Add test case to memory-tests.factor for recent large object allocation bug

diff --git a/core/memory/memory-tests.factor b/core/memory/memory-tests.factor
index 45e6090e773877981357c1bcae3ed312b3ab3ac3..74e002faa9f9013cdd04b3dc10ecbfacb68f203b 100644 (file)
--- a/core/memory/memory-tests.factor
+++ b/core/memory/memory-tests.factor
@@ -1,34 +1,56 @@
-USING: generic kernel kernel.private math memory prettyprint io
-sequences tools.test words namespaces layouts classes
-classes.builtin arrays quotations io.launcher system ;
+USING: accessors kernel kernel.private math memory prettyprint
+io sequences tools.test words namespaces layouts classes
+classes.builtin arrays quotations system ;
+FROM: tools.memory => data-room code-room ;
 IN: memory.tests
 
-[ ] [ { } { } become ] unit-test
-
-! LOL
-[ ] [
-    vm
-    "-i=" image append
-    "-generations=2"
-    "-e=USING: memory io prettyprint system ; input-stream gc . 0 exit"
-    4array try-process
-] unit-test
+[ save-image-and-exit ] must-fail
 
+! Tests for 'instances'
 [ [ ] instances ] must-infer
+2 [ [ [ 3 throw ] instances ] must-fail ] times
+
+! Tests for 'become'
+[ ] [ { } { } become ] unit-test
 
-! Code GC wasn't kicking in when needed
+! Bug found on Windows build box, having too many words in the
+! image breaks 'become'
+[ ] [ 100000 [ f <uninterned-word> ] replicate { } { } become drop ] unit-test
+
+! Bug: code heap collection had to be done when data heap was
+! full, not just when code heap was full. If the code heap
+! contained dead code blocks referring to large data heap
+! objects, those large objects would continue to live on even
+! if the code blocks were not reachable, as long as the code
+! heap did not fill up.
 : leak-step ( -- ) 800000 f <array> 1quotation call( -- obj ) drop ;
 
 : leak-loop ( -- ) 100 [ leak-step ] times ;
 
 [ ] [ leak-loop ] unit-test
 
-TUPLE: testing x y z ;
-
-[ save-image-and-exit ] must-fail
-
-! Erg's bug
-2 [ [ [ 3 throw ] instances ] must-fail ] times
+! Bug: allocation of large objects directly into tenured space
+! can proceed past the high water mark.
+!
+! Suppose the nursery and aging spaces are mostly comprised of
+! reachable objects. When doing a full GC, objects from young
+! generations ere promoted *before* unreachable objects in
+! tenured space are freed by the sweep phase. So if large object
+! allocation filled up the heap past the high water mark, this
+! promotion might trigger heap growth, even if most of those
+! large objects are unreachable.
+SYMBOL: foo
 
-! Bug found on Windows build box, having too many words in the image breaks 'become'
-[ ] [ 100000 [ f <uninterned-word> ] replicate { } { } become drop ] unit-test
+[ ] [
+    gc
+
+    data-room tenured>> size>>
+    
+    10 [
+        4 [ 120 1024 * f <array> ] replicate foo set-global
+        100 [ 256 1024 * f <array> drop ] times
+    ] times
+    
+    data-room tenured>> size>>
+    assert=
+] unit-test