Introduce FIELD_ADDR and use it for TLS statics and instance class fields
#77353
Conversation
ghost
added
the
community-contribution
dotnet-issue-labeler
bot
added
the
area-CodeGen-coreclr
|
Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch Issue DetailsRemoving two uses of We expect a couple diffs in tests where we now fail to hoist a vector constant because we don't resolve struct handles (thus recognize SIMD types) for
|
SingleAccretion
changed the title
FIELD_ADDR and use it for TLS statics and class instance fieldsFIELD_ADDR and use it for TLS statics and instance class fields
SingleAccretion
force-pushed
the
Field-Addr-Upstream
branch
4 times, most recently
from
e7d5f78 to
d22eede
Compare
SingleAccretion
force-pushed
the
Field-Addr-Upstream
branch
from
d22eede to
a6d0afe
Compare
Tested manually to work as well as it did before.
SingleAccretion
force-pushed
the
Field-Addr-Upstream
branch
3 times, most recently
from
71a410f to
a872f5a
Compare
SingleAccretion
force-pushed
the
Field-Addr-Upstream
branch
from
a872f5a to
b621b7b
Compare
With NativeAOT, we can have FIELD_ADDR nodes that are effectively NOPs (those for the method table pointer field). This means not all operators that the expansion produces will be simple (in fact, they can be more or less arbitrary). This means we cannot simply call "fgMorphSmpOp" as we used to. Unfortunately, we cannot just call "fgMorphTree" either, because it propagates assertions on newly created (or, I suppose, bashed) IND nodes. This creates lots of new cases where GTF_ORDER_SIDEEFF is applied to these nodes, affecting CQ. Work around this by calling "fgMorphTree" for non-simple operators only.
|
@dotnet/jit-contrib |
|
/azp run runtime-coreclr jitstress, runtime-coreclr libraries-jitstress, Fuzzlyn |
|
Azure Pipelines successfully started running 3 pipeline(s). |
|
The unreduced BadResult failures in the Fuzzlyn run looks like a known intermittent failure that I have seen before, I don't think it is related to this PR. |
| // Return Value: | ||
| // The expanded tree - a GT_ADD. | ||
| // | ||
| GenTree* Compiler::fgMorphExpandTlsFieldAddr(GenTree* tree) |
There was a problem hiding this comment.
I assume there's no actual changes here except constructing the address part of this? The diff is quite hard to read.
There was a problem hiding this comment.
There is one change: tlsRef = gtNewIndir(TYP_I_IMPL, tlsRef, GTF_IND_NONFAULTING | GTF_IND_INVARIANT); - for the indir off of the TLS handle.
(Not doing this was making fgDebugCheckFlags fail)
|
The stress failures were #78023 |
Removing two uses of
GT_ADDRin the process.We expect a couple diffs in tests where we now fail to hoist a vector constant because we don't resolve struct handles (thus recognize SIMD types) for
FIELD_ADDRs, as well as one case of new substitution due to smaller tree size. We also expect a minor TP improvement in the optimized case.