The code is now quadratic in the number of statements in a basic block,
whereas before it was quadratic in the number of _control-flow nodes_ in
a basic block.
The `iDominates` relation is directly on control-flow nodes, and its
transitive closure is far too large. It got compiled into a recursion
rather than `fastTC`, and I've observed that recursion to take about an
hour on a medium-size customer snapshot.
The fix is to check for dominance at the basic-block level.
The `cpp/missing-case-in-switch` performed badly on some snapshots, to
the extent where it was as slow as the most expensive IR stages
(example: ChakraCore). This commit makes it faster, removing a
`pragma[noopt]` along the way.
The intermediate tuple counts on a customer codebase drop from 84M to
3M, while the content hash of `getAMissingCase` is the same.
Before:
(124s) Tuple counts for Stmt::EnumSwitch::getAMissingCase#ff#antijoin_rhs:
20867789 ~0% {3} r1 = JOIN Stmt::SwitchStmt::getASwitchCase_dispred#ff AS L WITH Stmt::EnumSwitch::getAMissingCase#ff#shared AS R ON FIRST 1 OUTPUT L.<1>, R.<0>, R.<1>
20122830 ~0% {3} r2 = JOIN r1 WITH Stmt::SwitchCase::getExpr_dispred#ff AS R ON FIRST 1 OUTPUT R.<1>, r1.<1>, r1.<2>
20122830 ~0% {3} r3 = JOIN r2 WITH Expr::Expr::getValue_dispred#ff AS R ON FIRST 1 OUTPUT r2.<2>, r2.<1>, R.<1>
83961918 ~0% {4} r4 = JOIN r3 WITH Enum::EnumConstant::getInitializer_dispred#ff AS R ON FIRST 1 OUTPUT R.<1>, r3.<1>, r3.<0>, r3.<2>
83961918 ~0% {4} r5 = JOIN r4 WITH initialisers AS R ON FIRST 1 OUTPUT R.<2>, r4.<3>, r4.<1>, r4.<2>
234348 ~185% {2} r6 = JOIN r5 WITH Expr::Expr::getValue_dispred#ff AS R ON FIRST 2 OUTPUT r5.<2>, r5.<3>
return r6
...
(124s) Tuple counts for Stmt::EnumSwitch::getAMissingCase#ff:
663127 ~4% {2} r1 = Stmt::EnumSwitch::getAMissingCase#ff#shared AS L AND NOT Stmt::EnumSwitch::getAMissingCase#ff#antijoin_rhs AS R(L.<0>, L.<1>)
return r1
(124s) Registering Stmt::EnumSwitch::getAMissingCase#ff + [] with content 2060ff326cvhihcsvoph6k9divuv4
(124s) >>> Wrote relation Stmt::EnumSwitch::getAMissingCase#ff with 663127 rows and 2 columns.
After:
(5s) Tuple counts for Stmt::EnumSwitch::getAMissingCase_dispred#ff#antijoin_rhs:
746029 ~0% {2} r1 = JOIN Stmt::EnumSwitch::getAMissingCase_dispred#ff#shared AS L WITH Enum::Enum::getAnEnumConstant_dispred#ff AS R ON FIRST 1 OUTPUT R.<1>, L.<1>
3116197 ~2% {3} r2 = JOIN r1 WITH Enum::EnumConstant::getInitializer_dispred#ff AS R ON FIRST 1 OUTPUT R.<1>, r1.<1>, r1.<0>
3116197 ~0% {3} r3 = JOIN r2 WITH initialisers AS R ON FIRST 1 OUTPUT R.<2>, r2.<1>, r2.<2>
3116197 ~311% {3} r4 = JOIN r3 WITH Expr::Expr::getValue_dispred#ff AS R ON FIRST 1 OUTPUT r3.<1>, R.<1>, r3.<2>
234348 ~185% {2} r5 = JOIN r4 WITH Stmt::EnumSwitch::matchesValue#ff AS R ON FIRST 2 OUTPUT r4.<0>, r4.<2>
return r5
(5s) Registering Stmt::EnumSwitch::getAMissingCase_dispred#ff#antijoin_rhs + [] with content 173483d71508vl534mvlr1g0ehi12
(5s) >>> Wrote relation Stmt::EnumSwitch::getAMissingCase_dispred#ff#antijoin_rhs with 82902 rows and 2 columns.
(5s) Starting to evaluate predicate Stmt::EnumSwitch::getAMissingCase_dispred#ff/2@ae4c0b
(5s) Tuple counts for Stmt::EnumSwitch::getAMissingCase_dispred#ff:
746029 ~2% {2} r1 = JOIN Stmt::EnumSwitch::getAMissingCase_dispred#ff#shared AS L WITH Enum::Enum::getAnEnumConstant_dispred#ff AS R ON FIRST 1 OUTPUT L.<1>, R.<1>
663127 ~4% {2} r2 = r1 AND NOT Stmt::EnumSwitch::getAMissingCase_dispred#ff#antijoin_rhs AS R(r1.<0>, r1.<1>)
return r2
(5s) Registering Stmt::EnumSwitch::getAMissingCase_dispred#ff + [] with content 2060ff326cvhihcsvoph6k9divuv4
(5s) >>> Wrote relation Stmt::EnumSwitch::getAMissingCase_dispred#ff with 663127 rows and 2 columns.
This query used two fastTC operations that were already somewhat
inefficient on their own but could send the evaluator into an OOM loop
when run in parallel without enough RAM.
The fix is to recurse manually, starting just from the expressions that
are potential candidates for alerts.
Fixes issue https://github.com/github/codeql-c-analysis-team/issues/20
This change undoes the workaround in https://github.com/Semmle/ql/pull/2736, and replaces it with a fix for the underlying cause. The problem was that the IR construction code for side effects incorrectly assumed that `BufferAccessOpcode` included `SizedBufferAccessOpcode`. I think that was actually a perfectly reasonable assumption to make, so I changed the `Opcode` hierarchy to make it true.
