The `AnyError` placeholder error can cause tests to succeed for errors
other than what was expected. That is, the use of `AnyError` can mask
bugs in a unit test. So replace it with the specific error, or error type,
the test expects to be raised.
This does not remove the anyError structure because it is used in
the TestCase.DoesNotCompile() method. To keep the size of this change
as small as possible I want to defer updating that use to a separate
change. However, remove the public AnyError var so future test writers
don't attempt to use it.
This simple optimization improves the performance of pipelines containing only
one form drastically. Pipelines containing more than one form also execute a
little bit faster.
A subset of the benchmark results, on MacBook Air M1 2020:
| Benchmark | Before | After | Speedup |
| --------- | ----------- | ----------- | ------- |
| nop | 3398 ns/op | 934.1 ns/op | 3.6x |
| nop-nop | 13596 ns/op | 11423 ns/op | 1.2x |
| put-x | 4163 ns/op | 1611 ns/op | 2.6x |
This has always been the documented behavior, but up until this point, "set"
actually behaved like the legacy assignment form, which creates the variable if
it doesn't exist yet.
This fixes the discrepancy. Addresses #645.
For some reason this wrong sequencing was only discovered on FreeBSD, and the
test reliably pass on macOS (likely differences in scheduling algorithms).
Also make the bug a bit easier to reproduce on FreeBSD by using an external
command on the reader side. Builtin commands are too quick to execute.
Also replace (*Frame).OutputChan with (*Frame).ValueOutput, which returns a
small interface for writing to the value output that is also aware when the
reader is gone.
The `edit:after-command` hooks are called with a single argument:
a pseudo-map with these keys:
"command": the command line that was run
"duration": the execution duration in seconds
"error": any error that occurred ($nil if no error occurred)
The `edit:command-duration` variable is the elapsed seconds (as a
float64) of the most recently run interactive command.
Resolves#1029
This commit replaces scopeOp, the only remaining place that mutates *Ns in
place, with nsOp, which performs copy-on-write for *Ns.
As a result, the "eval" command no longer mutates the passed namespace. The
default namespace it uses is also changed to match the default of "-source" (an
amalgamated namespace from the local and upvalue scopes), making "-source $file"
equivalent to "eval (slurp <$file)", and formally deprecated.
Another result is that (*Evaler).Eval can now guard the mutation of the global
namespace with the mutex, making it concurrency-safe to execute multiple sources
that touch the global namespace.
This fixes#1137.
Also change the variable name used to keep the Exception returned from "err" to
"exc".
This uncovers several error scenarios that were not returning Exception, and
would result in the absense of stack traces when such errors occur.
This change is a preparation step for refining all *Op types to return Exception
as the error.
Keeping Exception as a struct type will make such a change error-prone, since
a (*Exception)(nil) != error(nil), so if an *Op returns a nil *Exception, it
is not nil if the return value is stored in an error-typed variable.
Doing something like the following is likely to result in too many open
files (assuming `ulimit -n` == 256) resulting in a panic:
> fn fact [n]{ if (== $n 0) { put 1 } else { put (* $n (fact (- $n 1))) } }
> fact 60
panic: interface conversion: error is *os.SyscallError, not
*eval.Exception
goroutine 24161 [running]:
github.com/elves/elvish/pkg/eval.(*pipelineOp).exec.func1(0x152a5a0,
0xc000dca210, 0xc000a44e70, 0xc00057b540, 0xc001030590, 0x203000)
github.com/elves/elvish/pkg/eval/compile_effect.go:153 +0x152
created by github.com/elves/elvish/pkg/eval.(*pipelineOp).exec
github.com/elves/elvish/pkg/eval/compile_effect.go:149 +0x225
Exception: elvish exited with 2
Fixes#1208
- Move NewEnvListVar to pkg/eval/vars.
- De-export GlobPattern, GlobFlag and ExternalCmd.
- Merge editor.go and chdir.go into eval.go, value_helper.go into compile_value.go.
- Remove eval_internal_test.go and replace it with a new test for $pid.
- Make Evaler mostly thread-safe. The only remaining thread-unsafe part is the
modules field, which is more tricky than other fields.
- Remove the state and evalerScopes type, and move their fields into Evaler.
- Expose valuePrefix via a get method, and change PortsFromFiles to take the
prefix instead of a *Evaler. Also expose a PortsFromStdFiles.
- Make Evaler a normal field of Frame, instead of an embedded field. This makes
access to global states more explicit.
- Remove reliance on scopeOp, concentrating all the scope for creating the local
scope in (*closure).call.
- Check options at the very beginning, and include all unsupported options in the
error.
Commit 734eb95 changed most uses of `fm.ports[x].` to a public method
that makes the intent clearer. This changes a couple of uses that were
overlooked by that prior change.
I stumbled across a comment that began with "XXX". It was clearly meant as
a "TODO" comment. This changes all such occurrences. However, a few "XXX"
comments are ambiguous and a better prefix might be "WARNING". The "TODO"
prefix at least ensures someone, eventually, looks into the situation
and either rewords the comment or fixes the problem. This change means
everyone can assume searching for "// TODO" will find all such comments
rather than requiring they also know to search for "// XXX".
