Set/restore rounding mode only when needed

The most common use case of math functions is with default rounding
mode, i.e. rounding to nearest.  Setting and restoring rounding mode
is an unnecessary overhead for this, so I've added support for a
context, which does the set/restore only if the FP status needs a
change.  The code is written such that only x86 uses these.  Other
architectures should be unaffected by it, but would definitely benefit
if the set/restore has as much overhead relative to the rest of the
code, as the x86 bits do.

Here's a summary of the performance improvement due to these
improvements; I've only mentioned functions that use the set/restore
and have benchmark inputs for x86_64:

Before:

cos(): ITERS:4.69335e+08: TOTAL:28884.6Mcy, MAX:4080.28cy, MIN:57.562cy, 16248.6 calls/Mcy
exp(): ITERS:4.47604e+08: TOTAL:28796.2Mcy, MAX:207.721cy, MIN:62.385cy, 15543.9 calls/Mcy
pow(): ITERS:1.63485e+08: TOTAL:28879.9Mcy, MAX:362.255cy, MIN:172.469cy, 5660.86 calls/Mcy
sin(): ITERS:3.89578e+08: TOTAL:28900Mcy, MAX:704.859cy, MIN:47.583cy, 13480.2 calls/Mcy
tan(): ITERS:7.0971e+07: TOTAL:28902.2Mcy, MAX:1357.79cy, MIN:388.58cy, 2455.55 calls/Mcy

After:

cos(): ITERS:6.0014e+08: TOTAL:28875.9Mcy, MAX:364.283cy, MIN:45.716cy, 20783.4 calls/Mcy
exp(): ITERS:5.48578e+08: TOTAL:28764.9Mcy, MAX:191.617cy, MIN:51.011cy, 19071.1 calls/Mcy
pow(): ITERS:1.70013e+08: TOTAL:28873.6Mcy, MAX:689.522cy, MIN:163.989cy, 5888.18 calls/Mcy
sin(): ITERS:4.64079e+08: TOTAL:28891.5Mcy, MAX:6959.3cy, MIN:36.189cy, 16062.8 calls/Mcy
tan(): ITERS:7.2354e+07: TOTAL:28898.9Mcy, MAX:1295.57cy, MIN:380.698cy, 2503.7 calls/Mcy

So the improvements are:

cos: 27.9089%
exp: 22.6919%
pow: 4.01564%
sin: 19.1585%
tan: 1.96086%

The downside of the change is that it will have an adverse performance
impact on non-default rounding modes, but I think the tradeoff is
justified.

2 files changed, 214 insertions, 14 deletions

diff --git a/sysdeps/generic/math_private.h b/sysdeps/generic/math_private.h
index e98360dd47..c0fc03d38d 100644
--- a/sysdeps/generic/math_private.h
+++ b/sysdeps/generic/math_private.h
@@ -553,35 +553,62 @@ default_libc_feupdateenv_test (fenv_t *e, int ex)
 # define libc_feresetround_noexl libc_fesetenvl
 #endif
 
+#if HAVE_RM_CTX
+/* Set/Restore Rounding Modes only when necessary.  If defined, these functions
+   set/restore floating point state only if the state needed within the lexical
+   block is different from the current state.  This saves a lot of time when
+   the floating point unit is much slower than the fixed point units.  */
+
+# ifndef libc_feresetround_noex_ctx
+#   define libc_feresetround_noex_ctx  libc_fesetenv_ctx
+# endif
+# ifndef libc_feresetround_noexf_ctx
+#   define libc_feresetround_noexf_ctx libc_fesetenvf_ctx
+# endif
+# ifndef libc_feresetround_noexl_ctx
+#   define libc_feresetround_noexl_ctx libc_fesetenvl_ctx
+# endif
+
+# ifndef libc_feholdsetround_53bit_ctx
+#   define libc_feholdsetround_53bit_ctx libc_feholdsetround_ctx
+# endif
+
+# ifndef libc_feresetround_53bit_ctx
+#   define libc_feresetround_53bit_ctx libc_feresetround_ctx
+# endif
+
+# define SET_RESTORE_ROUND_GENERIC(RM,ROUNDFUNC,CLEANUPFUNC) \
+  struct rm_ctx ctx __attribute__((cleanup(CLEANUPFUNC ## _ctx)));	      \
+  ROUNDFUNC ## _ctx (&ctx, (RM))
+#else
+# define SET_RESTORE_ROUND_GENERIC(RM, ROUNDFUNC, CLEANUPFUNC) \
+  fenv_t __libc_save_rm __attribute__((cleanup(CLEANUPFUNC)));	\
+  ROUNDFUNC (&__libc_save_rm, (RM))
+#endif
+
 /* Save and restore the rounding mode within a lexical block.  */
 
 #define SET_RESTORE_ROUND(RM) \
-  fenv_t __libc_save_rm __attribute__((cleanup(libc_feresetround)));	\
-  libc_feholdsetround (&__libc_save_rm, (RM))
+  SET_RESTORE_ROUND_GENERIC (RM, libc_feholdsetround, libc_feresetround)
 #define SET_RESTORE_ROUNDF(RM) \
-  fenv_t __libc_save_rm __attribute__((cleanup(libc_feresetroundf)));	\
-  libc_feholdsetroundf (&__libc_save_rm, (RM))
+  SET_RESTORE_ROUND_GENERIC (RM, libc_feholdsetroundf, libc_feresetroundf)
 #define SET_RESTORE_ROUNDL(RM) \
-  fenv_t __libc_save_rm __attribute__((cleanup(libc_feresetroundl)));	\
-  libc_feholdsetroundl (&__libc_save_rm, (RM))
+  SET_RESTORE_ROUND_GENERIC (RM, libc_feholdsetroundl, libc_feresetroundl)
 
 /* Save and restore the rounding mode within a lexical block, and also
    the set of exceptions raised within the block may be discarded.  */
 
 #define SET_RESTORE_ROUND_NOEX(RM) \
-  fenv_t __libc_save_rm __attribute__((cleanup(libc_feresetround_noex))); \
-  libc_feholdsetround (&__libc_save_rm, (RM))
+  SET_RESTORE_ROUND_GENERIC (RM, libc_feholdsetround, libc_feresetround_noex)
 #define SET_RESTORE_ROUND_NOEXF(RM) \
-  fenv_t __libc_save_rm __attribute__((cleanup(libc_feresetround_noexf))); \
-  libc_feholdsetroundf (&__libc_save_rm, (RM))
+  SET_RESTORE_ROUND_GENERIC (RM, libc_feholdsetroundf, libc_feresetround_noexf)
 #define SET_RESTORE_ROUND_NOEXL(RM) \
-  fenv_t __libc_save_rm __attribute__((cleanup(libc_feresetround_noexl))); \
-  libc_feholdsetroundl (&__libc_save_rm, (RM))
+  SET_RESTORE_ROUND_GENERIC (RM, libc_feholdsetroundl, libc_feresetround_noexl)
 
 /* Like SET_RESTORE_ROUND, but also set rounding precision to 53 bits.  */
 #define SET_RESTORE_ROUND_53BIT(RM) \
-  fenv_t __libc_save_rm __attribute__((cleanup(libc_feresetround_53bit))); \
-  libc_feholdsetround_53bit (&__libc_save_rm, (RM))
+  SET_RESTORE_ROUND_GENERIC (RM, libc_feholdsetround_53bit,	      \
+			     libc_feresetround_53bit)
 
 #define __nan(str) \
   (__builtin_constant_p (str) && str[0] == '\0' ? NAN : __nan (str))
diff --git a/sysdeps/i386/fpu/fenv_private.h b/sysdeps/i386/fpu/fenv_private.h
index 1f8336cf96..3998387c57 100644
--- a/sysdeps/i386/fpu/fenv_private.h
+++ b/sysdeps/i386/fpu/fenv_private.h
@@ -322,6 +322,179 @@ libc_feresetround_387 (fenv_t *e)
 # define libc_feholdsetround_53bit	libc_feholdsetround_387_53bit
 #endif
 
+/* We have support for rounding mode context.  */
+#define HAVE_RM_CTX 1
+
+static __always_inline void
+libc_feholdexcept_setround_sse_ctx (struct rm_ctx *ctx, int r)
+{
+  unsigned int mxcsr, new_mxcsr;
+  asm (STMXCSR " %0" : "=m" (*&mxcsr));
+  new_mxcsr = ((mxcsr | 0x1f80) & ~0x603f) | (r << 3);
+
+  ctx->env.__mxcsr = mxcsr;
+  if (__glibc_unlikely (mxcsr != new_mxcsr))
+    {
+      asm volatile (LDMXCSR " %0" : : "m" (*&new_mxcsr));
+      ctx->updated_status = true;
+    }
+  else
+    ctx->updated_status = false;
+}
+
+/* Unconditional since we want to overwrite any exceptions that occurred in the
+   context.  This is also why all fehold* functions unconditionally write into
+   ctx->env.  */
+static __always_inline void
+libc_fesetenv_sse_ctx (struct rm_ctx *ctx)
+{
+  libc_fesetenv_sse (&ctx->env);
+}
+
+static __always_inline void
+libc_feupdateenv_sse_ctx (struct rm_ctx *ctx)
+{
+  if (__glibc_unlikely (ctx->updated_status))
+    libc_feupdateenv_test_sse (&ctx->env, 0);
+}
+
+static __always_inline void
+libc_feholdexcept_setround_387_prec_ctx (struct rm_ctx *ctx, int r)
+{
+  libc_feholdexcept_387 (&ctx->env);
+
+  fpu_control_t cw = ctx->env.__control_word;
+  fpu_control_t old_cw = cw;
+  cw &= ~(_FPU_RC_ZERO | _FPU_EXTENDED);
+  cw |= r | 0x3f;
+
+  if (__glibc_unlikely (old_cw != cw))
+    {
+      _FPU_SETCW (cw);
+      ctx->updated_status = true;
+    }
+  else
+    ctx->updated_status = false;
+}
+
+static __always_inline void
+libc_feholdexcept_setround_387_ctx (struct rm_ctx *ctx, int r)
+{
+  libc_feholdexcept_setround_387_prec_ctx (ctx, r | _FPU_EXTENDED);
+}
+
+static __always_inline void
+libc_feholdexcept_setround_387_53bit_ctx (struct rm_ctx *ctx, int r)
+{
+  libc_feholdexcept_setround_387_prec_ctx (ctx, r | _FPU_DOUBLE);
+}
+
+static __always_inline void
+libc_feholdsetround_387_prec_ctx (struct rm_ctx *ctx, int r)
+{
+  fpu_control_t cw, new_cw;
+
+  _FPU_GETCW (cw);
+  new_cw = cw;
+  new_cw &= ~(_FPU_RC_ZERO | _FPU_EXTENDED);
+  new_cw |= r;
+
+  ctx->env.__control_word = cw;
+  if (__glibc_unlikely (new_cw != cw))
+    {
+      _FPU_SETCW (new_cw);
+      ctx->updated_status = true;
+    }
+  else
+    ctx->updated_status = false;
+}
+
+static __always_inline void
+libc_feholdsetround_387_ctx (struct rm_ctx *ctx, int r)
+{
+  libc_feholdsetround_387_prec_ctx (ctx, r | _FPU_EXTENDED);
+}
+
+static __always_inline void
+libc_feholdsetround_387_53bit_ctx (struct rm_ctx *ctx, int r)
+{
+  libc_feholdsetround_387_prec_ctx (ctx, r | _FPU_DOUBLE);
+}
+
+static __always_inline void
+libc_feholdsetround_sse_ctx (struct rm_ctx *ctx, int r)
+{
+  unsigned int mxcsr, new_mxcsr;
+
+  asm (STMXCSR " %0" : "=m" (*&mxcsr));
+  new_mxcsr = (mxcsr & ~0x6000) | (r << 3);
+
+  ctx->env.__mxcsr = mxcsr;
+  if (__glibc_unlikely (new_mxcsr != mxcsr))
+    {
+      asm volatile (LDMXCSR " %0" : : "m" (*&new_mxcsr));
+      ctx->updated_status = true;
+    }
+  else
+    ctx->updated_status = false;
+}
+
+static __always_inline void
+libc_feresetround_sse_ctx (struct rm_ctx *ctx)
+{
+  if (__glibc_unlikely (ctx->updated_status))
+    libc_feresetround_sse (&ctx->env);
+}
+
+static __always_inline void
+libc_feresetround_387_ctx (struct rm_ctx *ctx)
+{
+  if (__glibc_unlikely (ctx->updated_status))
+    _FPU_SETCW (ctx->env.__control_word);
+}
+
+static __always_inline void
+libc_feupdateenv_387_ctx (struct rm_ctx *ctx)
+{
+  if (__glibc_unlikely (ctx->updated_status))
+    libc_feupdateenv_test_387 (&ctx->env, 0);
+}
+
+#ifdef __SSE_MATH__
+# define libc_feholdexcept_setroundf_ctx libc_feholdexcept_setround_sse_ctx
+# define libc_fesetenvf_ctx		libc_fesetenv_sse_ctx
+# define libc_feupdateenvf_ctx		libc_feupdateenv_sse_ctx
+# define libc_feholdsetroundf_ctx	libc_feholdsetround_sse_ctx
+# define libc_feresetroundf_ctx		libc_feresetround_sse_ctx
+#else
+# define libc_feholdexcept_setroundf_ctx libc_feholdexcept_setround_387_ctx
+# define libc_feupdateenvf_ctx		libc_feupdateenv_387_ctx
+# define libc_feholdsetroundf_ctx	libc_feholdsetround_387_ctx
+# define libc_feresetroundf_ctx		libc_feresetround_387_ctx
+#endif /* __SSE_MATH__ */
+
+#ifdef __SSE2_MATH__
+# define libc_feholdexcept_setround_ctx	libc_feholdexcept_setround_sse_ctx
+# define libc_fesetenv_ctx		libc_fesetenv_sse_ctx
+# define libc_feupdateenv_ctx		libc_feupdateenv_sse_ctx
+# define libc_feholdsetround_ctx	libc_feholdsetround_sse_ctx
+# define libc_feresetround_ctx		libc_feresetround_sse_ctx
+#else
+# define libc_feholdexcept_setround_ctx	libc_feholdexcept_setround_387_ctx
+# define libc_feupdateenv_ctx		libc_feupdateenv_387_ctx
+# define libc_feresetround_ctx		libc_feresetround_387_ctx
+#endif /* __SSE2_MATH__ */
+
+#define libc_feholdexcept_setroundl_ctx	libc_feholdexcept_setround_387_ctx
+#define libc_feupdateenvl_ctx		libc_feupdateenv_387_ctx
+#define libc_feholdsetroundl_ctx	libc_feholdsetround_387_ctx
+#define libc_feresetroundl_ctx		libc_feresetround_387_ctx
+
+#ifndef __SSE2_MATH__
+# define libc_feholdsetround_53bit_ctx	libc_feholdsetround_387_53bit_ctx
+# define libc_feresetround_53bit_ctx	libc_feresetround_387_ctx
+#endif
+
 #undef __mxcsr
 
 #endif /* FENV_PRIVATE_H */