Attempt at less-than-full-byte-boundary B patching. #2
Conversation
bobmcwhirter
commented
Jan 28, 2019
bobmcwhirter
commented
|
Argh. My square bracket and the mask are interleaved poorly. |
|
Code updated. |
bobmcwhirter
force-pushed
the
svm/aarch64
branch
15 times, most recently
from
d0a2374 to
9d78bfa
Compare
We want to find a better way to implement array offset calculation.
The old pattern for an int array member loading is:
and x4, x3, #0xffffffff
lsl x4, x4, #2
add x4, x4, #0x10
ldr x0, [x2, x4]
And the optimized pattern is:
add x1, x2, w3, uxtw #2
ldr x0, [x1, #16]
Here is a jmh unit test for int array member loading:
private final int func(int[] arr) {
int ret = 0;
for (int i=0; i < field0; i++) {
ret += arr[i];
}
return ret;
}
And the jmh performance results:
without this patch with this patch units
1173501.595 948290.420 us/op
Here are the additional issues fixed related to this patch:
1. Fix an issue in MatchProcessor, so that we can process platform specific node
as a matchable node.
2. Close this optimization for substratevm because the derived reference is not
supported yet in SVM.
mx bootstrap passes.
mx gate -o -t 'CTW:hosted' passes.
Change-Id: Ia34f2bf044ca6863be075cf000b29524ca44c42d
|
Hello Bob McWhirter, thanks for contributing a PR to our project! We use the Oracle Contributor Agreement to make the copyright of contributions clear. We don't have a record of you having signed this yet, based on your email address bob -(at)- mcwhirter -(dot)- org. You can sign it at that link. If you think you've already signed it, please comment below and we'll check. |
|
Hello Bob McWhirter, thanks for contributing a PR to our project! We use the Oracle Contributor Agreement to make the copyright of contributions clear. We don't have a record of you having signed this yet, based on your email address bmcwhirt -(at)- redhat -(dot)- com. You can sign it at that link. If you think you've already signed it, please comment below and we'll check. |
2 participants
