[GR-52844] Add NI compiler configuration Os.

PullRequest: graal/17340

compiler/src/jdk.graal.compiler/src/jdk/graal/compiler/asm/amd64/AMD64Assembler.java

@@ -131,14 +131,14 @@ public AMD64Assembler(TargetDescription target) {
 
     public AMD64Assembler(TargetDescription target, OptionValues optionValues) {
         super(target);
-        useBranchesWithin32ByteBoundary = Options.UseBranchesWithin32ByteBoundary.getValue(optionValues);
+        useBranchesWithin32ByteBoundary = !GraalOptions.ReduceCodeSize.getValue(optionValues) && Options.UseBranchesWithin32ByteBoundary.getValue(optionValues);
         optimizeLongJumps = GraalOptions.OptimizeLongJumps.getValue(optionValues);
     }
 
     public AMD64Assembler(TargetDescription target, OptionValues optionValues, boolean hasIntelJccErratum) {
         super(target);
         if (Options.UseBranchesWithin32ByteBoundary.hasBeenSet(optionValues)) {
-            useBranchesWithin32ByteBoundary = Options.UseBranchesWithin32ByteBoundary.getValue(optionValues);
+            useBranchesWithin32ByteBoundary = !GraalOptions.ReduceCodeSize.getValue(optionValues) && Options.UseBranchesWithin32ByteBoundary.getValue(optionValues);
         } else {
             useBranchesWithin32ByteBoundary = hasIntelJccErratum;
         }

compiler/src/jdk.graal.compiler/src/jdk/graal/compiler/core/common/GraalOptions.java

@@ -38,6 +38,10 @@ public final class GraalOptions {
     @Option(help = "Uses compiler intrinsifications.", type = OptionType.Expert)
     public static final OptionKey<Boolean> Intrinsify = new OptionKey<>(true);
 
+
+    @Option(help = "", type = OptionType.Debug)
+    public static final OptionKey<Boolean> ReduceCodeSize = new OptionKey<>(true);
+
     @Option(help = "Rewrite signed comparisons to unsigned ones if the result is equal.", type = OptionType.Debug)
     public static final OptionKey<Boolean> PreferUnsignedComparison = new OptionKey<>(true);
 
@@ -317,4 +321,8 @@ public final class GraalOptions {
 
     @Option(help = "AMD64 only: Replace forward jumps (jmp, jcc) with equivalent but smaller instructions if the actual jump displacement fits in one byte.", type = OptionType.Expert)
     public static final OptionKey<Boolean> OptimizeLongJumps = new OptionKey<>(false);
+
+    @Option(help = "Optimize integer division operation by using various mathematical foundations to "
+                    + " express it in faster, equivalent, arithmetic.", type = OptionType.Expert)
+    public static final OptionKey<Boolean> OptimizeDiv = new OptionKey<>(true);
 }

compiler/src/jdk.graal.compiler/src/jdk/graal/compiler/core/phases/HighTier.java

@@ -87,7 +87,10 @@ public HighTier(OptionValues options) {
         }
 
         LoopPolicies loopPolicies = createLoopPolicies(options);
-        appendPhase(new LoopFullUnrollPhase(canonicalizer, loopPolicies));
+
+        if (GraalOptions.FullUnroll.getValue(options)) {
+            appendPhase(new LoopFullUnrollPhase(canonicalizer, loopPolicies));
+        }
 
         if (GraalOptions.LoopPeeling.getValue(options)) {
             appendPhase(new LoopPeelingPhase(loopPolicies, canonicalizer));

compiler/src/jdk.graal.compiler/src/jdk/graal/compiler/core/phases/MidTier.java

@@ -88,7 +88,10 @@ public MidTier(OptionValues options) {
             appendPhase(new VerifyHeapAtReturnPhase());
         }
 
-        appendPhase(new LoopFullUnrollPhase(canonicalizer, createLoopPolicies(options)));
+        if (GraalOptions.FullUnroll.getValue(options)) {
+            appendPhase(new LoopFullUnrollPhase(canonicalizer, createLoopPolicies(options)));
+        }
+
         appendPhase(new RemoveValueProxyPhase(canonicalizer));
 
         appendPhase(new LoopSafepointInsertionPhase());
@@ -99,7 +102,9 @@ public MidTier(OptionValues options) {
             appendPhase(new IterativeConditionalEliminationPhase(canonicalizer, false));
         }
 
-        appendPhase(new OptimizeDivPhase(canonicalizer));
+        if (GraalOptions.OptimizeDiv.getValue(options)) {
+            appendPhase(new OptimizeDivPhase(canonicalizer));
+        }
 
         appendPhase(new FrameStateAssignmentPhase());
 

compiler/src/jdk.graal.compiler/src/jdk/graal/compiler/nodes/PrefetchAllocateNode.java

@@ -29,6 +29,7 @@
 import static jdk.graal.compiler.nodeinfo.NodeCycles.CYCLES_2;
 import static jdk.graal.compiler.nodeinfo.NodeSize.SIZE_2;
 
+import jdk.graal.compiler.core.common.GraalOptions;
 import jdk.graal.compiler.core.common.type.StampFactory;
 import jdk.graal.compiler.graph.NodeClass;
 import jdk.graal.compiler.nodeinfo.NodeInfo;
@@ -50,7 +51,9 @@ public PrefetchAllocateNode(ValueNode address) {
 
     @Override
     public void generate(NodeLIRBuilderTool gen) {
-        gen.getLIRGeneratorTool().emitPrefetchAllocate(gen.operand(address));
+        if (!GraalOptions.ReduceCodeSize.getValue(this.getOptions())) {
+            gen.getLIRGeneratorTool().emitPrefetchAllocate(gen.operand(address));
+        }
     }
 
     @NodeIntrinsic

docs/reference-manual/native-image/BuildOptions.md

@@ -76,7 +76,7 @@ Run `native-image --help` for help on build options.
 * `--target`: select the compilation target for `native-image` (in the `<OS>-<architecture>` format). It defaults to host's OS-architecture pair.
 * `--trace-class-initialization`: provide a comma-separated list of fully-qualified class names that a class initialization is traced for
 * `--trace-object-instantiation`: provide a comma-separated list of fully-qualified class names that an object instantiation is traced for
-* `-O<level>`: control code optimizations where available variants are: `b` - optimize for fastest build time, `0` - no optimizations, `1` - basic optimizations, `2` - aggressive optimizations, `3` - all optimizations for best performance (enabled automatically with the Profile-guided optimization) 
+* `-O<level>`: control code optimizations where available variants are: `b` - optimize for fastest build time, `s` - optimize for size, `0` - no optimizations, `1` - basic optimizations, `2` - aggressive optimizations, `3` - all optimizations for best performance (enabled automatically with Profile-Guided Optimizations)
 * `-da`, `-da[:[packagename]|:[classname]`, `disableassertions[:[packagename]|:[classname]`: disable assertions with specified granularity at run time
 * `-dsa`, `-disablesystemassertions`: disable assertions in all system classes at run time
 * `-ea`, `-ea[:[packagename]|:[classname]`, `enableassertions[:[packagename]|:[classname]`: enable assertions with specified granularity at run time

docs/reference-manual/native-image/BuildOutput.md

@@ -91,7 +91,8 @@ Please report version and vendor when you [file issues](https://github.com/oracl
 The selected optimization level and targeted machine type used by the Graal compiler.
 The optimization level can be controlled with the `-O` option and defaults to `2`, which enables aggressive optimizations.
 Use `-Ob` to enable quick build mode, which speeds up the [compilation stage](#stage-compiling).
-This is useful during development, or when peak throughput is less important and you would like to optimize for size.
+This is useful during development to reduce image build time. 
+Use `-Os` to optimize for size.
 The targeted machine type can be selected with the `-march` option and defaults to `x86-64-v3` on AMD64 and `armv8-a` on AArch64.
 See [here](#recommendation-cpu) for recommendations on how to use this option.
 

docs/reference-manual/native-image/OptimizationsAndPerformance.md

@@ -18,6 +18,7 @@ The following table provides an overview of the different optimization levels an
 | Level | Optimizations | Use Cases |
 |:---:|:---:|---|
 | `-Ob` | Reduced | Quick build mode: Speeds up builds during development by avoiding time-consuming optimizations. This can also reduce file size sometimes. |
+| `-Os` | Reduced | Optimize for size: `-Os` enables all `-O2` optimizations except those that can increase code or image size significantly. Typically creates the smallest possible images at the cost of reduced performance. |
 | `-O0` | None | Typically used together with `-g` to improve the debugging experience. |
 | `-O1` | Basic | Trades performance for reduced file size and build time. Oracle GraalVM's `-O1` is somewhat comparable to `-O2` in GraalVM Community Edition. |
 | `-O2` | Advanced | **Default:** Aims for good performance at a reasonable file size. |

substratevm/CHANGELOG.md

@@ -24,6 +24,7 @@ This changelog summarizes major changes to GraalVM Native Image.
 * (GR-53359) Provide the `.debug_gdb_scripts` section that triggers auto-loading of `svmhelpers.py` in GDB. Remove single and double quotes from `ClassLoader.nameAndId` in the debuginfo.
 * (GR-47365) Include dynamic proxy metadata in the reflection metadata with the syntax `"type": { "proxy": [<interface list>] }`. This allows members of proxy classes to be accessed reflectively. `proxy-config.json` is now deprecated but will still be honored.
 * (GR-18214) In-place compacting garbage collection for the Serial GC old generation with `-H:+CompactingOldGen`.
+* (GR-52844) Add `Os` a new optimization mode to configure the optimizer in a way to get the smallest code size.
 
 ## GraalVM for JDK 22 (Internal Version 24.0.0)
 * (GR-48304) Red Hat added support for the JFR event ThreadAllocationStatistics.

substratevm/src/com.oracle.svm.core.graal.llvm/src/com/oracle/svm/core/graal/llvm/LLVMToolchainUtils.java

@@ -110,7 +110,7 @@ public static void llvmCompile(DebugContext debug, String outputPath, String inp
 
     private static int optimizationLevel() {
         return switch (SubstrateOptions.optimizationLevel()) {
-            case O0, BUILD_TIME -> 0;
+            case O0, BUILD_TIME, SIZE -> 0;
             case O1 -> 1;
             case O2 -> 2;
             case O3 -> 3;

substratevm/src/com.oracle.svm.core/src/com/oracle/svm/core/SubstrateOptions.java

@@ -71,6 +71,7 @@
 import jdk.graal.compiler.options.OptionStability;
 import jdk.graal.compiler.options.OptionType;
 import jdk.graal.compiler.options.OptionValues;
+import jdk.graal.compiler.phases.common.DeadCodeEliminationPhase;
 import jdk.internal.misc.Unsafe;
 
 public class SubstrateOptions {
@@ -233,7 +234,8 @@ public enum OptimizationLevel {
         O1("Basic optimizations", "1"),
         O2("Advanced optimizations", "2"),
         O3("All optimizations for best performance", "3"),
-        BUILD_TIME("Optimize for fastest build time", "b");
+        BUILD_TIME("Optimize for fastest build time", "b"),
+        SIZE("Optimize for size", "s");
 
         private final String description;
         private final String optionSwitch;
@@ -268,7 +270,7 @@ public boolean isOneOf(OptimizationLevel... levels) {
     }
 
     @APIOption(name = "-O", valueSeparator = APIOption.NO_SEPARATOR)//
-    @Option(help = "Control code optimizations: b - optimize for fastest build time, " +
+    @Option(help = "Control code optimizations: b - optimize for fastest build time, s - optimize for size, " +
                     "0 - no optimizations, 1 - basic optimizations, 2 - advanced optimizations, 3 - all optimizations for best performance.", type = OptionType.User)//
     public static final HostedOptionKey<String> Optimize = new HostedOptionKey<>("2") {
 
@@ -290,20 +292,77 @@ protected void onValueUpdate(EconomicMap<OptionKey<?>, Object> values, String ol
             SubstrateOptions.ReduceImplicitExceptionStackTraceInformation.update(values, newLevel == OptimizationLevel.O3);
 
             GraalOptions.OptimizeLongJumps.update(values, !newLevel.isOneOf(OptimizationLevel.O0, OptimizationLevel.BUILD_TIME));
+
+            if (newLevel == OptimizationLevel.SIZE) {
+                configureOs(values);
+            }
+
             if (optimizeValueUpdateHandler != null) {
                 optimizeValueUpdateHandler.onValueUpdate(values, newLevel);
             }
 
         }
     };
 
+    public static void configureOs(EconomicMap<OptionKey<?>, Object> values) {
+        enable(GraalOptions.ReduceCodeSize, values);
+        enable(ReduceImplicitExceptionStackTraceInformation, values);
+        enable(GraalOptions.OptimizeLongJumps, values);
+
+        /*
+         * Remove all loop optimizations that can increase code size, i.e., duplicate a loop body
+         * somehow.
+         */
+        disable(GraalOptions.LoopPeeling, values);
+        disable(GraalOptions.LoopUnswitch, values);
+        disable(GraalOptions.FullUnroll, values);
+        disable(GraalOptions.PartialUnroll, values);
+
+        /*
+         * Do not align loop headers to further reduce code size.
+         */
+        GraalOptions.LoopHeaderAlignment.update(values, 0);
+        GraalOptions.IsolatedLoopHeaderAlignment.update(values, 0);
+
+        /*
+         * Do not run PEA - it can fan out allocations too much.
+         */
+        disable(GraalOptions.PartialEscapeAnalysis, values);
+
+        /*
+         * Do not fan out division.
+         */
+        disable(GraalOptions.OptimizeDiv, values);
+
+        /*
+         * Do more conditional elimination.
+         */
+        GraalOptions.ConditionalEliminationMaxIterations.update(values, 10);
+
+        /*
+         * Every dead code elimination should be non-optional
+         */
+        disable(DeadCodeEliminationPhase.Options.ReduceDCE, values);
+    }
+
+    public static void disable(OptionKey<Boolean> key, EconomicMap<OptionKey<?>, Object> values) {
+        key.update(values, false);
+    }
+
+    public static void enable(OptionKey<Boolean> key, EconomicMap<OptionKey<?>, Object> values) {
+        key.update(values, true);
+    }
+
     /**
      * Only allows 'b' or positive numeric optimization levels, throws a user error otherwise.
      */
     private static OptimizationLevel parseOptimizationLevel(String value) {
         if (value.equals("b")) {
             return OptimizationLevel.BUILD_TIME;
         }
+        if (value.equals("s")) {
+            return OptimizationLevel.SIZE;
+        }
 
         int intLevel;
         try {
@@ -335,11 +394,20 @@ public static boolean useEconomyCompilerConfig(OptionValues options) {
         return "b".equals(Optimize.getValue(options));
     }
 
+    public static boolean useCodeSizeCompilerConfig(OptionValues options) {
+        return "s".equals(Optimize.getValue(options));
+    }
+
     @Fold
     public static boolean useEconomyCompilerConfig() {
         return useEconomyCompilerConfig(HostedOptionValues.singleton());
     }
 
+    @Fold
+    public static boolean useCodeSizeCompilerConfig() {
+        return useCodeSizeCompilerConfig(HostedOptionValues.singleton());
+    }
+
     @Fold
     public static boolean isMaximumOptimizationLevel() {
         return optimizationLevel() == OptimizationLevel.O3;

substratevm/src/com.oracle.svm.hosted/src/com/oracle/svm/hosted/NativeImageGenerator.java

@@ -1581,17 +1581,21 @@ private static boolean checkInvocationPluginMethods(SubstrateReplacements replac
         return true;
     }
 
-    public static Suites createSuites(FeatureHandler featureHandler, RuntimeConfiguration runtimeConfig, boolean hosted) {
+    public static Suites createSuites(FeatureHandler featureHandler, RuntimeConfiguration runtimeConfig, boolean hosted, OptionValues optionsToUse) {
         SubstrateBackend backend = runtimeConfig.getBackendForNormalMethod();
         Suites suites;
         if (hosted) {
-            suites = GraalConfiguration.hostedInstance().createSuites(HostedOptionValues.singleton(), hosted, ConfigurationValues.getTarget().arch);
+            suites = GraalConfiguration.hostedInstance().createSuites(optionsToUse == null ? HostedOptionValues.singleton() : optionsToUse, hosted, ConfigurationValues.getTarget().arch);
         } else {
-            suites = GraalConfiguration.runtimeInstance().createSuites(RuntimeOptionValues.singleton(), hosted, ConfigurationValues.getTarget().arch);
+            suites = GraalConfiguration.runtimeInstance().createSuites(optionsToUse == null ? RuntimeOptionValues.singleton() : optionsToUse, hosted, ConfigurationValues.getTarget().arch);
         }
         return modifySuites(backend, suites, featureHandler, hosted, false);
     }
 
+    public static Suites createSuites(FeatureHandler featureHandler, RuntimeConfiguration runtimeConfig, boolean hosted) {
+        return createSuites(featureHandler, runtimeConfig, hosted, null);
+    }
+
     public static Suites createFirstTierSuites(FeatureHandler featureHandler, RuntimeConfiguration runtimeConfig, boolean hosted) {
         SubstrateBackend backend = runtimeConfig.getBackendForNormalMethod();
         Suites suites;

substratevm/src/com.oracle.svm.hosted/src/com/oracle/svm/hosted/image/NativeImageViaCC.java

@@ -47,6 +47,7 @@
 import com.oracle.svm.core.SubstrateOptions;
 import com.oracle.svm.core.SubstrateUtil;
 import com.oracle.svm.core.util.InterruptImageBuilding;
+import com.oracle.svm.hosted.DeadlockWatchdog;
 import com.oracle.svm.hosted.FeatureImpl.BeforeImageWriteAccessImpl;
 import com.oracle.svm.hosted.NativeImageOptions;
 import com.oracle.svm.hosted.c.NativeLibraries;
@@ -138,6 +139,11 @@ private void runLinkerCommand(String imageName, LinkerInvocation inv, List<Strin
 
             List<String> lines;
             try (InputStream inputStream = linkerProcess.getInputStream()) {
+                /*
+                 * The linker can be slow, record activity just before so that we have the full
+                 * watchdog interval available.
+                 */
+                DeadlockWatchdog.singleton().recordActivity();
                 lines = FileUtils.readAllLines(inputStream);
                 FileUtils.traceCommandOutput(lines);
             }

vm/mx.vm/mx_vm_benchmark.py

@@ -552,7 +552,7 @@ def _configure_from_name(self, config_name):
         # This defines the allowed config names for NativeImageVM. The ones registered will be available via --jvm-config
         rule = r'^(?P<native_architecture>native-architecture-)?(?P<string_inlining>string-inlining-)?(?P<gate>gate-)?(?P<upx>upx-)?(?P<quickbuild>quickbuild-)?(?P<gc>g1gc-)?(?P<llvm>llvm-)?(?P<pgo>pgo-|pgo-ctx-insens-|pgo-sampler-)?(?P<inliner>inline-)?' \
                r'(?P<analysis_context_sensitivity>insens-|allocsens-|1obj-|2obj1h-|3obj2h-|4obj3h-)?(?P<no_inlining_before_analysis>no-inline-)?(?P<jdk_profiles>jdk-profiles-collect-|adopted-jdk-pgo-)?' \
-               r'(?P<profile_inference>profile-inference-feature-extraction-|profile-inference-pgo-)?(?P<sampler>safepoint-sampler-|async-sampler-)?(?P<optimization_level>O0-|O1-|O2-|O3-)?(?P<edition>ce-|ee-)?$'
+               r'(?P<profile_inference>profile-inference-feature-extraction-|profile-inference-pgo-)?(?P<sampler>safepoint-sampler-|async-sampler-)?(?P<optimization_level>O0-|O1-|O2-|O3-|Os-)?(?P<edition>ce-|ee-)?$'
 
         mx.logv(f"== Registering configuration: {config_name}")
         match_name = f"{config_name}-"  # adding trailing dash to simplify the regex
@@ -679,7 +679,7 @@ def generate_profiling_package_prefixes():
         if matching.group("optimization_level") is not None:
             olevel = matching.group("optimization_level")[:-1]
             mx.logv(f"GraalVM optimization level is set to: {olevel}")
-            if olevel in ["O0", "O1", "O2", "O3"]:
+            if olevel in ["O0", "O1", "O2", "O3", "Os"]:
                 self.optimization_level = olevel
             else:
                 mx.abort(f"Unknown configuration for optimization level: {olevel}")
@@ -1801,7 +1801,7 @@ def register_graalvm_vms():
             mx_polybenchmarks_benchmark.polybenchmark_vm_registry.add_vm(PolyBenchVm(host_vm_name, "native", [], ["--native"]))
             mx_polybenchmarks_benchmark.rules = polybenchmark_rules
 
-    optimization_levels = ['O0', 'O1', 'O2', 'O3']
+    optimization_levels = ['O0', 'O1', 'O2', 'O3', 'Os']
 
     # Inlining before analysis is done by default
     analysis_context_sensitivity = ['insens', 'allocsens', '1obj', '2obj1h', '3obj2h', '4obj3h']