Support SSE float environment in fenv.h functions.

	* cpu_features.c: New file.
	* cpu_features.h: New file.
	* crt1.c: Include "cpu_features.h".
	(__mingw_CRTStartup): Call cpu_features_init().
	* Makefile.in (MING_OBJS): Add cpu_features.c.
	(SRCDIST_FILES): Add cpu_features.c, cpu_features.h.
	* include/fenv,h ( fenv_t;): Append  __mxcsr field.
	(__MXCSR_EXCEPT_FLAG_SHIFT): New define.
	(__MXCSR_EXCEPT_MASK_SHIFT): New define.
	(__MXCSR_ROUND_FLAG_SHIFT): New define.
	* mingwex/feclearexcept.c: Include "cpu_features.h".
	Handle SSE environment.
	* mingwex/fegetenv.c: Likewise.
	* mingwex/feholdexcept.c: Likewise.
	* mingwex/fesetenv.c: Likewise.
	* mingwex/fesetexceptflag.c: Likewise.
	* mingwex/fesetround.c: Likewise.
	* mingwex/fetestexcept.c: Likewise.
	* mingwex/feupdateenv.c: Likewise.
	* mingwex/fegetround.c: Add comment.

15 files changed, 253 insertions, 18 deletions

diff --git a/winsup/mingw/ChangeLog b/winsup/mingw/ChangeLog
index cb4934969..ba8313e94 100644
--- a/winsup/mingw/ChangeLog
+++ b/winsup/mingw/ChangeLog
@@ -1,3 +1,6 @@
+2006-07-03  Danny Smith  <dannysmith@users.sourceforge.net>
+
+
 2006-06-25  Chris Sutcliffe  <ir0nh34d@users.sourceforge.net>
 
 	* Include/_mingw.h: Increment version to 3.10.
diff --git a/winsup/mingw/Makefile.in b/winsup/mingw/Makefile.in
index f439df773..1bbbf26aa 100644
--- a/winsup/mingw/Makefile.in
+++ b/winsup/mingw/Makefile.in
@@ -156,7 +156,7 @@ CRT0S = crt1.o dllcrt1.o crt2.o dllcrt2.o CRT_noglob.o crtmt.o crtst.o \
 	CRT_fp8.o CRT_fp10.o txtmode.o binmode.o
 MINGW_OBJS = CRTglob.o CRTfmode.o CRTinit.o  dllmain.o gccmain.o \
 	     main.o crtst.o mthr_stub.o CRT_fp10.o txtmode.o \
-	     pseudo-reloc.o pseudo-reloc-list.o
+	     pseudo-reloc.o pseudo-reloc-list.o cpu_features.o
 
 MOLD_OBJS = isascii.o iscsym.o iscsymf.o toascii.o \
 	strcasecmp.o strncasecmp.o wcscmpi.o
@@ -187,7 +187,7 @@ mthr.c mthr_init.c mthr_stub.c readme.txt \
 isascii.c iscsym.c iscsymf.c toascii.c \
 strcasecmp.c strncasecmp.c wcscmpi.c \
 CRT_fp8.c CRT_fp10.c test_headers.c txtmode.c binmode.c pseudo-reloc.c \
-pseudo-reloc-list.c \
+pseudo-reloc-list.c cpu_features.c cpu_features.h\
 DISCLAIMER CONTRIBUTORS
 
 
diff --git a/winsup/mingw/cpu_features.c b/winsup/mingw/cpu_features.c
new file mode 100755
index 000000000..0eccfbfe3
--- /dev/null
+++ b/winsup/mingw/cpu_features.c
@@ -0,0 +1,105 @@
+#include <stdbool.h>
+#include "cpu_features.h"
+
+/* level 1 edx bits */
+#define EDX_CX8 (1 << 8) /* CMPXCHG8B */
+#define EDX_CMOV (1 << 15)
+#define EDX_MMX (1 << 23)
+#define EDX_FXSR (1 << 24) /* FXSAVE and FXRSTOR */
+#define EDX_SSE (1 << 25)
+#define EDX_SSE2 (1 << 26) 
+
+/*  level 1 ecx bits */
+#define ECX_SSE3 (1 << 0)
+#define ECX_CX16 (1 << 13) /* CMPXCHG16B */
+
+/* extended level 0x80000001 edx bits */
+#define EDX_3DNOW (1 << 31)
+#define EDX_3DNOWP (1 << 30)
+#define EDX_LM (1 << 29) /*LONG MODE */
+
+#define __cpuid(level,a,b,c,d)			 		\
+  __asm__ __volatile__ ("cpuid;"				\
+			: "=a" (a), "=b" (b), "=c" (c), "=d" (d)\
+			: "0" (level))
+
+/* Combine the different cpuid flags into a single bitmap.  */ 
+
+unsigned int __cpu_features = 0;
+
+void  __cpu_features_init (void)
+{
+  unsigned int eax, ebx, ecx, edx;
+  /* Try to change the value of CPUID bit (bit 21) in EFLAGS.
+     If the bit can be toggled, CPUID is supported.  */
+  asm volatile ("pushfl; pushfl; popl %0;"
+		"movl %0,%1; xorl %2,%0;"
+		"pushl %0; popfl; pushfl; popl %0; popfl"
+		: "=&r" (eax), "=&r" (ebx)
+		: "i" (0x00200000));
+
+  if (((eax ^ ebx) & 0x00200000) == 0)
+    return;
+
+  __cpuid (0, eax, ebx, ecx, edx);
+  if (eax == 0)
+    return;
+
+  __cpuid (1, eax, ebx, ecx, edx);
+
+  if (edx & EDX_CX8)
+     __cpu_features |= _CRT_CMPXCHG8B;
+  if (edx & EDX_CMOV)
+     __cpu_features |= _CRT_CMOV;
+
+  if (edx & EDX_MMX)
+     __cpu_features |= _CRT_MMX;
+  if (edx & EDX_FXSR)
+     __cpu_features |= _CRT_FXSR; 
+  if (edx & EDX_SSE)
+     __cpu_features |= _CRT_SSE;
+  if (edx & EDX_SSE2)
+     __cpu_features |= _CRT_SSE2;
+
+
+  if (ecx & ECX_SSE3)
+     __cpu_features |= _CRT_SSE3;
+  if (ecx & ECX_CX16)
+     __cpu_features |= _CRT_CMPXCHG16B;
+
+  __cpuid (0x80000000, eax, ebx, ecx, edx);
+  if (eax < 0x80000001)
+    return;
+  __cpuid (0x80000001, eax, ebx, ecx, edx);
+  if (edx & EDX_3DNOW);
+    __cpu_features |= _CRT_3DNOW; 
+  if (edx & EDX_3DNOWP)
+    __cpu_features |= _CRT_3DNOWP; 
+
+  return;
+}
+
+#ifdef TEST
+
+#include <stdio.h>
+#define report(feature) \
+  if ((feature) & __cpu_features) printf( #feature " found\n")
+
+int main()
+{
+  __cpu_features_init();
+
+  report(_CRT_CMPXCHG8B);
+  report(_CRT_CMOV);
+  report(_CRT_MMX);
+  report(_CRT_FXSR);
+  report(_CRT_SSE);
+  report(_CRT_SSE2);
+  report(_CRT_SSE3);
+  report(_CRT_CMPXCHG16B);
+  report(_CRT_3DNOW);
+  report(_CRT_3DNOWP);
+ return 0;
+}
+
+#endif
diff --git a/winsup/mingw/cpu_features.h b/winsup/mingw/cpu_features.h
new file mode 100755
index 000000000..a19f875c8
--- /dev/null
+++ b/winsup/mingw/cpu_features.h
@@ -0,0 +1,23 @@
+#ifndef _CPU_FEATURES_H
+#define _CPU_FEATURES_H
+
+#include <stdbool.h>
+
+#define  _CRT_CMPXCHG8B		0x0001
+#define  _CRT_CMOV		0x0002
+#define  _CRT_MMX		0x0004
+#define  _CRT_FXSR		0x0008
+#define  _CRT_SSE		0x0010
+#define  _CRT_SSE2		0x0020
+#define  _CRT_SSE3		0x0040
+#define  _CRT_CMPXCHG16B	0x0080
+#define  _CRT_3DNOW		0x0100
+#define  _CRT_3DNOWP		0x0200
+
+extern unsigned int __cpu_features;
+
+/* Currently we use this in fpenv  functions */
+#define __HAS_SSE  __cpu_features & _CRT_SSE
+
+
+#endif
diff --git a/winsup/mingw/crt1.c b/winsup/mingw/crt1.c
index b9b2bd5b1..ad8270a04 100644
--- a/winsup/mingw/crt1.c
+++ b/winsup/mingw/crt1.c
@@ -27,6 +27,7 @@
  *       be manually synchronized, but it does lead to this not-generally-
  *       a-good-idea use of include. */
 #include "init.c"
+#include "cpu_features.h"
 
 extern void _pei386_runtime_relocator (void);
 
@@ -195,6 +196,7 @@ __mingw_CRTStartup (void)
   /*
    * Initialize floating point unit.
    */
+  __cpu_features_init ();	/* Do we have SSE, etc.*/
   _fpreset ();			/* Supplied by the runtime library. */
 
   /*
diff --git a/winsup/mingw/include/fenv.h b/winsup/mingw/include/fenv.h
index 43dfd338a..5a2270289 100644
--- a/winsup/mingw/include/fenv.h
+++ b/winsup/mingw/include/fenv.h
@@ -1,7 +1,6 @@
 #ifndef _FENV_H_
 #define _FENV_H_
 
-
 /* FPU status word exception flags */
 #define FE_INVALID	0x01
 #define FE_DENORMAL	0x02
@@ -18,6 +17,18 @@
 #define FE_UPWARD	0x0800
 #define FE_TOWARDZERO	0x0c00
 
+/* The MXCSR exception flags are the same as the
+   FE flags. */
+#define __MXCSR_EXCEPT_FLAG_SHIFT 0
+
+/* How much to shift FE status word exception flags
+   to get the MXCSR exeptions masks,  */
+#define __MXCSR_EXCEPT_MASK_SHIFT 7
+
+/* How much to shift FE control word rounding flags
+   to get MXCSR rounding flags,  */
+#define __MXCSR_ROUND_FLAG_SHIFT 3
+
 #ifndef RC_INVOKED
 /*
   For now, support only for the basic abstraction of flags that are
@@ -26,8 +37,10 @@
 */
 typedef unsigned short fexcept_t;
 
-/* This 28-byte struct represents the entire floating point
-   environment as stored by fnstenv or fstenv */
+/* This 32-byte struct represents the entire floating point
+   environment as stored by fnstenv or fstenv, augmented by
+   the  contents of the MXCSR register, as stored by stmxcsr
+   (if CPU supports it). */
 typedef struct
 {
   unsigned short __control_word;
@@ -40,8 +53,9 @@ typedef struct
   unsigned short __ip_selector;  
   unsigned short __opcode;
   unsigned int	 __data_offset;
-  unsigned short __data_selector;  
-  unsigned short __unused3;
+  unsigned short __data_selector;
+  unsigned short  __unused3;
+  unsigned int __mxcsr; /* contents of the MXCSR register  */
 } fenv_t;
 
 
diff --git a/winsup/mingw/mingwex/feclearexcept.c b/winsup/mingw/mingwex/feclearexcept.c
index 8c943893b..a68884f41 100644
--- a/winsup/mingw/mingwex/feclearexcept.c
+++ b/winsup/mingw/mingwex/feclearexcept.c
@@ -1,4 +1,5 @@
 #include <fenv.h>
+#include "cpu_features.h"
 
 /* 7.6.2.1
    The feclearexcept function clears the supported exceptions
@@ -7,9 +8,17 @@
 int feclearexcept (int excepts)
 {
   fenv_t _env;
+  excepts &= FE_ALL_EXCEPT;
   __asm__ volatile ("fnstenv %0;" : "=m" (_env)); /* get the env */
-  _env.__status_word &= ~(excepts & FE_ALL_EXCEPT); /* clear the except */
+  _env.__status_word &= ~excepts; /* clear the except */
   __asm__ volatile ("fldenv %0;" :: "m" (_env)); /*set the env */
 
+  if (__HAS_SSE)
+    {
+       unsigned _csr;
+       __asm__ volatile("stmxcsr %0" : "=m" (_csr)); /* get the register */
+       _csr &= ~excepts; /* clear the except */
+       __asm__ volatile("ldmxcsr %0" : : "m" (_csr)); /* set the register */
+    }
   return 0;
 }
diff --git a/winsup/mingw/mingwex/fegetenv.c b/winsup/mingw/mingwex/fegetenv.c
index 930105673..0c5d591e7 100644
--- a/winsup/mingw/mingwex/fegetenv.c
+++ b/winsup/mingw/mingwex/fegetenv.c
@@ -1,4 +1,5 @@
 #include <fenv.h>
+#include "cpu_features.h"
 
 /* 7.6.4.1
    The fegetenv function stores the current floating-point environment
@@ -10,5 +11,10 @@ int fegetenv (fenv_t * envp)
  /* fnstenv sets control word to non-stop for all exceptions, so we
     need to reload our env to restore the original mask.  */
   __asm__ ("fldenv %0" : : "m" (*envp));
+
+  /* And the SSE environment.  */
+  if (__HAS_SSE)
+    __asm__ ("stmxcsr %0" : "=m" (envp->__mxcsr));
+
   return 0;
 }
diff --git a/winsup/mingw/mingwex/fegetround.c b/winsup/mingw/mingwex/fegetround.c
index 076b1068c..01d42849f 100644
--- a/winsup/mingw/mingwex/fegetround.c
+++ b/winsup/mingw/mingwex/fegetround.c
@@ -1,4 +1,5 @@
 #include <fenv.h>
+#include "cpu_features.h"
 
 /* 7.6.3.1
    The fegetround function returns the value of the rounding direction
@@ -9,6 +10,10 @@ fegetround (void)
 {
   unsigned short _cw;
   __asm__ ("fnstcw %0;"	: "=m" (_cw));
+
+  /* If the MXCSR flag is different, there is no way to indicate, so just
+     report the FPU flag. */
   return _cw
           & (FE_TONEAREST | FE_DOWNWARD | FE_UPWARD | FE_TOWARDZERO);
+
 }
diff --git a/winsup/mingw/mingwex/feholdexcept.c b/winsup/mingw/mingwex/feholdexcept.c
index 8d4a3f67b..66912d85d 100644
--- a/winsup/mingw/mingwex/feholdexcept.c
+++ b/winsup/mingw/mingwex/feholdexcept.c
@@ -1,4 +1,5 @@
 #include <fenv.h>
+#include "cpu_features.h"
 
 /* 7.6.4.2
    The feholdexcept function saves the current floating-point
@@ -12,5 +13,18 @@ int feholdexcept (fenv_t * envp)
  /* fnstenv sets control word to non-stop for all exceptions, so all we
     need to do is clear the exception flags.  */
   __asm__ ("fnclex");
+
+  if (__HAS_SSE)
+    {
+       unsigned int _csr;
+       /* Save the SSE MXCSR register.  */
+       __asm__ ("stmxcsr %0" :  "=m" (envp->__mxcsr));  
+       /* Clear the exception flags. */
+       _csr = envp->__mxcsr & ~FE_ALL_EXCEPT;
+       /* Set exception mask to non-stop  */
+       _csr |= (FE_ALL_EXCEPT << __MXCSR_EXCEPT_MASK_SHIFT) /*= 0x1f80 */;
+       __asm__ volatile ("ldmxcsr %0" : : "m" (_csr));
+    } 
+
   return 0;
 }
diff --git a/winsup/mingw/mingwex/fesetenv.c b/winsup/mingw/mingwex/fesetenv.c
index 3d7b5604a..105a0af10 100644
--- a/winsup/mingw/mingwex/fesetenv.c
+++ b/winsup/mingw/mingwex/fesetenv.c
@@ -1,5 +1,6 @@
 #include <fenv.h>
 #include <float.h>
+#include "cpu_features.h"
 
 /* 7.6.4.3
    The fesetenv function establishes the floating-point environment
@@ -15,6 +16,11 @@ extern void (*_imp___fpreset)( void ) ;
 
 int fesetenv (const fenv_t * envp)
 {
+  /* Default mxcsr status is to mask all exceptions.  All other bits
+     are zero.  */
+     
+  unsigned int _csr = FE_ALL_EXCEPT << __MXCSR_EXCEPT_MASK_SHIFT /*= 0x1f80 */; 
+  
   if (envp == FE_PC64_ENV)
    /*
     *  fninit initializes the control register to 0x37f,
@@ -37,7 +43,15 @@ int fesetenv (const fenv_t * envp)
     _fpreset();
 
   else
-    __asm__ ("fldenv %0;" : : "m" (*envp));
-
+    {
+      __asm__ ("fldenv %0;" : : "m" (*envp));
+       /* Setting the reserved high order bits of MXCSR causes a segfault */
+       _csr = envp ->__mxcsr & 0xffff;
+    }
+
+  /* Set MXCSR */   
+   if (__HAS_SSE)
+     __asm__ volatile ("ldmxcsr %0" : : "m" (_csr));
+ 
   return 0;
 }
diff --git a/winsup/mingw/mingwex/fesetexceptflag.c b/winsup/mingw/mingwex/fesetexceptflag.c
index 7f4b8e562..876174b69 100644
--- a/winsup/mingw/mingwex/fesetexceptflag.c
+++ b/winsup/mingw/mingwex/fesetexceptflag.c
@@ -1,4 +1,5 @@
 #include <fenv.h> 
+#include "cpu_features.h"
 
 /* 7.6.2.4
    The fesetexceptflag function sets the complete status for those
@@ -18,5 +19,15 @@ int fesetexceptflag (const fexcept_t * flagp, int excepts)
   _env.__status_word &= ~excepts;
   _env.__status_word |= (*flagp & excepts);
   __asm__ volatile ("fldenv %0;" : : "m" (_env));
+
+  if (__HAS_SSE)
+   {
+      unsigned int _csr;
+      __asm__ __volatile__("stmxcsr %0" : "=m" (_csr));
+      _csr &= ~excepts;
+      _csr |= *flagp & excepts;
+      __asm__ volatile ("ldmxcsr %0" : : "m" (_csr));
+   }
+
   return 0;
 }
diff --git a/winsup/mingw/mingwex/fesetround.c b/winsup/mingw/mingwex/fesetround.c
index a8cef86a4..e90dbb4e7 100644
--- a/winsup/mingw/mingwex/fesetround.c
+++ b/winsup/mingw/mingwex/fesetround.c
@@ -1,4 +1,6 @@
 #include <fenv.h>
+#include "cpu_features.h"
+
  /* 7.6.3.2
     The fesetround function establishes the rounding direction
     represented by its argument round. If the argument is not equal
@@ -15,5 +17,14 @@ int fesetround (int mode)
   _cw &= ~(FE_TONEAREST | FE_DOWNWARD | FE_UPWARD | FE_TOWARDZERO);
   _cw |= mode;
   __asm__ volatile ("fldcw %0;" : : "m" (_cw));
+
+ if (__HAS_SSE)
+    {
+      __asm__ volatile ("stmxcsr %0" : "=m" (_cw));
+      _cw &= ~ 0x6000;
+      _cw |= (mode <<  __MXCSR_ROUND_FLAG_SHIFT);
+      __asm__ volatile ("ldmxcsr %0" : : "m" (_cw));
+    }
+
   return 0;
 }
diff --git a/winsup/mingw/mingwex/fetestexcept.c b/winsup/mingw/mingwex/fetestexcept.c
index 6934ed925..062d585eb 100644
--- a/winsup/mingw/mingwex/fetestexcept.c
+++ b/winsup/mingw/mingwex/fetestexcept.c
@@ -1,4 +1,5 @@
-#include <fenv.h> 
+#include <fenv.h>
+#include "cpu_features.h"
 /* 7.6.2.5 
    The fetestexcept function determines which of a specified subset of
    the exception flags are currently set. The excepts argument
@@ -9,7 +10,18 @@
 
 int fetestexcept (int excepts)
 {
-  unsigned short _sw;
-  __asm__ ("fnstsw %%ax" : "=a" (_sw));
-  return _sw & excepts & FE_ALL_EXCEPT;
+  
+  unsigned int _res;
+  __asm__ ("fnstsw %%ax" : "=a" (_res));
+
+
+  /* If SSE supported, return the union of the FPU and SSE flags.  */
+  if (__HAS_SSE)
+    {     
+        unsigned int _csr; 	
+      __asm__ volatile("stmxcsr %0" : "=m" (_csr));
+      _res |= _csr;    
+    }
+
+  return (_res & excepts & FE_ALL_EXCEPT);
 }
diff --git a/winsup/mingw/mingwex/feupdateenv.c b/winsup/mingw/mingwex/feupdateenv.c
index f414837f5..e37566b02 100644
--- a/winsup/mingw/mingwex/feupdateenv.c
+++ b/winsup/mingw/mingwex/feupdateenv.c
@@ -1,4 +1,5 @@
 #include <fenv.h>
+#include "cpu_features.h"
 
 /* 7.6.4.4
    The feupdateenv function saves the currently raised exceptions in
@@ -8,13 +9,18 @@
    set by a call to feholdexcept or fegetenv, or equal the macro
    FE_DFL_ENV or an implementation-defined environment macro. */
 
-/* FIXME: this works but surely there must be a better way.  */
 
 int feupdateenv (const fenv_t * envp)
 {
-  unsigned int _fexcept = fetestexcept (FE_ALL_EXCEPT); /*save excepts */
+  unsigned int _fexcept;
+  __asm__ ("fnstsw %%ax" : "=a" (_fexcept)); /*save excepts */
+  if (__HAS_SSE)
+    {
+      unsigned int  _csr;
+      __asm__ ("stmxcsr %0" : "=m" (_csr));
+      _fexcept |= _csr;
+    }
   fesetenv (envp); /* install the env  */
-  feraiseexcept (_fexcept); /* raise the execept */
+  feraiseexcept (_fexcept & FE_ALL_EXCEPT); /* raise the execeptions */
   return 0;
 }
-