Commit 7ff60235 authored by U-Maokai\andi's avatar U-Maokai\andi Committed by Ben Gamari

cmm: Use two equality checks for two alt switch with default

For code like:
f 1 = e1
f 7 = e2
f _ = e3

We can treat it as a sparse jump table, check if we are outside of the
range in one direction first and then start checking the values.

GHC currently does this by checking for x>7, then x <= 7 and at last x
== 1.

This patch changes this such that we only compare for equality against
the two values and jump to the default if non are equal.

The resulting code is both faster and smaller.
wheel-sieve1 improves by 4-8% depending on problem size.

This implements the idea from #14644

Reviewers: bgamari, simonmar, simonpj, nomeata

Reviewed By: simonpj, nomeata

Subscribers: nomeata, simonpj, rwbarton, thomie, carter

Differential Revision:
parent 59fa7b32
......@@ -251,6 +251,68 @@ data SwitchPlan
-- findSingleValues
-- 5. The thus collected pieces are assembled to a balanced binary tree.
Note [Two alts + default]
Discussion and a bit more info at #14644
When dealing with a switch of the form:
switch(e) {
case 1: goto l1;
case 3000: goto l2;
default: goto ldef;
If we treat it as a sparse jump table we would generate:
if (e > 3000) //Check if value is outside of the jump table.
goto ldef;
else {
if (e < 3000) { //Compare to upper value
if(e != 1) //Compare to remaining value
goto ldef;
goto l2;
goto l1;
Instead we special case this to :
if (e==1) goto l1;
else if (e==3000) goto l2;
else goto l3;
This means we have:
* Less comparisons for: 1,<3000
* Unchanged for 3000
* One more for >3000
This improves code in a few ways:
* One comparison less means smaller code which helps with cache.
* It exchanges a taken jump for two jumps no taken in the >range case.
Jumps not taken are cheaper (See Agner guides) making this about as fast.
* For all other cases the first range check is removed making it faster.
The end result is that the change is not measurably slower for the case
>3000 and faster for the other cases.
This makes running this kind of match in an inner loop cheaper by 10-20%
depending on the data.
In nofib this improves wheel-sieve1 by 4-9% depending on problem
We could also add a second conditional jump after the comparison to
keep the range check like this:
cmp 3000, rArgument
jg <default>
je <branch 2>
While this is fairly cheap it made no big difference for the >3000 case
and slowed down all other cases making it not worthwhile.
-- | Does the target support switch out of the box? Then leave this to the
-- target!
......@@ -272,6 +334,10 @@ createSwitchPlan (SwitchTargets _signed (lo,hi) Nothing m)
--Checking If |range| = 2 is enough if we have two unique literals
, hi - lo == 1
= IfEqual x1 l1 (Unconditionally l2)
-- See Note [Two alts + default]
createSwitchPlan (SwitchTargets _signed _range (Just defLabel) m)
| [(x1, l1), (x2,l2)] <- M.toAscList m
= IfEqual x1 l1 (IfEqual x2 l2 (Unconditionally defLabel))
createSwitchPlan (SwitchTargets signed range mbdef m) =
-- pprTrace "createSwitchPlan" (text (show ids) $$ text (show (range,m)) $$ text (show pieces) $$ text (show flatPlan) $$ text (show plan)) $
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment