RFR: 8143067: aarch64: guarantee failure in javac (adrp out of range in relocation)

[Vladimir Kozlov]

Looks good to me.

Thanks,
Vladimir

On 11/17/15 1:56 AM, Edward Nevill wrote:
> Hi,
>
> Please review the following webrev
>
> http://cr.openjdk.java.net/~enevill/8143067/webrev
>
> JIRA Issuel: https://bugs.openjdk.java.net/browse/JDK-8143067
>
> The issue is described in more detail in the JIRA issue but basically
> adrp instructions are becoming out of range during code relocation. This
> is happening with -Xmx4G (or greater).
>
> The problem is the existing code for adrp() does
>
>   if (offset out of range) {
>     generate other code to form address
>   } else {
>     generate adrp
>   }
>
> Unfortunately this not work if the adrp subsequently becomes out of
> range due to code relocation, and this cannot be fixed up in a
> relocation because it will require more than 1 instruction.
>
> The solution I have adopted is to add a new method far_adrp() which must
> be used if the adrp may be out of range of a single adrp instruction.
> The existing adrp() has been modified to guarantee that the offset is in
> range.
>
> far_adrp() generates either
>
>    adrp Rn, symbol
>    nop
>
> if the symbol is initially within range or
>
>    adrp Rn, <part symbol> ; fill in bits 0..31 in Rn
>    mov  Rn, <2nd part> << 32 ; fill in bits 32..47
>
> If the symbol subsequently becomes out of range because of a code
> relocation the nop in the first form is rewritten as a movk in the 2nd
> form.
>
> Tested with jtreg hotspot and langtools
>
> Hotspot before: Test results: passed: 927; failed: 6; error: 11
> Hotspot after: Test results: passed: 928; failed: 5; error: 11
>
> Langtools before and after: Test results: passed: 3,326; error: 2
>
> If have also tested it with jtreg/jdk to ensure there were no fatal
> errors and tested jtreg/hotspot with a fastdebug version to check the
> fix does not cause any assert failure.
>
> Thanks for your help,
> Ed.
>
>