Note: requires LLVM 7+ to generatic rsqrt intrinsics

Changes testset to avoid compiler heuristics (copyast) that disables inference.
And changes the allocated macro to rely less on inference to elid allocations for the machinary itself.

This lets us scan a datatype slightly easier,
and opens up a future possibility where we don't have a one-to-one
relationship between fields and contained pointers.

* get MulAddMul out of the BLAS way, promote alpha/beta coeffs

* fix ambiguity

* simplify promotion, :crossedfingers:

* remove promote_unless_bool, add symmetry check `syrk_wrapper!` (bugfix)

* give BLAS another chance

* extend coefficient promotion to sym(v/m)! and hem(v/m)!

* fix typo

On non-case-sensitive platforms, it is possible for `cd` and `pwd()` to
return values that are different from what `realpath()` of the same
would give back.  Therefore, we must adjust our tests to be more robust
against this variation.

This should hopefully fix some of the testing hangs on Windows

"Julia from `v0.5` forwards natively supports multi-threading"
became
"Julia forwards natively supports multi-threading", where
the "forwards" doesn't make much sense anymore.

delete no-op code

Add LIBEXECDIR relative folder constant

Fixes JuliaLang#33808

This was a hold-over from the old iteration protocol, which needed to maintain state between `done` and `next`.
The `iteration` function of `Dict` has since been re-written to be safe for concurrent deletions.

Replaces: JuliaLang#33756
Co-Authored-By: Hans-Peter Suter <[email protected]>

some small gc optimizations

… lengths (JuliaLang#33832)

Fix `realpath()` assumptions in REPL test suite

…-sqrt-fast

Add sqrt_llvm_fast intrinsic (see JuliaLang#33220)

…liaLang#33848)

- avoid allocating `gettask` closure
- stop using a non-const global for eventloop()
- remove some unused code

@KristofferC 

see JuliaLang#32651 (comment)

JuliaLang#33667)

* document return nothing in Functions man page

+ fix broken link to Markdown syntax in doc README

Instead of running a tight infinite loop.
Seems somewhat preferable.

simplify allocated macro

When edit_title_case was implemented, `uppercasefirst` (née `ucfirst`)
and `titlecase` were doing the same transformation on only one word.
But `titlecase` now transforms non-leading letters to lowercase,
which is more expected for the M-c keyboard combo, as it's the
standard readline behavior (I believe).

)

For consistency with base/gmp.jl and test/mpfr.jl.

…uliaLang#33839)

fix JuliaLang#32852

…Lang#33749)

Bind this shell-like feature to the usual "meta-.".

…aLang#33742)

* docs: disambiguate what rand((2, 3)) does (fix JuliaLang#33309)

* Update stdlib/Random/src/Random.jl

Co-Authored-By: Matt Bauman <[email protected]>

)

This adds an `iocontext::Dict{Symbol,Any}` field to `LineEditREPL.options`,
which can be initialized with `atreplinit`, and updated
interactively, e.g. `Base.active_repl.options.iocontext[:compact] = true`.

Fixes JuliaLang#20509.

Yanking twice was fooling `setmark` into believing
that `^Space`, the keybinding for `setmark`, had
been pressed twice, leading to activating the region.
So let's make `setmark` behave as a "command" (and
activate the region) only when called directly by
the user, otherwise treat it as a low-level function.

…uliaLang#33868)

[ci skip]

…ring-nothing

Switch Void to Cvoid in finalizer docstring

…iaLang#33736)

It's not unusual to have code that simply returns an argument.  When
used in generic code though, that currently might involve copying the
value to a new box. We sometimes define functions like
`identity(nospecialize x) = x` to work around that, (which is preferable
anyways for other reasons), but it's better for codegen to also just be
aware of this pattern.

datatype: reorganize layout calculation code

* correct genmatmul! for empty input matrices

* tests for getmatmul!

…aLang#33889)

Use `cache.julialang.org` to download `busybox.exe` at test time