The first repair (action source from block, onProfiler:, unproxifiedMethod) took 13 away. The 5 heavy GraphCallerTest tests give a count of 1. A round-2 draft (<work directory>, adds resetSharedState to clear the Microdown resource cache and a warm-up run) is not in the repo and has no report .md; cause is a draft, not proven.
Next step: Validate the round-2 draft: copy it to scripts/pkg-jit-test/fixes/jordanmontt-pharo-instrumentation.st, rerun row 146 on both VMs; if the VMs differ, treat as a VM defect.
| Load | Stock Cog | iospharo VM |
|---|---|---|
| ok | tests=17 P=12 F=5 E=0 S=0 T=0 X=0 | tests=17 P=12 F=5 E=0 S=0 T=0 X=0 |
| Test | Stock Cog | iospharo VM |
|---|---|---|
GraphCallerTest>>testFuzzyMatcher | FAIL | FAIL |
GraphCallerTest>>testGraphAlgos | FAIL | FAIL |
GraphCallerTest>>testMicrodown | FAIL | FAIL |
GraphCallerTest>>testProfStef | FAIL | FAIL |
GraphCallerTest>>testRewriteTool | FAIL | FAIL |
"jordanmontt/pharo-instrumentation: three defects of the package (the two VMs fail the same 17 - 4 tests the same way; the defects are in the package code and in its test, not in a VM). 1. SourceCodeInstrumentator builds the source text of an instrumented method from the two action blocks with 'aBlock asString'. In Pharo 14 BlockClosure>>printOn: prints '[ ] in Class>>selector', not the source of the block. The compiled method then holds the text 'runInMetaLevel: [ ] in Class>>setUp ]': the empty block [ ] (a ConstantBlockClosure) gets the unary message #in, and every call of an instrumented method fails with 'ConstantBlockClosure does not understand #in'. This gives the 6 errors (testInstrument, testProfile, testProfileMethodWithTempVariables, testProfileSeveralMethods, testProfileSeveralMethodsAndClasses, testAfterHasLocalReturn) and the 2 failures with a counter of 0 (testAfterRaisesError, testProfileWithException: the test only asks for 'raise: Error', and the DNU is an Error, so the call of the method fails before any action ran). The repair adds SourceCodeInstrumentator>>actionSource: (a String stays as it is; a block gives the source text of its AST node) and uses it in instrument: and wrapBeforeActionFor:. The tests that give strings (the prototype trap tests) get the same text as before. 2. GraphCallerTest>>setUp calls MpCallGraphInstrumentator new and RwiCallGraphInstrumentator new. Both classes refuse new with 'Use onProfiler: instead of new' since the refactoring of the package; the test was not changed. The repair makes the two profilers first and gives each profiler an instrumentator made with onProfiler:. 3. The instrumentators send #unproxifiedMethod to each CompiledMethod of the set of methods. No loaded package and no method of the image defines CompiledMethod>>unproxifiedMethod (the MethodProxies of the baseline, master, has no such method). The repair adds CompiledMethod>>unproxifiedMethod (answers the receiver: an ordinary method is not a proxy). 4. The five GraphCallerTest tests run whole test packages two times, with a proxy on each method of a package. The time limit of SUnit is 10 s a test. GraphCallerTest class>>defaultTimeLimit gives 10 minutes. Only a run shows whether the limit is needed." add value: #SourceCodeInstrumentator value: #actionSource: value: 'profiling' value: 'actionSource: anAction "Answer the Smalltalk source text of anAction. A String is the text already. A block gives the source of its AST node: BlockClosure>>asString prints ''[ ] in Class>>selector'' in Pharo 14, which is not code." ^ anAction isString ifTrue: [ anAction ] ifFalse: [ anAction sourceNode sourceCode ]'. fix value: #SourceCodeInstrumentator value: #instrument: value: [ :source | source includesSubstring: 'afterAction asString' ] value: 'instrument: aMethod | beforeActionString afterActionString pragmasString originalCodeString instrumentedMethodCode | beforeActionString := self wrapBeforeActionFor: aMethod. afterActionString := self wrapWithMetaSafeMechanism: (self actionSource: afterAction). originalCodeString := ''[ '' , aMethod methodNode body sourceCode , '' ]''. pragmasString := aMethod pragmas inject: '''' into: [ :sum :node | sum , node asString , String cr]. instrumentedMethodCode := String streamContents: [ :stream | stream << pragmasString << beforeActionString << String cr << '' '' << originalCodeString << '' ensure: ['' << String cr << '' '' << afterActionString << '' ]'' ]. instrumentedMethodCode := aMethod sourceNode selectorAndArgumentNames , String cr , instrumentedMethodCode. aMethod methodClass compile: instrumentedMethodCode'. fix value: #SourceCodeInstrumentator value: #wrapBeforeActionFor: value: [ :source | source includesSubstring: 'beforeAction asString' ] value: 'wrapBeforeActionFor: aMethod ^ self wrapWithMetaSafeMechanism: ''[ '' , (self actionSource: beforeAction) , '' cull: '' , aMethod nameForProfiler , '' ]'''. fix value: #GraphCallerTest value: #setUp value: [ :source | source includesSubstring: 'MpCallGraphInstrumentator new' ] value: 'setUp super setUp. mpProfiler := CallGraphProfiler new. mpProfiler instrumentator: (MpCallGraphInstrumentator onProfiler: mpProfiler). rwiProfiler := CallGraphProfiler new. rwiProfiler instrumentator: (RwiCallGraphInstrumentator onProfiler: rwiProfiler)'. add value: #CompiledMethod value: #unproxifiedMethod value: '*InstrumentationProfilers' value: 'unproxifiedMethod "An ordinary method is not a proxy: it is the unproxified method." ^ self'. add value: #'GraphCallerTest class' value: #defaultTimeLimit value: 'accessing' value: 'defaultTimeLimit "A test runs whole test packages twice, with a method proxy on each method of a package." ^ 10 minutes'.
State of 2026-11-20. Made from the test records of the iospharo project (a second Pharo VM). The fixes are proposals for review.