floating-point constant expressions are incorrectly rounded

[llvmbot]

Bugzilla Link	11406
Resolution	FIXED
Resolved on	Nov 25, 2011 21:38
Version	trunk
OS	FreeBSD
Attachments	fail.c
Reporter	LLVM Bugzilla Contributor
CC	@efriedma-quic

Extended Description

In the default floating-point rounding mode (IEEE round-to-nearest), halfway cases are supposed to round to the nearest EVEN representable number, but in clang, apparently they don't always do that.

The attached program prints 0x1.0000000000001p+0 when compiled with clang -O, but it prints 0x1.0p+0 when compiled with gcc or clang -O0, or when 'd' is declared volatile. 0x1.0p+0 is the correct answer.

[llvmbot]

Also note that this problem is not due to DBL_EPSILON being evaluated in extra precision. The same problem occurs if DBL_EPSILON is replaced with 0x1p-52.

[efriedma-quic]

r145141.