math: Add inputs that yield larger errors for float type (x86_64) [published under 2-clause BSD license]
New exp10f code: patch 1, patch 2, patch 3. Version 2: cumulated patch. Not faster on ia64 than previous assembly code but fixes some issues for directed rounding modes. Twice as fast on i386 than previous assembly code. Mean number of cycles goes down from 18.4 to 12.0 on x86_64.
Faster quadruple precision exponential (use main.c to experiment with). This is experimental code for 64-bit little endian platforms. On 105 random inputs in [-10,10], it returns the correct rounding in 99951 cases. On an AMD Ryzen 5 2400G, it takes on average 234 cycles (against 3142 for libquadmath 9.2.1); on an Intel Core i7-8750H, it takes on average 212 cycles (against 3156 for libquadmath 9.2.1). Accuracy/performance figures courtesy of Alexei Sibidanov. See this talk.