[PATCH v6] Add support for SoftFloat library on ARM (original) (raw)

David Holmes david.holmes at oracle.com
Sun Dec 16 23:43:03 UTC 2018

Jakub,

Please note that patches can not be accepted unless hosted on OpenJDK infrastructure ie cr.openjdk.java.net or included directly as an email attachment or inline, to the mailing lists.

Thanks, David

On 13/12/2018 11:21 pm, Jakub Vaněk wrote:

Hi,

I have made a mistake when editing the patch (I forgot to change the added line count in one header). https://raw.githubusercontent.com/ev3dev-lang-java/openjdk-ev3/52d38ea739fda826759f171313991717e477c31d/upstream/softfloat.patch Thanks, Jakub čt 13. 12. 2018 v 14:10 odesílatel Jakub Vaněk <linuxtardis at gmail.com> napsal:

Hi Boris, I have updated the patch: https://raw.githubusercontent.com/ev3dev-lang-java/openjdk-ev3/0887015e0dfcc45ffed03872a8ad42a609496459/upstream/softfloat.patch Now reinterpretcast is used and the URL is present only in the documentation for enabling flags. I didn't expect a performance penalty, I have seen somewhere that compilers are nowadays good in seeing through this type of operation: https://blog.regehr.org/archives/959 I tried compiling a similar code (with noinline functions) on https://godbolt.org/ and on GCC >= 4.6.4 memcpy is optimized away even on -O0. However I agree that the reinterpretcast way is more readable and conscise. Thanks, Jakub čt 13. 12. 2018 v 9:11 odesílatel Boris Ulasevich <boris.ulasevich at bell-sw.com> napsal:

Hi Jakub, Please see comments inline. And one more point: please remove links to mail threads from sources (http://mail.openjdk.java.net/pipermail/aarch32-port-dev/2016-November/000611.html). Boris 12.12.2018 17:36, Jakub Vaněk пишет: Hi Erik,

On 2018-12-12 at 15:41 +0300, Boris Ulasevich wrote: Hi Jakub,

I do not understand why you use memcpy in softfloatarm.cpp. If type cast is the only reason why don't you use reinterpretcast? regards, Boris I'm using memcpy there because float32t is a struct containing a single uint32t member. I think that the proper way of doing a byte- level conversion between different types (type punning) is via memcpy. Do not you expect performance penalty using type conversion this way? If I used a casted pointer, this would violate the rule that I can access a memory location through a pointer of only one type. -fno-strict-aliasing gcc option is used to disable compiler complains on this rule: jdk-jdk/make/hotspot/lib/JvmFeatures.gmk: JVMLDFLAGSFEATURES += -fno-strict-aliasing Using union for this seems to be also undefined in C++, as then I would access the value through a non-active member. Thanks, Jakub On 10.12.2018 23:23, Jakub Vaněk wrote: Hi Erik,

On 2018-12-10 at 09:39 -0800, Erik Joelsson wrote: The build changes look ok to me.

I do think --enable-softfloat is redundant as using any of the other parameters would imply it to be set, but it's ok to have it there. Thanks for the review. The motivation for this was that I was closely following how libffi is handled. The enable option was a workaround for the detection of softfloat in system paths. I don't think this is how the library is going to be used, so I removed this option. New patch can be found here: https://github.com/ev3dev-lang-java/openjdk-ev3/blob/78a9d35986ed51273d9a3bb9878cbfcbe2488bfb/upstream/softfloat.patch Regards, Jakub /Erik < raw patch at_ _https://raw.githubusercontent.com/ev3dev-lang-java/openjdk-ev3/78a9d35986ed51273d9a3bb9878cbfcbe2488bfb/upstream/softfloat.patch_ _>

More information about the build-dev mailing list