Use i386/crtfastmath.c on Solaris 2/x86

Use i386/crtfastmath.c on Solaris 2/x86

[Rainer Orth]

I had long meant to support -fast-math on Solaris 2/x86.  While working
on the Solaris toplevel libgcc move, I've done it with the following
patch.

The only complication is that I need to make sure that SSE insns are only
used if the host supports them.

Bootstrapped without regressions on i386-pc-solaris2.8,
i386-pc-solaris2.9, i386-pc-solaris2.11, and sparc-sun-solaris2.11.

The libgcc part depends on the toplevel libgcc patch, so actually
applying this patch will have to wait until that one is in.

Ok for mainline?

	Rainer


2011-05-28  Rainer Orth  <ro@CeBiTec.Uni-Bielefeld.DE>

	gcc:
	* config/i386/crtfastmath.c [!__x86_64__ && __sun__ && __svr4__]:
	Include <signal.h>, <ucontext.h>.
	(sigill_caught): Define.
	(sigill_hdlr): New function.
	(set_fast_math) [!__x86_64__ && __sun__ && __svr4__]: Check if SSE
	insns can be executed.
	* config/sol2.h (ENDFILE_SPEC): Use crtfastmath.o if -ffast-math
	etc.
	* config/sparc/sol2.h (ENDFILE_SPEC): Remove.

	libgcc:
	* config.host (i[34567]86-*-solaris2*): Add i386/t-crtfm to
	tmake_file.
	Add crtfastmath.o to extra_parts.

diff --git a/gcc/config/i386/crtfastmath.c b/gcc/config/i386/crtfastmath.c
--- a/gcc/config/i386/crtfastmath.c
+++ b/gcc/config/i386/crtfastmath.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2005, 2007, 2009 Free Software Foundation, Inc.
+ * Copyright (C) 2005, 2007, 2009, 2011 Free Software Foundation, Inc.
  *
  * This file is free software; you can redistribute it and/or modify it
  * under the terms of the GNU General Public License as published by the
@@ -30,6 +30,26 @@
 #include "cpuid.h"
 #endif
 
+#if !defined __x86_64 && defined __sun__ && defined __svr4__
+#include <signal.h>
+#include <ucontext.h>
+
+static volatile sig_atomic_t sigill_caught;
+
+static void
+sigill_hdlr (int sig __attribute((unused)),
+	     siginfo_t *sip __attribute__((unused)),
+	     ucontext_t *ucp)
+{
+  sigill_caught = 1;
+  /* Set PC to the instruction after the faulting one to skip over it,
+     otherwise we enter an infinite loop.  4 is the size of the stmxcsr
+     instruction.  */
+  ucp->uc_mcontext.gregs[EIP] += 4;
+  setcontext (ucp);
+}
+#endif
+
 static void __attribute__((constructor))
 #ifndef __x86_64__
 /* The i386 ABI only requires 4-byte stack alignment, so this is necessary
@@ -45,6 +65,32 @@ set_fast_math (void)
   if (!__get_cpuid (1, &eax, &ebx, &ecx, &edx))
     return;
 
+#if defined __sun__ && defined __svr4__
+  /* Solaris 2 before Solaris 9 4/04 cannot execute SSE instructions even
+     if the CPU supports them.  Programs receive SIGILL instead, so check
+     for that at runtime.  */
+
+  if (edx & bit_SSE)
+    {
+      struct sigaction act, oact;
+
+      act.sa_handler = sigill_hdlr;
+      sigemptyset (&act.sa_mask);
+      /* Need to set SA_SIGINFO so a ucontext_t * is passed to the handler.  */
+      act.sa_flags = SA_SIGINFO;
+      sigaction (SIGILL, &act, &oact);
+
+      /* We need a single SSE instruction here so the handler can safely skip
+	 over it.  */
+      __asm__ volatile ("movss %xmm2,%xmm1");
+
+      sigaction (SIGILL, &oact, NULL);
+
+      if (sigill_caught)
+	return;
+    }
+#endif /* __sun__ && __svr4__ */
+
   if (edx & bit_SSE)
     {
       unsigned int mxcsr = __builtin_ia32_stmxcsr ();
diff --git a/gcc/config/sol2.h b/gcc/config/sol2.h
--- a/gcc/config/sol2.h
+++ b/gcc/config/sol2.h
@@ -141,7 +141,9 @@ along with GCC; see the file COPYING3.  
      %{p|pg:-ldl} -lc}"
 
 #undef  ENDFILE_SPEC
-#define ENDFILE_SPEC "crtend.o%s crtn.o%s"
+#define ENDFILE_SPEC \
+  "%{Ofast|ffast-math|funsafe-math-optimizations:crtfastmath.o%s} \
+   crtend.o%s crtn.o%s"
 
 /* We don't use the standard svr4 STARTFILE_SPEC because it's wrong for us.  */
 #undef STARTFILE_SPEC
diff --git a/gcc/config/sparc/sol2.h b/gcc/config/sparc/sol2.h
--- a/gcc/config/sparc/sol2.h
+++ b/gcc/config/sparc/sol2.h
@@ -117,11 +117,6 @@ along with GCC; see the file COPYING3.  
 #define NO_DBX_BNSYM_ENSYM 1
 \f
 
-#undef  ENDFILE_SPEC
-#define ENDFILE_SPEC \
-  "%{Ofast|ffast-math|funsafe-math-optimizations:crtfastmath.o%s} \
-   crtend.o%s crtn.o%s"
-
 /* Select a format to encode pointers in exception handling data.  CODE
    is 0 for data, 1 for code labels, 2 for function pointers.  GLOBAL is
    true if the symbol may be affected by dynamic relocations.
diff --git a/libgcc/config.host b/libgcc/config.host
--- a/libgcc/config.host
+++ b/libgcc/config.host
@@ -338,6 +338,8 @@ i[34567]86-*-rtems*)
 	tmake_file="${tmake_file} t-crtin i386/t-softfp i386/t-crtstuff t-rtems"
 	;;
 i[34567]86-*-solaris2*)
+	tmake_file="$tmake_file i386/t-crtfm"
+	extra_parts="$extra_parts crtfastmath.o"
 	;;
 i[4567]86-wrs-vxworks|i[4567]86-wrs-vxworksae)
 	;;


-- 
-----------------------------------------------------------------------------
Rainer Orth, Center for Biotechnology, Bielefeld University

Re: Use i386/crtfastmath.c on Solaris 2/x86

[Uros Bizjak]

On Tue, May 31, 2011 at 2:25 PM, Rainer Orth
<ro@cebitec.uni-bielefeld.de> wrote:

> The only complication is that I need to make sure that SSE insns are only
> used if the host supports them.
>
> Bootstrapped without regressions on i386-pc-solaris2.8,
> i386-pc-solaris2.9, i386-pc-solaris2.11, and sparc-sun-solaris2.11.
>
> The libgcc part depends on the toplevel libgcc patch, so actually
> applying this patch will have to wait until that one is in.
>
> Ok for mainline?
>
>        Rainer
>
>
> 2011-05-28  Rainer Orth  <ro@CeBiTec.Uni-Bielefeld.DE>
>
>        gcc:
>        * config/i386/crtfastmath.c [!__x86_64__ && __sun__ && __svr4__]:
>        Include <signal.h>, <ucontext.h>.
>        (sigill_caught): Define.
>        (sigill_hdlr): New function.
>        (set_fast_math) [!__x86_64__ && __sun__ && __svr4__]: Check if SSE
>        insns can be executed.
>        * config/sol2.h (ENDFILE_SPEC): Use crtfastmath.o if -ffast-math
>        etc.
>        * config/sparc/sol2.h (ENDFILE_SPEC): Remove.
>
>        libgcc:
>        * config.host (i[34567]86-*-solaris2*): Add i386/t-crtfm to
>        tmake_file.
>        Add crtfastmath.o to extra_parts.

Please just put "if (edx & bit_SSE)" part inside existing check. You
will need to split assignment of mxcsr from the declaration, though.

OK with this change.

Thanks,
Uros.

Re: Use i386/crtfastmath.c on Solaris 2/x86

[Rainer Orth]

Uros Bizjak <ubizjak@gmail.com> writes:

> Please just put "if (edx & bit_SSE)" part inside existing check. You
> will need to split assignment of mxcsr from the declaration, though.
>
> OK with this change.

Here's the patch I've actually comitted after a quick bootstrap on
i386-pc-solaris2.10.

Thanks.
        Rainer


2011-05-28  Rainer Orth  <ro@CeBiTec.Uni-Bielefeld.DE>

	gcc:
	* config/i386/crtfastmath.c [!__x86_64__ && __sun__ && __svr4__]:
	Include <signal.h>, <ucontext.h>.
	(sigill_caught): Define.
	(sigill_hdlr): New function.
	(set_fast_math) [!__x86_64__ && __sun__ && __svr4__]: Check if SSE
	insns can be executed.
	* config/sol2.h (ENDFILE_SPEC): Use crtfastmath.o if -ffast-math
	etc.
	* config/sparc/sol2.h (ENDFILE_SPEC): Remove.

	libgcc:
	* config.host (i[34567]86-*-solaris2*): Add i386/t-crtfm to
	tmake_file.
	Add crtfastmath.o to extra_parts.

diff --git a/gcc/config/i386/crtfastmath.c b/gcc/config/i386/crtfastmath.c
--- a/gcc/config/i386/crtfastmath.c
+++ b/gcc/config/i386/crtfastmath.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2005, 2007, 2009 Free Software Foundation, Inc.
+ * Copyright (C) 2005, 2007, 2009, 2011 Free Software Foundation, Inc.
  *
  * This file is free software; you can redistribute it and/or modify it
  * under the terms of the GNU General Public License as published by the
@@ -30,6 +30,26 @@
 #include "cpuid.h"
 #endif
 
+#if !defined __x86_64 && defined __sun__ && defined __svr4__
+#include <signal.h>
+#include <ucontext.h>
+
+static volatile sig_atomic_t sigill_caught;
+
+static void
+sigill_hdlr (int sig __attribute((unused)),
+	     siginfo_t *sip __attribute__((unused)),
+	     ucontext_t *ucp)
+{
+  sigill_caught = 1;
+  /* Set PC to the instruction after the faulting one to skip over it,
+     otherwise we enter an infinite loop.  4 is the size of the stmxcsr
+     instruction.  */
+  ucp->uc_mcontext.gregs[EIP] += 4;
+  setcontext (ucp);
+}
+#endif
+
 static void __attribute__((constructor))
 #ifndef __x86_64__
 /* The i386 ABI only requires 4-byte stack alignment, so this is necessary
@@ -47,9 +67,31 @@ set_fast_math (void)
 
   if (edx & bit_SSE)
     {
-      unsigned int mxcsr = __builtin_ia32_stmxcsr ();
+      unsigned int mxcsr;
   
-      mxcsr |= MXCSR_FTZ;
+#if defined __sun__ && defined __svr4__
+      /* Solaris 2 before Solaris 9 4/04 cannot execute SSE instructions even
+	 if the CPU supports them.  Programs receive SIGILL instead, so check
+	 for that at runtime.  */
+      struct sigaction act, oact;
+
+      act.sa_handler = sigill_hdlr;
+      sigemptyset (&act.sa_mask);
+      /* Need to set SA_SIGINFO so a ucontext_t * is passed to the handler.  */
+      act.sa_flags = SA_SIGINFO;
+      sigaction (SIGILL, &act, &oact);
+
+      /* We need a single SSE instruction here so the handler can safely skip
+	 over it.  */
+      __asm__ volatile ("movss %xmm2,%xmm1");
+
+      sigaction (SIGILL, &oact, NULL);
+
+      if (sigill_caught)
+	return;
+#endif /* __sun__ && __svr4__ */
+
+      mxcsr = __builtin_ia32_stmxcsr () | MXCSR_FTZ;
 
       if (edx & bit_FXSAVE)
 	{
diff --git a/gcc/config/sol2.h b/gcc/config/sol2.h
--- a/gcc/config/sol2.h
+++ b/gcc/config/sol2.h
@@ -141,7 +141,9 @@ along with GCC; see the file COPYING3.  
      %{p|pg:-ldl} -lc}"
 
 #undef  ENDFILE_SPEC
-#define ENDFILE_SPEC "crtend.o%s crtn.o%s"
+#define ENDFILE_SPEC \
+  "%{Ofast|ffast-math|funsafe-math-optimizations:crtfastmath.o%s} \
+   crtend.o%s crtn.o%s"
 
 /* We don't use the standard svr4 STARTFILE_SPEC because it's wrong for us.  */
 #undef STARTFILE_SPEC
diff --git a/gcc/config/sparc/sol2.h b/gcc/config/sparc/sol2.h
--- a/gcc/config/sparc/sol2.h
+++ b/gcc/config/sparc/sol2.h
@@ -117,11 +117,6 @@ along with GCC; see the file COPYING3.  
 #define NO_DBX_BNSYM_ENSYM 1
 \f
 
-#undef  ENDFILE_SPEC
-#define ENDFILE_SPEC \
-  "%{Ofast|ffast-math|funsafe-math-optimizations:crtfastmath.o%s} \
-   crtend.o%s crtn.o%s"
-
 /* Select a format to encode pointers in exception handling data.  CODE
    is 0 for data, 1 for code labels, 2 for function pointers.  GLOBAL is
    true if the symbol may be affected by dynamic relocations.
diff --git a/libgcc/config.host b/libgcc/config.host
--- a/libgcc/config.host
+++ b/libgcc/config.host
@@ -338,6 +338,8 @@ i[34567]86-*-rtems*)
 	tmake_file="${tmake_file} t-crtin i386/t-softfp i386/t-crtstuff t-rtems"
 	;;
 i[34567]86-*-solaris2*)
+	tmake_file="$tmake_file i386/t-crtfm"
+	extra_parts="$extra_parts crtfastmath.o"
 	;;
 i[4567]86-wrs-vxworks|i[4567]86-wrs-vxworksae)
 	;;


-- 
-----------------------------------------------------------------------------
Rainer Orth, Center for Biotechnology, Bielefeld University

Re: Use i386/crtfastmath.c on Solaris 2/x86

[Richard Henderson]

On 06/01/2011 07:51 AM, Rainer Orth wrote:
> +  /* Set PC to the instruction after the faulting one to skip over it,
> +     otherwise we enter an infinite loop.  4 is the size of the stmxcsr
> +     instruction.  */
...
> +      /* We need a single SSE instruction here so the handler can safely skip
> +	 over it.  */
> +      __asm__ volatile ("movss %xmm2,%xmm1");

The comment referencing stmxcsr doesn't match the movss code.
It's still a 4 byte opcode, so the code still works.

I do wonder if using "movaps %xmm0,%xmm0" might be cleaner,
to avoid clobbering a register, even if that register is
surely dead anyway.  That's a 3 byte opcode though, so the
handler would need updating.


r~

Re: Use i386/crtfastmath.c on Solaris 2/x86

[Rainer Orth]

Richard Henderson <rth@redhat.com> writes:

> On 06/01/2011 07:51 AM, Rainer Orth wrote:
>> +  /* Set PC to the instruction after the faulting one to skip over it,
>> +     otherwise we enter an infinite loop.  4 is the size of the stmxcsr
>> +     instruction.  */
> ...
>> +      /* We need a single SSE instruction here so the handler can safely skip
>> +	 over it.  */
>> +      __asm__ volatile ("movss %xmm2,%xmm1");
>
> The comment referencing stmxcsr doesn't match the movss code.
> It's still a 4 byte opcode, so the code still works.

Copy-and-paste error ;-(  We already have the same code in
libgfortran/config/fpu-387.h and (without the comment) in
gcc/testsuite/lib/target-supports.exp.  I still mean to fix
driver-i386.c to correcly handle -march=native on Solaris 8 and 9 which
cannot in general execute SSE insns.  I wonder if there's a better place
to share this code?

> I do wonder if using "movaps %xmm0,%xmm0" might be cleaner,
> to avoid clobbering a register, even if that register is
> surely dead anyway.  That's a 3 byte opcode though, so the
> handler would need updating.

I'll give it a try.

Thanks.
        Rainer

-- 
-----------------------------------------------------------------------------
Rainer Orth, Center for Biotechnology, Bielefeld University

Re: Use i386/crtfastmath.c on Solaris 2/x86

[Richard Henderson]

On 06/01/2011 10:29 AM, Rainer Orth wrote:
> I still mean to fix
> driver-i386.c to correcly handle -march=native on Solaris 8 and 9 which
> cannot in general execute SSE insns.  I wonder if there's a better place
> to share this code?

I can't think of a good place.  :-(


r~

Re: Use i386/crtfastmath.c on Solaris 2/x86

[Rainer Orth]

Richard Henderson <rth@redhat.com> writes:

> On 06/01/2011 07:51 AM, Rainer Orth wrote:
>> +  /* Set PC to the instruction after the faulting one to skip over it,
>> +     otherwise we enter an infinite loop.  4 is the size of the stmxcsr
>> +     instruction.  */
> ...
>> +      /* We need a single SSE instruction here so the handler can safely skip
>> +	 over it.  */
>> +      __asm__ volatile ("movss %xmm2,%xmm1");
>
> The comment referencing stmxcsr doesn't match the movss code.
> It's still a 4 byte opcode, so the code still works.
>
> I do wonder if using "movaps %xmm0,%xmm0" might be cleaner,
> to avoid clobbering a register, even if that register is
> surely dead anyway.  That's a 3 byte opcode though, so the
> handler would need updating.

Here's what I installed after regtesting completed on
i386-pc-solaris2.{8, 11}.

Thanks.
	Rainer


2011-06-02  Richard Henderson  <rth@redhat.com>
	    Rainer Orth  <ro@CeBiTec.Uni-Bielefeld.DE>

	gcc:
	* config/i386/crtfastmath.c [!__x86_64__ && __sun__ && __svr4__]
	(sigill_hdlr): Correct insn, insn size.
	(set_fast_math) [!__x86_64__ && __sun__ && __svr4__]: Use movaps.

	gcc/testsuite:
	* lib/target-supports.exp (check_sse_os_support_available): Use
	movaps.

	libgfortran.
	* config/fpu-387.h [__sun__ && __svr4__] (sigill_hdlr): Correct
	insn, insn size.
	(has_sse) [!__x86_64__ && __sun__ && __svr4__]: Use movaps.

diff --git a/gcc/config/i386/crtfastmath.c b/gcc/config/i386/crtfastmath.c
--- a/gcc/config/i386/crtfastmath.c
+++ b/gcc/config/i386/crtfastmath.c
@@ -30,7 +30,7 @@
 #include "cpuid.h"
 #endif
 
-#if !defined __x86_64 && defined __sun__ && defined __svr4__
+#if !defined __x86_64__ && defined __sun__ && defined __svr4__
 #include <signal.h>
 #include <ucontext.h>
 
@@ -43,9 +43,9 @@ sigill_hdlr (int sig __attribute((unused
 {
   sigill_caught = 1;
   /* Set PC to the instruction after the faulting one to skip over it,
-     otherwise we enter an infinite loop.  4 is the size of the stmxcsr
+     otherwise we enter an infinite loop.  3 is the size of the movaps
      instruction.  */
-  ucp->uc_mcontext.gregs[EIP] += 4;
+  ucp->uc_mcontext.gregs[EIP] += 3;
   setcontext (ucp);
 }
 #endif
@@ -83,7 +83,7 @@ set_fast_math (void)
 
       /* We need a single SSE instruction here so the handler can safely skip
 	 over it.  */
-      __asm__ volatile ("movss %xmm2,%xmm1");
+      __asm__ volatile ("movaps %xmm0,%xmm0");
 
       sigaction (SIGILL, &oact, NULL);
 
diff --git a/gcc/testsuite/lib/target-supports.exp b/gcc/testsuite/lib/target-supports.exp
--- a/gcc/testsuite/lib/target-supports.exp
+++ b/gcc/testsuite/lib/target-supports.exp
@@ -1059,7 +1059,7 @@ proc check_sse_os_support_available { } 
 	    check_runtime_nocache sse_os_support_available {
 		int main ()
 		{
-		    __asm__ volatile ("movss %xmm2,%xmm1");
+		    __asm__ volatile ("movaps %xmm0,%xmm0");
 		    return 0;
 		}
 	    } "-msse"
diff --git a/libgfortran/config/fpu-387.h b/libgfortran/config/fpu-387.h
--- a/libgfortran/config/fpu-387.h
+++ b/libgfortran/config/fpu-387.h
@@ -1,5 +1,5 @@
 /* FPU-related code for x86 and x86_64 processors.
-   Copyright 2005, 2007, 2009, 2010 Free Software Foundation, Inc.
+   Copyright 2005, 2007, 2009, 2010, 2011 Free Software Foundation, Inc.
    Contributed by Francois-Xavier Coudert <coudert@clipper.ens.fr>
 
 This file is part of the GNU Fortran 95 runtime library (libgfortran).
@@ -40,9 +40,9 @@ sigill_hdlr (int sig __attribute((unused
 {
   sigill_caught = 1;
   /* Set PC to the instruction after the faulting one to skip over it,
-     otherwise we enter an infinite loop.  4 is the size of the stmxcsr
+     otherwise we enter an infinite loop.  3 is the size of the movaps
      instruction.  */
-  ucp->uc_mcontext.gregs[EIP] += 4;
+  ucp->uc_mcontext.gregs[EIP] += 3;
   setcontext (ucp);
 }
 #endif
@@ -73,7 +73,7 @@ has_sse (void)
 
       /* We need a single SSE instruction here so the handler can safely skip
 	 over it.  */
-      __asm__ volatile ("movss %xmm2,%xmm1");
+      __asm__ volatile ("movaps %xmm0,%xmm0");
 
       sigaction (SIGILL, &oact, NULL);
 


-- 
-----------------------------------------------------------------------------
Rainer Orth, Center for Biotechnology, Bielefeld University