[PATCH] Patchpoint - support symbolic targets.
Philip Reames
listmail at philipreames.com
Wed Apr 8 15:27:07 PDT 2015
LGTM. Since this is very similar to the statepoint code I used that for
comparison sake.
In our local code we've got extra handling for ExternalSymbolSDNode.
I'm not actually sure if that's required though. :)
Here's the line we have:
else if (ExternalSymbolSDNode *ES =
dyn_cast<ExternalSymbolSDNode>(CallTarget)) {
CallTarget = DAG.getTargetExternalSymbol(
ES->getSymbol(), CallTarget.getValueType(), ES->getTargetFlags());
}
One slight tweak you could make is to use a pc relative call rather than
a register call for symbolic targets. This avoids the need for an extra
register, but does require that the target be within a fixed offset.
(Or maybe this can be handled via relocations? Not sure.)
Philip
On 03/31/2015 05:25 PM, Lang Hames wrote:
> Hi All,
>
> At the moment llvm.patchpoint call targets must be integer constants
> (e.g. 0xDEADBEEF). This patch adds support for symbolic targets like
> @foo, which addresses a couple of FIXMEs.
>
> Making this work just involved teaching FastISel and SelectionDAG to
> construct the appropriate MI/SDNodes, and teaching the Targets how to
> lower these to MCInsts. Target support for x86 is included in this
> patch. Support for other targets should be easy to derive from that.
>
> With this patch applied, you can write patchpoints of the following form:
>
> tail call i64 (i64, i32, i8*, i32, ...)*
> @llvm.experimental.patchpoint.void(i64 9, i32 15,
> i8* bitcast (i64 (i64, i64)* @foo to i8*), ; <- Call target
> i32 2, i64 %p1, i64 %p2)
>
> and this will generate:
>
> movabsq $_foo, %r11
> callq *%r11
> // <nop-padding>
>
> For integer targets this would have been something like:
>
> movabsq $0xDEADBEEF, %r11
> callq *%r11
> // <nop-padding>
>
> The advantage of symbolic targets, beyond improved readability, is
> that you can cache the IR or compiled objects and re-use them in
> contexts where the target address may have changed. For example,
> objects that use symbolic patchpoints can be cached in Orc/MCJIT
> object-caches and re-used across JIT invocations.
>
> Immediate-address targets are still fully supported.
>
> I don't really see any downside to this patch, but I thought I'd throw
> it out for general discussion before I go and widen what patchpoint is
> supposed to support.
>
> Does anyone see any problem with this idea?
>
> For the curious, my motivation is that I'd like (eventually) to
> support patchpoints for JIT re-entry in Orc as an alternative to
> indirect calls, and I want that transformation to be straightforward. E.g.
>
> call @not_yet_compiled, ...
> to
> call llvm.patchpoint ..., @not_yet_compiled, ...
>
> Cheers,
> Lang.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20150408/41c4926d/attachment.html>
More information about the llvm-commits
mailing list