Add optimized sincos() function (ported from linaro_android_4.2.2)
Change-Id: I77f84d7ebc2dd7d5aa352b108b05e2651b5605c1
Signed-off-by: Bernhard Rosenkränzer <Bernhard.Rosenkranzer@linaro.org>
diff --git a/libm/sincos.c b/libm/sincos.c
index 891aa31..59b2301 100644
--- a/libm/sincos.c
+++ b/libm/sincos.c
@@ -1,4 +1,8 @@
/*-
+ * Written by Bernhard Rosenkränzer <Bernhard.Rosenkranzer@linaro.org>
+ * based on the implementations of sin() and cos() found in other Bionic
+ * files.
+ *
* Copyright (c) 2010 The Android Open Source Project
* All rights reserved.
*
@@ -26,26 +30,155 @@
*/
#define _GNU_SOURCE 1
#include <math.h>
+#define INLINE_KERNEL_COSDF
+#define INLINE_KERNEL_SINDF
+#include "src/math_private.h"
+#include "src/k_cosf.c"
+#include "src/k_sinf.c"
-// Disable sincos optimization for all functions in this file,
-// otherwise gcc would generate infinite calls.
-// Refer to gcc PR46926.
-// -fno-builtin-sin or -fno-builtin-cos can disable sincos optimization,
-// but these two options do not work inside optimize pragma in-file.
-// Thus we just enforce -O0 when compiling this file.
+/* Small multiples of pi/2 rounded to double precision. */
+static const double
+ s1pio2 = 1*M_PI_2, /* 0x3FF921FB, 0x54442D18 */
+ s2pio2 = 2*M_PI_2, /* 0x400921FB, 0x54442D18 */
+ s3pio2 = 3*M_PI_2, /* 0x4012D97C, 0x7F3321D2 */
+ s4pio2 = 4*M_PI_2; /* 0x401921FB, 0x54442D18 */
+
+/* For implementation details, see src/s_sin.c, src/s_cos.c */
+void sincos(double x, double *psin, double *pcos)
+{
+ double y[2], z=0.0;
+ int32_t n, ix;
+
+ /* High word of x. */
+ GET_HIGH_WORD(ix, x);
+
+ /* |x| ~< pi/4 */
+ ix &= 0x7fffffff;
+ if(ix <= 0x3fe921fb) {
+ if(ix < 0x3e400000) { /* \x\ < 2**-27 */
+ if((int)x==0) { /* generate inexact */
+ *psin = x;
+ *pcos = 1.0;
+ return;
+ }
+ }
+ *psin = __kernel_sin(x, z, 0);
+ *pcos = __kernel_cos(x, z);
+ return;
+ } else if(ix>=0x7ff00000) { /* sin(Inf or NaN) and cos(Inf or NaN) is NaN */
+ *psin = *pcos = x-x;
+ return;
+ } else {
+ n = __ieee754_rem_pio2(x, y);
+ switch(n&3) {
+ case 0:
+ *psin = __kernel_sin(y[0],y[1],1);
+ *pcos = __kernel_cos(y[0],y[1]);
+ return;
+ case 1:
+ *psin = __kernel_cos(y[0],y[1]);
+ *pcos = -__kernel_sin(y[0],y[1],1);
+ return;
+ case 2:
+ *psin = -__kernel_sin(y[0],y[1],1);
+ *pcos = -__kernel_cos(y[0],y[1]);
+ return;
+ default:
+ *psin = -__kernel_cos(y[0],y[1]);
+ *pcos = __kernel_sin(y[0],y[1],1);
+ return;
+ }
+ }
+}
+
+/* For implementation details, see src/s_sinf.c, src/s_cosf.c */
+void sincosf(float x, float *psin, float *pcos)
+{
+ float y[2];
+ int32_t n, hx, ix;
+
+ GET_FLOAT_WORD(hx, x);
+ ix = hx & 0x7fffffff;
+
+ if(ix <= 0x3f490fda) { /* |x| ~<= pi/4 */
+ if(ix < 0x39800000) { /* |x| < 2**-12 */
+ if(((int)x)==0) { /* x with inexact if x != 0 */
+ *psin = x;
+ *pcos = 1.0;
+ return;
+ }
+ }
+ *psin = __kernel_sindf(x);
+ *pcos = __kernel_cosdf(x);
+ return;
+ } else if(ix <= 0x407b53d1) { /* |x| ~<= 5*pi/4 */
+ if(ix <= 0x4016cbe3) { /* |x| ~<= 3pi/4 */
+ if(hx>0) {
+ *psin = __kernel_cosdf(x - s1pio2);
+ *pcos = __kernel_sindf(s1pio2 - x);
+ return;
+ } else {
+ *psin = -__kernel_cosdf(x + s1pio2);
+ *pcos = __kernel_sindf(x + s1pio2);
+ return;
+ }
+ } else {
+ *psin = __kernel_sindf((hx > 0 ? s2pio2 : -s2pio2) - x);
+ *pcos = -__kernel_cosdf(x + (hx > 0 ? -s2pio2 : s2pio2));
+ return;
+ }
+ } else if(ix <= 0x40e231d5) { /* |x| ~<= 9*pi/4 */
+ if(ix <= 0x40afeddf) { /* |x| ~<= 7*pi/4 */
+ if(hx>0) {
+ *psin = -__kernel_cosdf(x - s3pio2);
+ *pcos = __kernel_sindf(x - s3pio2);
+ return;
+ } else {
+ *psin = __kernel_cosdf(x + s3pio2);
+ *pcos = __kernel_sindf(-s3pio2 - x);
+ return;
+ }
+ } else {
+ *psin = __kernel_sindf(x + (hx > 0 ? -s4pio2 : s4pio2));
+ *pcos = __kernel_cosdf(x + (hx > 0 ? -s4pio2 : s4pio2));
+ return;
+ }
+ } else if(ix>=0x7f800000) { /* sin and cos (Inf or NaN) is NaN */
+ *psin = *pcos = x-x;
+ return;
+ } else {
+ n = __ieee754_rem_pio2f(x,y);
+ switch(n&3) {
+ case 0:
+ *psin = __kernel_sindf((double)y[0]+y[1]);
+ *pcos = __kernel_cosdf((double)y[0]+y[1]);
+ return;
+ case 1:
+ *psin = __kernel_cosdf((double)y[0]+y[1]);
+ *pcos = __kernel_sindf(-(double)y[0]-y[1]);
+ return;
+ case 2:
+ *psin = __kernel_sindf(-(double)y[0]-y[1]);
+ *pcos = -__kernel_cosdf((double)y[0]+y[1]);
+ return;
+ default:
+ *psin = -__kernel_cosdf((double)y[0]+y[1]);
+ *pcos = __kernel_sindf((double)y[0]+y[1]);
+ return;
+ }
+ }
+}
+
+/* We can't use a properly optimized version here since Android fakes
+ * long double bits (see fake_long_double.c)
+ * Also, can't allow gcc to optimize because sinl() followed by cosl()
+ * might be merged into sincosl(), causing an infinite loop.
+ * Unfortunately, we can't just pass -fno-builtin-sin -fno-builtin-cos
+ * in a #pragma.
+ */
#pragma GCC optimize ("O0")
-
-void sincos(double x, double* p_sin, double* p_cos) {
- *p_sin = sin(x);
- *p_cos = cos(x);
-}
-
-void sincosf(float x, float* p_sinf, float* p_cosf) {
- *p_sinf = sinf(x);
- *p_cosf = cosf(x);
-}
-
-void sincosl(long double x, long double* p_sinl, long double* p_cosl) {
- *p_sinl = sinl(x);
- *p_cosl = cosl(x);
+void sincosl(long double x, long double *psin, long double *pcos)
+{
+ *psin = sinl(x);
+ *pcos = cosl(x);
}