cmd/compile/internal/escape: avoid reading ir.Node during inner loop of walkOne

Broadly speaking, escape analysis has two main phases. First, it
traverses the AST while building a data-flow graph of locations and
edges. Second, during "solve", it repeatedly walks the data-flow graph
while carefully propagating information about each location, including
whether a location's address reaches the heap.

Once escape analysis is in the solve phase and repeatedly walking the
data-flow graph, almost all the information it needs is within the
location graph, with a notable exception being the ir.Class of an
ir.Name, which currently must be checked by following a pointer from
the location to its ir.Node.

For typical graphs, that does not matter much, but if the graph becomes
large enough, cache misses in the inner solve loop start to matter more,
and the class is checked many times in the inner loop.

We therefore store the class information on the location in the graph
to reduce how much memory we need to load in the inner loop.

The package github.com/microsoft/typescript-go/internal/checker
has many locations, and compilation currently spends most of its time
in escape analysis.

This CL gives roughly a 30% speedup for wall clock compilation time
for the checker package:

  go1.24.0:      91.79s
  this CL:       64.98s

Linux perf shows a healthy reduction for example in l2_request.miss and
dTLB-load-misses on an amd64 test VM.

We could tweak things a bit more, though initial review feedback
has suggested it would be good to get this in as it stands.

Subsequent CLs in this stack give larger improvements.

Updates #72815

Change-Id: I3117430dff684c99e6da1e0d7763869873379238
Reviewed-on: https://go-review.googlesource.com/c/go/+/657295
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Keith Randall <khr@google.com>
Reviewed-by: Jake Bailey <jacob.b.bailey@gmail.com>
Reviewed-by: David Chase <drchase@google.com>

diff --git a/src/cmd/compile/internal/escape/graph.go b/src/cmd/compile/internal/escape/graph.go
index 75e2546a7b7a0604195e69e06cf057fe2c908b52..cd800bc4d65ba68ca72f0eaa9796d7537b445447 100644 (file)
--- a/src/cmd/compile/internal/escape/graph.go
+++ b/src/cmd/compile/internal/escape/graph.go
@@ -75,6 +75,8 @@ type location struct {
        captured   bool // has a closure captured this variable?
        reassigned bool // has this variable been reassigned?
        addrtaken  bool // has this variable's address been taken?
+       param      bool // is this variable a parameter (ONAME of class ir.PPARAM)?
+       paramOut   bool // is this variable an out parameter (ONAME of class ir.PPARAMOUT)?
 }
 
 type locAttr uint8
@@ -281,6 +283,11 @@ func (e *escape) newLoc(n ir.Node, persists bool) *location {
                curfn:     e.curfn,
                loopDepth: e.loopDepth,
        }
+       if loc.isName(ir.PPARAM) {
+               loc.param = true
+       } else if loc.isName(ir.PPARAMOUT) {
+               loc.paramOut = true
+       }
        if persists {
                loc.attrs |= attrPersists
        }

diff --git a/src/cmd/compile/internal/escape/solve.go b/src/cmd/compile/internal/escape/solve.go
index 2002f2fbe41f3e30b8a613d605c165920092d487..4b0db1884d7a2cff74d4f63e3016ee2c36d4963d 100644 (file)
--- a/src/cmd/compile/internal/escape/solve.go
+++ b/src/cmd/compile/internal/escape/solve.go
@@ -126,7 +126,7 @@ func (b *batch) walkOne(root *location, walkgen uint32, enqueue func(*location))
                // corresponding result parameter, then record
                // that value flow for tagging the function
                // later.
-               if l.isName(ir.PPARAM) {
+               if l.param {
                        if b.outlives(root, l) {
                                if !l.hasAttr(attrEscapes) && (logopt.Enabled() || base.Flag.LowerM >= 2) {
                                        if base.Flag.LowerM >= 2 {
@@ -270,7 +270,7 @@ func (b *batch) outlives(l, other *location) bool {
        // We don't know what callers do with returned values, so
        // pessimistically we need to assume they flow to the heap and
        // outlive everything too.
-       if l.isName(ir.PPARAMOUT) {
+       if l.paramOut {
                // Exception: Closures can return locations allocated outside of
                // them without forcing them to the heap, if we can statically
                // identify all call sites. For example: