Currently, our intrinsics are generated as functions that have the
usual setup, which means an alloca, and therefore also a jump, for
those intrinsics that return an immediate value. This is especially bad
for unoptimized builds because it means that an intrinsic like
"contains_managed" that should be just "ret 0" or "ret 1" actually ends
up allocating stack space, doing a jump and a store/load sequence
before it finally returns the value.
To fix that, we need a way to stop the generic function declaration
mechanism from allocating stack space for the return value. This
implicitly also kills the jump, because the block for static allocas
isn't required anymore.
Additionally, trans_intrinsic needs to build the return itself instead
of calling finish_fn, because the latter relies on the availability of
the return value pointer.
With these changes, we get the bare minimum code required for our
intrinsics, which makes them small enough that inlining them makes the
resulting code smaller, so we can mark them as "always inline" to get
better performing unoptimized builds.
Optimized builds also benefit slightly from this change as there's less
code for LLVM to translate and the smaller intrinsics help it to make
better inlining decisions for a few code paths.
Building stage2 librustc gets ~1% faster for the optimized version and 5% for
the unoptimized version.
Most arms of the huge match contain the same code, differing only in
small details like the name of the llvm intrinsic that is to be called.
Thus the duplicated code can be factored out into a few functions that
take some parameters to handle the differences.
Whenever a lang_item is required, some relevant message is displayed, often with
a span of what triggered the usage of the lang item.
Now "hello word" is as small as:
```rust
#[no_std];
extern {
fn puts(s: *u8);
}
extern "rust-intrinsic" {
fn transmute<T, U>(t: T) -> U;
}
#[start]
fn main(_: int, _: **u8, _: *u8) -> int {
unsafe {
let (ptr, _): (*u8, uint) = transmute("Hello!");
puts(ptr);
}
return 0;
}
```
Allowing them in type signatures is a significant amount of extra work, unfortunately. This also doesn't apply to static values, which takes a different code path.
As per @pcwalton's request, `debug!(..)` is only activated when the `debug` cfg is set; that is, for `RUST_LOG=some_module=4 ./some_program` to work, it needs to be compiled with `rustc --cfg debug some_program.rs`. (Although, there is the sneaky `__debug!(..)` macro that is always active, if you *really* need it.)
It functions by making `debug!` expand to `if false { __debug!(..) }` (expanding to an `if` like this is required to make sure `debug!` statements are typechecked and to avoid unused variable warnings), and adjusting trans to skip the pointless branches in `if true ...` and `if false ...`.
The conditional expansion change also required moving the inject-std-macros step into a new pass, and makes it actually insert them at the top of the crate; this means that the cfg stripping traverses over the macros and so filters out the unused ones.
This appears to takes an unoptimised build of `librustc` from 65s to 59s; and the full bootstrap from 18m41s to 18m26s on my computer (with general background use).
`./configure --enable-debug` will enable `debug!` statements in the bootstrap build.
That is, the `b` branch in `if true { a } else { b }` will not be
trans'd, and that expression will be exactly the same as `a`. This
means that, for example, macros conditionally expanding to `if false
{ .. }` (like debug!) will not waste time in LLVM (or trans).
An alloca in an unreachable block would shortcircuit with Undef, but with type
`Type`, rather than type `*Type` (i.e. a plain value, not a pointer) but it is
expected to return a pointer into the stack, leading to confusion and LLVM
asserts later.
Similarly, attaching the range metadata to a Load in an unreachable block
makes LLVM unhappy, since the Load returns Undef.
Fixes#7344.
Macros can be conditionally defined because stripping occurs before macro
expansion, but, the built-in macros were only added as part of the actual
expansion process and so couldn't be stripped to have definitions conditional
on cfg flags.
debug! is defined conditionally in terms of the debug config, expanding to
nothing unless the --cfg debug flag is passed (to be precise it expands to
`if false { normal_debug!(...) }` so that they are still type checked, and
to avoid unused variable lints).
If the TLS key is 0-sized, then the linux linker is apparently smart enough to
put everything at the same pointer. OSX on the other hand, will reserve some
space for all of them. To get around this, the TLS key now actuall consumes
space to ensure that it gets a unique pointer
We used to have concrete types in glue functions, but the way we used
to implement that broke inlining of those functions. To fix that, we
converted all glue to just take an i8* and always casted to that type.
The problem with the old implementation was that we made a wrong
assumption about the glue functions, taking it for granted that they
always take an i8*, because that's the function type expected by the
TyDesc fields. Therefore, we always ended up with some kind of cast.
But actually, we can initially have the glue with concrete types and
only cast the functions to the generic type once we actually emit the
TyDesc data.
That means that for glue calls that can be statically resolved, we don't
need any casts, unless the glue uses a simplified type. In that case we
cast the argument. And for glue calls that are resolved at runtime, we
cast the argument to i8*, because that's what the glue function in the
TyDesc expects.
Since most of out glue calls are static, this saves a lot of bitcasts.
The size of the unoptimized librustc.ll goes down by 240k lines.
Turns out this was a more subtle bug than I originally thought. My analysis can be found in #7732, but I also tried to put descriptive info into the comments.
Closes#7732
We used to have concrete types in glue functions, but the way we used
to implement that broke inlining of those functions. To fix that, we
converted all glue to just take an i8* and always casted to that type.
The problem with the old implementation was that we made a wrong
assumption about the glue functions, taking it for granted that they
always take an i8*, because that's the function type expected by the
TyDesc fields. Therefore, we always ended up with some kind of cast.
But actually, we can initially have the glue with concrete types and
only cast the functions to the generic type once we actually emit the
TyDesc data.
That means that for glue calls that can be statically resolved, we don't
need any casts, unless the glue uses a simplified type. In that case we
cast the argument. And for glue calls that are resolved at runtime, we
cast the argument to i8*, because that's what the glue function in the
TyDesc expects.
Since most of out glue calls are static, this saves a lot of bitcasts.
The size of the unoptimized librustc.ll goes down by 240k lines.
Currently, we always create a dedicated "return" basic block, but when
there's only a single predecessor for that block, it can be merged with
that predecessor. We can achieve that merge by only creating the return
block on demand, avoiding its creation when its not required.
Reduces the pre-optimization size of librustc.ll created with --passes ""
by about 90k lines which equals about 4%.
Currently, immediate values are copied into an alloca only to have an
addressable storage so that it can be used with memcpy. Obviously we
can skip the memcpy in this case.
cc #6004 and #3273
This is a rewrite of TLS to get towards not requiring `@` when using task local storage. Most of the rewrite is straightforward, although there are two caveats:
1. Changing `local_set` to not require `@` is blocked on #7673
2. The code in `local_pop` is some of the most unsafe code I've written. A second set of eyes should definitely scrutinize it...
The public-facing interface currently hasn't changed, although it will have to change because `local_data::get` cannot return `Option<T>`, nor can it return `Option<&T>` (the lifetime isn't known). This will have to be changed to be given a closure which yield `&T` (or as an Option). I didn't do this part of the api rewrite in this pull request as I figured that it could wait until when `@` is fully removed.
This also doesn't deal with the issue of using something other than functions as keys, but I'm looking into using static slices (as mentioned in the issues).
Currently, immediate values are copied into an alloca only to have an
addressable storage so that it can be used with memcpy. Obviously we
can skip the memcpy in this case.
