bjoernager/rust
bjoernager/rust
1
Fork 0
Code Issues Pull requests Activity

Change signature of target_features_cfg.

Browse source 
Currently it is called twice, once with `allow_unstable` set to true and
once with it set to false. This results in some duplicated work. Most
notably, for the LLVM backend, `LLVMRustHasFeature` is called twice for
every feature, and it's moderately slow. For very short running
compilations on platforms with many features (e.g. a `check` build of
hello-world on x86) this is a significant fraction of runtime.

This commit changes `target_features_cfg` so it is only called once, and
it now returns a pair of feature sets. This halves the number of
`LLVMRustHasFeature` calls.
This commit is contained in:
Nicholas Nethercote 2025-02-25 15:25:54 +11:00
parent 2df8e657f2
commit 936a8232df
6 changed files with 85 additions and 71 deletions
Download patch file Download diff file Expand all files Collapse all files

13
compiler/rustc_codegen_cranelift/src/lib.rs
View file

@ -176,13 +176,9 @@ impl CodegenBackend for CraneliftCodegenBackend {
}
}
fn target_features_cfg(
&self,
sess: &Session,
_allow_unstable: bool,
) -> Vec<rustc_span::Symbol> {
fn target_features_cfg(&self, sess: &Session) -> (Vec<Symbol>, Vec<Symbol>) {
// FIXME return the actually used target features. this is necessary for #[cfg(target_feature)]
if sess.target.arch == "x86_64" && sess.target.os != "none" {
let target_features = if sess.target.arch == "x86_64" && sess.target.os != "none" {
// x86_64 mandates SSE2 support and rustc requires the x87 feature to be enabled
vec![sym::fsxr, sym::sse, sym::sse2, Symbol::intern("x87")]
} else if sess.target.arch == "aarch64" {
@ -196,7 +192,10 @@ impl CodegenBackend for CraneliftCodegenBackend {
}
} else {
vec![]
}
};
// FIXME do `unstable_target_features` properly
let unstable_target_features = target_features.clone();
(target_features, unstable_target_features)
}
fn print_version(&self) {

68
compiler/rustc_codegen_gcc/src/lib.rs
View file

@ -259,8 +259,8 @@ impl CodegenBackend for GccCodegenBackend {
.join(sess)
}
fn target_features_cfg(&self, sess: &Session, allow_unstable: bool) -> Vec<Symbol> {
target_features_cfg(sess, allow_unstable, &self.target_info)
fn target_features_cfg(&self, sess: &Session) -> (Vec<Symbol>, Vec<Symbol>) {
target_features_cfg(sess, &self.target_info)
}
}
@ -486,35 +486,41 @@ fn to_gcc_opt_level(optlevel: Option<OptLevel>) -> OptimizationLevel {
/// Returns the features that should be set in `cfg(target_feature)`.
fn target_features_cfg(
sess: &Session,
allow_unstable: bool,
target_info: &LockedTargetInfo,
) -> Vec<Symbol> {
) -> (Vec<Symbol>, Vec<Symbol>) {
// TODO(antoyo): use global_gcc_features.
sess.target
.rust_target_features()
.iter()
.filter_map(|&(feature, gate, _)| {
if allow_unstable
|| (gate.in_cfg() && (sess.is_nightly_build() || gate.requires_nightly().is_none()))
{
Some(feature)
} else {
None
}
})
.filter(|feature| {
// TODO: we disable Neon for now since we don't support the LLVM intrinsics for it.
if *feature == "neon" {
return false;
}
target_info.cpu_supports(feature)
/*
adx, aes, avx, avx2, avx512bf16, avx512bitalg, avx512bw, avx512cd, avx512dq, avx512er, avx512f, avx512fp16, avx512ifma,
avx512pf, avx512vbmi, avx512vbmi2, avx512vl, avx512vnni, avx512vp2intersect, avx512vpopcntdq,
bmi1, bmi2, cmpxchg16b, ermsb, f16c, fma, fxsr, gfni, lzcnt, movbe, pclmulqdq, popcnt, rdrand, rdseed, rtm,
sha, sse, sse2, sse3, sse4.1, sse4.2, sse4a, ssse3, tbm, vaes, vpclmulqdq, xsave, xsavec, xsaveopt, xsaves
*/
})
.map(Symbol::intern)
.collect()
let f = |allow_unstable| {
sess.target
.rust_target_features()
.iter()
.filter_map(|&(feature, gate, _)| {
if allow_unstable
|| (gate.in_cfg()
&& (sess.is_nightly_build() || gate.requires_nightly().is_none()))
{
Some(feature)
} else {
None
}
})
.filter(|feature| {
// TODO: we disable Neon for now since we don't support the LLVM intrinsics for it.
if *feature == "neon" {
return false;
}
target_info.cpu_supports(feature)
/*
adx, aes, avx, avx2, avx512bf16, avx512bitalg, avx512bw, avx512cd, avx512dq, avx512er, avx512f, avx512fp16, avx512ifma,
avx512pf, avx512vbmi, avx512vbmi2, avx512vl, avx512vnni, avx512vp2intersect, avx512vpopcntdq,
bmi1, bmi2, cmpxchg16b, ermsb, f16c, fma, fxsr, gfni, lzcnt, movbe, pclmulqdq, popcnt, rdrand, rdseed, rtm,
sha, sse, sse2, sse3, sse4.1, sse4.2, sse4a, ssse3, tbm, vaes, vpclmulqdq, xsave, xsavec, xsaveopt, xsaves
*/
})
.map(Symbol::intern)
.collect()
};
let target_features = f(false);
let unstable_target_features = f(true);
(target_features, unstable_target_features)
}

4
compiler/rustc_codegen_llvm/src/lib.rs
View file

@ -338,8 +338,8 @@ impl CodegenBackend for LlvmCodegenBackend {
llvm_util::print_version();
}
fn target_features_cfg(&self, sess: &Session, allow_unstable: bool) -> Vec<Symbol> {
target_features_cfg(sess, allow_unstable)
fn target_features_cfg(&self, sess: &Session) -> (Vec<Symbol>, Vec<Symbol>) {
target_features_cfg(sess)
}
fn codegen_crate<'tcx>(

54
compiler/rustc_codegen_llvm/src/llvm_util.rs
View file

@ -306,7 +306,7 @@ pub(crate) fn to_llvm_features<'a>(sess: &Session, s: &'a str) -> Option<LLVMFea
/// Must express features in the way Rust understands them.
///
/// We do not have to worry about RUSTC_SPECIFIC_FEATURES here, those are handled outside codegen.
pub(crate) fn target_features_cfg(sess: &Session, allow_unstable: bool) -> Vec<Symbol> {
pub(crate) fn target_features_cfg(sess: &Session) -> (Vec<Symbol>, Vec<Symbol>) {
let mut features: FxHashSet<Symbol> = Default::default();
// Add base features for the target.
@ -331,6 +331,9 @@ pub(crate) fn target_features_cfg(sess: &Session, allow_unstable: bool) -> Vec<S
if let Some(feat) = to_llvm_features(sess, feature) {
for llvm_feature in feat {
let cstr = SmallCStr::new(llvm_feature);
// `LLVMRustHasFeature` is moderately expensive. On targets with many
// features (e.g. x86) these calls take a non-trivial fraction of runtime
// when compiling very small programs.
if !unsafe { llvm::LLVMRustHasFeature(target_machine.raw(), cstr.as_ptr()) }
{
return false;
@ -371,11 +374,7 @@ pub(crate) fn target_features_cfg(sess: &Session, allow_unstable: bool) -> Vec<S
// `features.contains` below.
#[allow(rustc::potential_query_instability)]
features.retain(|f| {
if sess
.target
.implied_target_features(f.as_str())
.contains(&feature.as_str())
{
if sess.target.implied_target_features(f.as_str()).contains(&feature.as_str()) {
// If `f` if implies `feature`, then `!feature` implies `!f`, so we have to
// remove `f`. (This is the standard logical contraposition principle.)
false
@ -387,24 +386,31 @@ pub(crate) fn target_features_cfg(sess: &Session, allow_unstable: bool) -> Vec<S
}
}
// Filter enabled features based on feature gates
sess.target
.rust_target_features()
.iter()
.filter_map(|(feature, gate, _)| {
// The `allow_unstable` set is used by rustc internally to determined which target
// features are truly available, so we want to return even perma-unstable "forbidden"
// features.
if allow_unstable
|| (gate.in_cfg() && (sess.is_nightly_build() || gate.requires_nightly().is_none()))
{
Some(Symbol::intern(feature))
} else {
None
}
})
.filter(|feature| features.contains(&feature))
.collect()
// Filter enabled features based on feature gates.
let f = |allow_unstable| {
sess.target
.rust_target_features()
.iter()
.filter_map(|(feature, gate, _)| {
// The `allow_unstable` set is used by rustc internally to determined which target
// features are truly available, so we want to return even perma-unstable
// "forbidden" features.
if allow_unstable
|| (gate.in_cfg()
&& (sess.is_nightly_build() || gate.requires_nightly().is_none()))
{
Some(Symbol::intern(feature))
} else {
None
}
})
.filter(|feature| features.contains(&feature))
.collect()
};
let target_features = f(false);
let unstable_target_features = f(true);
(target_features, unstable_target_features)
}
pub(crate) fn print_version() {

9
compiler/rustc_codegen_ssa/src/traits/backend.rs
View file

@ -45,10 +45,13 @@ pub trait CodegenBackend {
fn print(&self, _req: &PrintRequest, _out: &mut String, _sess: &Session) {}
/// Returns the features that should be set in `cfg(target_features)`.
/// Returns two feature sets:
/// - The first has the features that should be set in `cfg(target_features)`.
/// - The second is like the first, but also includes unstable features.
///
/// RUSTC_SPECIFIC_FEATURES should be skipped here, those are handled outside codegen.
fn target_features_cfg(&self, _sess: &Session, _allow_unstable: bool) -> Vec<Symbol> {
vec![]
fn target_features_cfg(&self, _sess: &Session) -> (Vec<Symbol>, Vec<Symbol>) {
(vec![], vec![])
}
fn print_passes(&self) {}

8
compiler/rustc_interface/src/util.rs
View file

@ -39,11 +39,11 @@ pub(crate) fn add_configuration(
) {
let tf = sym::target_feature;
let unstable_target_features = codegen_backend.target_features_cfg(sess, true);
sess.unstable_target_features.extend(unstable_target_features.iter().cloned());
let (target_features, unstable_target_features) = codegen_backend.target_features_cfg(sess);
let target_features = codegen_backend.target_features_cfg(sess, false);
sess.target_features.extend(target_features.iter().cloned());
sess.unstable_target_features.extend(unstable_target_features.iter().copied());
sess.target_features.extend(target_features.iter().copied());
cfg.extend(target_features.into_iter().map(|feat| (tf, Some(feat))));