[libc-commits] [libc] [libc][fuzzing] Improve printf long double fuzzing (PR #172113)

[Michael Jones via libc-commits]

================
@@ -101,16 +103,19 @@ extern "C" int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
       prec = (prec << 8) + data[cur];
     } else if (cur < sizeof(raw_num) + sizeof(prec) + sizeof(width)) {
       width = (width << 8) + data[cur];
+    } else if (cur < sizeof(raw_num) + sizeof(prec) + sizeof(width) +
+                         sizeof(ld_raw_num)) {
+      ld_raw_num = (ld_raw_num << 8) + data[cur];
     }
   }
 
   num = LIBC_NAMESPACE::fputil::FPBits<double>(raw_num).get_val();
+  ld_num = LIBC_NAMESPACE::fputil::FPBits<long double>(ld_raw_num).get_val();
 
-  // While we could create a "ld_raw_num" from additional bytes, it's much
-  // easier to stick with simply casting num to long double. This avoids the
-  // issues around 80 bit long doubles, especially unnormal and pseudo-denormal
-  // numbers, which MPFR doesn't handle well.
-  long double ld_num = static_cast<long double>(num);
+  // checking the same value in double and long double could help find
+  // mismatches. Mostly this is here to match previous behavior where this was
----------------
michaelrj-google wrote:

updated the comment.

https://github.com/llvm/llvm-project/pull/172113