[powerpc] __fesetround_inline optimizations

On POWER9, use more efficient means to update the 2-bit rounding mode
via the 'mffscrn' instruction (instead of two 'mtfsb0/1' instructions
or one 'mtfsfi' instruction that modifies 4 bits).

Suggested-by: Paul E. Murphy  <murphyp@linux.ibm.com>
Reviewed-By: Paul E Murphy <murphyp@linux.ibm.com>

diff --git a/ChangeLog b/ChangeLog
index 5fde58a298229bfb6a31ab1488e2a879c1bc4e68..5a53b922cb8c377c467d133e9152d9aa630a1624 100644 (file)
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,3 +1,9 @@
+2019-09-27  Paul A. Clarke  <pc@us.ibm.com>
+
+       * sysdeps/powerpc/fpu/fenv_libc.h (__fesetround_inline): Use
+       'mffscrn' instruction on POWER9.
+       (__fesetround_inline_nocheck): Likewise.
+
 2019-09-27  Paul A. Clarke  <pc@us.ibm.com>
 
        * sysdeps/powerpc/fpu/fenv_libc.h (FPSCR_EXCEPTIONS_MASK):  New.

diff --git a/sysdeps/powerpc/fpu/fenv_libc.h b/sysdeps/powerpc/fpu/fenv_libc.h
index 38ab6cd3cb94b811c303ba3d4963b20f783f5ee4..4aad7923930500d6f77744f90b3e2320be00a0ab 100644 (file)
--- a/sysdeps/powerpc/fpu/fenv_libc.h
+++ b/sysdeps/powerpc/fpu/fenv_libc.h
@@ -148,7 +148,12 @@ typedef union
 static inline int
 __fesetround_inline (int round)
 {
-  if ((unsigned int) round < 2)
+#ifdef _ARCH_PWR9
+  __fe_mffscrn (round);
+#else
+  if (__glibc_likely (GLRO(dl_hwcap2) & PPC_FEATURE2_ARCH_3_00))
+    __fe_mffscrn (round);
+  else if ((unsigned int) round < 2)
     {
        asm volatile ("mtfsb0 30");
        if ((unsigned int) round == 0)
@@ -164,7 +169,7 @@ __fesetround_inline (int round)
        else
          asm volatile ("mtfsb1 31");
     }
-
+#endif
   return 0;
 }
 
@@ -173,7 +178,14 @@ __fesetround_inline (int round)
 static inline void
 __fesetround_inline_nocheck (const int round)
 {
-  asm volatile ("mtfsfi 7,%0" : : "i" (round));
+#ifdef _ARCH_PWR9
+  __fe_mffscrn (round);
+#else
+  if (__glibc_likely (GLRO(dl_hwcap2) & PPC_FEATURE2_ARCH_3_00))
+    __fe_mffscrn (round);
+  else
+    asm volatile ("mtfsfi 7,%0" : : "i" (round));
+#endif
 }
 
 #define FPSCR_MASK(bit) (1 << (31 - (bit)))