This previous change broke DoubleToIStub on armv6:
https://code.google.com/p/v8/source/detail?r=16322
The problem is that DoubleToIStub::Generate assumed that it could safely use the ip register, but on armv6 the ubfx implementation will clobber any previous value stored there. So instead, pick another register.
Test case:
for (var i=0; i<2; i++) {
v = 4294967295;
v &= -2;
print(v)
}
This should print -2 twice, but on armv6 without this patch, it prints -2 followed by 2046.
This problem causes sunspider's bitops-nsieve-bit, crypto-md5 and crypto-sha1 tests to generate incorrect results (but the results are not checked for validity in sunspider-1.0 as available in chromium, but are checked and reported as incorrect in sunspider-1.0.2).
Thanks to Tomasz Kilarski for helping out with this.
R=bmeurer@chromium.org, rmcilroy@chromium.org
Review URL: https://codereview.chromium.org/131823004
Patch from Mostyn Bramley-Moore <mostynb@opera.com>.
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@18688 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
6eb9e02b00