On Mon, 17 Sep 2012, Diego Novillo wrote: > On 2012-09-17 09:25 , Richard Guenther wrote: > > > > This makes use of the new builtin functions for FILE, LINE and FUNCTION > > to turn gimple_build_assign_with_ops/gimple_build_assign_with_ops3 into > > two overloads of gimple_build_assign_with_ops (in theory the _stats > > function can go and we could move the inlines to gimple.c instead, > > eventually removing the assert and simply calling the 3-op overload > > from the 2-op overload?) > > Sounds like a good next step, yes. > > > The patch merely serves as an example on how to transform mem-stat > > code to non-macros. > > > > Quickly build-tested with --disable-gather-detailed-mem-stats, > > stage1 with --enable-gather-detailed-mem-stats and host GCC 4.6 > > prints > > > > gimple.h:766 ((null)) 112: 0.0% > > 0: 0.0% 0: 0.0% 0: 0.0% > > > > for a former gimple_build_assign_with_ops3 call, stage2 (or stage1 > > with a GCC 4.8 host compiler) prints > > > > tree-ssa-math-opts.c:2610 (convert_mult_to_fma) 112: 0.0% > > 0: 0.0% 0: 0.0% 0: 0.0% 1 > > > > so it effectively cripples -fmem-report when not compiled with > > a compiler supporting the builtins. > > But for a bootstrapped 4.8+ compiler, this won't matter, right? It's only > when using a host compiler that doesn't support the builtins.
Yes. Though as it's mostly used for development in which case non-bootstrapped compilers are used it makes --enable-gather-detailed-mem-stats less useful unless you know of this fact. It will also actively break installed pre-release 4.8 compilers used as host compilers ... > > Any comments/objections? > > Looks good to me. Thanks. Thanks, currently bootstrapping / reg-testing on x86_64-unknown-linux-gnu, I'll apply it tomorrow. Richard.
