[master] RFR: 8368755: [Lilliput] Fix CompressedClassSpaceSize handling

[Thomas Stuefe]

On Fri, 26 Sep 2025 09:35:38 GMT, Roman Kennke <rkennke at openjdk.org> wrote:

> We currently set the default CompressedClassSpaceSize to 128M - much lower than the current upstream default of 1G, and still much lower than the possible encoding range with 4-byte-headers of 512M. I think I did this earlier to make some tests pass and to address a problem with CDS which also wants a smallish share of it, but I have in-fact only hidden a bug.
> 
> The way it should work is to set the CompressedClassSpaceSize to 512M and let CDS take its share and use the rest for class-space. This is infact implemented by [JDK-8332514](https://bugs.openjdk.org/browse/JDK-8332514) since a while. However, there is a bug there that prevents it from working with 4-byte-headers: it has 4G encoding range hard-coded.
> 
> This change fixes:
> - It reverts the default CCSS to 1G
> - It sets CCSS to 512M when running with COH
> - It makes the auto-sizing work by setting the correct encoding range
> - It fixes a test that tries to use CCSS*2. I changed it to use CCSS/2, I think this is still in the spirit of the test ('use a different CCSS and see if it still works')
> 
> Testing:
> - [x] tier1
> - [x] tier2
> - [x] gc/arguments/TestUseCompressedOopsErgoTools.java (affected by the change)
> - [x] runtime/cds/appcds/dynamicArchive/DynamicLotsOfClasses.java (failed before, fixed by this change)
> - [x] runtime/cds/appcds/LotsOfJRTClasses.java (failed before, fixed by this change)

src/hotspot/share/runtime/arguments.cpp line 3772:

> 3770:   }
> 3771:   if (UseCompactObjectHeaders && FLAG_IS_DEFAULT(CompressedClassSpaceSize)) {
> 3772:     FLAG_SET_DEFAULT(CompressedClassSpaceSize, 512 * M);

This may be a bit less ugly and hard-coded:

Suggestion:

    FLAG_SET_DEFAULT(CompressedClassSpaceSize, CompressedKlassPointers::max_klass_range_size_coh);


with this patch in addition:


diff --git a/src/hotspot/share/oops/compressedKlass.cpp b/src/hotspot/share/oops/compressedKlass.cpp
index d7c97d3c8d5..bf45f4ecd78 100644
--- a/src/hotspot/share/oops/compressedKlass.cpp
+++ b/src/hotspot/share/oops/compressedKlass.cpp
@@ -47,6 +47,7 @@ size_t CompressedKlassPointers::_protection_zone_size = 0;
 size_t CompressedKlassPointers::max_klass_range_size() {
 #ifdef _LP64
   const size_t encoding_allows = nth_bit(narrow_klass_pointer_bits() + max_shift());
+  assert(!UseCompactObjectHeaders || max_klass_range_size_coh == encoding_allows, "Sanity");
   constexpr size_t cap = 4 * G;
   return MIN2(encoding_allows, cap);
 #else
diff --git a/src/hotspot/share/oops/compressedKlass.hpp b/src/hotspot/share/oops/compressedKlass.hpp
index 64b9fcf9c82..4f94675d8ed 100644
--- a/src/hotspot/share/oops/compressedKlass.hpp
+++ b/src/hotspot/share/oops/compressedKlass.hpp
@@ -192,6 +192,10 @@ class CompressedKlassPointers : public AllStatic {
   // resulting from the current encoding settings (base, shift), capped to a certain max. value.
   static size_t max_klass_range_size();
 
+  // For use before pre-initialization
+  static constexpr size_t max_klass_range_size_coh =
+      nth_bit(narrow_klass_pointer_bits_coh + max_shift_coh);
+
   // On 64-bit, we need the class space to confine Klass structures to the encoding range, which is determined
   // by bit size of narrowKlass IDs and the shift. On 32-bit, we support compressed class pointer only
   // "pro-forma": narrowKlass have the same size as addresses (32 bits), and therefore the encoding range is

-------------

PR Review Comment: https://git.openjdk.org/lilliput/pull/203#discussion_r2381975977