Patchwork [AArch64] ILP32 math changes

login
register
mail settings
Submitter Steve Ellcey
Date Aug. 29, 2017, 9:16 p.m.
Message ID <1504041387.5204.20.camel@cavium.com>
Download mbox | patch
Permalink /patch/22414/
State New
Headers show

Comments

Steve Ellcey - Aug. 29, 2017, 9:16 p.m.
Since Szabolcs has expressed interest in getting some aarch64 ILP32
changes into glibc mainline (if they don't affect the kernel or glibc
ABI) I am resubmitting this patch for approval in advance of the main
ILP32 support patches.  I updated it with Richard Henderson's code (but
only doing it when the incoming argument is large).  I tested it on
aarch64 ILP32 and LP64 with no regressions.

Is this something that can be checked in now withnout waiting for the
kernel ILP32 support?

Steve Ellcey
sellcey@cavium.com


2017-08-29  Steve Ellcey  <sellcey@cavium.com>
	    Richard Henderson <rth@twiddle.net>

	* sysdeps/aarch64/fpu/s_llrint.c (OREG_SIZE): New macro.
	* sysdeps/aarch64/fpu/s_llround.c (OREG_SIZE): Likewise.
	* sysdeps/aarch64/fpu/s_llrintf.c (OREGS, IREGS): Remove.
	(IREG_SIZE, OREG_SIZE): New macros.
	* sysdeps/aarch64/fpu/s_llroundf.c: (OREGS, IREGS): Remove.
	(IREG_SIZE, OREG_SIZE): New macros.
	* sysdeps/aarch64/fpu/s_lrintf.c (IREGS): Remove.
	(IREG_SIZE): New macro.
	* sysdeps/aarch64/fpu/s_lroundf.c (IREGS): Remove.
	(IREG_SIZE): New macro.
	* sysdeps/aarch64/fpu/s_lrint.c (get-rounding-mode.h, stdint.h):
	New includes.
	(IREG_SIZE, OREG_SIZE): Initialize if not already set.
	(OREGS, IREGS): Set based on IREG_SIZE and OREG_SIZE.
	(__CONCATX): Handle exceptions correctly on large values that may
	set FE_INVALID.
	* sysdeps/aarch64/fpu/s_lround.c (IREG_SIZE, OREG_SIZE):
	Initialize if not already set.
        (OREGS, IREGS): Set based on IREG_SIZE and OREG_SIZE.
Szabolcs Nagy - Aug. 31, 2017, 4:22 p.m.
On 29/08/17 22:16, Steve Ellcey wrote:
> Since Szabolcs has expressed interest in getting some aarch64 ILP32
> changes into glibc mainline (if they don't affect the kernel or glibc
> ABI) I am resubmitting this patch for approval in advance of the main
> ILP32 support patches.  I updated it with Richard Henderson's code (but
> only doing it when the incoming argument is large).  I tested it on
> aarch64 ILP32 and LP64 with no regressions.
> 
> Is this something that can be checked in now withnout waiting for the
> kernel ILP32 support?
> 

OK.

(we plan to change the math functions to use gcc builtins
in some places instead of asm which will simplify things,
but that can go in later on top of this.)

> Steve Ellcey
> sellcey@cavium.com
> 
> 
> 2017-08-29  Steve Ellcey  <sellcey@cavium.com>
> 	    Richard Henderson <rth@twiddle.net>
> 
> 	* sysdeps/aarch64/fpu/s_llrint.c (OREG_SIZE): New macro.
> 	* sysdeps/aarch64/fpu/s_llround.c (OREG_SIZE): Likewise.
> 	* sysdeps/aarch64/fpu/s_llrintf.c (OREGS, IREGS): Remove.
> 	(IREG_SIZE, OREG_SIZE): New macros.
> 	* sysdeps/aarch64/fpu/s_llroundf.c: (OREGS, IREGS): Remove.
> 	(IREG_SIZE, OREG_SIZE): New macros.
> 	* sysdeps/aarch64/fpu/s_lrintf.c (IREGS): Remove.
> 	(IREG_SIZE): New macro.
> 	* sysdeps/aarch64/fpu/s_lroundf.c (IREGS): Remove.
> 	(IREG_SIZE): New macro.
> 	* sysdeps/aarch64/fpu/s_lrint.c (get-rounding-mode.h, stdint.h):
> 	New includes.
> 	(IREG_SIZE, OREG_SIZE): Initialize if not already set.
> 	(OREGS, IREGS): Set based on IREG_SIZE and OREG_SIZE.
> 	(__CONCATX): Handle exceptions correctly on large values that may
> 	set FE_INVALID.
> 	* sysdeps/aarch64/fpu/s_lround.c (IREG_SIZE, OREG_SIZE):
> 	Initialize if not already set.
>         (OREGS, IREGS): Set based on IREG_SIZE and OREG_SIZE.
>

Patch

diff --git a/sysdeps/aarch64/fpu/s_llrint.c b/sysdeps/aarch64/fpu/s_llrint.c
index c0d0d0e..57821c0 100644
--- a/sysdeps/aarch64/fpu/s_llrint.c
+++ b/sysdeps/aarch64/fpu/s_llrint.c
@@ -18,4 +18,5 @@ 
 
 #define FUNC llrint
 #define OTYPE long long int
+#define OREG_SIZE 64
 #include <s_lrint.c>
diff --git a/sysdeps/aarch64/fpu/s_llrintf.c b/sysdeps/aarch64/fpu/s_llrintf.c
index 67724c6..98ed4f8 100644
--- a/sysdeps/aarch64/fpu/s_llrintf.c
+++ b/sysdeps/aarch64/fpu/s_llrintf.c
@@ -18,6 +18,7 @@ 
 
 #define FUNC llrintf
 #define ITYPE float
-#define IREGS "s"
+#define IREG_SIZE 32
 #define OTYPE long long int
+#define OREG_SIZE 64
 #include <s_lrint.c>
diff --git a/sysdeps/aarch64/fpu/s_llround.c b/sysdeps/aarch64/fpu/s_llround.c
index ed4b192..ef7aedf 100644
--- a/sysdeps/aarch64/fpu/s_llround.c
+++ b/sysdeps/aarch64/fpu/s_llround.c
@@ -18,4 +18,5 @@ 
 
 #define FUNC llround
 #define OTYPE long long int
+#define OREG_SIZE 64
 #include <s_lround.c>
diff --git a/sysdeps/aarch64/fpu/s_llroundf.c b/sysdeps/aarch64/fpu/s_llroundf.c
index 360ce8b..294f0f4 100644
--- a/sysdeps/aarch64/fpu/s_llroundf.c
+++ b/sysdeps/aarch64/fpu/s_llroundf.c
@@ -18,6 +18,7 @@ 
 
 #define FUNC llroundf
 #define ITYPE float
-#define IREGS "s"
+#define IREG_SIZE 32
 #define OTYPE long long int
+#define OREG_SIZE 64
 #include <s_lround.c>
diff --git a/sysdeps/aarch64/fpu/s_lrint.c b/sysdeps/aarch64/fpu/s_lrint.c
index 8c61a03..6ef64e2 100644
--- a/sysdeps/aarch64/fpu/s_lrint.c
+++ b/sysdeps/aarch64/fpu/s_lrint.c
@@ -17,6 +17,8 @@ 
    <http://www.gnu.org/licenses/>.  */
 
 #include <math.h>
+#include <get-rounding-mode.h>
+#include <stdint.h>
 
 #ifndef FUNC
 # define FUNC lrint
@@ -24,18 +26,37 @@ 
 
 #ifndef ITYPE
 # define ITYPE double
-# define IREGS "d"
+# define IREG_SIZE 64
 #else
-# ifndef IREGS
-#  error IREGS not defined
+# ifndef IREG_SIZE
+#  error IREG_SIZE not defined
 # endif
 #endif
 
 #ifndef OTYPE
 # define OTYPE long int
+# ifdef __ILP32__
+#  define OREG_SIZE 32
+# else
+#  define OREG_SIZE 64
+# endif
+#else
+# ifndef OREG_SIZE
+#  error OREG_SIZE not defined
+# endif
 #endif
 
-#define OREGS "x"
+#if IREG_SIZE == 32
+# define IREGS "s"
+#else
+# define IREGS "d"
+#endif
+
+#if OREG_SIZE == 32
+# define OREGS "w"
+#else
+# define OREGS "x"
+#endif
 
 #define __CONCATX(a,b) __CONCAT(a,b)
 
@@ -44,6 +65,37 @@  __CONCATX(__,FUNC) (ITYPE x)
 {
   OTYPE result;
   ITYPE temp;
+
+#if IREG_SIZE == 64 && OREG_SIZE == 32
+  if (__builtin_fabs (x) > INT32_MAX)
+    {
+      /* Converting large values to a 32 bit int may cause the frintx/fcvtza
+	 sequence to set both FE_INVALID and FE_INEXACT.  To avoid this
+	 check the rounding mode and do a single instruction with the
+	 appropriate rounding mode.  */
+
+      switch (get_rounding_mode ())
+	{
+	case FE_TONEAREST:
+	  asm volatile ("fcvtns" "\t%" OREGS "0, %" IREGS "1"
+			: "=r" (result) : "w" (x));
+	  break;
+	case FE_UPWARD:
+	  asm volatile ("fcvtps" "\t%" OREGS "0, %" IREGS "1"
+			: "=r" (result) : "w" (x));
+	  break;
+	case FE_DOWNWARD:
+	  asm volatile ("fcvtms" "\t%" OREGS "0, %" IREGS "1"
+			: "=r" (result) : "w" (x));
+	  break;
+	case FE_TOWARDZERO:
+	default:
+	  asm volatile ("fcvtzs" "\t%" OREGS "0, %" IREGS "1"
+			: "=r" (result) : "w" (x));
+	}
+      return result;
+    }
+#endif
   asm ( "frintx" "\t%" IREGS "1, %" IREGS "2\n\t"
         "fcvtzs" "\t%" OREGS "0, %" IREGS "1"
         : "=r" (result), "=w" (temp) : "w" (x) );
diff --git a/sysdeps/aarch64/fpu/s_lrintf.c b/sysdeps/aarch64/fpu/s_lrintf.c
index a995e4b..2e73271 100644
--- a/sysdeps/aarch64/fpu/s_lrintf.c
+++ b/sysdeps/aarch64/fpu/s_lrintf.c
@@ -18,5 +18,5 @@ 
 
 #define FUNC lrintf
 #define ITYPE float
-#define IREGS "s"
+#define IREG_SIZE 32
 #include <s_lrint.c>
diff --git a/sysdeps/aarch64/fpu/s_lround.c b/sysdeps/aarch64/fpu/s_lround.c
index 9be9e7f..1f77d82 100644
--- a/sysdeps/aarch64/fpu/s_lround.c
+++ b/sysdeps/aarch64/fpu/s_lround.c
@@ -24,18 +24,37 @@ 
 
 #ifndef ITYPE
 # define ITYPE double
-# define IREGS "d"
+# define IREG_SIZE 64
 #else
-# ifndef IREGS
-#  error IREGS not defined
+# ifndef IREG_SIZE
+#  error IREG_SIZE not defined
 # endif
 #endif
 
 #ifndef OTYPE
 # define OTYPE long int
+# ifdef __ILP32__
+#  define OREG_SIZE 32
+# else
+#  define OREG_SIZE 64
+# endif
+#else
+# ifndef OREG_SIZE
+#  error OREG_SIZE not defined
+# endif
+#endif
+
+#if IREG_SIZE == 32
+# define IREGS "s"
+#else
+# define IREGS "d"
 #endif
 
-#define OREGS "x"
+#if OREG_SIZE == 32
+# define OREGS "w"
+#else
+# define OREGS "x"
+#endif
 
 #define __CONCATX(a,b) __CONCAT(a,b)
 
diff --git a/sysdeps/aarch64/fpu/s_lroundf.c b/sysdeps/aarch64/fpu/s_lroundf.c
index 4a066d4..b30ddb6 100644
--- a/sysdeps/aarch64/fpu/s_lroundf.c
+++ b/sysdeps/aarch64/fpu/s_lroundf.c
@@ -18,5 +18,5 @@ 
 
 #define FUNC lroundf
 #define ITYPE float
-#define IREGS "s"
+#define IREG_SIZE 32
 #include <s_lround.c>