https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64036
--- Comment #4 from Oleg Endo <olegendo at gcc dot gnu.org> ---
(In reply to Oleg Endo from comment #2)
> An example function, compiling with -O2 -m4:
>
> int test_0 (unsigned short* x, int y, int z)
> {
> return
> (x[0] + x[1] + x[2] + x[3] + x[4] + x[5] + x[6]
> + x[7] + x[8] + x[9] + x[10]) ? y : z;
> }
>
> Without sched1, there are lots of dependencies on the results of memory
> loads.
>
> With sched1, there is generally more code generated and variable live ranges
> are longer. The above function will use r8 and r9, which is not really
> necessary. Memory load dependencies are reduced and more LS/EX/MT
> instructions can be executed in parallel. Code size for the test function
> increases from ~37 insns to ~50 insns. Approximated cycles on SH4 pipeline
> should be ~37 cycles without sched1 and ~33 cycles with sched1. On SH4A the
> latency of a load is 1 cycle, so without sched1 it should be ~28 cycles.
BTW, with AMS there is no difference of sched1 or no-sched1 in code size,
because it uses post-inc loads. With AMS + sched1 the example above compiles
to 38 insns and there are no dependencies/stalls on the mem loads anymore. The
r8,r9 usage issue is also gone.
