This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
Re: [PATCH] Fix i386 memmove issue [BZ #22644]
- From: Andrew Senkevich <andrew dot n dot senkevich at gmail dot com>
- To: "H.J. Lu" <hjl dot tools at gmail dot com>
- Cc: Florian Weimer <fweimer at redhat dot com>, Szabolcs Nagy <szabolcs dot nagy at arm dot com>, Andreas Schwab <schwab at suse dot de>, nd <nd at arm dot com>, libc-alpha <libc-alpha at sourceware dot org>, Max Horn <max at quendi dot de>, thomas at grindinggear dot com
- Date: Mon, 19 Mar 2018 21:32:27 +0100
- Subject: Re: [PATCH] Fix i386 memmove issue [BZ #22644]
- References: <CAMXFM3s5b2a6o_CMS0cJ35oE3SuJUHssqLAUY94LSZhEEtG71A@mail.gmail.com> <mvm606ts2w6.fsf@suse.de> <CAMXFM3vqw7eiUT1XYVZWyW=eJkgEpjNc74A8mbmdKuR_hywN3w@mail.gmail.com> <CAMe9rOo4Ys=ScEKKnuv5XvYqTHjvFxSkXjch8ya8sX7dfP+q4Q@mail.gmail.com> <CAMXFM3vA=-LSK5VtKhhbs0g9faBdG0GjoAbNb6OPEbSt8dAA7A@mail.gmail.com> <mvmlgeojjmi.fsf@suse.de> <597954a1-85da-e524-1454-27c46af57413@redhat.com> <e7d81f18-315c-4434-a734-41029c0d387f@arm.com> <8c8ab285-bddc-90e6-a50e-7be89703367a@redhat.com> <CAMXFM3s39wwjHt=5PBp=Np8Rp=mXHfPwMq6_evtfx9MV08-mtQ@mail.gmail.com> <CAMe9rOoCe+qiOFK8=dwxm+eB=_wSoHbXdid0RKSUHqOtM7Wwfg@mail.gmail.com> <CAMXFM3vWVVbe0unZaueF+u62KJ-Sxa6ttMdZRaf_UesxMU=ZKA@mail.gmail.com> <CAMe9rOqsVpCV24f7g=kpFeOAQ_rkHuj-ZN5TDqs7BZ814BDQkg@mail.gmail.com>
2018-03-19 20:38 GMT+01:00 H.J. Lu <hjl.tools@gmail.com>:
> On Mon, Mar 19, 2018 at 12:29 PM, Andrew Senkevich
> <andrew.n.senkevich@gmail.com> wrote:
>> 2018-03-19 18:57 GMT+01:00 H.J. Lu <hjl.tools@gmail.com>:
>>> On Mon, Mar 19, 2018 at 10:52 AM, Andrew Senkevich
>>> <andrew.n.senkevich@gmail.com> wrote:
>>>> 2018-03-19 16:33 GMT+01:00 Florian Weimer <fweimer@redhat.com>:
>>>>> On 03/19/2018 03:25 PM, Szabolcs Nagy wrote:
>>>>>>
>>>>>> i thought not using MAP_FIXED is the 'non-overriding MAP_FIXED variant'
>>>>
>>>> Hmm, I tried so and had got much bigger address. But now I see it
>>>> works (with size just 0x20000000 and 0x70000000 as a hint) in both 64
>>>> and 32 bit cases, so proposing the following:
>>>>
>>>> diff --git a/string/test-memmove.c b/string/test-memmove.c
>>>> index edc7a4c..fa4037d 100644
>>>> --- a/string/test-memmove.c
>>>> +++ b/string/test-memmove.c
>>>> @@ -24,6 +24,7 @@
>>>> # define TEST_NAME "memmove"
>>>> #endif
>>>> #include "test-string.h"
>>>> +#include <support/test-driver.h>
>>>>
>>>> char *simple_memmove (char *, const char *, size_t);
>>>>
>>>> @@ -245,6 +246,59 @@ do_random_tests (void)
>>>> }
>>>> }
>>>>
>>>> +static void
>>>> +do_test2 (void)
>>>> +{
>>>> + size_t size = 0x20000000;
>>>> + uint32_t * large_buf;
>>>> +
>>>> + large_buf = mmap ((void*)0x70000000, size, PROT_READ | PROT_WRITE,
>>>> + MAP_PRIVATE | MAP_ANON, -1, 0);
>>>> +
>>>> + if (large_buf == MAP_FAILED)
>>>> + error (EXIT_UNSUPPORTED, errno, "Large mmap failed");
>>>> +
>>>> + if ((uintptr_t)large_buf > 0x80000000 - 128)
>>>
>>> Don't you need to test if address is too low such that
>>> 0x80000000 - (uintptr_t)large_buf > 0x20000000?
>>
>> Indeed, thanks. Updated below.
>>
>> diff --git a/string/test-memmove.c b/string/test-memmove.c
>> index edc7a4c..9f1437d 100644
>> --- a/string/test-memmove.c
>> +++ b/string/test-memmove.c
>> @@ -24,6 +24,7 @@
>> # define TEST_NAME "memmove"
>> #endif
>> #include "test-string.h"
>> +#include <support/test-driver.h>
>>
>> char *simple_memmove (char *, const char *, size_t);
>>
>> @@ -245,6 +246,60 @@ do_random_tests (void)
>> }
>> }
>>
>> +static void
>> +do_test2 (void)
>> +{
>> + size_t size = 0x20000000;
>> + uint32_t * large_buf;
>> +
>> + large_buf = mmap ((void*)0x70000000, size, PROT_READ | PROT_WRITE,
>> + MAP_PRIVATE | MAP_ANON, -1, 0);
>> +
>> + if (large_buf == MAP_FAILED)
>> + error (EXIT_UNSUPPORTED, errno, "Large mmap failed");
>> +
>> + if ((uintptr_t)large_buf > 0x80000000 - 128
>> + || 0x80000000 - (uintptr_t)large_buf > 0x20000000)
>> + {
>> + error (0, 0,"Large mmap allocated improperly");
>> + ret = EXIT_UNSUPPORTED;
>> + munmap((void *)large_buf, size);
>> + return;
>> + }
>> +
>> + size_t bytes_move = 0x80000000 - (uintptr_t)large_buf;
>> + size_t arr_size = bytes_move / sizeof(uint32_t);
>> + size_t i;
>> +
>> + FOR_EACH_IMPL (impl, 0)
>> + {
>> + for (i = 0; i < arr_size; i++)
>> + large_buf[i] = i;
>> +
>> + uint32_t * dst = &large_buf[33];
>> +
>> +#ifdef TEST_BCOPY
>> + CALL (impl, (char *)large_buf, (char *)dst, bytes_move);
>> +#else
>> + CALL (impl, (char *)dst, (char *)large_buf, bytes_move);
>> +#endif
>> +
>> + for (i = 0; i < arr_size; i++)
>> + {
>> + if (dst[i] != i)
>> + {
>> + error (0, 0,
>> + "Wrong result in function %s dst \"%p\" src \"%p\" offset \"%d\"",
>> + impl->name, dst, large_buf, i);
>> + ret = 1;
>> + break;
>> + }
>> + }
>> + }
>> +
>> + munmap((void *)large_buf, size);
>> +}
>> +
>> int
>> test_main (void)
>> {
>> @@ -284,6 +339,9 @@ test_main (void)
>> }
>>
>> do_random_tests ();
>> +
>> + do_test2 ();
>> +
>> return ret;
>> }
>>
>>
>
> Look good.
>
> Please submit the updated patch with the proper commit message.
[BZ #22644] Fix i386 memmove issue.
* sysdeps/i386/i686/multiarch/memcpy-sse2-unaligned.S: Fixed
branch conditions.
* string/test-memmove.c (do_test2): New testcase.
diff --git a/ChangeLog b/ChangeLog
index 4eb0632..5a0e335 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,3 +1,11 @@
+2018-03-19 Andrew Senkevich <andrew.senkevich@intel.com>
+ Max Horn <max@quendi.de>
+
+ [BZ #22644]
+ * sysdeps/i386/i686/multiarch/memcpy-sse2-unaligned.S: Fixed
+ branch conditions.
+ * string/test-memmove.c (do_test2): New testcase.
+
2018-03-19 Wilco Dijkstra <wdijkstr@arm.com>
* benchtests/bench-timing.h (attribute_hidden): Undefine.
diff --git a/string/test-memmove.c b/string/test-memmove.c
index edc7a4c..9f1437d 100644
--- a/string/test-memmove.c
+++ b/string/test-memmove.c
@@ -24,6 +24,7 @@
# define TEST_NAME "memmove"
#endif
#include "test-string.h"
+#include <support/test-driver.h>
char *simple_memmove (char *, const char *, size_t);
@@ -245,6 +246,60 @@ do_random_tests (void)
}
}
+static void
+do_test2 (void)
+{
+ size_t size = 0x20000000;
+ uint32_t * large_buf;
+
+ large_buf = mmap ((void*)0x70000000, size, PROT_READ | PROT_WRITE,
+ MAP_PRIVATE | MAP_ANON, -1, 0);
+
+ if (large_buf == MAP_FAILED)
+ error (EXIT_UNSUPPORTED, errno, "Large mmap failed");
+
+ if ((uintptr_t)large_buf > 0x80000000 - 128
+ || 0x80000000 - (uintptr_t)large_buf > 0x20000000)
+ {
+ error (0, 0,"Large mmap allocated improperly");
+ ret = EXIT_UNSUPPORTED;
+ munmap((void *)large_buf, size);
+ return;
+ }
+
+ size_t bytes_move = 0x80000000 - (uintptr_t)large_buf;
+ size_t arr_size = bytes_move / sizeof(uint32_t);
+ size_t i;
+
+ FOR_EACH_IMPL (impl, 0)
+ {
+ for (i = 0; i < arr_size; i++)
+ large_buf[i] = i;
+
+ uint32_t * dst = &large_buf[33];
+
+#ifdef TEST_BCOPY
+ CALL (impl, (char *)large_buf, (char *)dst, bytes_move);
+#else
+ CALL (impl, (char *)dst, (char *)large_buf, bytes_move);
+#endif
+
+ for (i = 0; i < arr_size; i++)
+ {
+ if (dst[i] != i)
+ {
+ error (0, 0,
+ "Wrong result in function %s dst \"%p\" src \"%p\" offset \"%d\"",
+ impl->name, dst, large_buf, i);
+ ret = 1;
+ break;
+ }
+ }
+ }
+
+ munmap((void *)large_buf, size);
+}
+
int
test_main (void)
{
@@ -284,6 +339,9 @@ test_main (void)
}
do_random_tests ();
+
+ do_test2 ();
+
return ret;
}
diff --git a/sysdeps/i386/i686/multiarch/memcpy-sse2-unaligned.S
b/sysdeps/i386/i686/multiarch/memcpy-sse2-unaligned.S
index 9c3bbe7..9aa17de 100644
--- a/sysdeps/i386/i686/multiarch/memcpy-sse2-unaligned.S
+++ b/sysdeps/i386/i686/multiarch/memcpy-sse2-unaligned.S
@@ -72,7 +72,7 @@ ENTRY (MEMCPY)
cmp %edx, %eax
# ifdef USE_AS_MEMMOVE
- jg L(check_forward)
+ ja L(check_forward)
L(mm_len_0_or_more_backward):
/* Now do checks for lengths. We do [0..16], [16..32], [32..64], [64..128]
@@ -81,7 +81,7 @@ L(mm_len_0_or_more_backward):
jbe L(mm_len_0_16_bytes_backward)
cmpl $32, %ecx
- jg L(mm_len_32_or_more_backward)
+ ja L(mm_len_32_or_more_backward)
/* Copy [0..32] and return. */
movdqu (%eax), %xmm0
@@ -92,7 +92,7 @@ L(mm_len_0_or_more_backward):
L(mm_len_32_or_more_backward):
cmpl $64, %ecx
- jg L(mm_len_64_or_more_backward)
+ ja L(mm_len_64_or_more_backward)
/* Copy [0..64] and return. */
movdqu (%eax), %xmm0
@@ -107,7 +107,7 @@ L(mm_len_32_or_more_backward):
L(mm_len_64_or_more_backward):
cmpl $128, %ecx
- jg L(mm_len_128_or_more_backward)
+ ja L(mm_len_128_or_more_backward)
/* Copy [0..128] and return. */
movdqu (%eax), %xmm0
@@ -132,7 +132,7 @@ L(mm_len_128_or_more_backward):
add %ecx, %eax
cmp %edx, %eax
movl SRC(%esp), %eax
- jle L(forward)
+ jbe L(forward)
PUSH (%esi)
PUSH (%edi)
PUSH (%ebx)
@@ -269,7 +269,7 @@ L(check_forward):
add %edx, %ecx
cmp %eax, %ecx
movl LEN(%esp), %ecx
- jle L(forward)
+ jbe L(forward)
/* Now do checks for lengths. We do [0..16], [0..32], [0..64], [0..128]
separately. */
--
WBR,
Andrew