|
 |
5de29b |
#
|
|
|
5de29b |
# AVX-512 support for glibc:
|
|
|
5de29b |
#
|
|
|
5de29b |
# Notes: Renamed configure.ac changes to configure.in.
|
|
|
5de29b |
#
|
|
|
5de29b |
# commit aa4de9cea5c07d43caeaca9722c2d417e9a2919c
|
|
|
5de29b |
# Author: H.J. Lu <hjl.tools@gmail.com>
|
|
|
5de29b |
# Date: Fri Mar 14 08:51:25 2014 -0700
|
|
|
5de29b |
#
|
|
|
5de29b |
# Check AVX-512 assembler support first
|
|
|
5de29b |
#
|
|
|
5de29b |
# It checks AVX-512 assembler support first and sets libc_cv_cc_avx512 to
|
|
|
5de29b |
# $libc_cv_asm_avx512, instead of yes. GCC won't support AVX-512 if
|
|
|
5de29b |
# assembler doesn't support it.
|
|
|
5de29b |
#
|
|
|
5de29b |
# * sysdeps/x86_64/configure.ac: Check AVX-512 assembler support
|
|
|
5de29b |
# first. Disable AVX-512 GCC support if assembler doesn't support
|
|
|
5de29b |
# it.
|
|
|
5de29b |
# * sysdeps/x86_64/configure: Regenerated.
|
|
|
5de29b |
#
|
|
|
5de29b |
# commit 2d63a517e4084ec80403cd9f278690fa8b676cc4
|
|
|
5de29b |
# Author: Igor Zamyatin <igor.zamyatin@intel.com>
|
|
|
5de29b |
# Date: Thu Mar 13 11:10:22 2014 -0700
|
|
|
5de29b |
#
|
|
|
5de29b |
# Save and restore AVX-512 zmm registers to x86-64 ld.so
|
|
|
5de29b |
#
|
|
|
5de29b |
# AVX-512 ISA adds 512-bit zmm registers. This patch updates
|
|
|
5de29b |
# _dl_runtime_profile to pass zmm registers to run-time audit. It also
|
|
|
5de29b |
# changes _dl_x86_64_save_sse and _dl_x86_64_restore_sse to upport zmm
|
|
|
5de29b |
# registers, which are called when only when RTLD_PREPARE_FOREIGN_CALL
|
|
|
5de29b |
# is used. Its performance impact is minimum.
|
|
|
5de29b |
#
|
|
|
5de29b |
# * config.h.in (HAVE_AVX512_SUPPORT): New #undef.
|
|
|
5de29b |
# (HAVE_AVX512_ASM_SUPPORT): Likewise.
|
|
|
5de29b |
# * sysdeps/x86_64/bits/link.h (La_x86_64_zmm): New.
|
|
|
5de29b |
# (La_x86_64_vector): Add zmm.
|
|
|
5de29b |
# * sysdeps/x86_64/Makefile (tests): Add tst-audit10.
|
|
|
5de29b |
# (modules-names): Add tst-auditmod10a and tst-auditmod10b.
|
|
|
5de29b |
# ($(objpfx)tst-audit10): New target.
|
|
|
5de29b |
# ($(objpfx)tst-audit10.out): Likewise.
|
|
|
5de29b |
# (tst-audit10-ENV): New.
|
|
|
5de29b |
# (AVX512-CFLAGS): Likewise.
|
|
|
5de29b |
# (CFLAGS-tst-audit10.c): Likewise.
|
|
|
5de29b |
# (CFLAGS-tst-auditmod10a.c): Likewise.
|
|
|
5de29b |
# (CFLAGS-tst-auditmod10b.c): Likewise.
|
|
|
5de29b |
# * sysdeps/x86_64/configure.ac: Set config-cflags-avx512,
|
|
|
5de29b |
# HAVE_AVX512_SUPPORT and HAVE_AVX512_ASM_SUPPORT.
|
|
|
5de29b |
# * sysdeps/x86_64/configure: Regenerated.
|
|
|
5de29b |
# * sysdeps/x86_64/dl-trampoline.S (_dl_runtime_profile): Add
|
|
|
5de29b |
# AVX-512 zmm register support.
|
|
|
5de29b |
# (_dl_x86_64_save_sse): Likewise.
|
|
|
5de29b |
# (_dl_x86_64_restore_sse): Likewise.
|
|
|
5de29b |
# * sysdeps/x86_64/dl-trampoline.h: Updated to support different
|
|
|
5de29b |
# size vector registers.
|
|
|
5de29b |
# * sysdeps/x86_64/link-defines.sym (YMM_SIZE): New.
|
|
|
5de29b |
# (ZMM_SIZE): Likewise.
|
|
|
5de29b |
# * sysdeps/x86_64/tst-audit10.c: New file.
|
|
|
5de29b |
# * sysdeps/x86_64/tst-auditmod10a.c: Likewise.
|
|
|
5de29b |
# * sysdeps/x86_64/tst-auditmod10b.c: Likewise.
|
|
|
5de29b |
#
|
|
|
5de29b |
# In addition adds:
|
|
|
5de29b |
# https://sourceware.org/ml/libc-alpha/2014-09/msg00228.html
|
|
|
5de29b |
# To extend zmm register checking.
|
|
|
5de29b |
#
|
|
|
12745e |
diff -urN glibc-2.17-c758a686/config.h.in glibc-2.17-c758a686/config.h.in
|
|
|
5de29b |
--- glibc-2.17-c758a686/config.h.in 2014-09-10 23:11:14.605787816 -0400
|
|
|
12745e |
+++ glibc-2.17-c758a686/config.h.in 2014-09-10 23:16:36.331167056 -0400
|
|
|
5de29b |
@@ -101,6 +101,12 @@
|
|
|
5de29b |
/* Define if gcc supports VEX encoding. */
|
|
|
5de29b |
#undef HAVE_SSE2AVX_SUPPORT
|
|
|
5de29b |
|
|
|
5de29b |
+/* Define if compiler supports AVX512. */
|
|
|
5de29b |
+#undef HAVE_AVX512_SUPPORT
|
|
|
5de29b |
+
|
|
|
5de29b |
+/* Define if assembler supports AVX512. */
|
|
|
5de29b |
+#undef HAVE_AVX512_ASM_SUPPORT
|
|
|
5de29b |
+
|
|
|
5de29b |
/* Define if gcc supports FMA4. */
|
|
|
5de29b |
#undef HAVE_FMA4_SUPPORT
|
|
|
5de29b |
|
|
|
12745e |
diff -urN glibc-2.17-c758a686/sysdeps/x86/bits/link.h glibc-2.17-c758a686/sysdeps/x86/bits/link.h
|
|
|
5de29b |
--- glibc-2.17-c758a686/sysdeps/x86/bits/link.h 2012-12-24 22:02:13.000000000 -0500
|
|
|
12745e |
+++ glibc-2.17-c758a686/sysdeps/x86/bits/link.h 2014-09-10 23:16:36.331167056 -0400
|
|
|
5de29b |
@@ -66,6 +66,8 @@
|
|
|
5de29b |
typedef float La_x86_64_xmm __attribute__ ((__vector_size__ (16)));
|
|
|
5de29b |
typedef float La_x86_64_ymm
|
|
|
5de29b |
__attribute__ ((__vector_size__ (32), __aligned__ (16)));
|
|
|
5de29b |
+typedef double La_x86_64_zmm
|
|
|
5de29b |
+ __attribute__ ((__vector_size__ (64), __aligned__ (16)));
|
|
|
5de29b |
# else
|
|
|
5de29b |
typedef float La_x86_64_xmm __attribute__ ((__mode__ (__V4SF__)));
|
|
|
5de29b |
# endif
|
|
|
5de29b |
@@ -74,6 +76,7 @@
|
|
|
5de29b |
{
|
|
|
5de29b |
# if __GNUC_PREREQ (4,0)
|
|
|
5de29b |
La_x86_64_ymm ymm[2];
|
|
|
5de29b |
+ La_x86_64_zmm zmm[1];
|
|
|
5de29b |
# endif
|
|
|
5de29b |
La_x86_64_xmm xmm[4];
|
|
|
5de29b |
} La_x86_64_vector __attribute__ ((__aligned__ (16)));
|
|
|
12745e |
diff -urN glibc-2.17-c758a686/sysdeps/x86_64/configure glibc-2.17-c758a686/sysdeps/x86_64/configure
|
|
|
5de29b |
--- glibc-2.17-c758a686/sysdeps/x86_64/configure 2014-09-10 23:11:15.000787061 -0400
|
|
|
12745e |
+++ glibc-2.17-c758a686/sysdeps/x86_64/configure 2014-09-10 23:16:36.338167042 -0400
|
|
|
5de29b |
@@ -91,6 +91,59 @@
|
|
|
5de29b |
|
|
|
5de29b |
fi
|
|
|
5de29b |
|
|
|
5de29b |
+{ $as_echo "$as_me:${as_lineno-$LINENO}: checking for AVX512 support in assembler" >&5
|
|
|
5de29b |
+$as_echo_n "checking for AVX512 support in assembler... " >&6; }
|
|
|
5de29b |
+if ${libc_cv_asm_avx512+:} false; then :
|
|
|
5de29b |
+ $as_echo_n "(cached) " >&6
|
|
|
5de29b |
+else
|
|
|
5de29b |
+ cat > conftest.s <<\EOF
|
|
|
5de29b |
+ vmovdqu64 %zmm0, (%rsp)
|
|
|
5de29b |
+EOF
|
|
|
5de29b |
+if { ac_try='${CC-cc} -c $ASFLAGS conftest.s 1>&5'
|
|
|
5de29b |
+ { { eval echo "\"\$as_me\":${as_lineno-$LINENO}: \"$ac_try\""; } >&5
|
|
|
5de29b |
+ (eval $ac_try) 2>&5
|
|
|
5de29b |
+ ac_status=$?
|
|
|
5de29b |
+ $as_echo "$as_me:${as_lineno-$LINENO}: \$? = $ac_status" >&5
|
|
|
5de29b |
+ test $ac_status = 0; }; }; then
|
|
|
5de29b |
+ libc_cv_asm_avx512=yes
|
|
|
5de29b |
+else
|
|
|
5de29b |
+ libc_cv_asm_avx512=no
|
|
|
5de29b |
+fi
|
|
|
5de29b |
+rm -f conftest*
|
|
|
5de29b |
+fi
|
|
|
5de29b |
+{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $libc_cv_asm_avx512" >&5
|
|
|
5de29b |
+$as_echo "$libc_cv_asm_avx512" >&6; }
|
|
|
5de29b |
+if test $libc_cv_asm_avx512 == yes; then
|
|
|
5de29b |
+ $as_echo "#define HAVE_AVX512_ASM_SUPPORT 1" >>confdefs.h
|
|
|
5de29b |
+
|
|
|
5de29b |
+fi
|
|
|
5de29b |
+
|
|
|
5de29b |
+{ $as_echo "$as_me:${as_lineno-$LINENO}: checking for AVX512 support" >&5
|
|
|
5de29b |
+$as_echo_n "checking for AVX512 support... " >&6; }
|
|
|
5de29b |
+if ${libc_cv_cc_avx512+:} false; then :
|
|
|
5de29b |
+ $as_echo_n "(cached) " >&6
|
|
|
5de29b |
+else
|
|
|
5de29b |
+ if { ac_try='${CC-cc} -mavx512f -xc /dev/null -S -o /dev/null'
|
|
|
5de29b |
+ { { eval echo "\"\$as_me\":${as_lineno-$LINENO}: \"$ac_try\""; } >&5
|
|
|
5de29b |
+ (eval $ac_try) 2>&5
|
|
|
5de29b |
+ ac_status=$?
|
|
|
5de29b |
+ $as_echo "$as_me:${as_lineno-$LINENO}: \$? = $ac_status" >&5
|
|
|
5de29b |
+ test $ac_status = 0; }; }; then :
|
|
|
5de29b |
+ libc_cv_cc_avx512=$libc_cv_asm_avx512
|
|
|
5de29b |
+else
|
|
|
5de29b |
+ libc_cv_cc_avx512=no
|
|
|
5de29b |
+fi
|
|
|
5de29b |
+
|
|
|
5de29b |
+fi
|
|
|
5de29b |
+{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $libc_cv_cc_avx512" >&5
|
|
|
5de29b |
+$as_echo "$libc_cv_cc_avx512" >&6; }
|
|
|
5de29b |
+if test $libc_cv_cc_avx512 = yes; then
|
|
|
5de29b |
+ $as_echo "#define HAVE_AVX512_SUPPORT 1" >>confdefs.h
|
|
|
5de29b |
+
|
|
|
5de29b |
+fi
|
|
|
5de29b |
+config_vars="$config_vars
|
|
|
5de29b |
+config-cflags-avx512 = $libc_cv_cc_avx512"
|
|
|
5de29b |
+
|
|
|
5de29b |
{ $as_echo "$as_me:${as_lineno-$LINENO}: checking for AVX encoding of SSE instructions" >&5
|
|
|
5de29b |
$as_echo_n "checking for AVX encoding of SSE instructions... " >&6; }
|
|
|
5de29b |
if ${libc_cv_cc_sse2avx+:} false; then :
|
|
|
12745e |
diff -urN glibc-2.17-c758a686/sysdeps/x86_64/configure.in glibc-2.17-c758a686/sysdeps/x86_64/configure.in
|
|
|
5de29b |
--- glibc-2.17-c758a686/sysdeps/x86_64/configure.in 2012-12-24 22:02:13.000000000 -0500
|
|
|
12745e |
+++ glibc-2.17-c758a686/sysdeps/x86_64/configure.in 2014-09-10 23:16:36.338167042 -0400
|
|
|
5de29b |
@@ -21,6 +21,30 @@
|
|
|
5de29b |
AC_DEFINE(HAVE_AVX_SUPPORT)
|
|
|
5de29b |
fi
|
|
|
5de29b |
|
|
|
5de29b |
+dnl Check if asm supports AVX512.
|
|
|
5de29b |
+AC_CACHE_CHECK(for AVX512 support in assembler, libc_cv_asm_avx512, [dnl
|
|
|
5de29b |
+cat > conftest.s <<\EOF
|
|
|
5de29b |
+ vmovdqu64 %zmm0, (%rsp)
|
|
|
5de29b |
+EOF
|
|
|
5de29b |
+if AC_TRY_COMMAND(${CC-cc} -c $ASFLAGS conftest.s 1>&AS_MESSAGE_LOG_FD); then
|
|
|
5de29b |
+ libc_cv_asm_avx512=yes
|
|
|
5de29b |
+else
|
|
|
5de29b |
+ libc_cv_asm_avx512=no
|
|
|
5de29b |
+fi
|
|
|
5de29b |
+rm -f conftest*])
|
|
|
5de29b |
+if test $libc_cv_asm_avx512 == yes; then
|
|
|
5de29b |
+ AC_DEFINE(HAVE_AVX512_ASM_SUPPORT)
|
|
|
5de29b |
+fi
|
|
|
5de29b |
+
|
|
|
5de29b |
+dnl Check if -mavx512f works.
|
|
|
5de29b |
+AC_CACHE_CHECK(for AVX512 support, libc_cv_cc_avx512, [dnl
|
|
|
5de29b |
+LIBC_TRY_CC_OPTION([-mavx512f], [libc_cv_cc_avx512=$libc_cv_asm_avx512], [libc_cv_cc_avx512=no])
|
|
|
5de29b |
+])
|
|
|
5de29b |
+if test $libc_cv_cc_avx512 = yes; then
|
|
|
5de29b |
+ AC_DEFINE(HAVE_AVX512_SUPPORT)
|
|
|
5de29b |
+fi
|
|
|
5de29b |
+LIBC_CONFIG_VAR([config-cflags-avx512], [$libc_cv_cc_avx512])
|
|
|
5de29b |
+
|
|
|
5de29b |
dnl Check if -msse2avx works.
|
|
|
5de29b |
AC_CACHE_CHECK(for AVX encoding of SSE instructions, libc_cv_cc_sse2avx, [dnl
|
|
|
5de29b |
LIBC_TRY_CC_OPTION([-msse2avx],
|
|
|
12745e |
diff -urN glibc-2.17-c758a686/sysdeps/x86_64/dl-trampoline.h glibc-2.17-c758a686/sysdeps/x86_64/dl-trampoline.h
|
|
|
5de29b |
--- glibc-2.17-c758a686/sysdeps/x86_64/dl-trampoline.h 2012-12-24 22:02:13.000000000 -0500
|
|
|
12745e |
+++ glibc-2.17-c758a686/sysdeps/x86_64/dl-trampoline.h 2014-09-10 23:16:36.334167050 -0400
|
|
|
5de29b |
@@ -19,14 +19,14 @@
|
|
|
5de29b |
|
|
|
5de29b |
#ifdef RESTORE_AVX
|
|
|
5de29b |
/* This is to support AVX audit modules. */
|
|
|
5de29b |
- vmovdqu %ymm0, (LR_VECTOR_OFFSET)(%rsp)
|
|
|
5de29b |
- vmovdqu %ymm1, (LR_VECTOR_OFFSET + VECTOR_SIZE)(%rsp)
|
|
|
5de29b |
- vmovdqu %ymm2, (LR_VECTOR_OFFSET + VECTOR_SIZE*2)(%rsp)
|
|
|
5de29b |
- vmovdqu %ymm3, (LR_VECTOR_OFFSET + VECTOR_SIZE*3)(%rsp)
|
|
|
5de29b |
- vmovdqu %ymm4, (LR_VECTOR_OFFSET + VECTOR_SIZE*4)(%rsp)
|
|
|
5de29b |
- vmovdqu %ymm5, (LR_VECTOR_OFFSET + VECTOR_SIZE*5)(%rsp)
|
|
|
5de29b |
- vmovdqu %ymm6, (LR_VECTOR_OFFSET + VECTOR_SIZE*6)(%rsp)
|
|
|
5de29b |
- vmovdqu %ymm7, (LR_VECTOR_OFFSET + VECTOR_SIZE*7)(%rsp)
|
|
|
5de29b |
+ VMOV %VEC(0), (LR_VECTOR_OFFSET)(%rsp)
|
|
|
5de29b |
+ VMOV %VEC(1), (LR_VECTOR_OFFSET + VECTOR_SIZE)(%rsp)
|
|
|
5de29b |
+ VMOV %VEC(2), (LR_VECTOR_OFFSET + VECTOR_SIZE*2)(%rsp)
|
|
|
5de29b |
+ VMOV %VEC(3), (LR_VECTOR_OFFSET + VECTOR_SIZE*3)(%rsp)
|
|
|
5de29b |
+ VMOV %VEC(4), (LR_VECTOR_OFFSET + VECTOR_SIZE*4)(%rsp)
|
|
|
5de29b |
+ VMOV %VEC(5), (LR_VECTOR_OFFSET + VECTOR_SIZE*5)(%rsp)
|
|
|
5de29b |
+ VMOV %VEC(6), (LR_VECTOR_OFFSET + VECTOR_SIZE*6)(%rsp)
|
|
|
5de29b |
+ VMOV %VEC(7), (LR_VECTOR_OFFSET + VECTOR_SIZE*7)(%rsp)
|
|
|
5de29b |
|
|
|
5de29b |
/* Save xmm0-xmm7 registers to detect if any of them are
|
|
|
5de29b |
changed by audit module. */
|
|
|
5de29b |
@@ -72,7 +72,7 @@
|
|
|
5de29b |
je 2f
|
|
|
5de29b |
vmovdqa %xmm0, (LR_VECTOR_OFFSET)(%rsp)
|
|
|
5de29b |
jmp 1f
|
|
|
5de29b |
-2: vmovdqu (LR_VECTOR_OFFSET)(%rsp), %ymm0
|
|
|
5de29b |
+2: VMOV (LR_VECTOR_OFFSET)(%rsp), %VEC(0)
|
|
|
5de29b |
vmovdqa %xmm0, (LR_XMM_OFFSET)(%rsp)
|
|
|
5de29b |
|
|
|
5de29b |
1: vpcmpeqq (LR_SIZE + XMM_SIZE)(%rsp), %xmm1, %xmm8
|
|
|
5de29b |
@@ -81,7 +81,7 @@
|
|
|
5de29b |
je 2f
|
|
|
5de29b |
vmovdqa %xmm1, (LR_VECTOR_OFFSET + VECTOR_SIZE)(%rsp)
|
|
|
5de29b |
jmp 1f
|
|
|
5de29b |
-2: vmovdqu (LR_VECTOR_OFFSET + VECTOR_SIZE)(%rsp), %ymm1
|
|
|
5de29b |
+2: VMOV (LR_VECTOR_OFFSET + VECTOR_SIZE)(%rsp), %VEC(1)
|
|
|
5de29b |
vmovdqa %xmm1, (LR_XMM_OFFSET + XMM_SIZE)(%rsp)
|
|
|
5de29b |
|
|
|
5de29b |
1: vpcmpeqq (LR_SIZE + XMM_SIZE*2)(%rsp), %xmm2, %xmm8
|
|
|
5de29b |
@@ -90,7 +90,7 @@
|
|
|
5de29b |
je 2f
|
|
|
5de29b |
vmovdqa %xmm2, (LR_VECTOR_OFFSET + VECTOR_SIZE*2)(%rsp)
|
|
|
5de29b |
jmp 1f
|
|
|
5de29b |
-2: vmovdqu (LR_VECTOR_OFFSET + VECTOR_SIZE*2)(%rsp), %ymm2
|
|
|
5de29b |
+2: VMOV (LR_VECTOR_OFFSET + VECTOR_SIZE*2)(%rsp), %VEC(2)
|
|
|
5de29b |
vmovdqa %xmm2, (LR_XMM_OFFSET + XMM_SIZE*2)(%rsp)
|
|
|
5de29b |
|
|
|
5de29b |
1: vpcmpeqq (LR_SIZE + XMM_SIZE*3)(%rsp), %xmm3, %xmm8
|
|
|
5de29b |
@@ -99,7 +99,7 @@
|
|
|
5de29b |
je 2f
|
|
|
5de29b |
vmovdqa %xmm3, (LR_VECTOR_OFFSET + VECTOR_SIZE*3)(%rsp)
|
|
|
5de29b |
jmp 1f
|
|
|
5de29b |
-2: vmovdqu (LR_VECTOR_OFFSET + VECTOR_SIZE*3)(%rsp), %ymm3
|
|
|
5de29b |
+2: VMOV (LR_VECTOR_OFFSET + VECTOR_SIZE*3)(%rsp), %VEC(3)
|
|
|
5de29b |
vmovdqa %xmm3, (LR_XMM_OFFSET + XMM_SIZE*3)(%rsp)
|
|
|
5de29b |
|
|
|
5de29b |
1: vpcmpeqq (LR_SIZE + XMM_SIZE*4)(%rsp), %xmm4, %xmm8
|
|
|
5de29b |
@@ -108,7 +108,7 @@
|
|
|
5de29b |
je 2f
|
|
|
5de29b |
vmovdqa %xmm4, (LR_VECTOR_OFFSET + VECTOR_SIZE*4)(%rsp)
|
|
|
5de29b |
jmp 1f
|
|
|
5de29b |
-2: vmovdqu (LR_VECTOR_OFFSET + VECTOR_SIZE*4)(%rsp), %ymm4
|
|
|
5de29b |
+2: VMOV (LR_VECTOR_OFFSET + VECTOR_SIZE*4)(%rsp), %VEC(4)
|
|
|
5de29b |
vmovdqa %xmm4, (LR_XMM_OFFSET + XMM_SIZE*4)(%rsp)
|
|
|
5de29b |
|
|
|
5de29b |
1: vpcmpeqq (LR_SIZE + XMM_SIZE*5)(%rsp), %xmm5, %xmm8
|
|
|
5de29b |
@@ -117,7 +117,7 @@
|
|
|
5de29b |
je 2f
|
|
|
5de29b |
vmovdqa %xmm5, (LR_VECTOR_OFFSET + VECTOR_SIZE*5)(%rsp)
|
|
|
5de29b |
jmp 1f
|
|
|
5de29b |
-2: vmovdqu (LR_VECTOR_OFFSET + VECTOR_SIZE*5)(%rsp), %ymm5
|
|
|
5de29b |
+2: VMOV (LR_VECTOR_OFFSET + VECTOR_SIZE*5)(%rsp), %VEC(5)
|
|
|
5de29b |
vmovdqa %xmm5, (LR_XMM_OFFSET + XMM_SIZE*5)(%rsp)
|
|
|
5de29b |
|
|
|
5de29b |
1: vpcmpeqq (LR_SIZE + XMM_SIZE*6)(%rsp), %xmm6, %xmm8
|
|
|
5de29b |
@@ -126,7 +126,7 @@
|
|
|
5de29b |
je 2f
|
|
|
5de29b |
vmovdqa %xmm6, (LR_VECTOR_OFFSET + VECTOR_SIZE*6)(%rsp)
|
|
|
5de29b |
jmp 1f
|
|
|
5de29b |
-2: vmovdqu (LR_VECTOR_OFFSET + VECTOR_SIZE*6)(%rsp), %ymm6
|
|
|
5de29b |
+2: VMOV (LR_VECTOR_OFFSET + VECTOR_SIZE*6)(%rsp), %VEC(6)
|
|
|
5de29b |
vmovdqa %xmm6, (LR_XMM_OFFSET + XMM_SIZE*6)(%rsp)
|
|
|
5de29b |
|
|
|
5de29b |
1: vpcmpeqq (LR_SIZE + XMM_SIZE*7)(%rsp), %xmm7, %xmm8
|
|
|
5de29b |
@@ -135,7 +135,7 @@
|
|
|
5de29b |
je 2f
|
|
|
5de29b |
vmovdqa %xmm7, (LR_VECTOR_OFFSET + VECTOR_SIZE*7)(%rsp)
|
|
|
5de29b |
jmp 1f
|
|
|
5de29b |
-2: vmovdqu (LR_VECTOR_OFFSET + VECTOR_SIZE*7)(%rsp), %ymm7
|
|
|
5de29b |
+2: VMOV (LR_VECTOR_OFFSET + VECTOR_SIZE*7)(%rsp), %VEC(7)
|
|
|
5de29b |
vmovdqa %xmm7, (LR_XMM_OFFSET + XMM_SIZE*7)(%rsp)
|
|
|
5de29b |
|
|
|
5de29b |
1:
|
|
|
5de29b |
@@ -213,8 +213,8 @@
|
|
|
5de29b |
|
|
|
5de29b |
#ifdef RESTORE_AVX
|
|
|
5de29b |
/* This is to support AVX audit modules. */
|
|
|
5de29b |
- vmovdqu %ymm0, LRV_VECTOR0_OFFSET(%rcx)
|
|
|
5de29b |
- vmovdqu %ymm1, LRV_VECTOR1_OFFSET(%rcx)
|
|
|
5de29b |
+ VMOV %VEC(0), LRV_VECTOR0_OFFSET(%rcx)
|
|
|
5de29b |
+ VMOV %VEC(1), LRV_VECTOR1_OFFSET(%rcx)
|
|
|
5de29b |
|
|
|
5de29b |
/* Save xmm0/xmm1 registers to detect if they are changed
|
|
|
5de29b |
by audit module. */
|
|
|
5de29b |
@@ -243,13 +243,13 @@
|
|
|
5de29b |
vpmovmskb %xmm2, %esi
|
|
|
5de29b |
cmpl $0xffff, %esi
|
|
|
5de29b |
jne 1f
|
|
|
5de29b |
- vmovdqu LRV_VECTOR0_OFFSET(%rsp), %ymm0
|
|
|
5de29b |
+ VMOV LRV_VECTOR0_OFFSET(%rsp), %VEC(0)
|
|
|
5de29b |
|
|
|
5de29b |
1: vpcmpeqq (LRV_SIZE + XMM_SIZE)(%rsp), %xmm1, %xmm2
|
|
|
5de29b |
vpmovmskb %xmm2, %esi
|
|
|
5de29b |
cmpl $0xffff, %esi
|
|
|
5de29b |
jne 1f
|
|
|
5de29b |
- vmovdqu LRV_VECTOR1_OFFSET(%rsp), %ymm1
|
|
|
5de29b |
+ VMOV LRV_VECTOR1_OFFSET(%rsp), %VEC(1)
|
|
|
5de29b |
|
|
|
5de29b |
1:
|
|
|
5de29b |
#endif
|
|
|
12745e |
diff -urN glibc-2.17-c758a686/sysdeps/x86_64/dl-trampoline.S glibc-2.17-c758a686/sysdeps/x86_64/dl-trampoline.S
|
|
|
5de29b |
--- glibc-2.17-c758a686/sysdeps/x86_64/dl-trampoline.S 2012-12-24 22:02:13.000000000 -0500
|
|
|
12745e |
+++ glibc-2.17-c758a686/sysdeps/x86_64/dl-trampoline.S 2014-09-10 23:16:36.334167050 -0400
|
|
|
5de29b |
@@ -96,7 +96,7 @@
|
|
|
5de29b |
|
|
|
5de29b |
/* Actively align the La_x86_64_regs structure. */
|
|
|
5de29b |
andq $0xfffffffffffffff0, %rsp
|
|
|
5de29b |
-# ifdef HAVE_AVX_SUPPORT
|
|
|
5de29b |
+# if defined HAVE_AVX_SUPPORT || defined HAVE_AVX512_ASM_SUPPORT
|
|
|
5de29b |
/* sizeof(La_x86_64_regs). Need extra space for 8 SSE registers
|
|
|
5de29b |
to detect if any xmm0-xmm7 registers are changed by audit
|
|
|
5de29b |
module. */
|
|
|
5de29b |
@@ -130,7 +130,7 @@
|
|
|
5de29b |
movaps %xmm6, (LR_XMM_OFFSET + XMM_SIZE*6)(%rsp)
|
|
|
5de29b |
movaps %xmm7, (LR_XMM_OFFSET + XMM_SIZE*7)(%rsp)
|
|
|
5de29b |
|
|
|
5de29b |
-# ifdef HAVE_AVX_SUPPORT
|
|
|
5de29b |
+# if defined HAVE_AVX_SUPPORT || defined HAVE_AVX512_ASM_SUPPORT
|
|
|
5de29b |
.data
|
|
|
5de29b |
L(have_avx):
|
|
|
5de29b |
.zero 4
|
|
|
5de29b |
@@ -138,7 +138,7 @@
|
|
|
5de29b |
.previous
|
|
|
5de29b |
|
|
|
5de29b |
cmpl $0, L(have_avx)(%rip)
|
|
|
5de29b |
- jne 1f
|
|
|
5de29b |
+ jne L(defined)
|
|
|
5de29b |
movq %rbx, %r11 # Save rbx
|
|
|
5de29b |
movl $1, %eax
|
|
|
5de29b |
cpuid
|
|
|
5de29b |
@@ -147,18 +147,54 @@
|
|
|
5de29b |
// AVX and XSAVE supported?
|
|
|
5de29b |
andl $((1 << 28) | (1 << 27)), %ecx
|
|
|
5de29b |
cmpl $((1 << 28) | (1 << 27)), %ecx
|
|
|
5de29b |
- jne 2f
|
|
|
5de29b |
+ jne 10f
|
|
|
5de29b |
+# ifdef HAVE_AVX512_ASM_SUPPORT
|
|
|
5de29b |
+ // AVX512 supported in processor?
|
|
|
5de29b |
+ movq %rbx, %r11 # Save rbx
|
|
|
5de29b |
+ xorl %ecx, %ecx
|
|
|
5de29b |
+ mov $0x7, %eax
|
|
|
5de29b |
+ cpuid
|
|
|
5de29b |
+ andl $(1 << 16), %ebx
|
|
|
5de29b |
+# endif
|
|
|
5de29b |
xorl %ecx, %ecx
|
|
|
5de29b |
// Get XFEATURE_ENABLED_MASK
|
|
|
5de29b |
xgetbv
|
|
|
5de29b |
- andl $0x6, %eax
|
|
|
5de29b |
-2: subl $0x5, %eax
|
|
|
5de29b |
+# ifdef HAVE_AVX512_ASM_SUPPORT
|
|
|
5de29b |
+ test %ebx, %ebx
|
|
|
5de29b |
+ movq %r11, %rbx # Restore rbx
|
|
|
5de29b |
+ je 20f
|
|
|
5de29b |
+ // Verify that XCR0[7:5] = '111b' and
|
|
|
5de29b |
+ // XCR0[2:1] = '11b' which means
|
|
|
5de29b |
+ // that zmm state is enabled
|
|
|
5de29b |
+ andl $0xe6, %eax
|
|
|
5de29b |
+ cmpl $0xe6, %eax
|
|
|
5de29b |
+ jne 20f
|
|
|
5de29b |
+ movl %eax, L(have_avx)(%rip)
|
|
|
5de29b |
+L(avx512):
|
|
|
5de29b |
+# define RESTORE_AVX
|
|
|
5de29b |
+# define VMOV vmovdqu64
|
|
|
5de29b |
+# define VEC(i) zmm##i
|
|
|
5de29b |
+# define MORE_CODE
|
|
|
5de29b |
+# include "dl-trampoline.h"
|
|
|
5de29b |
+# undef VMOV
|
|
|
5de29b |
+# undef VEC
|
|
|
5de29b |
+# undef RESTORE_AVX
|
|
|
5de29b |
+# endif
|
|
|
5de29b |
+20: andl $0x6, %eax
|
|
|
5de29b |
+10: subl $0x5, %eax
|
|
|
5de29b |
movl %eax, L(have_avx)(%rip)
|
|
|
5de29b |
cmpl $0, %eax
|
|
|
5de29b |
|
|
|
5de29b |
-1: js L(no_avx)
|
|
|
5de29b |
+L(defined):
|
|
|
5de29b |
+ js L(no_avx)
|
|
|
5de29b |
+# ifdef HAVE_AVX512_ASM_SUPPORT
|
|
|
5de29b |
+ cmpl $0xe6, L(have_avx)(%rip)
|
|
|
5de29b |
+ je L(avx512)
|
|
|
5de29b |
+# endif
|
|
|
5de29b |
|
|
|
5de29b |
# define RESTORE_AVX
|
|
|
5de29b |
+# define VMOV vmovdqu
|
|
|
5de29b |
+# define VEC(i) ymm##i
|
|
|
5de29b |
# define MORE_CODE
|
|
|
5de29b |
# include "dl-trampoline.h"
|
|
|
5de29b |
|
|
|
5de29b |
@@ -180,9 +216,9 @@
|
|
|
5de29b |
.align 16
|
|
|
5de29b |
cfi_startproc
|
|
|
5de29b |
_dl_x86_64_save_sse:
|
|
|
5de29b |
-# ifdef HAVE_AVX_SUPPORT
|
|
|
5de29b |
+# if defined HAVE_AVX_SUPPORT || defined HAVE_AVX512_ASM_SUPPORT
|
|
|
5de29b |
cmpl $0, L(have_avx)(%rip)
|
|
|
5de29b |
- jne 1f
|
|
|
5de29b |
+ jne L(defined_5)
|
|
|
5de29b |
movq %rbx, %r11 # Save rbx
|
|
|
5de29b |
movl $1, %eax
|
|
|
5de29b |
cpuid
|
|
|
5de29b |
@@ -191,21 +227,43 @@
|
|
|
5de29b |
// AVX and XSAVE supported?
|
|
|
5de29b |
andl $((1 << 28) | (1 << 27)), %ecx
|
|
|
5de29b |
cmpl $((1 << 28) | (1 << 27)), %ecx
|
|
|
5de29b |
- jne 2f
|
|
|
5de29b |
+ jne 1f
|
|
|
5de29b |
+# ifdef HAVE_AVX512_ASM_SUPPORT
|
|
|
5de29b |
+ // AVX512 supported in a processor?
|
|
|
5de29b |
+ movq %rbx, %r11 # Save rbx
|
|
|
5de29b |
+ xorl %ecx,%ecx
|
|
|
5de29b |
+ mov $0x7,%eax
|
|
|
5de29b |
+ cpuid
|
|
|
5de29b |
+ andl $(1 << 16), %ebx
|
|
|
5de29b |
+# endif
|
|
|
5de29b |
xorl %ecx, %ecx
|
|
|
5de29b |
// Get XFEATURE_ENABLED_MASK
|
|
|
5de29b |
xgetbv
|
|
|
5de29b |
- andl $0x6, %eax
|
|
|
5de29b |
- cmpl $0x6, %eax
|
|
|
5de29b |
- // Nonzero if SSE and AVX state saving is enabled.
|
|
|
5de29b |
- sete %al
|
|
|
5de29b |
-2: leal -1(%eax,%eax), %eax
|
|
|
5de29b |
+# ifdef HAVE_AVX512_ASM_SUPPORT
|
|
|
5de29b |
+ test %ebx, %ebx
|
|
|
5de29b |
+ movq %r11, %rbx # Restore rbx
|
|
|
5de29b |
+ je 2f
|
|
|
5de29b |
+ // Verify that XCR0[7:5] = '111b' and
|
|
|
5de29b |
+ // XCR0[2:1] = '11b' which means
|
|
|
5de29b |
+ // that zmm state is enabled
|
|
|
5de29b |
+ andl $0xe6, %eax
|
|
|
5de29b |
+ movl %eax, L(have_avx)(%rip)
|
|
|
5de29b |
+ cmpl $0xe6, %eax
|
|
|
5de29b |
+ je L(avx512_5)
|
|
|
5de29b |
+# endif
|
|
|
5de29b |
+
|
|
|
5de29b |
+2: andl $0x6, %eax
|
|
|
5de29b |
+1: subl $0x5, %eax
|
|
|
5de29b |
movl %eax, L(have_avx)(%rip)
|
|
|
5de29b |
cmpl $0, %eax
|
|
|
5de29b |
|
|
|
5de29b |
-1: js L(no_avx5)
|
|
|
5de29b |
+L(defined_5):
|
|
|
5de29b |
+ js L(no_avx5)
|
|
|
5de29b |
+# ifdef HAVE_AVX512_ASM_SUPPORT
|
|
|
5de29b |
+ cmpl $0xe6, L(have_avx)(%rip)
|
|
|
5de29b |
+ je L(avx512_5)
|
|
|
5de29b |
+# endif
|
|
|
5de29b |
|
|
|
5de29b |
-# define YMM_SIZE 32
|
|
|
5de29b |
vmovdqa %ymm0, %fs:RTLD_SAVESPACE_SSE+0*YMM_SIZE
|
|
|
5de29b |
vmovdqa %ymm1, %fs:RTLD_SAVESPACE_SSE+1*YMM_SIZE
|
|
|
5de29b |
vmovdqa %ymm2, %fs:RTLD_SAVESPACE_SSE+2*YMM_SIZE
|
|
|
5de29b |
@@ -215,6 +273,18 @@
|
|
|
5de29b |
vmovdqa %ymm6, %fs:RTLD_SAVESPACE_SSE+6*YMM_SIZE
|
|
|
5de29b |
vmovdqa %ymm7, %fs:RTLD_SAVESPACE_SSE+7*YMM_SIZE
|
|
|
5de29b |
ret
|
|
|
5de29b |
+# ifdef HAVE_AVX512_ASM_SUPPORT
|
|
|
5de29b |
+L(avx512_5):
|
|
|
5de29b |
+ vmovdqu64 %zmm0, %fs:RTLD_SAVESPACE_SSE+0*ZMM_SIZE
|
|
|
5de29b |
+ vmovdqu64 %zmm1, %fs:RTLD_SAVESPACE_SSE+1*ZMM_SIZE
|
|
|
5de29b |
+ vmovdqu64 %zmm2, %fs:RTLD_SAVESPACE_SSE+2*ZMM_SIZE
|
|
|
5de29b |
+ vmovdqu64 %zmm3, %fs:RTLD_SAVESPACE_SSE+3*ZMM_SIZE
|
|
|
5de29b |
+ vmovdqu64 %zmm4, %fs:RTLD_SAVESPACE_SSE+4*ZMM_SIZE
|
|
|
5de29b |
+ vmovdqu64 %zmm5, %fs:RTLD_SAVESPACE_SSE+5*ZMM_SIZE
|
|
|
5de29b |
+ vmovdqu64 %zmm6, %fs:RTLD_SAVESPACE_SSE+6*ZMM_SIZE
|
|
|
5de29b |
+ vmovdqu64 %zmm7, %fs:RTLD_SAVESPACE_SSE+7*ZMM_SIZE
|
|
|
5de29b |
+ ret
|
|
|
5de29b |
+# endif
|
|
|
5de29b |
L(no_avx5):
|
|
|
5de29b |
# endif
|
|
|
5de29b |
movdqa %xmm0, %fs:RTLD_SAVESPACE_SSE+0*XMM_SIZE
|
|
|
5de29b |
@@ -235,9 +305,13 @@
|
|
|
5de29b |
.align 16
|
|
|
5de29b |
cfi_startproc
|
|
|
5de29b |
_dl_x86_64_restore_sse:
|
|
|
5de29b |
-# ifdef HAVE_AVX_SUPPORT
|
|
|
5de29b |
+# if defined HAVE_AVX_SUPPORT || defined HAVE_AVX512_ASM_SUPPORT
|
|
|
5de29b |
cmpl $0, L(have_avx)(%rip)
|
|
|
5de29b |
js L(no_avx6)
|
|
|
5de29b |
+# ifdef HAVE_AVX512_ASM_SUPPORT
|
|
|
5de29b |
+ cmpl $0xe6, L(have_avx)(%rip)
|
|
|
5de29b |
+ je L(avx512_6)
|
|
|
5de29b |
+# endif
|
|
|
5de29b |
|
|
|
5de29b |
vmovdqa %fs:RTLD_SAVESPACE_SSE+0*YMM_SIZE, %ymm0
|
|
|
5de29b |
vmovdqa %fs:RTLD_SAVESPACE_SSE+1*YMM_SIZE, %ymm1
|
|
|
5de29b |
@@ -248,6 +322,18 @@
|
|
|
5de29b |
vmovdqa %fs:RTLD_SAVESPACE_SSE+6*YMM_SIZE, %ymm6
|
|
|
5de29b |
vmovdqa %fs:RTLD_SAVESPACE_SSE+7*YMM_SIZE, %ymm7
|
|
|
5de29b |
ret
|
|
|
5de29b |
+# ifdef HAVE_AVX512_ASM_SUPPORT
|
|
|
5de29b |
+L(avx512_6):
|
|
|
5de29b |
+ vmovdqu64 %fs:RTLD_SAVESPACE_SSE+0*ZMM_SIZE, %zmm0
|
|
|
5de29b |
+ vmovdqu64 %fs:RTLD_SAVESPACE_SSE+1*ZMM_SIZE, %zmm1
|
|
|
5de29b |
+ vmovdqu64 %fs:RTLD_SAVESPACE_SSE+2*ZMM_SIZE, %zmm2
|
|
|
5de29b |
+ vmovdqu64 %fs:RTLD_SAVESPACE_SSE+3*ZMM_SIZE, %zmm3
|
|
|
5de29b |
+ vmovdqu64 %fs:RTLD_SAVESPACE_SSE+4*ZMM_SIZE, %zmm4
|
|
|
5de29b |
+ vmovdqu64 %fs:RTLD_SAVESPACE_SSE+5*ZMM_SIZE, %zmm5
|
|
|
5de29b |
+ vmovdqu64 %fs:RTLD_SAVESPACE_SSE+6*ZMM_SIZE, %zmm6
|
|
|
5de29b |
+ vmovdqu64 %fs:RTLD_SAVESPACE_SSE+7*ZMM_SIZE, %zmm7
|
|
|
5de29b |
+ ret
|
|
|
5de29b |
+# endif
|
|
|
5de29b |
L(no_avx6):
|
|
|
5de29b |
# endif
|
|
|
5de29b |
movdqa %fs:RTLD_SAVESPACE_SSE+0*XMM_SIZE, %xmm0
|
|
|
12745e |
diff -urN glibc-2.17-c758a686/sysdeps/x86_64/link-defines.sym glibc-2.17-c758a686/sysdeps/x86_64/link-defines.sym
|
|
|
5de29b |
--- glibc-2.17-c758a686/sysdeps/x86_64/link-defines.sym 2012-12-24 22:02:13.000000000 -0500
|
|
|
12745e |
+++ glibc-2.17-c758a686/sysdeps/x86_64/link-defines.sym 2014-09-10 23:16:36.335167048 -0400
|
|
|
5de29b |
@@ -4,6 +4,8 @@
|
|
|
5de29b |
--
|
|
|
5de29b |
VECTOR_SIZE sizeof (La_x86_64_vector)
|
|
|
5de29b |
XMM_SIZE sizeof (La_x86_64_xmm)
|
|
|
5de29b |
+YMM_SIZE sizeof (La_x86_64_ymm)
|
|
|
5de29b |
+ZMM_SIZE sizeof (La_x86_64_zmm)
|
|
|
5de29b |
|
|
|
5de29b |
LR_SIZE sizeof (struct La_x86_64_regs)
|
|
|
5de29b |
LR_RDX_OFFSET offsetof (struct La_x86_64_regs, lr_rdx)
|
|
|
12745e |
diff -urN glibc-2.17-c758a686/sysdeps/x86_64/Makefile glibc-2.17-c758a686/sysdeps/x86_64/Makefile
|
|
|
5de29b |
--- glibc-2.17-c758a686/sysdeps/x86_64/Makefile 2012-12-24 22:02:13.000000000 -0500
|
|
|
12745e |
+++ glibc-2.17-c758a686/sysdeps/x86_64/Makefile 2014-09-10 23:22:04.269518711 -0400
|
|
|
5de29b |
@@ -37,6 +37,20 @@
|
|
|
5de29b |
|
|
|
5de29b |
$(objpfx)tst-quad1pie: $(objpfx)tst-quadmod1pie.o
|
|
|
5de29b |
$(objpfx)tst-quad2pie: $(objpfx)tst-quadmod2pie.o
|
|
|
5de29b |
+
|
|
|
5de29b |
+tests += tst-audit10
|
|
|
5de29b |
+modules-names += tst-auditmod10a tst-auditmod10b
|
|
|
5de29b |
+
|
|
|
5de29b |
+$(objpfx)tst-audit10: $(objpfx)tst-auditmod10a.so
|
|
|
5de29b |
+$(objpfx)tst-audit10.out: $(objpfx)tst-auditmod10b.so
|
|
|
5de29b |
+tst-audit10-ENV = LD_AUDIT=$(objpfx)tst-auditmod10b.so
|
|
|
5de29b |
+
|
|
|
5de29b |
+ifeq (yes,$(config-cflags-avx512))
|
|
|
5de29b |
+AVX512-CFLAGS = -mavx512f
|
|
|
5de29b |
+CFLAGS-tst-audit10.c += $(AVX512-CFLAGS)
|
|
|
5de29b |
+CFLAGS-tst-auditmod10a.c += $(AVX512-CFLAGS)
|
|
|
5de29b |
+CFLAGS-tst-auditmod10b.c += $(AVX512-CFLAGS)
|
|
|
5de29b |
+endif
|
|
|
5de29b |
endif
|
|
|
5de29b |
|
|
|
5de29b |
ifeq ($(subdir),csu)
|
|
|
12745e |
diff -urN glibc-2.17-c758a686/sysdeps/x86_64/tst-audit10.c glibc-2.17-c758a686/sysdeps/x86_64/tst-audit10.c
|
|
|
5de29b |
--- glibc-2.17-c758a686/sysdeps/x86_64/tst-audit10.c 1969-12-31 19:00:00.000000000 -0500
|
|
|
12745e |
+++ glibc-2.17-c758a686/sysdeps/x86_64/tst-audit10.c 2014-09-10 23:16:36.335167048 -0400
|
|
|
5de29b |
@@ -0,0 +1,70 @@
|
|
|
5de29b |
+/* Copyright (C) 2012-2014 Free Software Foundation, Inc.
|
|
|
5de29b |
+ This file is part of the GNU C Library.
|
|
|
5de29b |
+
|
|
|
5de29b |
+ The GNU C Library is free software; you can redistribute it and/or
|
|
|
5de29b |
+ modify it under the terms of the GNU Lesser General Public
|
|
|
5de29b |
+ License as published by the Free Software Foundation; either
|
|
|
5de29b |
+ version 2.1 of the License, or (at your option) any later version.
|
|
|
5de29b |
+
|
|
|
5de29b |
+ The GNU C Library is distributed in the hope that it will be useful,
|
|
|
5de29b |
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
|
5de29b |
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
|
|
5de29b |
+ Lesser General Public License for more details.
|
|
|
5de29b |
+
|
|
|
5de29b |
+ You should have received a copy of the GNU Lesser General Public
|
|
|
5de29b |
+ License along with the GNU C Library; if not, see
|
|
|
5de29b |
+ <http://www.gnu.org/licenses/>. */
|
|
|
5de29b |
+
|
|
|
5de29b |
+/* Test case for x86-64 preserved registers in dynamic linker. */
|
|
|
5de29b |
+
|
|
|
5de29b |
+#ifdef __AVX512F__
|
|
|
5de29b |
+#include <stdlib.h>
|
|
|
5de29b |
+#include <string.h>
|
|
|
5de29b |
+#include <cpuid.h>
|
|
|
5de29b |
+#include <immintrin.h>
|
|
|
5de29b |
+
|
|
|
5de29b |
+static int
|
|
|
5de29b |
+avx512_enabled (void)
|
|
|
5de29b |
+{
|
|
|
5de29b |
+ unsigned int eax, ebx, ecx, edx;
|
|
|
5de29b |
+
|
|
|
5de29b |
+ if (__get_cpuid (1, &eax, &ebx, &ecx, &edx) == 0
|
|
|
5de29b |
+ || (ecx & (bit_AVX | bit_OSXSAVE)) != (bit_AVX | bit_OSXSAVE))
|
|
|
5de29b |
+ return 0;
|
|
|
5de29b |
+
|
|
|
5de29b |
+ __cpuid_count (7, 0, eax, ebx, ecx, edx);
|
|
|
5de29b |
+ if (!(ebx & bit_AVX512F))
|
|
|
5de29b |
+ return 0;
|
|
|
5de29b |
+
|
|
|
5de29b |
+ asm ("xgetbv" : "=a" (eax), "=d" (edx) : "c" (0));
|
|
|
5de29b |
+
|
|
|
5de29b |
+ /* Verify that ZMM, YMM and XMM states are enabled. */
|
|
|
5de29b |
+ return (eax & 0xe6) == 0xe6;
|
|
|
5de29b |
+}
|
|
|
5de29b |
+
|
|
|
5de29b |
+
|
|
|
5de29b |
+extern __m512i audit_test (__m512i, __m512i, __m512i, __m512i,
|
|
|
5de29b |
+ __m512i, __m512i, __m512i, __m512i);
|
|
|
5de29b |
+int
|
|
|
5de29b |
+main (void)
|
|
|
5de29b |
+{
|
|
|
5de29b |
+ /* Run AVX512 test only if AVX512 is supported. */
|
|
|
5de29b |
+ if (avx512_enabled ())
|
|
|
5de29b |
+ {
|
|
|
5de29b |
+ __m512i zmm = _mm512_setzero_si512 ();
|
|
|
5de29b |
+ __m512i ret = audit_test (zmm, zmm, zmm, zmm, zmm, zmm, zmm, zmm);
|
|
|
5de29b |
+
|
|
|
5de29b |
+ zmm = _mm512_set1_epi64 (0x12349876);
|
|
|
5de29b |
+
|
|
|
5de29b |
+ if (memcmp (&zmm, &ret, sizeof (ret)))
|
|
|
5de29b |
+ abort ();
|
|
|
5de29b |
+ }
|
|
|
5de29b |
+ return 0;
|
|
|
5de29b |
+}
|
|
|
5de29b |
+#else
|
|
|
5de29b |
+int
|
|
|
5de29b |
+main (void)
|
|
|
5de29b |
+{
|
|
|
5de29b |
+ return 0;
|
|
|
5de29b |
+}
|
|
|
5de29b |
+#endif
|
|
|
12745e |
diff -urN glibc-2.17-c758a686/sysdeps/x86_64/tst-auditmod10a.c glibc-2.17-c758a686/sysdeps/x86_64/tst-auditmod10a.c
|
|
|
5de29b |
--- glibc-2.17-c758a686/sysdeps/x86_64/tst-auditmod10a.c 1969-12-31 19:00:00.000000000 -0500
|
|
|
12745e |
+++ glibc-2.17-c758a686/sysdeps/x86_64/tst-auditmod10a.c 2014-09-10 23:16:36.335167048 -0400
|
|
|
5de29b |
@@ -0,0 +1,65 @@
|
|
|
5de29b |
+/* Copyright (C) 2012-2014 Free Software Foundation, Inc.
|
|
|
5de29b |
+ This file is part of the GNU C Library.
|
|
|
5de29b |
+
|
|
|
5de29b |
+ The GNU C Library is free software; you can redistribute it and/or
|
|
|
5de29b |
+ modify it under the terms of the GNU Lesser General Public
|
|
|
5de29b |
+ License as published by the Free Software Foundation; either
|
|
|
5de29b |
+ version 2.1 of the License, or (at your option) any later version.
|
|
|
5de29b |
+
|
|
|
5de29b |
+ The GNU C Library is distributed in the hope that it will be useful,
|
|
|
5de29b |
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
|
5de29b |
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
|
|
5de29b |
+ Lesser General Public License for more details.
|
|
|
5de29b |
+
|
|
|
5de29b |
+ You should have received a copy of the GNU Lesser General Public
|
|
|
5de29b |
+ License along with the GNU C Library; if not, see
|
|
|
5de29b |
+ <http://www.gnu.org/licenses/>. */
|
|
|
5de29b |
+
|
|
|
5de29b |
+/* Test case for x86-64 preserved registers in dynamic linker. */
|
|
|
5de29b |
+
|
|
|
5de29b |
+#ifdef __AVX512F__
|
|
|
5de29b |
+#include <stdlib.h>
|
|
|
5de29b |
+#include <string.h>
|
|
|
5de29b |
+#include <immintrin.h>
|
|
|
5de29b |
+
|
|
|
5de29b |
+__m512i
|
|
|
5de29b |
+audit_test (__m512i x0, __m512i x1, __m512i x2, __m512i x3,
|
|
|
5de29b |
+ __m512i x4, __m512i x5, __m512i x6, __m512i x7)
|
|
|
5de29b |
+{
|
|
|
5de29b |
+ __m512i zmm;
|
|
|
5de29b |
+
|
|
|
5de29b |
+ zmm = _mm512_set1_epi64 (1);
|
|
|
5de29b |
+ if (memcmp (&zmm, &x0, sizeof (zmm)))
|
|
|
5de29b |
+ abort ();
|
|
|
5de29b |
+
|
|
|
5de29b |
+ zmm = _mm512_set1_epi64 (2);
|
|
|
5de29b |
+ if (memcmp (&zmm, &x1, sizeof (zmm)))
|
|
|
5de29b |
+ abort ();
|
|
|
5de29b |
+
|
|
|
5de29b |
+ zmm = _mm512_set1_epi64 (3);
|
|
|
5de29b |
+ if (memcmp (&zmm, &x2, sizeof (zmm)))
|
|
|
5de29b |
+ abort ();
|
|
|
5de29b |
+
|
|
|
5de29b |
+ zmm = _mm512_set1_epi64 (4);
|
|
|
5de29b |
+ if (memcmp (&zmm, &x3, sizeof (zmm)))
|
|
|
5de29b |
+ abort ();
|
|
|
5de29b |
+
|
|
|
5de29b |
+ zmm = _mm512_set1_epi64 (5);
|
|
|
5de29b |
+ if (memcmp (&zmm, &x4, sizeof (zmm)))
|
|
|
5de29b |
+ abort ();
|
|
|
5de29b |
+
|
|
|
5de29b |
+ zmm = _mm512_set1_epi64 (6);
|
|
|
5de29b |
+ if (memcmp (&zmm, &x5, sizeof (zmm)))
|
|
|
5de29b |
+ abort ();
|
|
|
5de29b |
+
|
|
|
5de29b |
+ zmm = _mm512_set1_epi64 (7);
|
|
|
5de29b |
+ if (memcmp (&zmm, &x6, sizeof (zmm)))
|
|
|
5de29b |
+ abort ();
|
|
|
5de29b |
+
|
|
|
5de29b |
+ zmm = _mm512_set1_epi64 (8);
|
|
|
5de29b |
+ if (memcmp (&zmm, &x7, sizeof (zmm)))
|
|
|
5de29b |
+ abort ();
|
|
|
5de29b |
+
|
|
|
5de29b |
+ return _mm512_setzero_si512 ();
|
|
|
5de29b |
+}
|
|
|
5de29b |
+#endif
|
|
|
12745e |
diff -urN glibc-2.17-c758a686/sysdeps/x86_64/tst-auditmod10b.c glibc-2.17-c758a686/sysdeps/x86_64/tst-auditmod10b.c
|
|
|
5de29b |
--- glibc-2.17-c758a686/sysdeps/x86_64/tst-auditmod10b.c 1969-12-31 19:00:00.000000000 -0500
|
|
|
12745e |
+++ glibc-2.17-c758a686/sysdeps/x86_64/tst-auditmod10b.c 2014-09-10 23:16:36.336167046 -0400
|
|
|
5de29b |
@@ -0,0 +1,219 @@
|
|
|
5de29b |
+/* Copyright (C) 2012-2014 Free Software Foundation, Inc.
|
|
|
5de29b |
+ This file is part of the GNU C Library.
|
|
|
5de29b |
+
|
|
|
5de29b |
+ The GNU C Library is free software; you can redistribute it and/or
|
|
|
5de29b |
+ modify it under the terms of the GNU Lesser General Public
|
|
|
5de29b |
+ License as published by the Free Software Foundation; either
|
|
|
5de29b |
+ version 2.1 of the License, or (at your option) any later version.
|
|
|
5de29b |
+
|
|
|
5de29b |
+ The GNU C Library is distributed in the hope that it will be useful,
|
|
|
5de29b |
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
|
5de29b |
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
|
|
5de29b |
+ Lesser General Public License for more details.
|
|
|
5de29b |
+
|
|
|
5de29b |
+ You should have received a copy of the GNU Lesser General Public
|
|
|
5de29b |
+ License along with the GNU C Library; if not, see
|
|
|
5de29b |
+ <http://www.gnu.org/licenses/>. */
|
|
|
5de29b |
+
|
|
|
5de29b |
+/* Verify that changing AVX512 registers in audit library won't affect
|
|
|
5de29b |
+ function parameter passing/return. */
|
|
|
5de29b |
+
|
|
|
5de29b |
+#include <dlfcn.h>
|
|
|
5de29b |
+#include <stdint.h>
|
|
|
5de29b |
+#include <stdio.h>
|
|
|
5de29b |
+#include <stdlib.h>
|
|
|
5de29b |
+#include <string.h>
|
|
|
5de29b |
+#include <unistd.h>
|
|
|
5de29b |
+#include <bits/wordsize.h>
|
|
|
5de29b |
+#include <gnu/lib-names.h>
|
|
|
5de29b |
+
|
|
|
5de29b |
+unsigned int
|
|
|
5de29b |
+la_version (unsigned int v)
|
|
|
5de29b |
+{
|
|
|
5de29b |
+ setlinebuf (stdout);
|
|
|
5de29b |
+
|
|
|
5de29b |
+ printf ("version: %u\n", v);
|
|
|
5de29b |
+
|
|
|
5de29b |
+ char buf[20];
|
|
|
5de29b |
+ sprintf (buf, "%u", v);
|
|
|
5de29b |
+
|
|
|
5de29b |
+ return v;
|
|
|
5de29b |
+}
|
|
|
5de29b |
+
|
|
|
5de29b |
+void
|
|
|
5de29b |
+la_activity (uintptr_t *cookie, unsigned int flag)
|
|
|
5de29b |
+{
|
|
|
5de29b |
+ if (flag == LA_ACT_CONSISTENT)
|
|
|
5de29b |
+ printf ("activity: consistent\n");
|
|
|
5de29b |
+ else if (flag == LA_ACT_ADD)
|
|
|
5de29b |
+ printf ("activity: add\n");
|
|
|
5de29b |
+ else if (flag == LA_ACT_DELETE)
|
|
|
5de29b |
+ printf ("activity: delete\n");
|
|
|
5de29b |
+ else
|
|
|
5de29b |
+ printf ("activity: unknown activity %u\n", flag);
|
|
|
5de29b |
+}
|
|
|
5de29b |
+
|
|
|
5de29b |
+char *
|
|
|
5de29b |
+la_objsearch (const char *name, uintptr_t *cookie, unsigned int flag)
|
|
|
5de29b |
+{
|
|
|
5de29b |
+ char buf[100];
|
|
|
5de29b |
+ const char *flagstr;
|
|
|
5de29b |
+ if (flag == LA_SER_ORIG)
|
|
|
5de29b |
+ flagstr = "LA_SET_ORIG";
|
|
|
5de29b |
+ else if (flag == LA_SER_LIBPATH)
|
|
|
5de29b |
+ flagstr = "LA_SER_LIBPATH";
|
|
|
5de29b |
+ else if (flag == LA_SER_RUNPATH)
|
|
|
5de29b |
+ flagstr = "LA_SER_RUNPATH";
|
|
|
5de29b |
+ else if (flag == LA_SER_CONFIG)
|
|
|
5de29b |
+ flagstr = "LA_SER_CONFIG";
|
|
|
5de29b |
+ else if (flag == LA_SER_DEFAULT)
|
|
|
5de29b |
+ flagstr = "LA_SER_DEFAULT";
|
|
|
5de29b |
+ else if (flag == LA_SER_SECURE)
|
|
|
5de29b |
+ flagstr = "LA_SER_SECURE";
|
|
|
5de29b |
+ else
|
|
|
5de29b |
+ {
|
|
|
5de29b |
+ sprintf (buf, "unknown flag %d", flag);
|
|
|
5de29b |
+ flagstr = buf;
|
|
|
5de29b |
+ }
|
|
|
5de29b |
+ printf ("objsearch: %s, %s\n", name, flagstr);
|
|
|
5de29b |
+
|
|
|
5de29b |
+ return (char *) name;
|
|
|
5de29b |
+}
|
|
|
5de29b |
+
|
|
|
5de29b |
+unsigned int
|
|
|
5de29b |
+la_objopen (struct link_map *l, Lmid_t lmid, uintptr_t *cookie)
|
|
|
5de29b |
+{
|
|
|
5de29b |
+ printf ("objopen: %ld, %s\n", lmid, l->l_name);
|
|
|
5de29b |
+
|
|
|
5de29b |
+ return 3;
|
|
|
5de29b |
+}
|
|
|
5de29b |
+
|
|
|
5de29b |
+void
|
|
|
5de29b |
+la_preinit (uintptr_t *cookie)
|
|
|
5de29b |
+{
|
|
|
5de29b |
+ printf ("preinit\n");
|
|
|
5de29b |
+}
|
|
|
5de29b |
+
|
|
|
5de29b |
+unsigned int
|
|
|
5de29b |
+la_objclose (uintptr_t *cookie)
|
|
|
5de29b |
+{
|
|
|
5de29b |
+ printf ("objclose\n");
|
|
|
5de29b |
+ return 0;
|
|
|
5de29b |
+}
|
|
|
5de29b |
+
|
|
|
5de29b |
+uintptr_t
|
|
|
5de29b |
+la_symbind64 (Elf64_Sym *sym, unsigned int ndx, uintptr_t *refcook,
|
|
|
5de29b |
+ uintptr_t *defcook, unsigned int *flags, const char *symname)
|
|
|
5de29b |
+{
|
|
|
5de29b |
+ printf ("symbind64: symname=%s, st_value=%#lx, ndx=%u, flags=%u\n",
|
|
|
5de29b |
+ symname, (long int) sym->st_value, ndx, *flags);
|
|
|
5de29b |
+
|
|
|
5de29b |
+ return sym->st_value;
|
|
|
5de29b |
+}
|
|
|
5de29b |
+
|
|
|
5de29b |
+#include <tst-audit.h>
|
|
|
5de29b |
+
|
|
|
5de29b |
+#ifdef __AVX512F__
|
|
|
5de29b |
+#include <immintrin.h>
|
|
|
5de29b |
+#include <cpuid.h>
|
|
|
5de29b |
+
|
|
|
5de29b |
+static int
|
|
|
5de29b |
+check_avx512 (void)
|
|
|
5de29b |
+{
|
|
|
5de29b |
+ unsigned int eax, ebx, ecx, edx;
|
|
|
5de29b |
+
|
|
|
5de29b |
+ if (__get_cpuid (1, &eax, &ebx, &ecx, &edx) == 0
|
|
|
5de29b |
+ || (ecx & (bit_AVX | bit_OSXSAVE)) != (bit_AVX | bit_OSXSAVE))
|
|
|
5de29b |
+ return 0;
|
|
|
5de29b |
+
|
|
|
5de29b |
+ __cpuid_count (7, 0, eax, ebx, ecx, edx);
|
|
|
5de29b |
+ if (!(ebx & bit_AVX512F))
|
|
|
5de29b |
+ return 0;
|
|
|
5de29b |
+
|
|
|
5de29b |
+ asm ("xgetbv" : "=a" (eax), "=d" (edx) : "c" (0));
|
|
|
5de29b |
+
|
|
|
5de29b |
+ /* Verify that ZMM, YMM and XMM states are enabled. */
|
|
|
5de29b |
+ return (eax & 0xe6) == 0xe6;
|
|
|
5de29b |
+}
|
|
|
5de29b |
+
|
|
|
5de29b |
+#else
|
|
|
5de29b |
+#include <emmintrin.h>
|
|
|
5de29b |
+#endif
|
|
|
5de29b |
+
|
|
|
5de29b |
+ElfW(Addr)
|
|
|
5de29b |
+pltenter (ElfW(Sym) *sym, unsigned int ndx, uintptr_t *refcook,
|
|
|
5de29b |
+ uintptr_t *defcook, La_regs *regs, unsigned int *flags,
|
|
|
5de29b |
+ const char *symname, long int *framesizep)
|
|
|
5de29b |
+{
|
|
|
5de29b |
+ printf ("pltenter: symname=%s, st_value=%#lx, ndx=%u, flags=%u\n",
|
|
|
5de29b |
+ symname, (long int) sym->st_value, ndx, *flags);
|
|
|
5de29b |
+
|
|
|
5de29b |
+#ifdef __AVX512F__
|
|
|
5de29b |
+ if (check_avx512 () && strcmp (symname, "audit_test") == 0)
|
|
|
5de29b |
+ {
|
|
|
5de29b |
+ __m512i zero = _mm512_setzero_si512 ();
|
|
|
5de29b |
+ if (memcmp (®s->lr_vector[0], &zero, sizeof (zero))
|
|
|
5de29b |
+ || memcmp (®s->lr_vector[1], &zero, sizeof (zero))
|
|
|
5de29b |
+ || memcmp (®s->lr_vector[2], &zero, sizeof (zero))
|
|
|
5de29b |
+ || memcmp (®s->lr_vector[3], &zero, sizeof (zero))
|
|
|
5de29b |
+ || memcmp (®s->lr_vector[4], &zero, sizeof (zero))
|
|
|
5de29b |
+ || memcmp (®s->lr_vector[5], &zero, sizeof (zero))
|
|
|
5de29b |
+ || memcmp (®s->lr_vector[6], &zero, sizeof (zero))
|
|
|
5de29b |
+ || memcmp (®s->lr_vector[7], &zero, sizeof (zero)))
|
|
|
5de29b |
+ abort ();
|
|
|
5de29b |
+
|
|
|
5de29b |
+ for (int i = 0; i < 8; i++)
|
|
|
5de29b |
+ regs->lr_vector[i].zmm[0]
|
|
|
5de29b |
+ = (La_x86_64_zmm) _mm512_set1_epi64 (i + 1);
|
|
|
5de29b |
+
|
|
|
5de29b |
+ __m512i zmm = _mm512_set1_epi64 (-1);
|
|
|
5de29b |
+ asm volatile ("vmovdqa64 %0, %%zmm0" : : "x" (zmm) : "xmm0" );
|
|
|
5de29b |
+ asm volatile ("vmovdqa64 %0, %%zmm1" : : "x" (zmm) : "xmm1" );
|
|
|
5de29b |
+ asm volatile ("vmovdqa64 %0, %%zmm2" : : "x" (zmm) : "xmm2" );
|
|
|
5de29b |
+ asm volatile ("vmovdqa64 %0, %%zmm3" : : "x" (zmm) : "xmm3" );
|
|
|
5de29b |
+ asm volatile ("vmovdqa64 %0, %%zmm4" : : "x" (zmm) : "xmm4" );
|
|
|
5de29b |
+ asm volatile ("vmovdqa64 %0, %%zmm5" : : "x" (zmm) : "xmm5" );
|
|
|
5de29b |
+ asm volatile ("vmovdqa64 %0, %%zmm6" : : "x" (zmm) : "xmm6" );
|
|
|
5de29b |
+ asm volatile ("vmovdqa64 %0, %%zmm7" : : "x" (zmm) : "xmm7" );
|
|
|
5de29b |
+
|
|
|
5de29b |
+ *framesizep = 1024;
|
|
|
5de29b |
+ }
|
|
|
5de29b |
+#endif
|
|
|
5de29b |
+
|
|
|
5de29b |
+ return sym->st_value;
|
|
|
5de29b |
+}
|
|
|
5de29b |
+
|
|
|
5de29b |
+unsigned int
|
|
|
5de29b |
+pltexit (ElfW(Sym) *sym, unsigned int ndx, uintptr_t *refcook,
|
|
|
5de29b |
+ uintptr_t *defcook, const La_regs *inregs, La_retval *outregs,
|
|
|
5de29b |
+ const char *symname)
|
|
|
5de29b |
+{
|
|
|
5de29b |
+ printf ("pltexit: symname=%s, st_value=%#lx, ndx=%u, retval=%tu\n",
|
|
|
5de29b |
+ symname, (long int) sym->st_value, ndx,
|
|
|
5de29b |
+ (ptrdiff_t) outregs->int_retval);
|
|
|
5de29b |
+
|
|
|
5de29b |
+#ifdef __AVX512F__
|
|
|
5de29b |
+ if (check_avx512 () && strcmp (symname, "audit_test") == 0)
|
|
|
5de29b |
+ {
|
|
|
5de29b |
+ __m512i zero = _mm512_setzero_si512 ();
|
|
|
5de29b |
+ if (memcmp (&outregs->lrv_vector0, &zero, sizeof (zero)))
|
|
|
5de29b |
+ abort ();
|
|
|
5de29b |
+
|
|
|
5de29b |
+ for (int i = 0; i < 8; i++)
|
|
|
5de29b |
+ {
|
|
|
5de29b |
+ __m512i zmm = _mm512_set1_epi64 (i + 1);
|
|
|
5de29b |
+ if (memcmp (&inregs->lr_vector[i], &zmm, sizeof (zmm)) != 0)
|
|
|
5de29b |
+ abort ();
|
|
|
5de29b |
+ }
|
|
|
5de29b |
+
|
|
|
5de29b |
+ outregs->lrv_vector0.zmm[0]
|
|
|
5de29b |
+ = (La_x86_64_zmm) _mm512_set1_epi64 (0x12349876);
|
|
|
5de29b |
+
|
|
|
5de29b |
+ __m512i zmm = _mm512_set1_epi64 (-1);
|
|
|
5de29b |
+ asm volatile ("vmovdqa64 %0, %%zmm0" : : "x" (zmm) : "xmm0" );
|
|
|
5de29b |
+ asm volatile ("vmovdqa64 %0, %%zmm1" : : "x" (zmm) : "xmm1" );
|
|
|
5de29b |
+ }
|
|
|
5de29b |
+#endif
|
|
|
5de29b |
+
|
|
|
5de29b |
+ return 0;
|
|
|
5de29b |
+}
|
|
|
12745e |
diff -urN glibc-2.17-c758a686/sysdeps/x86/Makefile glibc-2.17-c758a686/sysdeps/x86/Makefile
|
|
|
5de29b |
--- glibc-2.17-c758a686/sysdeps/x86/Makefile 2012-12-24 22:02:13.000000000 -0500
|
|
|
12745e |
+++ glibc-2.17-c758a686/sysdeps/x86/Makefile 2014-09-11 16:06:03.121319867 -0400
|
|
|
5de29b |
@@ -2,8 +2,8 @@
|
|
|
5de29b |
CFLAGS-.os += $(if $(filter $(@F),$(patsubst %,%.os,$(all-rtld-routines))),\
|
|
|
5de29b |
-mno-sse -mno-mmx)
|
|
|
5de29b |
|
|
|
5de29b |
-tests: $(objpfx)tst-xmmymm.out
|
|
|
5de29b |
-$(objpfx)tst-xmmymm.out: ../sysdeps/x86/tst-xmmymm.sh $(objpfx)ld.so
|
|
|
5de29b |
+tests: $(objpfx)tst-xmmymmzmm.out
|
|
|
5de29b |
+$(objpfx)tst-xmmymmzmm.out: ../sysdeps/x86/tst-xmmymmzmm.sh $(objpfx)ld.so
|
|
|
5de29b |
@echo "Checking ld.so for SSE register use. This will take a few seconds..."
|
|
|
5de29b |
$(SHELL) $< $(objpfx) '$(NM)' '$(OBJDUMP)' '$(READELF)' > $@
|
|
|
5de29b |
endif
|
|
|
12745e |
diff -urN glibc-2.17-c758a686/sysdeps/x86/tst-xmmymm.sh glibc-2.17-c758a686/sysdeps/x86/tst-xmmymm.sh
|
|
|
5de29b |
--- glibc-2.17-c758a686/sysdeps/x86/tst-xmmymm.sh 2012-12-24 22:02:13.000000000 -0500
|
|
|
12745e |
+++ glibc-2.17-c758a686/sysdeps/x86/tst-xmmymm.sh 1969-12-31 19:00:00.000000000 -0500
|
|
|
5de29b |
@@ -1,103 +0,0 @@
|
|
|
5de29b |
-#! /bin/bash
|
|
|
5de29b |
-# Make sure no code in ld.so uses xmm/ymm registers on x86-64.
|
|
|
5de29b |
-# Copyright (C) 2009-2012 Free Software Foundation, Inc.
|
|
|
5de29b |
-# This file is part of the GNU C Library.
|
|
|
5de29b |
-
|
|
|
5de29b |
-# The GNU C Library is free software; you can redistribute it and/or
|
|
|
5de29b |
-# modify it under the terms of the GNU Lesser General Public
|
|
|
5de29b |
-# License as published by the Free Software Foundation; either
|
|
|
5de29b |
-# version 2.1 of the License, or (at your option) any later version.
|
|
|
5de29b |
-
|
|
|
5de29b |
-# The GNU C Library is distributed in the hope that it will be useful,
|
|
|
5de29b |
-# but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
|
5de29b |
-# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
|
|
5de29b |
-# Lesser General Public License for more details.
|
|
|
5de29b |
-
|
|
|
5de29b |
-# You should have received a copy of the GNU Lesser General Public
|
|
|
5de29b |
-# License along with the GNU C Library; if not, see
|
|
|
5de29b |
-# <http://www.gnu.org/licenses/>.
|
|
|
5de29b |
-
|
|
|
5de29b |
-set -e
|
|
|
5de29b |
-
|
|
|
5de29b |
-objpfx="$1"
|
|
|
5de29b |
-NM="$2"
|
|
|
5de29b |
-OBJDUMP="$3"
|
|
|
5de29b |
-READELF="$4"
|
|
|
5de29b |
-
|
|
|
5de29b |
-tmp=$(mktemp ${objpfx}tst-xmmymm.XXXXXX)
|
|
|
5de29b |
-trap 'rm -f "$tmp"' 1 2 3 15
|
|
|
5de29b |
-
|
|
|
5de29b |
-# List of object files we have to test
|
|
|
5de29b |
-rtldobjs=$($READELF -W -wi ${objpfx}dl-allobjs.os |
|
|
|
5de29b |
- awk '/^ </ { if ($5 == "(DW_TAG_compile_unit)") c=1; else c=0 } $2 == "DW_AT_name" { if (c == 1) print $NF }' |
|
|
|
5de29b |
- sed 's,\(.*/\|\)\([_[:alnum:]-]*[.]\).$,\2os,')
|
|
|
5de29b |
-rtldobjs="$rtldobjs $(ar t ${objpfx}rtld-libc.a)"
|
|
|
5de29b |
-
|
|
|
5de29b |
-# OBJECT symbols can be ignored.
|
|
|
5de29b |
-$READELF -sW ${objpfx}dl-allobjs.os ${objpfx}rtld-libc.a |
|
|
|
5de29b |
-egrep " OBJECT *GLOBAL " |
|
|
|
5de29b |
-awk '{if ($7 != "ABS") print $8 }' |
|
|
|
5de29b |
-sort -u > "$tmp"
|
|
|
5de29b |
-declare -a objects
|
|
|
5de29b |
-objects=($(cat "$tmp"))
|
|
|
5de29b |
-
|
|
|
5de29b |
-objs="dl-runtime.os"
|
|
|
5de29b |
-tocheck="dl-runtime.os"
|
|
|
5de29b |
-
|
|
|
5de29b |
-while test -n "$objs"; do
|
|
|
5de29b |
- this="$objs"
|
|
|
5de29b |
- objs=""
|
|
|
5de29b |
-
|
|
|
5de29b |
- for f in $this; do
|
|
|
5de29b |
- undef=$($NM -u "$objpfx"../*/"$f" | awk '{print $2}')
|
|
|
5de29b |
- if test -n "$undef"; then
|
|
|
5de29b |
- for s in $undef; do
|
|
|
5de29b |
- for obj in ${objects[*]} "_GLOBAL_OFFSET_TABLE_"; do
|
|
|
5de29b |
- if test "$obj" = "$s"; then
|
|
|
5de29b |
- continue 2
|
|
|
5de29b |
- fi
|
|
|
5de29b |
- done
|
|
|
5de29b |
- for o in $rtldobjs; do
|
|
|
5de29b |
- ro=$(echo "$objpfx"../*/"$o")
|
|
|
5de29b |
- if $NM -g --defined-only "$ro" | egrep -qs " $s\$"; then
|
|
|
5de29b |
- if ! (echo "$tocheck $objs" | fgrep -qs "$o"); then
|
|
|
5de29b |
- echo "$o needed for $s"
|
|
|
5de29b |
- objs="$objs $o"
|
|
|
5de29b |
- fi
|
|
|
5de29b |
- break;
|
|
|
5de29b |
- fi
|
|
|
5de29b |
- done
|
|
|
5de29b |
- done
|
|
|
5de29b |
- fi
|
|
|
5de29b |
- done
|
|
|
5de29b |
- tocheck="$tocheck$objs"
|
|
|
5de29b |
-done
|
|
|
5de29b |
-
|
|
|
5de29b |
-echo
|
|
|
5de29b |
-echo
|
|
|
5de29b |
-echo "object files needed: $tocheck"
|
|
|
5de29b |
-
|
|
|
5de29b |
-cp /dev/null "$tmp"
|
|
|
5de29b |
-for f in $tocheck; do
|
|
|
5de29b |
- $OBJDUMP -d "$objpfx"../*/"$f" |
|
|
|
5de29b |
- awk 'BEGIN { last="" } /^[[:xdigit:]]* <[_[:alnum:]]*>:$/ { fct=substr($2, 2, length($2)-3) } /,%[xy]mm[[:digit:]]*$/ { if (last != fct) { print fct; last=fct} }' |
|
|
|
5de29b |
- while read fct; do
|
|
|
5de29b |
- if test "$fct" = "_dl_runtime_profile" -o "$fct" = "_dl_x86_64_restore_sse"; then
|
|
|
5de29b |
- continue;
|
|
|
5de29b |
- fi
|
|
|
5de29b |
- echo "function $fct in $f modifies xmm/ymm" >> "$tmp"
|
|
|
5de29b |
- result=1
|
|
|
5de29b |
- done
|
|
|
5de29b |
-done
|
|
|
5de29b |
-
|
|
|
5de29b |
-if test -s "$tmp"; then
|
|
|
5de29b |
- echo
|
|
|
5de29b |
- echo
|
|
|
5de29b |
- cat "$tmp"
|
|
|
5de29b |
- result=1
|
|
|
5de29b |
-else
|
|
|
5de29b |
- result=0
|
|
|
5de29b |
-fi
|
|
|
5de29b |
-
|
|
|
5de29b |
-rm "$tmp"
|
|
|
5de29b |
-exit $result
|
|
|
12745e |
diff -urN glibc-2.17-c758a686/sysdeps/x86/tst-xmmymmzmm.sh glibc-2.17-c758a686/sysdeps/x86/tst-xmmymmzmm.sh
|
|
|
5de29b |
--- glibc-2.17-c758a686/sysdeps/x86/tst-xmmymmzmm.sh 1969-12-31 19:00:00.000000000 -0500
|
|
|
12745e |
+++ glibc-2.17-c758a686/sysdeps/x86/tst-xmmymmzmm.sh 2014-09-11 16:05:10.073426623 -0400
|
|
|
5de29b |
@@ -0,0 +1,103 @@
|
|
|
5de29b |
+#! /bin/bash
|
|
|
5de29b |
+# Make sure no code in ld.so uses xmm/ymm/zmm registers on x86-64.
|
|
|
5de29b |
+# Copyright (C) 2009-2012 Free Software Foundation, Inc.
|
|
|
5de29b |
+# This file is part of the GNU C Library.
|
|
|
5de29b |
+
|
|
|
5de29b |
+# The GNU C Library is free software; you can redistribute it and/or
|
|
|
5de29b |
+# modify it under the terms of the GNU Lesser General Public
|
|
|
5de29b |
+# License as published by the Free Software Foundation; either
|
|
|
5de29b |
+# version 2.1 of the License, or (at your option) any later version.
|
|
|
5de29b |
+
|
|
|
5de29b |
+# The GNU C Library is distributed in the hope that it will be useful,
|
|
|
5de29b |
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
|
5de29b |
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
|
|
5de29b |
+# Lesser General Public License for more details.
|
|
|
5de29b |
+
|
|
|
5de29b |
+# You should have received a copy of the GNU Lesser General Public
|
|
|
5de29b |
+# License along with the GNU C Library; if not, see
|
|
|
5de29b |
+# <http://www.gnu.org/licenses/>.
|
|
|
5de29b |
+
|
|
|
5de29b |
+set -e
|
|
|
5de29b |
+
|
|
|
5de29b |
+objpfx="$1"
|
|
|
5de29b |
+NM="$2"
|
|
|
5de29b |
+OBJDUMP="$3"
|
|
|
5de29b |
+READELF="$4"
|
|
|
5de29b |
+
|
|
|
5de29b |
+tmp=$(mktemp ${objpfx}tst-xmmymmzmm.XXXXXX)
|
|
|
5de29b |
+trap 'rm -f "$tmp"' 1 2 3 15
|
|
|
5de29b |
+
|
|
|
5de29b |
+# List of object files we have to test
|
|
|
5de29b |
+rtldobjs=$($READELF -W -wi ${objpfx}dl-allobjs.os |
|
|
|
5de29b |
+ awk '/^ </ { if ($5 == "(DW_TAG_compile_unit)") c=1; else c=0 } $2 == "DW_AT_name" { if (c == 1) print $NF }' |
|
|
|
5de29b |
+ sed 's,\(.*/\|\)\([_[:alnum:]-]*[.]\).$,\2os,')
|
|
|
5de29b |
+rtldobjs="$rtldobjs $(ar t ${objpfx}rtld-libc.a)"
|
|
|
5de29b |
+
|
|
|
5de29b |
+# OBJECT symbols can be ignored.
|
|
|
5de29b |
+$READELF -sW ${objpfx}dl-allobjs.os ${objpfx}rtld-libc.a |
|
|
|
5de29b |
+egrep " OBJECT *GLOBAL " |
|
|
|
5de29b |
+awk '{if ($7 != "ABS") print $8 }' |
|
|
|
5de29b |
+sort -u > "$tmp"
|
|
|
5de29b |
+declare -a objects
|
|
|
5de29b |
+objects=($(cat "$tmp"))
|
|
|
5de29b |
+
|
|
|
5de29b |
+objs="dl-runtime.os"
|
|
|
5de29b |
+tocheck="dl-runtime.os"
|
|
|
5de29b |
+
|
|
|
5de29b |
+while test -n "$objs"; do
|
|
|
5de29b |
+ this="$objs"
|
|
|
5de29b |
+ objs=""
|
|
|
5de29b |
+
|
|
|
5de29b |
+ for f in $this; do
|
|
|
5de29b |
+ undef=$($NM -u "$objpfx"../*/"$f" | awk '{print $2}')
|
|
|
5de29b |
+ if test -n "$undef"; then
|
|
|
5de29b |
+ for s in $undef; do
|
|
|
5de29b |
+ for obj in ${objects[*]} "_GLOBAL_OFFSET_TABLE_"; do
|
|
|
5de29b |
+ if test "$obj" = "$s"; then
|
|
|
5de29b |
+ continue 2
|
|
|
5de29b |
+ fi
|
|
|
5de29b |
+ done
|
|
|
5de29b |
+ for o in $rtldobjs; do
|
|
|
5de29b |
+ ro=$(echo "$objpfx"../*/"$o")
|
|
|
5de29b |
+ if $NM -g --defined-only "$ro" | egrep -qs " $s\$"; then
|
|
|
5de29b |
+ if ! (echo "$tocheck $objs" | fgrep -qs "$o"); then
|
|
|
5de29b |
+ echo "$o needed for $s"
|
|
|
5de29b |
+ objs="$objs $o"
|
|
|
5de29b |
+ fi
|
|
|
5de29b |
+ break;
|
|
|
5de29b |
+ fi
|
|
|
5de29b |
+ done
|
|
|
5de29b |
+ done
|
|
|
5de29b |
+ fi
|
|
|
5de29b |
+ done
|
|
|
5de29b |
+ tocheck="$tocheck$objs"
|
|
|
5de29b |
+done
|
|
|
5de29b |
+
|
|
|
5de29b |
+echo
|
|
|
5de29b |
+echo
|
|
|
5de29b |
+echo "object files needed: $tocheck"
|
|
|
5de29b |
+
|
|
|
5de29b |
+cp /dev/null "$tmp"
|
|
|
5de29b |
+for f in $tocheck; do
|
|
|
5de29b |
+ $OBJDUMP -d "$objpfx"../*/"$f" |
|
|
|
5de29b |
+ awk 'BEGIN { last="" } /^[[:xdigit:]]* <[_[:alnum:]]*>:$/ { fct=substr($2, 2, length($2)-3) } /,%[xyz]mm[[:digit:]]*$/ { if (last != fct) { print fct; last=fct} }' |
|
|
|
5de29b |
+ while read fct; do
|
|
|
5de29b |
+ if test "$fct" = "_dl_runtime_profile" -o "$fct" = "_dl_x86_64_restore_sse"; then
|
|
|
5de29b |
+ continue;
|
|
|
5de29b |
+ fi
|
|
|
5de29b |
+ echo "function $fct in $f modifies xmm/ymm/zmm" >> "$tmp"
|
|
|
5de29b |
+ result=1
|
|
|
5de29b |
+ done
|
|
|
5de29b |
+done
|
|
|
5de29b |
+
|
|
|
5de29b |
+if test -s "$tmp"; then
|
|
|
5de29b |
+ echo
|
|
|
5de29b |
+ echo
|
|
|
5de29b |
+ cat "$tmp"
|
|
|
5de29b |
+ result=1
|
|
|
5de29b |
+else
|
|
|
5de29b |
+ result=0
|
|
|
5de29b |
+fi
|
|
|
5de29b |
+
|
|
|
5de29b |
+rm "$tmp"
|
|
|
5de29b |
+exit $result
|