diff options
| author | Matthew Dempsky <mdempsky@google.com> | 2018-03-06 14:36:49 -0800 |
|---|---|---|
| committer | Matthew Dempsky <mdempsky@google.com> | 2018-03-06 23:33:28 +0000 |
| commit | 2c0c68d621415f0c2dffb71b20fdc0fdbbcfb5fe (patch) | |
| tree | 8e3f475bbed8297dd131e63f7f88ae6e1a3e3227 /src/androidtest.bash | |
| parent | 558769a61b246464f979f3f4d370dad96cee2109 (diff) | |
| download | go-2c0c68d621415f0c2dffb71b20fdc0fdbbcfb5fe.tar.xz | |
cmd/compile: fix miscompilation of "defer delete(m, k)"
Previously, for slow map key types (i.e., any type other than a 32-bit
or 64-bit plain memory type), we would rewrite
defer delete(m, k)
into
ktmp := k
defer delete(m, &ktmp)
However, if the defer statement was inside a loop, we would end up
reusing the same ktmp value for all of the deferred deletes.
We already rewrite
defer print(x, y, z)
into
defer func(a1, a2, a3) {
print(a1, a2, a3)
}(x, y, z)
This CL generalizes this rewrite to also apply for slow map deletes.
This could be extended to apply even more generally to other builtins,
but as discussed on #24259, there are cases where we must *not* do
this (e.g., "defer recover()"). However, if we elect to do this more
generally, this CL should still make that easier.
Lastly, while here, fix a few isues in wrapCall (nee walkprintfunc):
1) lookupN appends the generation number to the symbol anyway, so "%d"
was being literally included in the generated function names.
2) walkstmt will be called when the function is compiled later anyway,
so no need to do it now.
Fixes #24259.
Change-Id: I70286867c64c69c18e9552f69e3f4154a0fc8b04
Reviewed-on: https://go-review.googlesource.com/99017
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Diffstat (limited to 'src/androidtest.bash')
0 files changed, 0 insertions, 0 deletions
