https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114729
--- Comment #19 from Vineet Gupta <vineetg at gcc dot gnu.org> ---
For last several weeks I've been working on various tweaks to address the model
schedule issue. It seems conceptually really simple, something along the lines
of following:
for (;;)
{
FOR_EACH_DEP (insn->insn, SD_LIST_HARD_BACK, sd_it, dep)
if (pro->insn
&& pro->model_priority < insn->model_priority
&& QUEUE_INDEX (pro->insn) != QUEUE_SCHEDULED)
{
- pro->model_priority = model_next_priority;
if (QUEUE_INDEX (pro->insn) == QUEUE_READY)
{
+ if (DEP_TYPE (dep) == REG_DEP_TRUE)
+ pro->model_priority = model_new_priority;
However adding the priority distinction has been really tricky: subtle changes
tend to lead to infinite looping.
I've tried numerous approaches:
- tracking true deps separately than existing overall unscheduled deps.
- Preventing recursion from adding same list multiple times.
A working version which bootstrapped the toolchain, fixed the reduced test,
caused severe spills regression in actual cactu build.
