Because it's hot and the number of entries can reach the 1000s, so
linear insertion and search is bad.
This reduces runtime for `sqlite` and `UE4Game-HTML5-Shipping` by 3-4%,
and a couple of other benchmarks (`sqlite`, `godot`, `clang`) by smaller
amounts.
It also increases runtime for `mono` and `tanks` by about 1%; this seems
to be due to incidental changes in which functions are inlined more than
algorithmic changes.
Also, say "guard-offset pages" rather than just "guard pages" to describe the
region of a heap which is never accessible and which exists to support
optimizations for heap accesses with offsets.
And, introduce a `Uimm64` immediate type, and make all heap fields use
`Uimm64` instead of `Imm64` since they really are unsigned.
* Fix verifier printing to print instruction encodings consistently.
Use `FuncWriter::write_instruction` for all instructions so that
encodings are printed consistently.
* Make use-before-def errors mention the relevant value.
* When there are verifier errors, print a message at the end.
* Make verifier errors prettier.
Fix the length of the "^~~~~" to match the printed entity, and print the
error messsage on its own line.
* Clean up "test verifier" failure messages.
* Tidy the uses-value-from-itself error.
The use instruction is the same as the def instruction, so don't print
both. Also, the use instruction is already being printed at the
beginning, so don't print it again at the end.
Memory access instructions which took the GPR_ZERO_DEREF_SAFE register
class (that was removed in #600) should check for the need of either an
offset (r13/rbp) or the SIB byte (r12/rsp). Some load/store instructions
would already take an index, thus already contain the SIB byte in this
case (see instructions which have a comment telling that the else branch
already contains an SIB byte). Non-indexed memory accesses lacked the
SIB byte check, which this patch adds.
This is a followup to af2a952aabd82cf401cc664d0262b139ff92d86b. It
teaches the spilling pass to use the is_ghost() property to test whether
to visit instructions. This fixes a bug handling multiple return values
with fallthrough_return.
It was the caller's responsibility to call TargetIsa::check() before;
now one can't manipulate a TargetIsa without calling the
TargetIsaBuilder::finish() method, which is less error-prone and more in
line with what's coming for other things we're going to generate in the
meta crate.
Also splits the construction of a RegClass in two parts: a prototype is
made first when declaring the RegClass, and missing bits are filled in
when adding it to the TargetIsa(Builder). This avoids an awkward passing
of the isa to the RegClass ctor.
Add an explicit "is_ghost" property to selected instructions, and use
that to determine whether reload and coloring should visit instructions.
This allows them to visit fallthrough_return instructions and insert
fills and register moves as needed.
