[Bug target/83240] New: x86_64 vectorized sqrt of denormal yields -inf when DAZ=0
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83240 Bug ID: 83240 Summary: x86_64 vectorized sqrt of denormal yields -inf when DAZ=0 Product: gcc Version: 8.0 Status: UNCONFIRMED Severity: normal Priority: P3 Component: target Assignee: unassigned at gcc dot gnu.org Reporter: gson at gson dot org Target Milestone: --- Created attachment 42766 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=42766&action=edit Preprocessed source When compiling for x95_64 with -O3 -ffast-math, gcc vectorizes square roots of single-precision floats into SSE code using the rsqrtps instruction followed by a Newton-Raphson step to calculate four square roots in one go. This code returns incorrect results (-infinity rather than zero) for denormal inputs when the DAZ (Denormals Are Zero) flag is not set. On Linux, if the -ffast-math option is also given at the link stage, the DAZ flag is set by crtfastmath.o, and the problem does not occur. However, if -ffast-math is used when compiling library code, it is difficult to guarantee that the link stage, which may be under the control of an entirely different software project, also uses the -ffast-math option. Also, some systems do not currently support crtfastmath.o at all (see http://gnats.netbsd.org/50940). Therefore, it seems to me that code built with -ffast-math ought to work correctly (even if not with optimal performance) whether DAZ is set or not. Or conversely, if there is an intentional requirement that DAZ must be set when executing code compiled with -ffast-math, this requirement ought to be clearly and prominently documented. The attached test program demonstrates the issue. Compile and run it as follows (using a separate compile and link stage to suppress the use of crtfastmath.o): gcc -c -O3 -ffast-math test_denormal.i gcc test_denormal.o -lm -o test_denormal ./test_denormal This issue affects a wide range of gcc versions, including 6.3.0 and the current SVN trunk (r255277).
[Bug target/83240] x86_64 vectorized sqrt of denormal yields -inf when DAZ=0
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83240 --- Comment #2 from Andreas Gustafsson --- Re "simply don't use -ffast-math if you are dealing with stuff like denormals", the documentation for -ffinite-math-only (which is implied by -ffast-math) says "Allow optimizations for floating-point arithmetic that assume that arguments and results are not NaNs or +-Infs". It does not mention denormals, which to me suggests that optimizations that assume arguments are not denormals are *not* allowed. Also, since we are discussing the case where DAZ is 0, it is likely that FTZ (Flush To Zero) is also 0. This means that even if all the inputs to the computation compiled with -ffast-math are guaranteed not to be denormals, denormals can arise within the computation itself, for example any time two inputs are added, subtracted, multiplied, or divided. In other words, if you can't use -ffast-math "when dealing with denormals", you can't use it for any code that does arithmetic on its inputs, which is pretty much the same thing as saying you can't use it ever.
[Bug target/83240] x86_64 vectorized sqrt of denormal yields -inf when DAZ=0
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83240 --- Comment #5 from Andreas Gustafsson --- Re "may violate IEEE or ANSI standards", that argument could be used to justify any behavior, even making 1 + 1 yield 42. Clearly the intent of -funsafe-math-optimizations is not to allow arbitrary incorrect results, but minor deviations from the standard such as differences in rounding or evaluation order that result in reasonable approximations of the correct result even if they differ from the exact result required by the standard. When taking the square root of a very small positive number, a result of 0 is a reasonable approximation of the correct result, even though it violates the standard, but a result of minus infinity is not.
