On 09/22/2017 03:28 AM, Rainer Orth wrote:
> Hi Daniel,
>
>> On 09/22/2017 02:18 AM, Rainer Orth wrote:
>>> Hi Daniel,
>>>
>>>> On 09/21/2017 05:18 PM, Daniel Santos wrote:
>>>>> So libgcc doesn't use a config.in. :(
>>>> Scratch that, I forgot that we're using gcc/config.in via auto-host.h.
>>>> So I only have to add this to gcc/configure.ac and it will be available
>>>> for my libgcc header -- this is what I used to sniff out support for the
>>>> .hidden directive.
>>> Please don't go that route: it's totally the wrong direction. There's
>>> work going on to further decouple libgcc from gcc-private headers and
>>> configure results. libgcc already has its own configure tests for
>>> assembler features, and its own config.in. What's wrong with adapting
>>> libitm's avx test in libitm/acinclude.m4 (LIBITM_CHECK_AS_AVX) for
>>> libgcc? Should be trivial...
>>>
>>> Rainer
>>>
>> Oops, I just saw your email after submitting my other patch. Yes, I am
>> mistaken about config.in, sorry about that. I didn't see a config.h
>> file, but examining further it looks like it outputs to auto-target.h.
>> Also, I was looking for some HAVE_AS* macros, but they are named
>> differently.
> Right: though some are for assembler features, the macros are named
> differently.
>
>> I had previously included gcc's auto-host.h since it was in the include
>> path in order to use HAVE_AS_HIDDEN, so in order to decouple this I'll
> HAVE_GAS_HIDDEN actually ;-)
>
>> need to add that check into libgcc/configure.ac as well. Again,
>> shouldn't be that much code. Sound sane to you?
> You could do that, but it was already used before your patches, so
> please separate it from the current issue if you go down that route.
> libgcc is still full of cleanup possibilities :-)
>
> Rainer
OK, so I'm just adding HAVE_AS_AVX mostly as-is from libitm (we don't
have $target_cpu so I'm using $target). I do have minor concerns about
how this test will work on a cross-build -- I'm not an autotools expert
and I don't understand which assembler it will invoke, but the results
of the test failing only means we use .byte instead of the real
mnemonic, so it really shouldn't be a problem.
I've got tests started again, so presuming that *this* one passes, is it
OK for the trunk?
gcc/testsuite:
* gcc.target/i386/pr82196-1.c: Simplify so that it doesn't break
on Solaris or with -mno-omit-frame-pointer.
* gcc.target/i386/pr82196-2.c: Likewise.
libgcc:
* configure.ac: Add check for HAVE_AS_AVX.
* config.in: Regenerate.
* configure: Likewise.
* config/i386/i386-asm.h: Include auto-target.h from libgcc.
(SSE_SAVE, SSE_RESTORE): Sniff HAVE_AS_AVX and directly emit raw
.byte code when assembler doesn't support avx, correct
out-of-date comments.
gcc/testsuite/gcc.target/i386/pr82196-1.c | 5 ++-
gcc/testsuite/gcc.target/i386/pr82196-2.c | 5 ++-
libgcc/config.in | 3 ++
libgcc/config/i386/i386-asm.h | 45 ++++++++++++++++++++++-----
libgcc/configure | 39 +++++++++++++++++++++++
libgcc/configure.ac | 16 ++++++++++
6 files changed, 100 insertions(+), 13 deletions(-)
Thanks,
Daniel
diff --git a/gcc/testsuite/gcc.target/i386/pr82196-1.c b/gcc/testsuite/gcc.target/i386/pr82196-1.c
index ef858328f00..541d975480d 100644
--- a/gcc/testsuite/gcc.target/i386/pr82196-1.c
+++ b/gcc/testsuite/gcc.target/i386/pr82196-1.c
@@ -1,7 +1,7 @@
/* { dg-do compile { target lp64 } } */
/* { dg-options "-msse -mcall-ms2sysv-xlogues -O2" } */
-/* { dg-final { scan-assembler "call.*__sse_savms64_18" } } */
-/* { dg-final { scan-assembler "jmp.*__sse_resms64x_18" } } */
+/* { dg-final { scan-assembler "call.*__sse_savms64f?_12" } } */
+/* { dg-final { scan-assembler "jmp.*__sse_resms64f?x_12" } } */
void __attribute__((sysv_abi)) a() {
}
@@ -9,6 +9,5 @@ void __attribute__((sysv_abi)) a() {
static void __attribute__((sysv_abi)) (*volatile a_noinfo)() = a;
void __attribute__((ms_abi)) b() {
- __asm__ __volatile__ ("" :::"rbx", "rbp", "r12", "r13", "r14", "r15");
a_noinfo ();
}
diff --git a/gcc/testsuite/gcc.target/i386/pr82196-2.c b/gcc/testsuite/gcc.target/i386/pr82196-2.c
index 8fe58411d5e..7166d068bc1 100644
--- a/gcc/testsuite/gcc.target/i386/pr82196-2.c
+++ b/gcc/testsuite/gcc.target/i386/pr82196-2.c
@@ -1,7 +1,7 @@
/* { dg-do compile { target lp64 } } */
/* { dg-options "-mavx -mcall-ms2sysv-xlogues -O2" } */
-/* { dg-final { scan-assembler "call.*__avx_savms64_18" } } */
-/* { dg-final { scan-assembler "jmp.*__avx_resms64x_18" } } */
+/* { dg-final { scan-assembler "call.*__avx_savms64f?_12" } } */
+/* { dg-final { scan-assembler "jmp.*__avx_resms64f?x_12" } } */
void __attribute__((sysv_abi)) a() {
}
@@ -9,6 +9,5 @@ void __attribute__((sysv_abi)) a() {
static void __attribute__((sysv_abi)) (*volatile a_noinfo)() = a;
void __attribute__((ms_abi)) b() {
- __asm__ __volatile__ ("" :::"rbx", "rbp", "r12", "r13", "r14", "r15");
a_noinfo ();
}
diff --git a/libgcc/config.in b/libgcc/config.in
index 7de22ee0a72..f9fb253874f 100644
--- a/libgcc/config.in
+++ b/libgcc/config.in
@@ -1,5 +1,8 @@
/* config.in. Generated from configure.ac by autoheader. */
+/* Define to 1 if the assembler supports AVX. */
+#undef HAVE_AS_AVX
+
/* Define to 1 if the target assembler supports thread-local storage. */
#undef HAVE_CC_TLS
diff --git a/libgcc/config/i386/i386-asm.h b/libgcc/config/i386/i386-asm.h
index 424e0f72aac..b417bfb93d6 100644
--- a/libgcc/config/i386/i386-asm.h
+++ b/libgcc/config/i386/i386-asm.h
@@ -26,6 +26,7 @@ see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
#ifndef I386_ASM_H
#define I386_ASM_H
+#include "auto-target.h"
#include "auto-host.h"
#define PASTE2(a, b) PASTE2a(a, b)
@@ -69,13 +70,15 @@ ASMNAME(fn):
#ifdef MS2SYSV_STUB_AVX
# define MS2SYSV_STUB_PREFIX __avx_
-# define MOVAPS vmovaps
+# ifdef HAVE_AS_AVX
+# define MOVAPS vmovaps
+# endif
#elif defined(MS2SYSV_STUB_SSE)
# define MS2SYSV_STUB_PREFIX __sse_
# define MOVAPS movaps
#endif
-#if defined (MS2SYSV_STUB_PREFIX) && defined (MOVAPS)
+#if defined (MS2SYSV_STUB_PREFIX)
# define MS2SYSV_STUB_BEGIN(base_name) \
HIDDEN_FUNC(PASTE2(MS2SYSV_STUB_PREFIX, base_name))
@@ -83,8 +86,10 @@ ASMNAME(fn):
# define MS2SYSV_STUB_END(base_name) \
FUNC_END(PASTE2(MS2SYSV_STUB_PREFIX, base_name))
-/* Save SSE registers 6-15. off is the offset of rax to get to xmm6. */
-# define SSE_SAVE \
+/* If expanding for sse or avx and we have assembler support. */
+# ifdef MOVAPS
+/* Save SSE registers 6-15 using rax as the base address. */
+# define SSE_SAVE \
MOVAPS %xmm15,-0x30(%rax); \
MOVAPS %xmm14,-0x20(%rax); \
MOVAPS %xmm13,-0x10(%rax); \
@@ -96,8 +101,8 @@ ASMNAME(fn):
MOVAPS %xmm7, 0x50(%rax); \
MOVAPS %xmm6, 0x60(%rax)
-/* Restore SSE registers 6-15. off is the offset of rsi to get to xmm6. */
-# define SSE_RESTORE \
+/* Restore SSE registers 6-15 using rsi as the base address. */
+# define SSE_RESTORE \
MOVAPS -0x30(%rsi), %xmm15; \
MOVAPS -0x20(%rsi), %xmm14; \
MOVAPS -0x10(%rsi), %xmm13; \
@@ -108,6 +113,32 @@ ASMNAME(fn):
MOVAPS 0x40(%rsi), %xmm8 ; \
MOVAPS 0x50(%rsi), %xmm7 ; \
MOVAPS 0x60(%rsi), %xmm6
-
+# else /* MOVAPS */
+/* If the assembler doesn't support AVX then directly emit machine code
+ for the instructions above directly. */
+# define BYTE .byte
+# define SSE_SAVE \
+ BYTE 0xc5, 0x78, 0x29, 0x78, 0xd0; /* vmovaps %xmm15,-0x30(%rax) */ \
+ BYTE 0xc5, 0x78, 0x29, 0x70, 0xe0; /* vmovaps %xmm14,-0x20(%rax) */ \
+ BYTE 0xc5, 0x78, 0x29, 0x68, 0xf0; /* vmovaps %xmm13,-0x10(%rax) */ \
+ BYTE 0xc5, 0x78, 0x29, 0x20; /* vmovaps %xmm12, (%rax) */ \
+ BYTE 0xc5, 0x78, 0x29, 0x58, 0x10; /* vmovaps %xmm11, 0x10(%rax) */ \
+ BYTE 0xc5, 0x78, 0x29, 0x50, 0x20; /* vmovaps %xmm10, 0x20(%rax) */ \
+ BYTE 0xc5, 0x78, 0x29, 0x48, 0x30; /* vmovaps %xmm9, 0x30(%rax) */ \
+ BYTE 0xc5, 0x78, 0x29, 0x40, 0x40; /* vmovaps %xmm8, 0x40(%rax) */ \
+ BYTE 0xc5, 0xf8, 0x29, 0x78, 0x50; /* vmovaps %xmm7, 0x50(%rax) */ \
+ BYTE 0xc5, 0xf8, 0x29, 0x70, 0x60; /* vmovaps %xmm6, 0x60(%rax) */
+# define SSE_RESTORE \
+ BYTE 0xc5, 0x78, 0x28, 0x7e, 0xd0; /* vmovaps -0x30(%rsi),%xmm15 */ \
+ BYTE 0xc5, 0x78, 0x28, 0x76, 0xe0; /* vmovaps -0x20(%rsi),%xmm14 */ \
+ BYTE 0xc5, 0x78, 0x28, 0x6e, 0xf0; /* vmovaps -0x10(%rsi),%xmm13 */ \
+ BYTE 0xc5, 0x78, 0x28, 0x26; /* vmovaps (%rsi),%xmm12 */ \
+ BYTE 0xc5, 0x78, 0x28, 0x5e, 0x10; /* vmovaps 0x10(%rsi),%xmm11 */ \
+ BYTE 0xc5, 0x78, 0x28, 0x56, 0x20; /* vmovaps 0x20(%rsi),%xmm10 */ \
+ BYTE 0xc5, 0x78, 0x28, 0x4e, 0x30; /* vmovaps 0x30(%rsi),%xmm9 */ \
+ BYTE 0xc5, 0x78, 0x28, 0x46, 0x40; /* vmovaps 0x40(%rsi),%xmm8 */ \
+ BYTE 0xc5, 0xf8, 0x28, 0x7e, 0x50; /* vmovaps 0x50(%rsi),%xmm7 */ \
+ BYTE 0xc5, 0xf8, 0x28, 0x76, 0x60; /* vmovaps 0x60(%rsi),%xmm6 */
+# endif /* MOVAPS */
#endif /* defined (MS2SYSV_STUB_ISA) && defined (MOVAPS) */
#endif /* I386_ASM_H */
diff --git a/libgcc/configure b/libgcc/configure
index 15d34b27cc0..63c50c0c7e2 100644
--- a/libgcc/configure
+++ b/libgcc/configure
@@ -5212,6 +5212,45 @@ if test "$enable_tls $gcc_cv_use_emutls" = "yes yes"; then
fi
+
+
+case "${target}" in
+i[34567]86-*-* | x86_64-*-*)
+ { $as_echo "$as_me:${as_lineno-$LINENO}: checking if the assembler supports AVX" >&5
+$as_echo_n "checking if the assembler supports AVX... " >&6; }
+if test "${libgcc_cv_as_avx+set}" = set; then :
+ $as_echo_n "(cached) " >&6
+else
+
+ cat confdefs.h - <<_ACEOF >conftest.$ac_ext
+/* end confdefs.h. */
+
+int
+main ()
+{
+asm("vzeroupper");
+ ;
+ return 0;
+}
+_ACEOF
+if ac_fn_c_try_compile "$LINENO"; then :
+ libgcc_cv_as_avx=yes
+else
+ libgcc_cv_as_avx=no
+fi
+rm -f core conftest.err conftest.$ac_objext conftest.$ac_ext
+
+fi
+{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $libgcc_cv_as_avx" >&5
+$as_echo "$libgcc_cv_as_avx" >&6; }
+ if test x$libgcc_cv_as_avx = xyes; then
+
+$as_echo "#define HAVE_AS_AVX 1" >>confdefs.h
+
+ fi
+ ;;
+esac
+
{ $as_echo "$as_me:${as_lineno-$LINENO}: checking for init priority support" >&5
$as_echo_n "checking for init priority support... " >&6; }
if test "${libgcc_cv_init_priority+set}" = set; then :
diff --git a/libgcc/configure.ac b/libgcc/configure.ac
index da499716f8c..dd60b01d2f8 100644
--- a/libgcc/configure.ac
+++ b/libgcc/configure.ac
@@ -506,6 +506,22 @@ if test "$enable_tls $gcc_cv_use_emutls" = "yes yes"; then
fi
AC_SUBST(set_use_emutls)
+dnl Check if as supports AVX instructions.
+AC_DEFUN([LIBGCC_CHECK_AS_AVX], [
+case "${target}" in
+i[[34567]]86-*-* | x86_64-*-*)
+ AC_CACHE_CHECK([if the assembler supports AVX], libgcc_cv_as_avx, [
+ AC_TRY_COMPILE([], [asm("vzeroupper");],
+ [libgcc_cv_as_avx=yes], [libgcc_cv_as_avx=no])
+ ])
+ if test x$libgcc_cv_as_avx = xyes; then
+ AC_DEFINE(HAVE_AS_AVX, 1, [Define to 1 if the assembler supports AVX.])
+ fi
+ ;;
+esac])
+LIBGCC_CHECK_AS_AVX
+
+dnl Check if as supports RTM instructions.
AC_CACHE_CHECK(for init priority support, libgcc_cv_init_priority, [
AC_COMPILE_IFELSE([AC_LANG_PROGRAM(,
[[void ip (void) __attribute__ ((constructor (1)));]])],
