Author: Edd Barrett <vex...@gmail.com>
Branch: jit_hint_docs
Changeset: r76961:8d151b2df127
Date: 2015-05-01 15:36 +0100
http://bitbucket.org/pypy/pypy/changeset/8d151b2df127/
Log: Add more detail to @jit.elidable and @jit.promote.
diff --git a/rpython/rlib/jit.py b/rpython/rlib/jit.py
--- a/rpython/rlib/jit.py
+++ b/rpython/rlib/jit.py
@@ -34,6 +34,26 @@
side effect, but those side effects are idempotent (ie caching).
If a particular call to this function ends up raising an exception, then it
is handled like a normal function call (this decorator is ignored).
+
+ Note also that this optimisation will only take effect if the arguments
+ to the function are "provably constant". By this we mean each argument
+ is either:
+
+ 1) literally constant in the RPython source
+ 2) easily shown to be constant by the tracer
+ 3) a promoted variable (see @jit.promote)
+
+ Examples of condition 2:
+
+ * i1 = int_eq(i0, 0), guard_true(i1)
+ * i1 = getfield_pc_pure(<constant>, "immutable_field")
+
+ In both cases, the tracer will deduce that i1 is constant.
+
+ Failing the above conditions, the function is not traced into (as if the
+ function were decorated with @jit.dont_look_inside). Generally speaking,
+ it is a bad idea to liberally sprinkle @jit.elidable without a concrete
+ need.
"""
if DEBUG_ELIDABLE_FUNCTIONS:
cache = {}
@@ -78,6 +98,25 @@
@specialize.argtype(0)
def promote(x):
+ """
+ Promotes a variable in a trace to a constant.
+
+ When a variable is promoted, a guard is inserted that assumes the value
+ of the variable is constant. In other words, the value of the variable
+ is checked to be the same as it was at trace collection time. Once the
+ variable is assumed constant, more aggressive constant folding may be
+ possible.
+
+ If however, the guard fails frequently, a bridge will be generated
+ this time assuming the constancy of the variable under its new value.
+ This optimisation should be used carefully, as in extreme cases, where
+ the promoted variable is not very constant at all, code explosion can
+ occur. In turn this leads to poor performance.
+
+ Overpromotion is characterised by a cascade of bridges branching from
+ very similar guard_value opcodes, each guarding the same variable under
+ a different value.
+ """
return hint(x, promote=True)
def promote_string(x):
_______________________________________________
pypy-commit mailing list
pypy-commit@python.org
https://mail.python.org/mailman/listinfo/pypy-commit
