[PATCH 7/7] Optimize private 387 fenv access; share code between i386 and x86_64.

[Richard Henderson <rth at twiddle dot net>]

	* sysdeps/i386/fpu/fenv_private.h: New file.
	* sysdeps/i386/fpu/math_private.h: Use it.
	(math_opt_barrier, math_force_eval): Remove.
	(libc_feholdexcept_setround_53bit): Remove.
	(libc_feupdateenv_53bit): Remove.
	* sysdeps/x86_64/fpu/math_private.h: Likewise.
	(math_opt_barrier, math_force_eval): Remove.
	(libc_feholdexcept): Remove.
	(libc_feholdexcept_setround): Remove.
	(libc_fetestexcept, libc_fesetenv): Remove.
	(libc_feupdateenv_test): Remove.
	(libc_feupdateenv, libc_feholdsetround): Remove.
	(libc_feresetround): Remove.
---
 sysdeps/i386/fpu/fenv_private.h   |  304 +++++++++++++++++++++++++++++++++++++
 sysdeps/i386/fpu/math_private.h   |   46 +------
 sysdeps/x86_64/fpu/math_private.h |  120 +---------------
 3 files changed, 307 insertions(+), 163 deletions(-)
 create mode 100644 sysdeps/i386/fpu/fenv_private.h

diff --git a/sysdeps/i386/fpu/fenv_private.h b/sysdeps/i386/fpu/fenv_private.h
new file mode 100644
index 0000000..f33f57c
--- /dev/null
+++ b/sysdeps/i386/fpu/fenv_private.h
@@ -0,0 +1,304 @@
+#ifndef FENV_PRIVATE_H
+#define FENV_PRIVATE_H 1
+
+#include <fenv.h>
+#include <fpu_control.h>
+
+#ifdef __SSE2_MATH__
+# define math_opt_barrier(x) \
+  ({ __typeof(x) __x;					\
+     if (sizeof (x) <= sizeof (double))			\
+       __asm ("" : "=x" (__x) : "0" (x));		\
+     else						\
+       __asm ("" : "=t" (__x) : "0" (x));		\
+     __x; })
+# define math_force_eval(x) \
+  do {							\
+    if (sizeof (x) <= sizeof (double))			\
+      __asm __volatile ("" : : "x" (x));		\
+    else						\
+      __asm __volatile ("" : : "f" (x));		\
+  } while (0)
+#else
+# define math_opt_barrier(x) \
+  ({ __typeof (x) __x;					\
+     __asm ("" : "=t" (__x) : "0" (x));			\
+     __x; })
+# define math_force_eval(x) \
+  do {							\
+    __typeof (x) __x = (x);				\
+    if (sizeof (x) <= sizeof (double))			\
+      __asm __volatile ("" : : "m" (__x));		\
+    else						\
+      __asm __volatile ("" : : "f" (__x));		\
+  } while (0)
+#endif
+
+/* This file is used by both the 32- and 64-bit ports.  The 64-bit port
+   has a field in the fenv_t for the mxcsr; the 32-bit port does not.
+   Instead, we (ab)use the only 32-bit field extant in the struct.  */
+#ifndef __x86_64__
+# define __mxcsr	__eip
+#endif
+
+
+/* All of these functions are private to libm, and are all used in pairs
+   to save+change the fp state and restore the original state.  Thus we
+   need not care for both the 387 and the sse unit, only the one we're
+   actually using.  */
+
+#if defined __AVX__ || defined SSE2AVX
+# define STMXCSR "vstmxcsr"
+# define LDMXCSR "vldmxcsr"
+#else
+# define STMXCSR "stmxcsr"
+# define LDMXCSR "ldmxcsr"
+#endif
+
+static __always_inline void
+libc_feholdexcept_sse (fenv_t *e)
+{
+  unsigned int mxcsr;
+  asm (STMXCSR " %0" : "=m" (*&mxcsr));
+  e->__mxcsr = mxcsr;
+  mxcsr = (mxcsr | 0x1f80) & ~0x3f;
+  asm volatile (LDMXCSR " %0" : : "m" (*&mxcsr));
+}
+
+static __always_inline void
+libc_feholdexcept_387 (fenv_t *e)
+{
+  /* Recall that fnstenv has a side-effect of masking exceptions.
+     Clobber all of the fp registers so that the TOS field is 0.  */
+  asm volatile ("fnstenv %0; fnclex"
+		: "=m"(*e)
+		: : "st", "st(1)", "st(2)", "st(3)",
+		    "st(4)", "st(5)", "st(6)", "st(7)");
+}
+
+static __always_inline void
+libc_feholdexcept_setround_sse (fenv_t *e, int r)
+{
+  unsigned int mxcsr;
+  asm (STMXCSR " %0" : "=m" (*&mxcsr));
+  e->__mxcsr = mxcsr;
+  mxcsr = ((mxcsr | 0x1f80) & ~0x603f) | (r << 3);
+  asm volatile (LDMXCSR " %0" : : "m" (*&mxcsr));
+}
+
+/* Set both rounding mode and precision.  A convenience function for use
+   by libc_feholdexcept_setround and libc_feholdexcept_setround_53bit. */
+static __always_inline void
+libc_feholdexcept_setround_387_prec (fenv_t *e, int r)
+{
+  libc_feholdexcept_387 (e);
+
+  fpu_control_t cw = e->__control_word;
+  cw &= ~(_FPU_RC_ZERO | _FPU_EXTENDED);
+  cw |= r | 0x3f;
+  _FPU_SETCW (cw);
+}
+
+static __always_inline void
+libc_feholdexcept_setround_387 (fenv_t *e, int r)
+{
+  libc_feholdexcept_setround_387_prec (e, r | _FPU_EXTENDED);
+}
+
+static __always_inline void
+libc_feholdexcept_setround_387_53bit (fenv_t *e, int r)
+{
+  libc_feholdexcept_setround_387_prec (e, r | _FPU_DOUBLE);
+}
+
+static __always_inline int
+libc_fetestexcept_sse (int e)
+{
+  unsigned int mxcsr;
+  asm volatile (STMXCSR " %0" : "=m" (*&mxcsr));
+  return mxcsr & e & FE_ALL_EXCEPT;
+}
+
+static __always_inline int
+libc_fetestexcept_387 (int ex)
+{
+  fexcept_t temp;
+  asm volatile ("fnstsw %0" : "=a" (temp));
+  return temp & ex & FE_ALL_EXCEPT;
+}
+
+static __always_inline void
+libc_fesetenv_sse (fenv_t *e)
+{
+  asm volatile (LDMXCSR " %0" : : "m" (e->__mxcsr));
+}
+
+static __always_inline void
+libc_fesetenv_387 (fenv_t *e)
+{
+  /* Clobber all fp registers so that the TOS value we saved earlier is
+     compatible with the current state of the compiler.  */
+  asm volatile ("fldenv %0"
+		: : "m" (*e)
+		: "st", "st(1)", "st(2)", "st(3)",
+		  "st(4)", "st(5)", "st(6)", "st(7)");
+}
+
+static __always_inline int
+libc_feupdateenv_test_sse (fenv_t *e, int ex)
+{
+  unsigned int mxcsr, old_mxcsr, cur_ex;
+  asm volatile (STMXCSR " %0" : "=m" (*&mxcsr));
+  cur_ex = mxcsr & FE_ALL_EXCEPT;
+
+  /* Merge current exceptions with the old environment.  */
+  old_mxcsr = e->__mxcsr;
+  mxcsr = old_mxcsr | cur_ex;
+  asm volatile (LDMXCSR " %0" : : "m" (*&mxcsr));
+
+  /* Raise SIGFPE for any new exceptions since the hold.  Expect that
+     the normal environment has all exceptions masked.  */
+  if (__builtin_expect ((old_mxcsr >> 7) & cur_ex, 0))
+    __feraiseexcept (cur_ex);
+
+  /* Test for exceptions raised since the hold.  */
+  return cur_ex & ex;
+}
+
+static __always_inline int
+libc_feupdateenv_test_387 (fenv_t *e, int ex)
+{
+  fexcept_t cur_ex;
+
+  /* Save current exceptions.  */
+  asm volatile ("fnstsw %0" : "=a" (cur_ex));
+  cur_ex &= FE_ALL_EXCEPT;
+
+  /* Reload original environment.  */
+  libc_fesetenv_387 (e);
+
+  /* Merge current exceptions.  */
+  __feraiseexcept (cur_ex);
+
+  /* Test for exceptions raised since the hold.  */
+  return cur_ex & ex;
+}
+
+static __always_inline void
+libc_feupdateenv_sse (fenv_t *e)
+{
+  libc_feupdateenv_test_sse (e, 0);
+}
+
+static __always_inline void
+libc_feupdateenv_387 (fenv_t *e)
+{
+  libc_feupdateenv_test_387 (e, 0);
+}
+
+static __always_inline void
+libc_feholdsetround_sse (fenv_t *e, int r)
+{
+  unsigned int mxcsr;
+  asm (STMXCSR " %0" : "=m" (*&mxcsr));
+  e->__mxcsr = mxcsr;
+  mxcsr = (mxcsr & ~0x6000) | (r << 3);
+  asm volatile (LDMXCSR " %0" : : "m" (*&mxcsr));
+}
+
+static __always_inline void
+libc_feholdsetround_387_prec (fenv_t *e, int r)
+{
+  fpu_control_t cw;
+
+  _FPU_GETCW (cw);
+  e->__control_word = cw;
+  cw &= ~(_FPU_RC_ZERO | _FPU_EXTENDED);
+  cw |= r;
+  _FPU_SETCW (cw);
+}
+
+static __always_inline void
+libc_feholdsetround_387 (fenv_t *e, int r)
+{
+  libc_feholdsetround_387_prec (e, r | _FPU_EXTENDED);
+}
+
+static __always_inline void
+libc_feholdsetround_387_53bit (fenv_t *e, int r)
+{
+  libc_feholdsetround_387_prec (e, r | _FPU_DOUBLE);
+}
+
+static __always_inline void
+libc_feresetround_sse (fenv_t *e)
+{
+  unsigned int mxcsr;
+  asm (STMXCSR " %0" : "=m" (*&mxcsr));
+  mxcsr = (mxcsr & ~0x6000) | (e->__mxcsr & 0x6000);
+  asm volatile (LDMXCSR " %0" : : "m" (*&mxcsr));
+}
+
+static __always_inline void
+libc_feresetround_387 (fenv_t *e)
+{
+  _FPU_SETCW (e->__control_word);
+}
+
+#ifdef __SSE_MATH__
+# define libc_feholdexceptf		libc_feholdexcept_sse
+# define libc_feholdexcept_setroundf	libc_feholdexcept_setround_sse
+# define libc_fetestexceptf		libc_fetestexcept_sse
+# define libc_fesetenvf			libc_fesetenv_sse
+# define libc_feupdateenv_testf		libc_feupdateenv_test_sse
+# define libc_feupdateenvf		libc_feupdateenv_sse
+# define libc_feholdsetroundf		libc_feholdsetround_sse
+# define libc_feresetroundf		libc_feresetround_sse
+#else
+# define libc_feholdexceptf		libc_feholdexcept_387
+# define libc_feholdexcept_setroundf	libc_feholdexcept_setround_387
+# define libc_fetestexceptf		libc_fetestexcept_387
+# define libc_fesetenvf			libc_fesetenv_387
+# define libc_feupdateenv_testf		libc_feupdateenv_test_387
+# define libc_feupdateenvf		libc_feupdateenv_387
+# define libc_feholdsetroundf		libc_feholdsetround_387
+# define libc_feresetroundf		libc_feresetround_387
+#endif /* __SSE_MATH__ */
+
+#ifdef __SSE2_MATH__
+# define libc_feholdexcept		libc_feholdexcept_sse
+# define libc_feholdexcept_setround	libc_feholdexcept_setround_sse
+# define libc_fetestexcept		libc_fetestexcept_sse
+# define libc_fesetenv			libc_fesetenv_sse
+# define libc_feupdateenv_test		libc_feupdateenv_test_sse
+# define libc_feupdateenv		libc_feupdateenv_sse
+# define libc_feholdsetround		libc_feholdsetround_sse
+# define libc_feresetround		libc_feresetround_sse
+#else
+# define libc_feholdexcept		libc_feholdexcept_387
+# define libc_feholdexcept_setround	libc_feholdexcept_setround_387
+# define libc_fetestexcept		libc_fetestexcept_387
+# define libc_fesetenv			libc_fesetenv_387
+# define libc_feupdateenv_test		libc_feupdateenv_test_387
+# define libc_feupdateenv		libc_feupdateenv_387
+# define libc_feholdsetround		libc_feholdsetround_387
+# define libc_feresetround		libc_feresetround_387
+#endif /* __SSE2_MATH__ */
+
+#define libc_feholdexceptl		libc_feholdexcept_387
+#define libc_feholdexcept_setroundl	libc_feholdexcept_setround_387
+#define libc_fetestexceptl		libc_fetestexcept_387
+#define libc_fesetenvl			libc_fesetenv_387
+#define libc_feupdateenv_testl		libc_feupdateenv_test_387
+#define libc_feupdateenvl		libc_feupdateenv_387
+#define libc_feholdsetroundl		libc_feholdsetround_387
+#define libc_feresetroundl		libc_feresetround_387
+
+#ifndef __SSE2_MATH__
+# define libc_feholdexcept_setround_53bit libc_feholdexcept_setround_387_53bit
+# define libc_feholdsetround_53bit	libc_feholdsetround_387_53bit
+#endif
+
+#undef __mxcsr
+
+#endif /* FENV_PRIVATE_H */
diff --git a/sysdeps/i386/fpu/math_private.h b/sysdeps/i386/fpu/math_private.h
index 9236448..541a7f8 100644
--- a/sysdeps/i386/fpu/math_private.h
+++ b/sysdeps/i386/fpu/math_private.h
@@ -1,48 +1,6 @@
 #ifndef _MATH_PRIVATE_H
 
-#include <fenv.h>
-#include <fpu_control.h>
-
-#define math_opt_barrier(x) \
-({ __typeof (x) __x;					\
-   __asm ("" : "=t" (__x) : "0" (x));			\
-   __x; })
-#define math_force_eval(x) \
-do							\
-  {							\
-    __typeof (x) __x = (x);				\
-    if (sizeof (x) <= sizeof (double))			\
-      __asm __volatile ("" : : "m" (__x));		\
-    else						\
-      __asm __volatile ("" : : "f" (__x));		\
-  }							\
-while (0)
-
-static __always_inline void
-libc_feholdexcept_setround_53bit (fenv_t *e, int r)
-{
-  feholdexcept (e);
-  fesetround (r);
-
-  fpu_control_t cw;
-  _FPU_GETCW (cw);
-  cw &= ~(fpu_control_t) _FPU_EXTENDED;
-  cw |= _FPU_DOUBLE;
-  _FPU_SETCW (cw);
-}
-#define libc_feholdexcept_setround_53bit libc_feholdexcept_setround_53bit
-
-static __always_inline void
-libc_feupdateenv_53bit (fenv_t *e)
-{
-  feupdateenv (e);
-
-  /* Unfortunately, feupdateenv fails to affect the rounding precision.
-     We can get that back by restoring the exact control word we saved.  */
-  _FPU_SETCW (e->__control_word);
-}
-#define libc_feupdateenv_53bit libc_feupdateenv_53bit
-
+#include "fenv_private.h"
 #include_next <math_private.h>
 
-#endif /* _MATH_PRIVATE_H */
+#endif
diff --git a/sysdeps/x86_64/fpu/math_private.h b/sysdeps/x86_64/fpu/math_private.h
index aa208b2..67c5f6a 100644
--- a/sysdeps/x86_64/fpu/math_private.h
+++ b/sysdeps/x86_64/fpu/math_private.h
@@ -1,33 +1,12 @@
 #ifndef X86_64_MATH_PRIVATE_H
 #define X86_64_MATH_PRIVATE_H 1
 
-#include <fenv.h>
-
-#define math_opt_barrier(x) \
-  ({ __typeof(x) __x;							      \
-     if (sizeof (x) <= sizeof (double))					      \
-       __asm ("" : "=x" (__x) : "0" (x));				      \
-     else								      \
-       __asm ("" : "=t" (__x) : "0" (x));				      \
-     __x; })
-#define math_force_eval(x) \
-  do {									      \
-    if (sizeof (x) <= sizeof (double))					      \
-      __asm __volatile ("" : : "x" (x));				      \
-    else								      \
-      __asm __volatile ("" : : "f" (x));				      \
-  } while (0)
-
 /* We can do a few things better on x86-64.  */
 
 #if defined __AVX__ || defined SSE2AVX
 # define MOVD "vmovd"
-# define STMXCSR "vstmxcsr"
-# define LDMXCSR "vldmxcsr"
 #else
 # define MOVD "movd"
-# define STMXCSR "stmxcsr"
-# define LDMXCSR "ldmxcsr"
 #endif
 
 /* Direct movement of float into integer register.  */
@@ -64,104 +43,7 @@
     f = f__;								      \
   } while (0)
 
-/* Specialized variants of the <fenv.h> interfaces which only handle
-   either the FPU or the SSE unit.  */
-static __always_inline void
-libc_feholdexcept (fenv_t *e)
-{
-  unsigned int mxcsr;
-  asm (STMXCSR " %0" : "=m" (*&mxcsr));
-  e->__mxcsr = mxcsr;
-  mxcsr = (mxcsr | 0x1f80) & ~0x3f;
-  asm volatile (LDMXCSR " %0" : : "m" (*&mxcsr));
-}
-#define libc_feholdexcept  libc_feholdexcept
-#define libc_feholdexceptf libc_feholdexcept
-
-static __always_inline void
-libc_feholdexcept_setround (fenv_t *e, int r)
-{
-  unsigned int mxcsr;
-  asm (STMXCSR " %0" : "=m" (*&mxcsr));
-  e->__mxcsr = mxcsr;
-  mxcsr = ((mxcsr | 0x1f80) & ~0x603f) | (r << 3);
-  asm volatile (LDMXCSR " %0" : : "m" (*&mxcsr));
-}
-#define libc_feholdexcept_setround  libc_feholdexcept_setround
-#define libc_feholdexcept_setroundf libc_feholdexcept_setround
-
-static __always_inline int
-libc_fetestexcept (int e)
-{
-  unsigned int mxcsr;
-  asm volatile (STMXCSR " %0" : "=m" (*&mxcsr));
-  return mxcsr & e & FE_ALL_EXCEPT;
-}
-#define libc_fetestexcept  libc_fetestexcept
-#define libc_fetestexceptf libc_fetestexcept
-
-static __always_inline void
-libc_fesetenv (fenv_t *e)
-{
-  asm volatile (LDMXCSR " %0" : : "m" (e->__mxcsr));
-}
-#define libc_fesetenv  libc_fesetenv
-#define libc_fesetenvf libc_fesetenv
-
-static __always_inline int
-libc_feupdateenv_test (fenv_t *e, int ex)
-{
-  unsigned int mxcsr, old_mxcsr, cur_ex;
-  asm volatile (STMXCSR " %0" : "=m" (*&mxcsr));
-  cur_ex = mxcsr & FE_ALL_EXCEPT;
-
-  /* Merge current exceptions with the old environment.  */
-  old_mxcsr = e->__mxcsr;
-  mxcsr = old_mxcsr | cur_ex;
-  asm volatile (LDMXCSR " %0" : : "m" (*&mxcsr));
-
-  /* Raise SIGFPE for any new exceptions since the hold.  Expect that
-     the normal environment has all exceptions masked.  */
-  if (__builtin_expect ((old_mxcsr >> 7) & cur_ex, 0))
-    __feraiseexcept (cur_ex);
-
-  /* Test for exceptions raised since the hold.  */
-  return cur_ex & ex;
-}
-#define libc_feupdateenv_test  libc_feupdateenv_test
-#define libc_feupdateenv_testf libc_feupdateenv_test
-
-static __always_inline void
-libc_feupdateenv (fenv_t *e)
-{
-  libc_feupdateenv_test (e, 0);
-}
-#define libc_feupdateenv  libc_feupdateenv
-#define libc_feupdateenvf libc_feupdateenv
-
-static __always_inline void
-libc_feholdsetround (fenv_t *e, int r)
-{
-  unsigned int mxcsr;
-  asm (STMXCSR " %0" : "=m" (*&mxcsr));
-  e->__mxcsr = mxcsr;
-  mxcsr = (mxcsr & ~0x6000) | (r << 3);
-  asm volatile (LDMXCSR " %0" : : "m" (*&mxcsr));
-}
-#define libc_feholdsetround  libc_feholdsetround
-#define libc_feholdsetroundf libc_feholdsetround
-
-static __always_inline void
-libc_feresetround (fenv_t *e)
-{
-  unsigned int mxcsr;
-  asm (STMXCSR " %0" : "=m" (*&mxcsr));
-  mxcsr = (mxcsr & ~0x6000) | (e->__mxcsr & 0x6000);
-  asm volatile (LDMXCSR " %0" : : "m" (*&mxcsr));
-}
-#define libc_feresetround  libc_feresetround
-#define libc_feresetroundf libc_feresetround
-
+#include <sysdeps/i386/fpu/fenv_private.h>
 #include_next <math_private.h>
 
 extern __always_inline double
-- 
1.7.7.6