It seems
>>>>> that AddressSpaceConversions though were only available when using
>>>>> OpenCL. I'm not familiar with OpenCL and not sure if this was
>>>>> intended, but it fixes the crash while still retaining the warning of
>>>>> using mismatched pointers and all the clang tests pass.
>>>> 
>>>> If address spaces 1 and 2 don't overlap — and absent target-specific
>>>> information, the assumption is that they don't — then neither the comparison
>>>> nor the conditional operator here should type-check.
>>>> 
>>>> John.
>>