Short-circuit alt_fail case in record_reg_classes

record_reg_classes is often the hottest function when generating
unoptimised output.  It seems typical for over 60% of the instructions
it handles to be moves, and of course moves tend to be the instructions
with the longest constraint strings.

Maybe we should avoid using move constraints to set costs in unoptimised
output and instead use the "natural" class for the mode being moved.
That's too invasive for stage 3 though.

However, seeing so many moves means that we see many "failing"
alternatives, usually because of '*' or because of hard registers
in function call sequences.  The frequency of alternatives that are
detected as failures after the first operand tends again to be more
than 60%.  Previously we would continue to process the other operands
of the alternative regardless.  This patch instead adds a short-cut.

As well as avoiding unnecessary work, it means that the alt_fail
variable can be jump-threaded away.

Tested on aach64-linux-gnu and x86_64-linux-gnu.  It reduces compile
time by about 1% on some tests with "-g -O0".

gcc/
	* ira-costs.c (record_reg_classes): Break from the inner loop
	early once alt_fail is known to be true.  Update outer loop
	handling accordingly.

From-SVN: r244446
This commit is contained in:
Richard Sandiford 2017-01-13 15:56:04 +00:00 committed by Richard Sandiford
parent 92daf2dedd
commit 3d8e492088
2 changed files with 21 additions and 1 deletions

View File

@ -1,3 +1,9 @@
2017-01-13 Richard Sandiford <richard.sandiford@arm.com>
* ira-costs.c (record_reg_classes): Break from the inner loop
early once alt_fail is known to be true. Update outer loop
handling accordingly.
2017-01-13 Jeff Law <law@redhat.com>
* tree-ssa-dse.c (decrement_count): New function.

View File

@ -820,6 +820,9 @@ record_reg_classes (int n_alts, int n_ops, rtx *ops,
constraints[i] = p;
if (alt_fail)
break;
/* How we account for this operand now depends on whether it
is a pseudo register or not. If it is, we first check if
any register classes are valid. If not, we ignore this
@ -999,10 +1002,21 @@ record_reg_classes (int n_alts, int n_ops, rtx *ops,
alt_cost += ira_memory_move_cost[mode][classes[i]][1];
else
alt_fail = 1;
if (alt_fail)
break;
}
if (alt_fail)
continue;
{
/* The loop above might have exited early once the failure
was seen. Skip over the constraints for the remaining
operands. */
i += 1;
for (; i < n_ops; ++i)
constraints[i] = skip_alternative (constraints[i]);
continue;
}
op_cost_add = alt_cost * frequency;
/* Finally, update the costs with the information we've