Author: Remi Meier <[email protected]>
Branch: stmgc-c8
Changeset: r76318:0cbb11adcde0
Date: 2015-03-11 10:19 +0100
http://bitbucket.org/pypy/pypy/changeset/0cbb11adcde0/
Log: update TODO
diff --git a/TODO b/TODO
--- a/TODO
+++ b/TODO
@@ -1,8 +1,18 @@
+------------------------------------------------------------
+
+better heuristic of when to break transactions? e.g., we should
+rarely break if there are not threads running in parallel.
+
------------------------------------------------------------
maybe statically optimize away some stm_become_inevitable(), there
are some loops that call it repeatedly (may be not relevant
-for performance)
+for performance)
+
+------------------------------------------------------------
+
+clang doesn't optimize multiple stm_write() in a row (unlike GCC).
+Optimize them manually...
------------------------------------------------------------
@@ -30,19 +40,6 @@
------------------------------------------------------------
-clang doesn't optimize multiple stm_write() in a row (unlike GCC).
-Optimize them manually...
-
-------------------------------------------------------------
-
-stm_read() should also be optimized (moved out of loops). E.g.
-quick_sort.py suffers a 30% slowdown because it uses deque.locate()
-extensively. This method happens to have a *very* tight loop
-where the read barrier has a big negative effect and could
-actually be moved out.
-
-------------------------------------------------------------
-
__pypy__.thread.getsegmentlimit():
XXX This limit is so far a compile time option (STM_NB_SEGMENTS in
@@ -52,13 +49,8 @@
------------------------------------------------------------
-JIT: add an artificial malloc if the loop is so small as to contain
-any!
+weakrefs stay alive longer than expected::
-------------------------------------------------------------
-
-weakrefs stay alive longer than expected::
-
y = some object that was already alive for a while
x = ref(y)
del y
@@ -122,7 +114,7 @@
hg diff -r b0339cb53372^2 -r dd8e2f69fe96^2 : these trunk changes are
missing in the stmgc-c7 branch of PyPy. In particular, reduce the diff
between stmgc-c7 and default (kill rstm.ThreadLocalRef, etc)
-
+
------------------------------------------------------------
there are in PyPy-STM crashes related to markers (with the JIT?).
@@ -130,10 +122,6 @@
------------------------------------------------------------
-GC: call __del__()
-
-------------------------------------------------------------
-
look at jit.elidables as a way to specify "this function can run
earlier, in a separate transaction". Useful to avoid pointless
conflicts in cases the jit.edliable update some cache, like with
_______________________________________________
pypy-commit mailing list
[email protected]
https://mail.python.org/mailman/listinfo/pypy-commit
