On Fri, 26 Apr 2024 21:16:03 GMT, Vladimir Kozlov <[email protected]> wrote:
> Move immutable nmethod's data from CodeCache to C heap. It includes
> `dependencies, nul_chk_table, handler_table, scopes_pcs, scopes_data,
> speculations, jvmci_data`. It amounts for about 30% (optimized VM) of space
> in CodeCache.
>
> Use HotSpot's `os::malloc()` to allocate memory in C heap for immutable
> nmethod's data. Bail out compilation if allocation failed.
>
> Shuffle fields order and change some fields size from 4 to 2 bytes to avoid
> nmethod's header size increase.
>
> Tested tier1-5, stress,xcomp
src/hotspot/share/code/nmethod.cpp line 117:
> 115: result = static_cast<T>(thing); \
> 116: assert(static_cast<int>(result) == thing, "failed: %d != %d",
> static_cast<int>(result), thing);
> 117:
I replaced `checked_cast<>()` with this macro because of next issues:
- The existing assert points to `utilities/checkedCast.hpp` file where this
method is located and not where failed cast. It does not help when it is used
several times in one method (for example, in `nmethod()` constructors).
- The existing assert does not print values
src/hotspot/share/code/nmethod.cpp line 1324:
> 1322:
> 1323: // native wrapper does not have read-only data but we need unique
> not null address
> 1324: _immutable_data = data_end();
I can't use nullptr because VM expects not null address when it checks, for
example, `dependencies_begin()` even so sizes are 0.
I used `data_end()` instead of nullptr in other places too.
src/hotspot/share/code/nmethod.hpp line 583:
> 581: int dependencies_size () const { return int(
> dependencies_end () - dependencies_begin ()); }
> 582: int handler_table_size () const { return int(
> handler_table_end() - handler_table_begin()); }
> 583: int nul_chk_table_size () const { return int(
> nul_chk_table_end() - nul_chk_table_begin()); }
Shift by one space to aline code.
test/hotspot/jtreg/compiler/c1/TestLinearScanOrderMain.java line 29:
> 27: * @compile TestLinearScanOrder.jasm
> 28: * @run main/othervm -Xcomp -XX:+TieredCompilation -XX:TieredStopAtLevel=1
> 29: * -XX:+IgnoreUnrecognizedVMOptions
> -XX:NMethodSizeLimit=655360
This test caught one `check_cast<>` issue during development but only on
aarch64.
On x64 the test bailed out compilation before that because default
`NMethodSizeLimit` was not big enough ((64*K)*wordSize = 524288).
-------------
PR Review Comment: https://git.openjdk.org/jdk/pull/18984#discussion_r1581561933
PR Review Comment: https://git.openjdk.org/jdk/pull/18984#discussion_r1581565762
PR Review Comment: https://git.openjdk.org/jdk/pull/18984#discussion_r1581558637
PR Review Comment: https://git.openjdk.org/jdk/pull/18984#discussion_r1581557756
