determine_ad_stack_size.cpp clips AdStack row claims by each range-for's trip count, but the loop that does it
iterates over a list that is always empty, so the optimisation never fires and autodiff heaps stay sized by the
unclipped reducer count.
The gather at the top of that block asks for a container statement type:
auto offloaded_tasks = irpass::analysis::gather_statements(root, [&](Stmt *s) { return s->is<OffloadedStmt>(); });gather_statements runs its predicate from StmtSearcher::visit(Stmt *), but BasicStmtVisitor claims
OffloadedStmt with a typed overload that calls preprocess_container_stmt and recurses into the body without
consulting the predicate. StmtSearcher does not override that hook, so an OffloadedStmt predicate matches
nothing. The same applies to MeshForStmt, StructForStmt, RangeForStmt, IfStmt and WhileStmt.
Impact
- The trip-count clipping is dead. Enabling it should reduce autodiff stack memory, but the code has never run, so
it needs to be validated rather than just switched on.
eliminate_recomputable_pushes.cpp builds a preorder index over "all" statements and its comment reasons about
"the container's preorder", which containers do not currently get. Relative order among non-container statements
is unaffected, so this looks latent rather than active.- Callers whose predicates test non-container types are unaffected, as is
auto_diff_common.h, which gathers
everything but only casts to GlobalStoreStmt/AtomicOpStmt/SNodeOpStmt.
Note
I hit this from a different direction: a mesh-for guard written as a gather_statements predicate silently matched
nothing, which sent mesh kernels down a path that produced IR that segfaulted at launch.
Rather than change what the helper returns to its existing callers, my PR adds an include_containers parameter
defaulting to the current behaviour and opts in at the three sites that need it. That deliberately leaves this
autodiff case as-is, since flipping it on is a behaviour change to autodiff that should be made on purpose and
tested by someone who owns that code, not as a side effect of a compile-caching change.
determine_ad_stack_size.cppclips AdStack row claims by each range-for's trip count, but the loop that does ititerates over a list that is always empty, so the optimisation never fires and autodiff heaps stay sized by the
unclipped reducer count.
The gather at the top of that block asks for a container statement type:
gather_statementsruns its predicate fromStmtSearcher::visit(Stmt *), butBasicStmtVisitorclaimsOffloadedStmtwith a typed overload that callspreprocess_container_stmtand recurses into the body withoutconsulting the predicate.
StmtSearcherdoes not override that hook, so anOffloadedStmtpredicate matchesnothing. The same applies to
MeshForStmt,StructForStmt,RangeForStmt,IfStmtandWhileStmt.Impact
it needs to be validated rather than just switched on.
eliminate_recomputable_pushes.cppbuilds a preorder index over "all" statements and its comment reasons about"the container's preorder", which containers do not currently get. Relative order among non-container statements
is unaffected, so this looks latent rather than active.
auto_diff_common.h, which gatherseverything but only casts to
GlobalStoreStmt/AtomicOpStmt/SNodeOpStmt.Note
I hit this from a different direction: a mesh-for guard written as a
gather_statementspredicate silently matchednothing, which sent mesh kernels down a path that produced IR that segfaulted at launch.
Rather than change what the helper returns to its existing callers, my PR adds an
include_containersparameterdefaulting to the current behaviour and opts in at the three sites that need it. That deliberately leaves this
autodiff case as-is, since flipping it on is a behaviour change to autodiff that should be made on purpose and
tested by someone who owns that code, not as a side effect of a compile-caching change.