[libc-commits] [libc] [llvm] [libc][math] reduce duplicated inv trig helpers (PR #191364)

[via libc-commits]

================
@@ -30,82 +30,96 @@ LIBC_INLINE_VAR constexpr DoubleDouble PI = {0x1.1a62633145c07p-53,
 LIBC_INLINE_VAR constexpr DoubleDouble PI_OVER_TWO = {0x1.1a62633145c07p-54,
                                                       0x1.921fb54442d18p0};
 
-#ifdef LIBC_MATH_HAS_SKIP_ACCURATE_PASS
+// Scalar double constants for pi and pi/2, used by float-precision
+// implementations (acosf etc.).
+LIBC_INLINE_VAR constexpr double M_MATH_PI = 0x1.921fb54442d18p+1;
+LIBC_INLINE_VAR constexpr double M_MATH_PI_2 = 0x1.921fb54442d18p+0;
 
-// When correct rounding is not needed, we use a degree-22 minimax polynomial to
-// approximate asin(x)/x on [0, 0.5] using Sollya with:
-// > P = fpminimax(asin(x)/x, [|0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22|],
+// > Q = fpminimax(asin(x)/x, [|0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24|],
 //                 [|1, D...|], [0, 0.5]);
-// > dirtyinfnorm(asin(x)/x - P, [0, 0.5]);
-// 0x1.1a71ef0a0f26a9fb7ed7e41dee788b13d1770db3dp-52
-
-LIBC_INLINE_VAR constexpr double ASIN_COEFFS[12] = {
-    0x1.0000000000000p0,  0x1.5555555556dcfp-3,  0x1.3333333082e11p-4,
-    0x1.6db6dd14099edp-5, 0x1.f1c69b35bf81fp-6,  0x1.6e97194225a67p-6,
-    0x1.1babddb82ce12p-6, 0x1.d55bd078600d6p-7,  0x1.33328959e63d6p-7,
-    0x1.2b5993bda1d9bp-6, -0x1.806aff270bf25p-7, 0x1.02614e5ed3936p-5,
+// > dirtyinfnorm((asin(x) - x*Q)/x, [0, 0.5]);
+//   0x1.feb2fcdba66447ccbe28a1a0f935b51678a718fb1p-59
----------------
lntue wrote:

Now that the degree of approximation and the evaluation scheme for double precision `asin`, `acos`, `atan` are different.  Can you retest whether the error bounds are changed for these functions when the accurate paths are skipped?

https://github.com/llvm/llvm-project/pull/191364