RFR: 8289422: Fix and re-enable vector conditional move
Fei Gao
fgao at openjdk.org
Wed Jul 27 06:26:59 UTC 2022
// float[] a, float[] b, float[] c;
for (int i = 0; i < a.length; i++) {
c[i] = (a[i] > b[i]) ? a[i] : b[i];
}
After [JDK-8139340](https://bugs.openjdk.org/browse/JDK-8139340) and [JDK-8192846](https://bugs.openjdk.org/browse/JDK-8192846), we hope to vectorize the case
above by enabling -XX:+UseCMoveUnconditionally and -XX:+UseVectorCmov.
But the transformation here[1] is going to optimize the BoolNode
with constant input to a constant and break the design logic of
cmove vector node[2]. We can't prevent all GVN transformation to
the BoolNode before matcher, so the patch keeps the condition input
as a constant while creating a cmove vector node, and then
restructures it into a binary tree before matching.
When the input order of original cmp node is different from the
input order of original cmove node, like:
// float[] a, float[] b, float[] c;
for (int i = 0; i < a.length; i++) {
c[i] = (a[i] < b[i]) ? a[i] : b[i];
}
the patch negates the mask of the BoolNode before creating the
cmove vector node in SuperWord::output().
We can also use VectorNode::implemented() to consult if vector
conditional move is supported in the backend. So, the patch cleans
the related code in SuperWord::implemented().
With the patch, the performance uplift is:
(The micro-benchmark functions are included in the file
test/micro/org/openjdk/bench/vm/compiler/TypeVectorOperations.java)
AArch64:
Benchmark (length) Mode Cnt uplift(ns/op)
cmoveD 523 avgt 15 68.89%
cmoveF 523 avgt 15 72.40%
X86:
Benchmark (length) Mode Cnt uplift(ns/op)
cmoveD 523 avgt 15 73.12%
cmoveF 523 avgt 15 85.45%
[1]https://github.com/openjdk/jdk/blob/779b4e1d1959bc15a27492b7e2b951678e39cca8/src/hotspot/share/opto/subnode.cpp#L1310
[2]https://github.com/openjdk/jdk/blob/779b4e1d1959bc15a27492b7e2b951678e39cca8/src/hotspot/share/opto/matcher.cpp#L2365
-------------
Commit messages:
- 8289422: Fix and re-enable vector conditional move
Changes: https://git.openjdk.org/jdk/pull/9652/files
Webrev: https://webrevs.openjdk.org/?repo=jdk&pr=9652&range=00
Issue: https://bugs.openjdk.org/browse/JDK-8289422
Stats: 290 lines in 9 files changed: 275 ins; 7 del; 8 mod
Patch: https://git.openjdk.org/jdk/pull/9652.diff
Fetch: git fetch https://git.openjdk.org/jdk pull/9652/head:pull/9652
PR: https://git.openjdk.org/jdk/pull/9652
More information about the hotspot-compiler-dev
mailing list