fmath: workaround for bug in sinf_approx

A more correct solutions is being investigated

Author: rasky

src/fmath.c

@@ -5,12 +5,10 @@
  */
 #include "fmath.h"
 #include "debug.h"
+#include "utils.h"
 #include <string.h>
 #include <stdint.h>
 
-/// Mark a branch as likely to be taken
-#define LIKELY(x)       __builtin_expect((x),1)
-
 static const float pi_hi            = 3.14159274e+00f; // 0x1.921fb6p+01
 static const float pi_lo            =-8.74227766e-08f; // -0x1.777a5cp-24
 static const float half_pi_hi       =  1.57079637e+0f; //  0x1.921fb6p+0
@@ -42,7 +40,12 @@ float fm_sinf_approx(float x, int approx) {
     // very accurate for large numbers, so it will introduce more error compared
     // to the 5 ULP figure.
     x = fm_fmodf(x+pi_hi, 2*pi_hi) - pi_hi;
-    return sinf_approx(x, approx);
+    x = sinf_approx(x, approx);
+    // FIXME: workaround for a bug in our sinf approximation. We found at least
+    // one input (0xbfc915a2 => -1.570973) that produces an out of bounds result
+    // -1.000000119209289551 (0xbf800001).
+    x = CLAMP(x, -1.0f, 1.0f);
+    return x;
 }
 
 float fm_sinf(float x) {