[PATCH] D11969: Avoid using of DataLayout::getPointerSize() without address space argument in DWARF info generation

Thu Aug 13 18:59:19 PDT 2015

> I think if you want to have disparate pointer sizes here you're going to need to change the DataLayout to encode a pointer size per address space or look at the places where we call getPointerSize() and actually somehow get the real size of the variable via something else. This might end up necessarily going back to DataLayout and a pointer size per address space.

Yes, the only issue is source variable is not available at these places, so needs to be passed. The current implementation always query for a pointer in address space zero, which is even more incorrect.

Stas

From: Eric Christopher [mailto:echristo at gmail.com]
Sent: Thursday, August 13, 2015 6:55 PM
To: Mekhanoshin, Stanislav <Stanislav.Mekhanoshin at amd.com>; Mehdi Amini <mehdi.amini at apple.com>
Cc: reviews+D11969+public+c218119e32184906 at reviews.llvm.org; Arsenault, Matthew <Matthew.Arsenault at amd.com>; llvm-commits at lists.llvm.org
Subject: Re: [PATCH] D11969: Avoid using of DataLayout::getPointerSize() without address space argument in DWARF info generation

On Thu, Aug 13, 2015 at 6:46 PM Mekhanoshin, Stanislav <Stanislav.Mekhanoshin at amd.com<mailto:Stanislav.Mekhanoshin at amd.com>> wrote:
> If it is to replace this with a getPointerSize() that has no clear attached semantic, I don’t see this as a progress.

It looks like the question now is in the correct name of the function. It really looks suspicious to call it CodePointerSize in this case. Largest seems to be better because that is what actually happens there - pointer size is used in different contexts, but largest possible size works for all of them.

The trick is that pointer size is going to be a dwarf type as pointer to <thing> and then used all over the place. It's a clever, but inaccurate workaround that "largest wins" here. It would stop you from correctly using something like say lldb to reconstruct function calls etc.

I think if you want to have disparate pointer sizes here you're going to need to change the DataLayout to encode a pointer size per address space or look at the places where we call getPointerSize() and actually somehow get the real size of the variable via something else. This might end up necessarily going back to DataLayout and a pointer size per address space.

-eric

Stas

-----Original Message-----
From: Mehdi Amini [mailto:mehdi.amini at apple.com<mailto:mehdi.amini at apple.com>]
Sent: Thursday, August 13, 2015 6:32 PM
To: Mekhanoshin, Stanislav <Stanislav.Mekhanoshin at amd.com<mailto:Stanislav.Mekhanoshin at amd.com>>
Cc: reviews+D11969+public+c218119e32184906 at reviews.llvm.org<mailto:reviews%2BD11969%2Bpublic%2Bc218119e32184906 at reviews.llvm.org>; Arsenault, Matthew <Matthew.Arsenault at amd.com<mailto:Matthew.Arsenault at amd.com>>; echristo at gmail.com<mailto:echristo at gmail.com>; llvm-commits at lists.llvm.org<mailto:llvm-commits at lists.llvm.org>
Subject: Re: [PATCH] D11969: Avoid using of DataLayout::getPointerSize() without address space argument in DWARF info generation

> On Aug 13, 2015, at 6:26 PM, Mekhanoshin, Stanislav <Stanislav.Mekhanoshin at amd.com<mailto:Stanislav.Mekhanoshin at amd.com>> wrote:
>
> Yes, I agree. A better longer term solution will require to dig in into each use of it.
> At the very least I think we need to remove TargetMachine.getPointerSize() and AsmPrinter.getPointerSize() not to increase a number of such functions without an argument.

If it is to replace this with a getPointerSize() that has no clear attached semantic, I don’t see this as a progress.

—
Mehdi

> I'm testing this conservative patch now.
>
> Stas
>
> -----Original Message-----
> From: Mehdi Amini [mailto:mehdi.amini at apple.com<mailto:mehdi.amini at apple.com>]
> Sent: Thursday, August 13, 2015 6:21 PM
> To: reviews+D11969+public+c218119e32184906 at reviews.llvm.org<mailto:reviews%2BD11969%2Bpublic%2Bc218119e32184906 at reviews.llvm.org>
> Cc: Mekhanoshin, Stanislav <Stanislav.Mekhanoshin at amd.com<mailto:Stanislav.Mekhanoshin at amd.com>>; Arsenault, Matthew <Matthew.Arsenault at amd.com<mailto:Matthew.Arsenault at amd.com>>; echristo at gmail.com<mailto:echristo at gmail.com>; llvm-commits at lists.llvm.org<mailto:llvm-commits at lists.llvm.org>
> Subject: Re: [PATCH] D11969: Avoid using of DataLayout::getPointerSize() without address space argument in DWARF info generation
>
>
>> On Aug 13, 2015, at 5:35 PM, Stanislav Mekhanoshin <stanislav.mekhanoshin at amd.com<mailto:stanislav.mekhanoshin at amd.com>> wrote:
>>
>> rampitec added a comment.
>>
>>> In http://reviews.llvm.org/D11969#223472, @joker.eph wrote:
>>
>>>
>>
>>>> - The DataLayout does not defined anything about which address space the code resides in
>>
>>>
>>
>>
>> I have checked with our engineer doing debug info, and it he confirmed these uses are not related to code addresses,
>
> Not nice, my local debug expert had a different interpretation :(
>
> It seems that every single use of this getPointerSize() will need to be studied to see what it really means and what is the “right” answer.
>
> Right now MCDwarf.cpp is the only user of MCAsmInfo::getPointerSize() as far I can see.
>
>> but generally to variable addresses. In fact in our case all code addresses are redirected, so we should not even hit it. Then a bigger type holding pointers is an inefficiency, but not a bug.
>>
>> I.e. it looks like the name "CodePointerSize" is not correct, a "PointerSize" makes more sense. Maybe a "getLargestPointerSize()" and a "LargestPointerSize" would be better here.
>> It can even be calculated from the DataLayout itself.
>
> It is not clear to me that it is the *right* answer to provide a “getLargestPointerSize()”, even though it seems that would be enough for your use case
>
> —
> Mehdi
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20150814/5660dff5/attachment.html>