mikros/rust
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Auto merge of #116409 - Zalathar:span-extraction, r=oli-obk

Browse Source 
coverage: Separate initial span extraction from span processing

One of the main subtasks of coverage instrumentation is looking through MIR to determine a list of source code spans that require coverage counters.

That task is in turn subdivided into a few main steps:
- Getting the initial spans from MIR statements/terminators
- Processing the list of spans to merge or truncate nearby spans as necessary
- Grouping the processed spans by their corresponding coverage graph node

---

This PR enforces a firmer separation between the first two steps (span extraction and span processing), which ends up slightly simplifying both steps, since they don't need to deal with state that is only meaningful for the other step.

---

`@rustbot` label +A-code-coverage
This commit is contained in:
bors 2023-10-10 09:55:15 +00:00
commit 5c37696f60
2 changed files with 215 additions and 237 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

268
compiler/rustc_mir_transform/src/coverage/spans.rs
View File

@ -1,15 +1,13 @@
use super::graph::{BasicCoverageBlock, BasicCoverageBlockData, CoverageGraph, START_BCB};
use std::cell::OnceCell;
use rustc_data_structures::graph::WithNumNodes;
use rustc_index::IndexVec;
use rustc_middle::mir::{
self, AggregateKind, BasicBlock, FakeReadCause, Rvalue, Statement, StatementKind, Terminator,
TerminatorKind,
};
use rustc_span::source_map::original_sp;
use rustc_middle::mir::{self, AggregateKind, Rvalue, Statement, StatementKind};
use rustc_span::{BytePos, ExpnKind, MacroKind, Span, Symbol};
use std::cell::OnceCell;
use super::graph::{BasicCoverageBlock, CoverageGraph, START_BCB};
mod from_mir;
pub(super) struct CoverageSpans {
/// Map from BCBs to their list of coverage spans.
@ -53,27 +51,13 @@ pub(super) fn bcbs_with_coverage_spans(
}
}
#[derive(Debug, Copy, Clone)]
pub(super) enum CoverageStatement {
Statement(BasicBlock, Span, usize),
Terminator(BasicBlock, Span),
}
impl CoverageStatement {
pub fn span(&self) -> Span {
match self {
Self::Statement(_, span, _) | Self::Terminator(_, span) => *span,
}
}
}
/// A BCB is deconstructed into one or more `Span`s. Each `Span` maps to a `CoverageSpan` that
/// references the originating BCB and one or more MIR `Statement`s and/or `Terminator`s.
/// Initially, the `Span`s come from the `Statement`s and `Terminator`s, but subsequent
/// transforms can combine adjacent `Span`s and `CoverageSpan` from the same BCB, merging the
/// `CoverageStatement` vectors, and the `Span`s to cover the extent of the combined `Span`s.
/// `merged_spans` vectors, and the `Span`s to cover the extent of the combined `Span`s.
///
/// Note: A `CoverageStatement` merged into another CoverageSpan may come from a `BasicBlock` that
/// Note: A span merged into another CoverageSpan may come from a `BasicBlock` that
/// is not part of the `CoverageSpan` bcb if the statement was included because it's `Span` matches
/// or is subsumed by the `Span` associated with this `CoverageSpan`, and it's `BasicBlock`
/// `dominates()` the `BasicBlock`s in this `CoverageSpan`.
@ -83,7 +67,9 @@ struct CoverageSpan {
pub expn_span: Span,
pub current_macro_or_none: OnceCell<Option<Symbol>>,
pub bcb: BasicCoverageBlock,
pub coverage_statements: Vec<CoverageStatement>,
/// List of all the original spans from MIR that have been merged into this
/// span. Mainly used to precisely skip over gaps when truncating a span.
pub merged_spans: Vec<Span>,
pub is_closure: bool,
}
@ -94,7 +80,7 @@ pub fn for_fn_sig(fn_sig_span: Span) -> Self {
expn_span: fn_sig_span,
current_macro_or_none: Default::default(),
bcb: START_BCB,
coverage_statements: vec![],
merged_spans: vec![],
is_closure: false,
}
}
@ -104,8 +90,6 @@ pub fn for_statement(
span: Span,
expn_span: Span,
bcb: BasicCoverageBlock,
bb: BasicBlock,
stmt_index: usize,
) -> Self {
let is_closure = match statement.kind {
StatementKind::Assign(box (_, Rvalue::Aggregate(box ref kind, _))) => {
@ -119,23 +103,18 @@ pub fn for_statement(
expn_span,
current_macro_or_none: Default::default(),
bcb,
coverage_statements: vec![CoverageStatement::Statement(bb, span, stmt_index)],
merged_spans: vec![span],
is_closure,
}
}
pub fn for_terminator(
span: Span,
expn_span: Span,
bcb: BasicCoverageBlock,
bb: BasicBlock,
) -> Self {
pub fn for_terminator(span: Span, expn_span: Span, bcb: BasicCoverageBlock) -> Self {
Self {
span,
expn_span,
current_macro_or_none: Default::default(),
bcb,
coverage_statements: vec![CoverageStatement::Terminator(bb, span)],
merged_spans: vec![span],
is_closure: false,
}
}
@ -143,15 +122,13 @@ pub fn for_terminator(
pub fn merge_from(&mut self, mut other: CoverageSpan) {
debug_assert!(self.is_mergeable(&other));
self.span = self.span.to(other.span);
self.coverage_statements.append(&mut other.coverage_statements);
self.merged_spans.append(&mut other.merged_spans);
}
pub fn cutoff_statements_at(&mut self, cutoff_pos: BytePos) {
self.coverage_statements.retain(|covstmt| covstmt.span().hi() <= cutoff_pos);
if let Some(highest_covstmt) =
self.coverage_statements.iter().max_by_key(|covstmt| covstmt.span().hi())
{
self.span = self.span.with_hi(highest_covstmt.span().hi());
self.merged_spans.retain(|span| span.hi() <= cutoff_pos);
if let Some(max_hi) = self.merged_spans.iter().map(|span| span.hi()).max() {
self.span = self.span.with_hi(max_hi);
}
}
@ -205,13 +182,7 @@ pub fn is_macro_expansion(&self) -> bool {
/// * Merge spans that represent continuous (both in source code and control flow), non-branching
/// execution
/// * Carve out (leave uncovered) any span that will be counted by another MIR (notably, closures)
struct CoverageSpansGenerator<'a, 'tcx> {
/// The MIR, used to look up `BasicBlockData`.
mir_body: &'a mir::Body<'tcx>,
/// A `Span` covering the signature of function for the MIR.
fn_sig_span: Span,
struct CoverageSpansGenerator<'a> {
/// A `Span` covering the function body of the MIR (typically from left curly brace to right
/// curly brace).
body_span: Span,
@ -221,7 +192,7 @@ struct CoverageSpansGenerator<'a, 'tcx> {
/// The initial set of `CoverageSpan`s, sorted by `Span` (`lo` and `hi`) and by relative
/// dominance between the `BasicCoverageBlock`s of equal `Span`s.
sorted_spans_iter: Option<std::vec::IntoIter<CoverageSpan>>,
sorted_spans_iter: std::vec::IntoIter<CoverageSpan>,
/// The current `CoverageSpan` to compare to its `prev`, to possibly merge, discard, force the
/// discard of the `prev` (and or `pending_dups`), or keep both (with `prev` moved to
@ -261,7 +232,7 @@ struct CoverageSpansGenerator<'a, 'tcx> {
refined_spans: Vec<CoverageSpan>,
}
impl<'a, 'tcx> CoverageSpansGenerator<'a, 'tcx> {
impl<'a> CoverageSpansGenerator<'a> {
/// Generate a minimal set of `CoverageSpan`s, each representing a contiguous code region to be
/// counted.
///
@ -284,17 +255,22 @@ impl<'a, 'tcx> CoverageSpansGenerator<'a, 'tcx> {
/// Note the resulting vector of `CoverageSpan`s may not be fully sorted (and does not need
/// to be).
pub(super) fn generate_coverage_spans(
mir_body: &'a mir::Body<'tcx>,
mir_body: &mir::Body<'_>,
fn_sig_span: Span, // Ensured to be same SourceFile and SyntaxContext as `body_span`
body_span: Span,
basic_coverage_blocks: &'a CoverageGraph,
) -> Vec<CoverageSpan> {
let mut coverage_spans = Self {
let sorted_spans = from_mir::mir_to_initial_sorted_coverage_spans(
mir_body,
fn_sig_span,
body_span,
basic_coverage_blocks,
sorted_spans_iter: None,
);
let coverage_spans = Self {
body_span,
basic_coverage_blocks,
sorted_spans_iter: sorted_spans.into_iter(),
refined_spans: Vec::with_capacity(basic_coverage_blocks.num_nodes() * 2),
some_curr: None,
curr_original_span: Span::with_root_ctxt(BytePos(0), BytePos(0)),
@ -304,48 +280,9 @@ pub(super) fn generate_coverage_spans(
pending_dups: Vec::new(),
};
let sorted_spans = coverage_spans.mir_to_initial_sorted_coverage_spans();
coverage_spans.sorted_spans_iter = Some(sorted_spans.into_iter());
coverage_spans.to_refined_spans()
}
fn mir_to_initial_sorted_coverage_spans(&self) -> Vec<CoverageSpan> {
let mut initial_spans =
Vec::<CoverageSpan>::with_capacity(self.mir_body.basic_blocks.len() * 2);
for (bcb, bcb_data) in self.basic_coverage_blocks.iter_enumerated() {
initial_spans.extend(self.bcb_to_initial_coverage_spans(bcb, bcb_data));
}
if initial_spans.is_empty() {
// This can happen if, for example, the function is unreachable (contains only a
// `BasicBlock`(s) with an `Unreachable` terminator).
return initial_spans;
}
initial_spans.push(CoverageSpan::for_fn_sig(self.fn_sig_span));
initial_spans.sort_by(|a, b| {
// First sort by span start.
Ord::cmp(&a.span.lo(), &b.span.lo())
// If span starts are the same, sort by span end in reverse order.
// This ensures that if spans A and B are adjacent in the list,
// and they overlap but are not equal, then either:
// - Span A extends further left, or
// - Both have the same start and span A extends further right
.then_with(|| Ord::cmp(&a.span.hi(), &b.span.hi()).reverse())
// If both spans are equal, sort the BCBs in dominator order,
// so that dominating BCBs come before other BCBs they dominate.
.then_with(|| self.basic_coverage_blocks.cmp_in_dominator_order(a.bcb, b.bcb))
// If two spans are otherwise identical, put closure spans first,
// as this seems to be what the refinement step expects.
.then_with(|| Ord::cmp(&a.is_closure, &b.is_closure).reverse())
});
initial_spans
}
/// Iterate through the sorted `CoverageSpan`s, and return the refined list of merged and
/// de-duplicated `CoverageSpan`s.
fn to_refined_spans(mut self) -> Vec<CoverageSpan> {
@ -485,48 +422,6 @@ fn check_invoked_macro_name_span(&mut self) {
}
}
// Generate a set of `CoverageSpan`s from the filtered set of `Statement`s and `Terminator`s of
// the `BasicBlock`(s) in the given `BasicCoverageBlockData`. One `CoverageSpan` is generated
// for each `Statement` and `Terminator`. (Note that subsequent stages of coverage analysis will
// merge some `CoverageSpan`s, at which point a `CoverageSpan` may represent multiple
// `Statement`s and/or `Terminator`s.)
fn bcb_to_initial_coverage_spans(
&self,
bcb: BasicCoverageBlock,
bcb_data: &'a BasicCoverageBlockData,
) -> Vec<CoverageSpan> {
bcb_data
.basic_blocks
.iter()
.flat_map(|&bb| {
let data = &self.mir_body[bb];
data.statements
.iter()
.enumerate()
.filter_map(move |(index, statement)| {
filtered_statement_span(statement).map(|span| {
CoverageSpan::for_statement(
statement,
function_source_span(span, self.body_span),
span,
bcb,
bb,
index,
)
})
})
.chain(filtered_terminator_span(data.terminator()).map(|span| {
CoverageSpan::for_terminator(
function_source_span(span, self.body_span),
span,
bcb,
bb,
)
}))
})
.collect()
}
fn curr(&self) -> &CoverageSpan {
self.some_curr
.as_ref()
@ -589,7 +484,7 @@ fn next_coverage_span(&mut self) -> bool {
self.some_prev = Some(curr);
self.prev_original_span = self.curr_original_span;
}
while let Some(curr) = self.sorted_spans_iter.as_mut().unwrap().next() {
while let Some(curr) = self.sorted_spans_iter.next() {
debug!("FOR curr={:?}", curr);
if self.some_prev.is_some() && self.prev_starts_after_next(&curr) {
debug!(
@ -757,7 +652,7 @@ fn cutoff_prev_at_overlapping_curr(&mut self) {
if self.pending_dups.is_empty() {
let curr_span = self.curr().span;
self.prev_mut().cutoff_statements_at(curr_span.lo());
if self.prev().coverage_statements.is_empty() {
if self.prev().merged_spans.is_empty() {
debug!(" ... no non-overlapping statements to add");
} else {
debug!(" ... adding modified prev={:?}", self.prev());
@ -774,104 +669,3 @@ fn span_bcb_dominates(&self, dom_covspan: &CoverageSpan, covspan: &CoverageSpan)
self.basic_coverage_blocks.dominates(dom_covspan.bcb, covspan.bcb)
}
}
/// If the MIR `Statement` has a span contributive to computing coverage spans,
/// return it; otherwise return `None`.
fn filtered_statement_span(statement: &Statement<'_>) -> Option<Span> {
match statement.kind {
// These statements have spans that are often outside the scope of the executed source code
// for their parent `BasicBlock`.
StatementKind::StorageLive(_)
| StatementKind::StorageDead(_)
// Coverage should not be encountered, but don't inject coverage coverage
| StatementKind::Coverage(_)
// Ignore `ConstEvalCounter`s
| StatementKind::ConstEvalCounter
// Ignore `Nop`s
| StatementKind::Nop => None,
// FIXME(#78546): MIR InstrumentCoverage - Can the source_info.span for `FakeRead`
// statements be more consistent?
//
// FakeReadCause::ForGuardBinding, in this example:
// match somenum {
// x if x < 1 => { ... }
// }...
// The BasicBlock within the match arm code included one of these statements, but the span
// for it covered the `1` in this source. The actual statements have nothing to do with that
// source span:
// FakeRead(ForGuardBinding, _4);
// where `_4` is:
// _4 = &_1; (at the span for the first `x`)
// and `_1` is the `Place` for `somenum`.
//
// If and when the Issue is resolved, remove this special case match pattern:
StatementKind::FakeRead(box (FakeReadCause::ForGuardBinding, _)) => None,
// Retain spans from all other statements
StatementKind::FakeRead(box (_, _)) // Not including `ForGuardBinding`
| StatementKind::Intrinsic(..)
| StatementKind::Assign(_)
| StatementKind::SetDiscriminant { .. }
| StatementKind::Deinit(..)
| StatementKind::Retag(_, _)
| StatementKind::PlaceMention(..)
| StatementKind::AscribeUserType(_, _) => {
Some(statement.source_info.span)
}
}
}
/// If the MIR `Terminator` has a span contributive to computing coverage spans,
/// return it; otherwise return `None`.
fn filtered_terminator_span(terminator: &Terminator<'_>) -> Option<Span> {
match terminator.kind {
// These terminators have spans that don't positively contribute to computing a reasonable
// span of actually executed source code. (For example, SwitchInt terminators extracted from
// an `if condition { block }` has a span that includes the executed block, if true,
// but for coverage, the code region executed, up to *and* through the SwitchInt,
// actually stops before the if's block.)
TerminatorKind::Unreachable // Unreachable blocks are not connected to the MIR CFG
| TerminatorKind::Assert { .. }
| TerminatorKind::Drop { .. }
| TerminatorKind::SwitchInt { .. }
// For `FalseEdge`, only the `real` branch is taken, so it is similar to a `Goto`.
| TerminatorKind::FalseEdge { .. }
| TerminatorKind::Goto { .. } => None,
// Call `func` operand can have a more specific span when part of a chain of calls
| TerminatorKind::Call { ref func, .. } => {
let mut span = terminator.source_info.span;
if let mir::Operand::Constant(box constant) = func {
if constant.span.lo() > span.lo() {
span = span.with_lo(constant.span.lo());
}
}
Some(span)
}
// Retain spans from all other terminators
TerminatorKind::UnwindResume
| TerminatorKind::UnwindTerminate(_)
| TerminatorKind::Return
| TerminatorKind::Yield { .. }
| TerminatorKind::GeneratorDrop
| TerminatorKind::FalseUnwind { .. }
| TerminatorKind::InlineAsm { .. } => {
Some(terminator.source_info.span)
}
}
}
/// Returns an extrapolated span (pre-expansion[^1]) corresponding to a range
/// within the function's body source. This span is guaranteed to be contained
/// within, or equal to, the `body_span`. If the extrapolated span is not
/// contained within the `body_span`, the `body_span` is returned.
///
/// [^1]Expansions result from Rust syntax including macros, syntactic sugar,
/// etc.).
#[inline]
fn function_source_span(span: Span, body_span: Span) -> Span {
let original_span = original_sp(span, body_span).with_ctxt(body_span.ctxt());
if body_span.contains(original_span) { original_span } else { body_span }
}

184
compiler/rustc_mir_transform/src/coverage/spans/from_mir.rs Normal file
View File

@ -0,0 +1,184 @@
use rustc_middle::mir::{
self, FakeReadCause, Statement, StatementKind, Terminator, TerminatorKind,
};
use rustc_span::Span;
use crate::coverage::graph::{BasicCoverageBlock, BasicCoverageBlockData, CoverageGraph};
use crate::coverage::spans::CoverageSpan;
pub(super) fn mir_to_initial_sorted_coverage_spans(
mir_body: &mir::Body<'_>,
fn_sig_span: Span,
body_span: Span,
basic_coverage_blocks: &CoverageGraph,
) -> Vec<CoverageSpan> {
let mut initial_spans = Vec::<CoverageSpan>::with_capacity(mir_body.basic_blocks.len() * 2);
for (bcb, bcb_data) in basic_coverage_blocks.iter_enumerated() {
initial_spans.extend(bcb_to_initial_coverage_spans(mir_body, body_span, bcb, bcb_data));
}
if initial_spans.is_empty() {
// This can happen if, for example, the function is unreachable (contains only a
// `BasicBlock`(s) with an `Unreachable` terminator).
return initial_spans;
}
initial_spans.push(CoverageSpan::for_fn_sig(fn_sig_span));
initial_spans.sort_by(|a, b| {
// First sort by span start.
Ord::cmp(&a.span.lo(), &b.span.lo())
// If span starts are the same, sort by span end in reverse order.
// This ensures that if spans A and B are adjacent in the list,
// and they overlap but are not equal, then either:
// - Span A extends further left, or
// - Both have the same start and span A extends further right
.then_with(|| Ord::cmp(&a.span.hi(), &b.span.hi()).reverse())
// If both spans are equal, sort the BCBs in dominator order,
// so that dominating BCBs come before other BCBs they dominate.
.then_with(|| basic_coverage_blocks.cmp_in_dominator_order(a.bcb, b.bcb))
// If two spans are otherwise identical, put closure spans first,
// as this seems to be what the refinement step expects.
.then_with(|| Ord::cmp(&a.is_closure, &b.is_closure).reverse())
});
initial_spans
}
// Generate a set of `CoverageSpan`s from the filtered set of `Statement`s and `Terminator`s of
// the `BasicBlock`(s) in the given `BasicCoverageBlockData`. One `CoverageSpan` is generated
// for each `Statement` and `Terminator`. (Note that subsequent stages of coverage analysis will
// merge some `CoverageSpan`s, at which point a `CoverageSpan` may represent multiple
// `Statement`s and/or `Terminator`s.)
fn bcb_to_initial_coverage_spans(
mir_body: &mir::Body<'_>,
body_span: Span,
bcb: BasicCoverageBlock,
bcb_data: &BasicCoverageBlockData,
) -> Vec<CoverageSpan> {
bcb_data
.basic_blocks
.iter()
.flat_map(|&bb| {
let data = &mir_body[bb];
data.statements
.iter()
.filter_map(move |statement| {
filtered_statement_span(statement).map(|span| {
CoverageSpan::for_statement(
statement,
function_source_span(span, body_span),
span,
bcb,
)
})
})
.chain(filtered_terminator_span(data.terminator()).map(|span| {
CoverageSpan::for_terminator(function_source_span(span, body_span), span, bcb)
}))
})
.collect()
}
/// If the MIR `Statement` has a span contributive to computing coverage spans,
/// return it; otherwise return `None`.
fn filtered_statement_span(statement: &Statement<'_>) -> Option<Span> {
match statement.kind {
// These statements have spans that are often outside the scope of the executed source code
// for their parent `BasicBlock`.
StatementKind::StorageLive(_)
| StatementKind::StorageDead(_)
// Coverage should not be encountered, but don't inject coverage coverage
| StatementKind::Coverage(_)
// Ignore `ConstEvalCounter`s
| StatementKind::ConstEvalCounter
// Ignore `Nop`s
| StatementKind::Nop => None,
// FIXME(#78546): MIR InstrumentCoverage - Can the source_info.span for `FakeRead`
// statements be more consistent?
//
// FakeReadCause::ForGuardBinding, in this example:
// match somenum {
// x if x < 1 => { ... }
// }...
// The BasicBlock within the match arm code included one of these statements, but the span
// for it covered the `1` in this source. The actual statements have nothing to do with that
// source span:
// FakeRead(ForGuardBinding, _4);
// where `_4` is:
// _4 = &_1; (at the span for the first `x`)
// and `_1` is the `Place` for `somenum`.
//
// If and when the Issue is resolved, remove this special case match pattern:
StatementKind::FakeRead(box (FakeReadCause::ForGuardBinding, _)) => None,
// Retain spans from all other statements
StatementKind::FakeRead(box (_, _)) // Not including `ForGuardBinding`
| StatementKind::Intrinsic(..)
| StatementKind::Assign(_)
| StatementKind::SetDiscriminant { .. }
| StatementKind::Deinit(..)
| StatementKind::Retag(_, _)
| StatementKind::PlaceMention(..)
| StatementKind::AscribeUserType(_, _) => {
Some(statement.source_info.span)
}
}
}
/// If the MIR `Terminator` has a span contributive to computing coverage spans,
/// return it; otherwise return `None`.
fn filtered_terminator_span(terminator: &Terminator<'_>) -> Option<Span> {
match terminator.kind {
// These terminators have spans that don't positively contribute to computing a reasonable
// span of actually executed source code. (For example, SwitchInt terminators extracted from
// an `if condition { block }` has a span that includes the executed block, if true,
// but for coverage, the code region executed, up to *and* through the SwitchInt,
// actually stops before the if's block.)
TerminatorKind::Unreachable // Unreachable blocks are not connected to the MIR CFG
| TerminatorKind::Assert { .. }
| TerminatorKind::Drop { .. }
| TerminatorKind::SwitchInt { .. }
// For `FalseEdge`, only the `real` branch is taken, so it is similar to a `Goto`.
| TerminatorKind::FalseEdge { .. }
| TerminatorKind::Goto { .. } => None,
// Call `func` operand can have a more specific span when part of a chain of calls
| TerminatorKind::Call { ref func, .. } => {
let mut span = terminator.source_info.span;
if let mir::Operand::Constant(box constant) = func {
if constant.span.lo() > span.lo() {
span = span.with_lo(constant.span.lo());
}
}
Some(span)
}
// Retain spans from all other terminators
TerminatorKind::UnwindResume
| TerminatorKind::UnwindTerminate(_)
| TerminatorKind::Return
| TerminatorKind::Yield { .. }
| TerminatorKind::GeneratorDrop
| TerminatorKind::FalseUnwind { .. }
| TerminatorKind::InlineAsm { .. } => {
Some(terminator.source_info.span)
}
}
}
/// Returns an extrapolated span (pre-expansion[^1]) corresponding to a range
/// within the function's body source. This span is guaranteed to be contained
/// within, or equal to, the `body_span`. If the extrapolated span is not
/// contained within the `body_span`, the `body_span` is returned.
///
/// [^1]Expansions result from Rust syntax including macros, syntactic sugar,
/// etc.).
#[inline]
fn function_source_span(span: Span, body_span: Span) -> Span {
use rustc_span::source_map::original_sp;
let original_span = original_sp(span, body_span).with_ctxt(body_span.ctxt());
if body_span.contains(original_span) { original_span } else { body_span }
}