[libc] use LLVM atof for fuzz precheck

[michaelrj]

Summary

The atof differential fuzz precheck skips all hexadecimal subnormals
since glibc's atof doesn't round them correctly. By using glibc's atof
to check if the number is a hexadecimal subnormal it was found that it
sometimes rounds down to zero when it should round up to the minimum
subnormal. This bypassed the check since 0 does not count as a
subnormal. This patch solves the issue by using the LLVM-libc
implementation of atof.

[sivachandra]

I am fine with this, but it is kind of reducing the validity of the fuzz test - we are ignoring the correctness check of a class of inputs using the libc's own results. Could we use MPFR and not depend on the system libc at all? MPFR has functions like mpfr_set_str and mpfr_strtofr which perform the atof operation.

[michaelrj]

Abandoned in favor of https://reviews.llvm.org/D150905

libc/fuzzing/stdlib/atof_differential_fuzz.cpp

bool has_hex_prefix(const uint8_t *str) {
   // Skip over sign
   if (str[index] == '-' || str[index] == '+') {
     ++index;
   }
   return str[index] == '0' && (tolower(str[index + 1])) == 'x';
 }
 
 bool should_be_skipped(const uint8_t *str) {
-  double init_result = ::atof(reinterpret_cast<const char *>(str));
+  // We can be more confident that our atof is correct than the system one, at
+  // least for this case.
+  double init_result = __llvm_libc::atof(reinterpret_cast<const char *>(str));
   if (init_result < 0) {
     init_result = -init_result;
   }
   if (init_result < MIN_NORMAL && init_result != 0) {
     return has_hex_prefix(str);
   }
   return false;
 }