Besides escaping points, C2 Partial Escape Analysis(PEA) can optionally materialize the object at the convergence block when any of its predecessors has already materialized it. It is referred to as passive materialization. It is optional because C2 currently skips it and programs are still correct. Here is an example. Object pt is escaped at line 6.
1 public double test(boolean cond, double x, double y) {
2 Point pt = new Point(x, y);
3 pt.x = 3.0d;
C2 is a speculative compiler. When it encounters a conditional branch and profiling data suggest that the control is highly likely to take one path, parser intends to prune the other path and leave a runtime stub uncommon_trap. We refer to it as “unstable_if trap” because the reason of this stub is unstable_if. If the rare case does happen, the trap triggers deoptimization. HotSpot realizes that current profiling data do not hold up and destroys this compilation. The execution returns to interpreter. Presumably, the upcoming compilation of this method will correct itself with updated profiling data.
Bci(byte code index) in JVM is analogous to the PC(program counter) of a physical processor. It points to the current bytecode in interpreter. Because an uncommon_trap stub punts to interpreter, it must restore the interpreter state in JVM. Restoring the interpreter state means that all variables live at the current bci must have cor
| // java -XX:CompileCommand=quiet -XX:CompileCommand=compileonly,RelockInDeopt::snippet2 -XX:-TieredCompilation -XX:-UseOnStackReplacement -XX:+PrintCompilation -XX:+TraceDeoptimization -XX:+PrintEliminateLocks -XX:+PrintDeoptimizationDetails -XX:-PrintInlining -Xbatch -XX:+PrintEliminateAllocations RelockInDeopt | |
| class RelockInDeopt { | |
| static int staticInt = 0; | |
| static Object staticObj = null; | |
| public static void snippet2() { | |
| Object escaped = new Object(); | |
| Object nonEscaped = new Object(); | |
| synchronized (nonEscaped) { |
➜ WebKit git:(main) ./WebKitBuild/Debug/jsc
dyld[8692]: Symbol not found: __ZN3JSC27retrieveTypeImportAttributeEPNS_14JSGlobalObjectERKN3WTF7HashMapINS2_6RefPtrINS2_17UniquedStringImplENS2_12RawPtrTraitsIS5_EENS2_21DefaultRefDerefTraitsIS5_EEEENS2_6StringENS2_11DefaultHashISA_EENS2_10HashTraitsISA_EENSE_ISB_EENS2_15HashTableTraitsEEE
Referenced from: <BD3D10DA-F74B-3755-8B54-29D1F10163E5> /Users/xxinliu/Devel/WebKit/WebKitBuild/Debug/jsc (built for macOS 14.0 which is newer than running OS)
Expected in: <957522FA-9B44-3C8F-9BD4-A209C728B133> /System/Library/Frameworks/JavaScriptCore.framework/Versions/A/JavaScriptCore
[1] 8692 abort ./WebKitBuild/Debug/jsc
The problem here is JSC depends on WTF(who would name its library that? it seems the acronym of webkit's template framework). Dynamic loader has difficulty to resolve this symbol in WTF. we need to teach dynamic laoder to get from what you built
| ; my handy scripts on MacOS | |
| function realpath() { | |
| [[ $1 = /* ]] && echo "$1" || echo "$PWD/${1#./}" | |
| } | |
| function jdk_select() { | |
| if [ -e $1/bin/java ]; then | |
| P=$(realpath $1) | |
| export JAVA_HOME=$P | |
| export PATH=$JAVA_HOME/bin:$PATH |
| // java -XX:CompileOnly='UnderProfiledSubprocedure.foo' -XX:+PrintInlining -XX:+PrintCompilation -XX:CompileCommand=quiet -Xbatch UnderProfiledSubprocedure | |
| import java.util.ArrayList; | |
| class UnderProfiledSubprocedure { | |
| private static int ODD = 100; | |
| public void foo(boolean cond) { | |
| var x = new ArrayList<Integer>(); | |
| if (cond) { // the branch is only taken by ODD | |
| x.add(0); // ArrayList::add(E) is the subprocedure. it will call ArrayList::add(E, Object[], int) |
| #!/usr/bin/env ruby | |
| lines = ARGF.readlines | |
| total = 0 | |
| lines.each do |line| | |
| if line =~ /(\d+)\smessages dropped due to async logging/ | |
| num = $1.to_i | |
| total = total + num | |
| #print line | |
| end | |
| end |
// ./build/linux-x86_64-server-fastdebug/jdk/bin/java -XX:+PrintCompilation -XX:CompileCommand=compileonly,EmptyString::main -XX:CompileCommand=quiet -Xbatch -XX:+PrintOptoAssembly -XX:CompileCommand=IGVPrintLevel,EmptyString::main,3 EmptyString
class EmptyString {
public static void main(String[] args) {
String msg = "";
int x = 0;
for (int i = 0; i < 300_000; ++i) {
if (msg.length() > 0) {
x++;
| module { | |
| func.func @local_sparse_attention_kernel_03_bf(%arg0: i32, %arg1: i32, %arg2: i32, %arg3: i32, %arg4: i32, %arg5: i32, %arg6: i32, %arg7: !tt.ptr<f16>, %arg8: !tt.ptr<f16>, %arg9: !tt.ptr<f16>, %arg10: !tt.ptr<f16>) { | |
| %c16_i32 = arith.constant 16 : i32 | |
| %cst = arith.constant 0.000000e+00 : f32 | |
| %cst_1 = arith.constant 0xFF800000 : f32 | |
| %c0_i32 = arith.constant 0 : i32 | |
| %c1_i32 = arith.constant 1 : i32 | |
| %alloc_5 = memref.alloc() {alignment = 64 : i64} : memref<1x128xf16> | |
| %alloc_7 = memref.alloc() {alignment = 64 : i64} : memref<1x128xf32> |