This helps in the case of field projections of the results of checked
binary operations. E.g.:
_1 = CheckedAdd(const 1i32, const 2i32);
assert(!(_1.1: bool), "attempt to add with overflow" -> bb1
Previously, the `_1.1` field projection lvalue would force_allocate `_1`
so it could read the memory in the old-style way. Now checked math with
its assertions will not allocate at all.
The oom2.rs compile-fail test had to be re-written, because the old
version of it no longer allocates _at all_ (yay!), so it would hit the
stack depth limit instead, from recursion.
Thanks to the `Value` locals refactoring, now primitive locals (ints,
floats, chars, bools, and the like) will not require `Allocation`s at
all, and locals that are never initialized at all because of conditional
control flow won't be wasting memory.
Previously ReturnPointer was just the first slot in the locals array,
which had type `Vec<Pointer>`. But after my recent refactoring, locals
is `Vec<Value>` and it became increasingly hacky to pull a pointer out
of the first slot to be the value. Besides, that hack wouldn't allow
ReturnPointer to ever be an `Lvalue::Local`, referring directly to a
local on a higher stack frame.
Now ReturnPointer has no presence in the locals array, instead being
upgraded to its own field on `Frame`.
This introduces a couple of new hacks, detailed by some of my FIXME
comments, so that I could get the tests passing again and commit. More
commits coming soon should clean up these hacks without much trouble,
and overall I feel that the code is converging on a cleaner, more
efficient design.
This reduces the number of allocations Miri makes drastically. The
`const_to_ptr` function was a lame hack that allocated for every since simple
constant, and all of those are avoided now, except for one extra allocation each
for string and bytestring literals which will be fixed in a followup commit.
There are a number of hacks such as `eval_operand_to_ptr` left over from this
commit, which will also be fixed in followup commits.
