[Pixman] Patch to fix test for GCC vector extensions

[Rob Tsuk]

From 662b17d0d691b567a3343af3439f5fbefe989584 Mon Sep 17 00:00:00 2001
From: Rob Tsuk <***@google.com>
Date: Mon, 7 Aug 2017 07:11:10 -0700
Subject: [PATCH 1/1] Fix test for GCC vector extensions

The previous test was getting a false positive for
clang version 4.
---
configure.ac | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/configure.ac b/configure.ac
index e833e45..cbebc82 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1101,7 +1101,7 @@ support_for_gcc_vector_extensions=no
AC_MSG_CHECKING(for GCC vector extensions)
AC_LINK_IFELSE([AC_LANG_SOURCE([[
unsigned int __attribute__ ((vector_size(16))) e, a, b;
-int main (void) { e = a - ((b << 27) + (b >> (32 - 27))) + 1; return e[0]; }
+int main (void) { __builtin_shuffle(a,b);e = a - ((b << 27) + (b >> (32 - 27))) + 1; return e[0]; }
]])], support_for_gcc_vector_extensions=yes)

if test x$support_for_gcc_vector_extensions = xyes; then

[Siarhei Siamashka]

On Mon, 7 Aug 2017 07:52:29 -0700

Post by Rob Tsuk
From 662b17d0d691b567a3343af3439f5fbefe989584 Mon Sep 17 00:00:00 2001
Date: Mon, 7 Aug 2017 07:11:10 -0700
Subject: [PATCH 1/1] Fix test for GCC vector extensions
The previous test was getting a false positive for
clang version 4.
---
configure.ac | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/configure.ac b/configure.ac
index e833e45..cbebc82 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1101,7 +1101,7 @@ support_for_gcc_vector_extensions=no
AC_MSG_CHECKING(for GCC vector extensions)
AC_LINK_IFELSE([AC_LANG_SOURCE([[
unsigned int __attribute__ ((vector_size(16))) e, a, b;
-int main (void) { e = a - ((b << 27) + (b >> (32 - 27))) + 1; return e[0]; }
+int main (void) { __builtin_shuffle(a,b);e = a - ((b << 27) + (b >> (32 - 27))) + 1; return e[0]; }
]])], support_for_gcc_vector_extensions=yes)
if test x$support_for_gcc_vector_extensions = xyes; then

Thanks, the patch looks mostly good. We indeed have to add
the use of __builtin_shuffle() to the configure check code
snippet because it is used by pixman and Clang implements
this functionality in a different and incompatible way.

We did not see this problem with older versions of Clang
because it did not support the shift vector by scalar operation
in its vector extensions.

Too bad that GCC and Clang parted their ways and we can't have
a single unified implementation. The whole reason to use GCC
vector extension was that we could avoid platform-dependent
intrinsics in this code.

About the patch. It would be great to use the result of
__builtin_shuffle() in the intermediate calculations just in
case. See
https://cgit.freedesktop.org/pixman/commit/?id=a566f627dbd6ea8f2cba70a446e62caaa2ecbd26

[Rob Tsuk]

Your proposed modification to my patch makes sense. I’m not certain when I”ll have time to perform it.

Post by Siarhei Siamashka
On Mon, 7 Aug 2017 07:52:29 -0700

Post by Rob Tsuk
From 662b17d0d691b567a3343af3439f5fbefe989584 Mon Sep 17 00:00:00 2001
Date: Mon, 7 Aug 2017 07:11:10 -0700
Subject: [PATCH 1/1] Fix test for GCC vector extensions
The previous test was getting a false positive for
clang version 4.
---
configure.ac | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/configure.ac b/configure.ac
index e833e45..cbebc82 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1101,7 +1101,7 @@ support_for_gcc_vector_extensions=no
AC_MSG_CHECKING(for GCC vector extensions)
AC_LINK_IFELSE([AC_LANG_SOURCE([[
unsigned int __attribute__ ((vector_size(16))) e, a, b;
-int main (void) { e = a - ((b << 27) + (b >> (32 - 27))) + 1; return e[0]; }
+int main (void) { __builtin_shuffle(a,b);e = a - ((b << 27) + (b >> (32 - 27))) + 1; return e[0]; }
]])], support_for_gcc_vector_extensions=yes)
if test x$support_for_gcc_vector_extensions = xyes; then

Thanks, the patch looks mostly good. We indeed have to add
the use of __builtin_shuffle() to the configure check code
snippet because it is used by pixman and Clang implements
this functionality in a different and incompatible way.
We did not see this problem with older versions of Clang
because it did not support the shift vector by scalar operation
in its vector extensions.
Too bad that GCC and Clang parted their ways and we can't have
a single unified implementation. The whole reason to use GCC
vector extension was that we could avoid platform-dependent
intrinsics in this code.
About the patch. It would be great to use the result of
__builtin_shuffle() in the intermediate calculations just in
case. See
https://cgit.freedesktop.org/pixman/commit/?id=a566f627dbd6ea8f2cba70a446e62caaa2ecbd26 <https://cgit.freedesktop.org/pixman/commit/?id=a566f627dbd6ea8f2cba70a446e62caaa2ecbd26>
--
Best regards,
Siarhei Siamashka

[Siarhei Siamashka]

On Mon, 18 Sep 2017 07:24:49 -0700

Post by Siarhei Siamashka
On Mon, 7 Aug 2017 07:52:29 -0700

Post by Rob Tsuk
From 662b17d0d691b567a3343af3439f5fbefe989584 Mon Sep 17 00:00:00 2001
Date: Mon, 7 Aug 2017 07:11:10 -0700
Subject: [PATCH 1/1] Fix test for GCC vector extensions
The previous test was getting a false positive for
clang version 4.
---
configure.ac | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/configure.ac b/configure.ac
index e833e45..cbebc82 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1101,7 +1101,7 @@ support_for_gcc_vector_extensions=no
AC_MSG_CHECKING(for GCC vector extensions)
AC_LINK_IFELSE([AC_LANG_SOURCE([[
unsigned int __attribute__ ((vector_size(16))) e, a, b;
-int main (void) { e = a - ((b << 27) + (b >> (32 - 27))) + 1; return e[0]; }
+int main (void) { __builtin_shuffle(a,b);e = a - ((b << 27) + (b >> (32 - 27))) + 1; return e[0]; }
]])], support_for_gcc_vector_extensions=yes)
if test x$support_for_gcc_vector_extensions = xyes; then

Thanks, the patch looks mostly good. We indeed have to add
the use of __builtin_shuffle() to the configure check code
snippet because it is used by pixman and Clang implements
this functionality in a different and incompatible way.
We did not see this problem with older versions of Clang
because it did not support the shift vector by scalar operation
in its vector extensions.
Too bad that GCC and Clang parted their ways and we can't have
a single unified implementation. The whole reason to use GCC
vector extension was that we could avoid platform-dependent
intrinsics in this code.
About the patch. It would be great to use the result of
__builtin_shuffle() in the intermediate calculations just in
case. See
https://cgit.freedesktop.org/pixman/commit/?id=a566f627dbd6ea8f2cba70a446e62caaa2ecbd26 <https://cgit.freedesktop.org/pixman/commit/?id=a566f627dbd6ea8f2cba70a446e62caaa2ecbd26>

Your proposed modification to my patch makes sense. I’m not certain when I”ll have time to perform it.

It does not really require much time. I just did that for you:
https://lists.freedesktop.org/archives/pixman/2017-September/004680.html

[Rob Tsuk]

Thanks! It had been so long since I’d submitted the patch that I’d forgotten on which machine I’d done the work and what I needed to do to get set up to modify it. I’m glad you found it easy to do for me.

Post by Siarhei Siamashka
On Mon, 18 Sep 2017 07:24:49 -0700

Post by Siarhei Siamashka
On Mon, 7 Aug 2017 07:52:29 -0700

Post by Rob Tsuk
From 662b17d0d691b567a3343af3439f5fbefe989584 Mon Sep 17 00:00:00 2001
Date: Mon, 7 Aug 2017 07:11:10 -0700
Subject: [PATCH 1/1] Fix test for GCC vector extensions
The previous test was getting a false positive for
clang version 4.
---
configure.ac | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/configure.ac b/configure.ac
index e833e45..cbebc82 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1101,7 +1101,7 @@ support_for_gcc_vector_extensions=no
AC_MSG_CHECKING(for GCC vector extensions)
AC_LINK_IFELSE([AC_LANG_SOURCE([[
unsigned int __attribute__ ((vector_size(16))) e, a, b;
-int main (void) { e = a - ((b << 27) + (b >> (32 - 27))) + 1; return e[0]; }
+int main (void) { __builtin_shuffle(a,b);e = a - ((b << 27) + (b >> (32 - 27))) + 1; return e[0]; }
]])], support_for_gcc_vector_extensions=yes)
if test x$support_for_gcc_vector_extensions = xyes; then

Thanks, the patch looks mostly good. We indeed have to add
the use of __builtin_shuffle() to the configure check code
snippet because it is used by pixman and Clang implements
this functionality in a different and incompatible way.
We did not see this problem with older versions of Clang
because it did not support the shift vector by scalar operation
in its vector extensions.
Too bad that GCC and Clang parted their ways and we can't have
a single unified implementation. The whole reason to use GCC
vector extension was that we could avoid platform-dependent
intrinsics in this code.
About the patch. It would be great to use the result of
__builtin_shuffle() in the intermediate calculations just in
case. See
https://cgit.freedesktop.org/pixman/commit/?id=a566f627dbd6ea8f2cba70a446e62caaa2ecbd26 <https://cgit.freedesktop.org/pixman/commit/?id=a566f627dbd6ea8f2cba70a446e62caaa2ecbd26>

Your proposed modification to my patch makes sense. I’m not certain when I”ll have time to perform it.

https://lists.freedesktop.org/archives/pixman/2017-September/004680.html
--
Best regards,
Siarhei Siamashka