Fixed the precision of expm1(float x) when x is small. Two versions are provided, one with 12.0 bits of precision, and another with 16.0 bits of precision. Tested all inputs on x86_64

Author: ZERICO2005

src/libc/expm1.c

@@ -1,8 +1,24 @@
 #include <math.h>
 
-float expm1f(float x)
+float M_expm1f(const float x)
 {
-    return expf(x) - 1; // FIXME: accuracy
+    #if 1
+        /**
+         * 0x1.0p-12f: At least 12.0bits of precision
+         */
+        if (fabsf(x) < 2.44140625e-4f /* 0x1.0p-12f */) {
+            return x;
+        }
+    #else
+        /**
+         * 0x1.0p-8f: At least 16.0bits of precision
+         */
+        if (fabsf(x) < 3.90625e-3f /* 0x1.0p-8f */) {
+            return x + 0.5f * (x * x);
+        }
+    #endif
+    return expf(x) - 1.0f;
 }
 
+
 double expm1(double) __attribute__((alias("expm1f")));