Hi Jozef, On Fri, Jul 24, 2020 at 12:50:48PM +0100, Jozef Lawrynowicz wrote: > On Thu, Jul 23, 2020 at 01:34:22PM -0500, Segher Boessenkool wrote: > > On Thu, Jul 23, 2020 at 04:56:14PM +0100, Jozef Lawrynowicz wrote: > > > + /* The returned cost must be relative to COSTS_N_INSNS (1). An insn > > > with a > > > + length of 2 bytes is the smallest possible size and so must be > > > equivalent > > > + to COSTS_N_INSNS (1). */ > > > + return COSTS_N_INSNS (cost) / (2 * COSTS_N_INSNS (1)); > > > > This is the same as "cost / 2", so "length / 2" here, which doesn't look > > right. The returned value should have the same "unit" as COSTS_N_INSNS > > does, so maybe you want COSTS_N_INSNS (length / 2) ? > > Indeed it looks like I made a thinko in that calculation in TARGET_INSN_COSTS; > trying to make it verbose to show the thought behind the calculation backfired > :) > > Fixing it to return "COSTS_N_INSNS (length / 2)" actually made codesize > noticeably worse for most of my benchmarks. > I had to define BRANCH_COST to indicate branches are not cheap. > > In the original patch a cheap instruction would have a cost of 1. > When using the default BRANCH_COST of 1 to calculate the cost of a branch > compared to an insn (e.g. in ifcvt.c), BRANCH_COST would be wrapped in > COSTS_N_INSNS, scaling the cost to 4, which suitably disparaged > it vs the cheap insn cost of 1. > > With the fixed insn_cost calculation, a cheap instruction would cost 4 > with the COSTS_N_INSNS scaling, and a branch would cost the same, which is not > right.
There isn't much you can do to battle the "default" cost of 4 -- this is pervasive throughout the compiler -- so it is much easier to go with the flow. > > It is already printed in the generated asm with -dp? Not sure if you > > want more detail than that. > > > > '-dp' > > Annotate the assembler output with a comment indicating which > > pattern and alternative is used. The length and cost of each > > instruction are also printed. > > > > During development I found it useful to see the insns in RTL format and their > costs alongside that. In hindsight, it doesn't really have much use in the > finalized patch, so I've removed it. There is -dP for that (capital P) :-) It isn't very pretty, not sure how that could be improved? > +/* The cost of a branch sequence is roughly 3 "cheap" instructions. */ > +#define BRANCH_COST(speed_p, predictable_p) 3 > + > +/* Override the default BRANCH_COST heuristic to indicate that it is > preferable > + to retain short-circuit operations, this results in significantly better > + codesize and performance. */ > +#define LOGICAL_OP_NON_SHORT_CIRCUIT 0 That looks just fine :-) Segher
