On Tue, Mar 27, 2012 at 8:01 AM, Bin.Cheng <amker.ch...@gmail.com> wrote: > On Mon, Mar 26, 2012 at 3:27 PM, Richard Guenther > <richard.guent...@gmail.com> wrote: >> On Sun, Mar 25, 2012 at 2:42 PM, Bin.Cheng <amker.ch...@gmail.com> wrote: >>> Hi, >>> In tree-complex.c's function expand_complex_comparison, gcc just >>> expand comparison on complex >>> operands into comparisons on inner type, like: >>> >>> D.5375_17 = REALPART_EXPR <g2>; >>> D.5376_18 = IMAGPART_EXPR <g2>; >>> g2.1_5 = COMPLEX_EXPR <D.5375_17, D.5376_18>; >>> D.5377_19 = REALPART_EXPR <g3>; >>> D.5378_20 = IMAGPART_EXPR <g3>; >>> g3.2_6 = COMPLEX_EXPR <D.5377_19, D.5378_20>; >>> D.5379_21 = D.5375_17 == D.5377_19; >>> D.5380_22 = D.5376_18 == D.5378_20; >>> D.5381_23 = D.5379_21 & D.5380_22; >>> if (D.5381_23 == 1) >>> goto <bb 3>; >>> else >>> goto <bb 4>; >>> >>> So is it possible to do shortcut operation for the "&" on the >>> real/imag part of complex data? >> >> Sure. Does the RTL expander not do that for your target? > > Yes, expand_gimple_cond decides how to expand such codes. > And, it depends on BRANCH_COST whether codes will be expanded into > shortcut operations. true for X86, false for arm at least. > I am wondering, should we take a more general strategy, rather than just > branch cost heuristic decision. > The new strategy should compare the cost of operation and branch, e.g., > On target has no hard float instructions, comparison of float point should be > shortcuted, since the helper function is very expensive. > > Any idea? Thanks
For the particluar case a better lowering from tree-complex.c is desired (I can imagine even using vector instructions ..). It probably was just easiest to not to have to alter the CFG at this point. Richard. > -- > Best Regards.
