[libc-commits] [libc] [libc][math] Improve performance of double precision trig functions. (PR #111793)

[Nick Desaulniers via libc-commits]

================
@@ -44,40 +41,47 @@ LLVM_LIBC_FUNCTION(void, sincos, (double x, double *sin_x, double *cos_x)) {
 
   DoubleDouble y;
   unsigned k;
-  generic::LargeRangeReduction<NO_FMA> range_reduction_large{};
+  LargeRangeReduction range_reduction_large{};
 
-  // |x| < 2^32 (with FMA) or |x| < 2^23 (w/o FMA)
+  // |x| < 2^16
   if (LIBC_LIKELY(x_e < FPBits::EXP_BIAS + FAST_PASS_EXPONENT)) {
-    // |x| < 2^-27
-    if (LIBC_UNLIKELY(x_e < FPBits::EXP_BIAS - 27)) {
-      // Signed zeros.
-      if (LIBC_UNLIKELY(x == 0.0)) {
-        *sin_x = x;
-        *cos_x = 1.0;
-        return;
-      }
-
-      // For |x| < 2^-27, max(|sin(x) - x|, |cos(x) - 1|) < ulp(x)/2.
+    // |x| < 2^-7
+    if (LIBC_UNLIKELY(x_e < FPBits::EXP_BIAS - 7)) {
+      // |x| < 2^-27
+      if (LIBC_UNLIKELY(x_e < FPBits::EXP_BIAS - 27)) {
+        // Signed zeros.
+        if (LIBC_UNLIKELY(x == 0.0)) {
+          *sin_x = x;
+          *cos_x = 1.0;
+          return;
+        }
----------------
nickdesaulniers wrote:

Perhaps in the future, we can have one implementation for computing sincos that BOTH sin and cos use, where once inlined we can skip computing the unneeded result. Currently, there's a fair amount of duplication between sincos, sin, and cos that I think can be avoided. Not necessary for this PR, but worth a thought.

https://github.com/llvm/llvm-project/pull/111793