Hi! On Thu, Mar 05, 2020 at 10:46:58AM +0800, Jiufu Guo wrote: > PR93709 mentioned regressions on maxlocval_4.f90 and minlocval_f.f90 which > relates to max of '-inf' and 'nan'. This regression occur on P9 which has > new instruction 'xsmaxcdp/xsmincdp'. > The similar issue also could be find on `a < b ? b : a` which is also > generated as `xsmaxcdp` under -O2 for P9. This instruction `xsmaxcdp` > more like C/C++ semantic (a>b?a:b). A testcase is added for this issue. > > The following patch improve code to check -+0 and NaN before 'smax/smin' to > be generated for those cases.
> - else if (rtx_equal_p (op1, true_cond) && rtx_equal_p (op0, false_cond)) > + /* Only when -fno-signed-zeros and -ffinite_math_only are in effect, > + `op0 < op1 ? op1 : op0` works like `op1 > op0 ? op1 : op0` which > + could use smax; > + `op0 > op1 ? op1 : op0` works like `op1 < op0 ? op1 : op0` which > + could use smin. */ > + else if (rtx_equal_p (op1, true_cond) && rtx_equal_p (op0, false_cond) > + && (flag_finite_math_only && !flag_signed_zeros)) > max_p = !max_p; I know I asked for it, but should this use HONOR_NANS (compare_mode) instead? Infinities will work fine? Just NaNs and zeros won't. Okay for trunk with that change (if it works :-) ) Thanks! Segher
