]> Cypherpunks.ru repositories - gostls13.git/commit
cmd/compile/internal/noder: shape-based stenciling for unified IR
authorMatthew Dempsky <mdempsky@google.com>
Sat, 6 Aug 2022 23:40:56 +0000 (16:40 -0700)
committerMatthew Dempsky <mdempsky@google.com>
Thu, 18 Aug 2022 13:16:21 +0000 (13:16 +0000)
commit38edd9bd8da9d7fc7beeba5fd4fd9d605457b04e
tree023cd736b039cfcb4cbb19a0695354092f7214f6
parenta2c2f06cad8aa722120cb73e965d168bfcb4d977
cmd/compile/internal/noder: shape-based stenciling for unified IR

This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.

For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.

For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.

Two notable distinctions from the non-unified frontend:

1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.

2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.

However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.

I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.

The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.

Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
src/cmd/compile/internal/devirtualize/devirtualize.go
src/cmd/compile/internal/noder/reader.go
src/cmd/compile/internal/noder/unified.go
src/cmd/compile/internal/noder/writer.go
src/cmd/compile/internal/ssa/debug_lines_test.go
src/cmd/compile/internal/typecheck/subr.go
src/cmd/compile/internal/walk/convert.go
test/typeparam/nested.go