S390: Do not clobber r13 with memcpy on 31bit with copies >1MB.
Commit Message
Hi,
If the default memcpy variant is called with a length of >1MB on 31bit,
r13 is clobbered as the algorithm is switching to mvcle. The mvcle code
returns without restoring r13. All other cases are restoring r13.
If memcpy is called from outside libc the ifunc resolver will only
select this variant if running on machines older than z10.
Otherwise or if memcpy is called from inside libc,
this default memcpy variant is called.
The testcase timezone/tst-tzset is triggering this issue in some
combinations of gcc versions and optimization levels.
This bug was introduced in commit 04bb21ac93e90d7696bcaf8febe2b2dd2d83585a
and thus is a regression compared to glibc 2.23 release. It should
be committed before glibc 2.24 release.
This patch removes the usage of r13 at all. Thus it is not saved
and restored. The base address for execute-instruction is now stored
in r5 which is obtained after r5 is not needed anymore as
256byte block counter.
Okay to commit before release 2.24?
Bye
Stefan
ChangeLog:
* sysdeps/s390/s390-32/memcpy.S (memcpy): Eliminate the usage
of r13 as it is not restored in mvcle case.
Comments
PING.
If there is no objection, I'll commit this patch the next days
as it is an s390 related regression compared to glibc 2.23 release.
On 07/13/2016 12:21 PM, Stefan Liebler wrote:
> Hi,
>
> If the default memcpy variant is called with a length of >1MB on 31bit,
> r13 is clobbered as the algorithm is switching to mvcle. The mvcle code
> returns without restoring r13. All other cases are restoring r13.
>
> If memcpy is called from outside libc the ifunc resolver will only
> select this variant if running on machines older than z10.
> Otherwise or if memcpy is called from inside libc,
> this default memcpy variant is called.
> The testcase timezone/tst-tzset is triggering this issue in some
> combinations of gcc versions and optimization levels.
>
> This bug was introduced in commit 04bb21ac93e90d7696bcaf8febe2b2dd2d83585a
> and thus is a regression compared to glibc 2.23 release. It should
> be committed before glibc 2.24 release.
>
> This patch removes the usage of r13 at all. Thus it is not saved
> and restored. The base address for execute-instruction is now stored
> in r5 which is obtained after r5 is not needed anymore as
> 256byte block counter.
>
> Okay to commit before release 2.24?
>
> Bye
> Stefan
>
> ChangeLog:
>
> * sysdeps/s390/s390-32/memcpy.S (memcpy): Eliminate the usage
> of r13 as it is not restored in mvcle case.
Sorry Stefan, for some reason I missed this thread. I think it is
ok for 2.24.
On 19/07/2016 03:26, Stefan Liebler wrote:
> PING.
>
> If there is no objection, I'll commit this patch the next days
> as it is an s390 related regression compared to glibc 2.23 release.
>
> On 07/13/2016 12:21 PM, Stefan Liebler wrote:
>> Hi,
>>
>> If the default memcpy variant is called with a length of >1MB on 31bit,
>> r13 is clobbered as the algorithm is switching to mvcle. The mvcle code
>> returns without restoring r13. All other cases are restoring r13.
>>
>> If memcpy is called from outside libc the ifunc resolver will only
>> select this variant if running on machines older than z10.
>> Otherwise or if memcpy is called from inside libc,
>> this default memcpy variant is called.
>> The testcase timezone/tst-tzset is triggering this issue in some
>> combinations of gcc versions and optimization levels.
>>
>> This bug was introduced in commit 04bb21ac93e90d7696bcaf8febe2b2dd2d83585a
>> and thus is a regression compared to glibc 2.23 release. It should
>> be committed before glibc 2.24 release.
>>
>> This patch removes the usage of r13 at all. Thus it is not saved
>> and restored. The base address for execute-instruction is now stored
>> in r5 which is obtained after r5 is not needed anymore as
>> 256byte block counter.
>>
>> Okay to commit before release 2.24?
>>
>> Bye
>> Stefan
>>
>> ChangeLog:
>>
>> * sysdeps/s390/s390-32/memcpy.S (memcpy): Eliminate the usage
>> of r13 as it is not restored in mvcle case.
>
On 07/19/2016 03:12 PM, Adhemerval Zanella wrote:
> Sorry Stefan, for some reason I missed this thread. I think it is
> ok for 2.24.
>
No problem. I've commited it.
commit a671883eb41bc3179e7a435e6c2857a913757bf1
Author: Stefan Liebler <stli@linux.vnet.ibm.com>
Date: Wed Jul 13 10:46:10 2016 +0200
S390: Do not clobber r13 with memcpy on 31bit with copies >1MB.
If the default memcpy variant is called with a length of >1MB on 31bit,
r13 is clobbered as the algorithm is switching to mvcle. The mvcle code
returns without restoring r13. All other cases are restoring r13.
If memcpy is called from outside libc the ifunc resolver will only select
this variant if running on machines older than z10.
Otherwise or if memcpy is called from inside libc, this default memcpy
variant is called.
The testcase timezone/tst-tzset is triggering this issue in some combinations
of gcc versions and optimization levels.
This bug was introduced in commit 04bb21ac93e90d7696bcaf8febe2b2dd2d83585a
and thus is a regression compared to glibc 2.23 release. It should
be committed before glibc 2.24 release.
This patch removes the usage of r13 at all. Thus it is not saved and restored.
The base address for execute-instruction is now stored in r5 which is obtained
after r5 is not needed anymore as 256byte block counter.
ChangeLog:
* sysdeps/s390/s390-32/memcpy.S (memcpy): Eliminate the usage
of r13 as it is not restored in mvcle case.
@@ -42,20 +42,18 @@ ENTRY(memcpy)
.machine "g5"
lr %r1,%r2 # r1: Use as dest ; r2: Return dest
.L_G5_start:
- st %r13,52(%r15)
- .cfi_offset 13, -44
- basr %r13,0
-.L_G5_16:
ltr %r4,%r4
- je .L_G5_4
+ je .L_G5_99
ahi %r4,-1
lr %r5,%r4
srl %r5,8
ltr %r5,%r5
jne .L_G5_13
- ex %r4,.L_G5_17-.L_G5_16(%r13)
.L_G5_4:
- l %r13,52(%r15)
+ basr %r5,0
+.L_G5_16:
+ ex %r4,.L_G5_17-.L_G5_16(%r5)
+.L_G5_99:
br %r14
.L_G5_13:
chi %r5,4096 # Switch to mvcle for copies >1MB
@@ -65,7 +63,6 @@ ENTRY(memcpy)
la %r1,256(%r1)
la %r3,256(%r3)
brct %r5,.L_G5_12
- ex %r4,.L_G5_17-.L_G5_16(%r13)
j .L_G5_4
.L_G5_17:
mvc 0(1,%r1),0(%r3)