[1/2] PR 63340: Avoid harmful union classes in ira-costs.c

This patch is the first of two to fix PR 63340, which is an ia64
regression caused by:

2014-09-22  Richard Sandiford  <richard.sandiford@arm.com>

	* hard-reg-set.h: Include hash-table.h.
	(target_hard_regs): Add a finalize method and a x_simplifiable_subregs
	field.
	* target-globals.c (target_globals::~target_globals): Call
	hard_regs->finalize.
	* rtl.h (subreg_shape): New structure.
	(shape_of_subreg): New function.
	(simplifiable_subregs): Declare.
	* reginfo.c (simplifiable_subreg): New structure.
	(simplifiable_subregs_hasher): Likewise.
	(simplifiable_subregs): New function.
	(invalid_mode_changes): Delete.
	(alid_mode_changes, valid_mode_changes_obstack): New variables.
	(record_subregs_of_mode): Remove subregs_of_mode parameter.
	Record valid mode changes in valid_mode_changes.
	(find_subregs_of_mode): Remove subregs_of_mode parameter.
	Update calls to record_subregs_of_mode.
	(init_subregs_of_mode): Remove invalid_mode_changes and bitmap
	handling.  Initialize new variables.  Update call to
	find_subregs_of_mode.
	(invalid_mode_change_p): Check new variables instead of
	invalid_mode_changes.
	(finish_subregs_of_mode): Finalize new variables instead of
	invalid_mode_changes.
	(target_hard_regs::finalize): New function.
	* ira-costs.c (print_allocno_costs): Call invalid_mode_change_p
	even when CLASS_CANNOT_CHANGE_MODE is undefined.

After that patch, we consider a hard register to be invalid for a pseudo
register if the hard register doesn't allow all the mode changes
required by the pseudo register.  A class is invalid if all hard
registers in it are invalid.  The problem was that this redefines the
behaviour for union classes like ia64's GR_AND_FR_REGS.  Before the
patch, validity was based entirely on CANNOT_CHANGE_MODE_CLASS, which
rejects any superset of FR_REGS if a mode change is invalid for FR_REGS.
After the patch, mode changes allowed by GR_REGS are also allowed by
GR_AND_FR_REGS.

As explained in the big block comment in the patch, there's an argument
whether the mode changes allowed by union classes should be the union or
the intersection of the mode changes allowed by subclasses.  Both are
right in some cases and wrong in others.  The upshot is that we really
shouldn't be using union classes if only one of the subclasses is valid.

This first patch lays the groundwork for the main fix by:

(a) taking the register mode as well as the "approximate" allocation
    class into account in setup_regno_cost_classes_by_aclass, so that
    a register's cost classes are always "correct" for its mode.

(b) checking contains_reg_of_mode when setting up the cost classes
    rather than when using them, so that this can be taken into account
    when deciding whether union classes are useful.

(c) excluding classes that allow the same set of registers as existing
    cost classes, such as GR_AND_FR_REGS if only GR_REGS or FR_REGS
    are valid.

(d) using the cost_classes "index" array to map excluded classes to the
    appropriate subclass, so that we can still take the subunion of
    two cost classes and expect it to have a valid index.  E.g. if:

      A = { 1          }
      B = {    2, 3, 4 }
      C = { 1, 2, 3    }
      D = { 1, 2, 3, 4 }

    and if 4 is invalid for a particular pseudo register, A, B and C
    are useful but D is redundant with C.  A \subunion B gives D,
    then the lookup will map it to C.

This significantly reduces the number of redundant classes in the
cost_classes structure, so it's also a minor compile-time improvement.
The time for -O0 fold-const.ii on x86_64 improved by ~0.5%.
(record_reg_classes was previously the hottest function in the
compilation, after the patch it goes down to number 2, though
it's still costly.)

I did a diff of the assembly output before and after the patch
on x86_64-linux-gnu, powerpc64-linux-gnu, s390x-linux-gnu and
aarch64-linux-gnu.  There were some minor register allocation
changes in a handful files, but nothing major.

Tested on x86_64-linux-gnu, powerpc64-linux-gnu and aarch64-elf.
OK to install?

Thanks,
Richard

gcc/
	PR rtl-optimization/63340 (part 1)
	* ira-costs.c (all_cost_classes): New variable.
	(complete_cost_classes): New function, split out from...
	(setup_cost_classes): ...here.
	(initiate_regno_cost_classes): Set up all_cost_classes.
	(restrict_cost_classes): New function.
	(setup_regno_cost_classes_by_aclass): Restrict the cost classes to
	registers that are valid for the register's mode.
	(setup_regno_cost_classes_by_mode): Model the mode cache as a
	restriction of all_cost_classes to a particular mode.
	(print_allocno_costs): Remove contains_reg_of_mode check.
	(print_pseudo_costs, find_costs_and_classes): Likewise.

[1/2] PR 63340: Avoid harmful union classes in ira-costs.c

Commit Message

Patch