On Fri, Apr 14, 2023 at 08:59:24PM +0200, Harald Anlauf via Fortran wrote: > > the compile-time simplification of intrinsic SET_EXPONENT was > broken since the early days of gfortran for argument X < 1 > (including negative X) and for I < 0. I identified and fixed > several issues in the implementation. The testcase explores > argument space comparing compile-time and runtime results and > is checked against Intel. > > Regtested on x86_64-pc-linux-gnu. OK for mainline? >
Yes, it is certainly better than the current situation. > This is not a regression, but can lead to wrong code. > Would it be OK to backport to open branches? Sure. Looks simply and fairly specific. I was wondering about the difference between set_exponent() and scale(), and found that set_exponent() talks about IEEE values while scale() doesn't. I'm wondering if we should add the IEEE special cases to the testsuite. Of particular note, I doubt that this is true: If X is an IEEE NaN, the result is the same NaN. program foo real x, y x = 1 y = x - x x = (x - x) / y print '(F4.0,1X,Z8.8)', x, x y = set_exponent(x,1) print '(F4.0,1X,Z8.8)', y, y end program foo gfcx -o z a.f90 && ./z NaN FFC00000 NaN 7FC00000 Those are not the same NaN. The second is a qNaN. The first looks like a qNaN with the sign bit set. -- Steve
