Re: [Xen-devel] [PATCH] x86/string: Use compiler builtins whenever possible

[Andrew Cooper]

On 09/08/16 15:39, Jan Beulich wrote:
 On 09.08.16 at 16:27,  wrote:
>> On 09/08/16 15:01, Jan Beulich wrote:
>> On 09.08.16 at 14:41,  wrote:
 The use of -fno-builtin inhibits this automatic transformation.  Manually
 tranform the callsites.  This causes constructs such as strlen("literal") 
 to
 be evaluated at compile time, and certain simple operations to be replaced
 with repeated string operations.
>>> Iirc this was specifically avoided in at least older Linux because of
>>> compiler issues. Are you reasonably certain that those issues
>>> predate gcc 4.1.x (and don't exist at all in clang)? Other than that
>>> I'm of course fine with this patch.
>> Do you have any reference to these issues?  I cant find any reference.
> This was when I barley started working with Linux, i.e. in late 2.4.x
> or early 2.6.x times. Querying Google seems to indicate that
> someone reported successful use in 2.6.3.
>
>> I notice that Linux doesn't use these builtins, even today, but does
>> have arch specific implementations which are more efficient than our
>> common ones.  Still, it doesn't appear to be able to do any compile-time
>> evaluation.
> Isn't it surprising that no-one has submitted Linux patches to
> change that? Anyway, please consider the patch ack-ed. We
> can always revert it in case we run into problems (and once
> my hardware setup will be fully functional again, I will test with
> a reasonably old gcc anyway).

It turns out that in its current form, it breaks the clang build, citing

string.c:45:5: error: definition of builtin function '__builtin_strcasecmp'
int strcasecmp(const char *s1, const char *s2)
^
/local/xen.git/xen/include/asm/string.h:18:28: note: expanded from macro
'strcasecmp'
#define strcasecmp(s1, s2) __builtin_strcasecmp(s1, s2)
   ^

which is something GCC clearly doesn't mind so much.

I will see about reworking it.

~Andrew

___
Xen-devel mailing list
Xen-devel@lists.xen.org
https://lists.xen.org/xen-devel

Re: [Xen-devel] [PATCH] x86/string: Use compiler builtins whenever possible

[Jan Beulich]

>>> On 09.08.16 at 16:27,  wrote:
> On 09/08/16 15:01, Jan Beulich wrote:
> On 09.08.16 at 14:41,  wrote:
>>> The use of -fno-builtin inhibits this automatic transformation.  Manually
>>> tranform the callsites.  This causes constructs such as strlen("literal") to
>>> be evaluated at compile time, and certain simple operations to be replaced
>>> with repeated string operations.
>> Iirc this was specifically avoided in at least older Linux because of
>> compiler issues. Are you reasonably certain that those issues
>> predate gcc 4.1.x (and don't exist at all in clang)? Other than that
>> I'm of course fine with this patch.
> 
> Do you have any reference to these issues?  I cant find any reference.

This was when I barley started working with Linux, i.e. in late 2.4.x
or early 2.6.x times. Querying Google seems to indicate that
someone reported successful use in 2.6.3.

> I notice that Linux doesn't use these builtins, even today, but does
> have arch specific implementations which are more efficient than our
> common ones.  Still, it doesn't appear to be able to do any compile-time
> evaluation.

Isn't it surprising that no-one has submitted Linux patches to
change that? Anyway, please consider the patch ack-ed. We
can always revert it in case we run into problems (and once
my hardware setup will be fully functional again, I will test with
a reasonably old gcc anyway).

Jan


___
Xen-devel mailing list
Xen-devel@lists.xen.org
https://lists.xen.org/xen-devel

Re: [Xen-devel] [PATCH] x86/string: Use compiler builtins whenever possible

[Andrew Cooper]

On 09/08/16 15:01, Jan Beulich wrote:
 On 09.08.16 at 14:41,  wrote:
>> The use of -fno-builtin inhibits this automatic transformation.  Manually
>> tranform the callsites.  This causes constructs such as strlen("literal") to
>> be evaluated at compile time, and certain simple operations to be replaced
>> with repeated string operations.
> Iirc this was specifically avoided in at least older Linux because of
> compiler issues. Are you reasonably certain that those issues
> predate gcc 4.1.x (and don't exist at all in clang)? Other than that
> I'm of course fine with this patch.

Do you have any reference to these issues?  I cant find any reference.

I notice that Linux doesn't use these builtins, even today, but does
have arch specific implementations which are more efficient than our
common ones.  Still, it doesn't appear to be able to do any compile-time
evaluation.

~Andrew

___
Xen-devel mailing list
Xen-devel@lists.xen.org
https://lists.xen.org/xen-devel

Re: [Xen-devel] [PATCH] x86/string: Use compiler builtins whenever possible

[Jan Beulich]

>>> On 09.08.16 at 14:41,  wrote:
> The use of -fno-builtin inhibits this automatic transformation.  Manually
> tranform the callsites.  This causes constructs such as strlen("literal") to
> be evaluated at compile time, and certain simple operations to be replaced
> with repeated string operations.

Iirc this was specifically avoided in at least older Linux because of
compiler issues. Are you reasonably certain that those issues
predate gcc 4.1.x (and don't exist at all in clang)? Other than that
I'm of course fine with this patch.

Jan


___
Xen-devel mailing list
Xen-devel@lists.xen.org
https://lists.xen.org/xen-devel