diff options
| author | Keith Randall <khr@golang.org> | 2021-09-10 15:24:16 -0700 |
|---|---|---|
| committer | Keith Randall <khr@golang.org> | 2021-09-27 20:42:34 +0000 |
| commit | 301f6c8019bea813b039c3d376a6464a0e117dce (patch) | |
| tree | 9f357af4032039785de2e13bfb63a71a01aad4cb /src/cmd/compile/internal/noder/transform.go | |
| parent | dfd875d015fb67671a2374c229e2159388d37693 (diff) | |
| download | go-301f6c8019bea813b039c3d376a6464a0e117dce.tar.xz | |
cmd/compile: keep methods on generic types from being deadcode eliminated
We currently make dictionaries contain a relocation pointing to
methods that generic code might use, so that those methods are not
deadcode eliminated. However, with inlining we can end up not using
the dictionary, making the reference from the dictionary to the method
no longer keep the method alive.
Fix this by keeping the dictionary alive at generic interface call sites.
It's a bit of overkill, as we only need to keep the dictionary statically
alive. We don't actually need it dynamically alive, which is what KeepAlive
does. But it works. It ends up generating a LEAQ + stack spill that aren't
necessary, but that's pretty low overhead.
To make this work, I needed to stop generating methods on shape types.
We should do this anyway, as we shouldn't ever need them. But currently
we do use them! issue44688.go has a test that only works because it calls
a method on a shape type. I've disabled that test for now, will work on it
in a subsequent CL.
Fixes #48047
Change-Id: I78968868d6486c1745f51b8b43be0898931432a2
Reviewed-on: https://go-review.googlesource.com/c/go/+/349169
Trust: Keith Randall <khr@golang.org>
Trust: Dan Scales <danscales@google.com>
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Dan Scales <danscales@google.com>
Reviewed-by: Than McIntosh <thanm@google.com>
Diffstat (limited to 'src/cmd/compile/internal/noder/transform.go')
| -rw-r--r-- | src/cmd/compile/internal/noder/transform.go | 15 |
1 files changed, 14 insertions, 1 deletions
diff --git a/src/cmd/compile/internal/noder/transform.go b/src/cmd/compile/internal/noder/transform.go index f7115904fe..9076db2822 100644 --- a/src/cmd/compile/internal/noder/transform.go +++ b/src/cmd/compile/internal/noder/transform.go @@ -132,7 +132,9 @@ func transformConvCall(n *ir.CallExpr) ir.Node { // transformCall transforms a normal function/method call. Corresponds to last half // (non-conversion, non-builtin part) of typecheck.tcCall. This code should work even // in the case of OCALL/OFUNCINST. -func transformCall(n *ir.CallExpr) { +// The dict parameter is used for OCALLINTER nodes to ensure that the called method +// is retained by the linker. +func transformCall(n *ir.CallExpr, dict *ir.Name) { // n.Type() can be nil for calls with no return value assert(n.Typecheck() == 1) transformArgs(n) @@ -142,6 +144,17 @@ func transformCall(n *ir.CallExpr) { switch l.Op() { case ir.ODOTINTER: n.SetOp(ir.OCALLINTER) + if n.X.(*ir.SelectorExpr).X.Type().HasShape() { + if dict == nil { + base.Fatalf("calls on shape interfaces need a dictionary reference") + } + dict.SetAddrtaken(true) + // KeepAlive isn't exactly the right thing here, as we only + // need to keep the dictionary live in the linker-deadcode + // sense, not the at-runtime sense. But the at-runtime sense + // is stronger, so it works. See issue 48047. + n.KeepAlive = append(n.KeepAlive, dict) + } case ir.ODOTMETH: l := l.(*ir.SelectorExpr) |
