Remove all #[inline] attributes (#196)

This commit adds a section on making builds with profile-guided
optimization to the README and also adds an example script that makes a
PGO build to the `.github/scripts` directory.

Author: k-sareen

.github/scripts/pgo-build.sh

@@ -0,0 +1,16 @@
+#!/bin/bash
+
+# Compile with profiling support
+RUSTFLAGS="-Cprofile-generate=/tmp/$USER/pgo-data" make CONF=linux-x86_64-normal-server-release THIRD_PARTY_HEAP=$PWD/../mmtk-openjdk/openjdk images
+
+# Remove extraneous profiling data
+rm -rf /tmp/$USER/pgo-data/*
+
+# Profile using fop
+MMTK_PLAN=GenImmix MMTK_STRESS_FACTOR=4194304 MMTK_PRECISE_STRESS=false ./build/linux-x86_64-normal-server-release/images/jdk/bin/java -XX:MetaspaceSize=500M -XX:+DisableExplicitGC -XX:-TieredCompilation -Xcomp -XX:+UseThirdPartyHeap -Xms60M -Xmx60M -jar /usr/share/benchmarks/dacapo/dacapo-evaluation-git-6e411f33.jar -n 5 fop
+
+# Merge profiling data
+/opt/rust/toolchains/1.66.1-x86_64-unknown-linux-gnu/lib/rustlib/x86_64-unknown-linux-gnu/bin/llvm-profdata merge -o /tmp/$USER/pgo-data/merged.profdata /tmp/$USER/pgo-data
+
+# Compile using profiling data
+RUSTFLAGS="-Cprofile-use=/tmp/$USER/pgo-data/merged.profdata -Cllvm-args=-pgo-warn-missing-function" make CONF=linux-x86_64-normal-server-release THIRD_PARTY_HEAP=$PWD/../mmtk-openjdk/openjdk images

README.md

@@ -126,6 +126,67 @@ $ make CONF=linux-x86_64-normal-server-release THIRD_PARTY_HEAP=$PWD/../mmtk-ope
 
 The output jdk is then found at `./build/linux-x86_64-normal-server-release/images/jdk`.
 
+### Profile-Guided Optimized Build
+
+In order to get the best performance, we recommend using a profile-guided
+optimized (PGO) build. Rust supports [PGO
+builds](https://doc.rust-lang.org/rustc/profile-guided-optimization.html) by
+directly hooking into the LLVM profiling infrastructure. In order to have the
+correct LLVM tools version, you should install the relevant `llvm-tools-preview`
+component using `rustup`:
+
+```console
+$ rustup component add llvm-tools-preview
+```
+
+In this example, we focus on the DaCapo benchmarks and the `GenImmix`
+collector. For best results, it is recommended to profile the workload you are
+interested in measuring. We use `fop` as it is a relatively small benchmark but
+also exercises the GC. In order to best tune our GC performance, we use a
+stress factor of 4 MB in order to trigger more GC events.
+
+First we compile MMTk with profiling support:
+
+```console
+$ RUSTFLAGS="-Cprofile-generate=/tmp/$USER/pgo-data" make CONF=linux-x86_64-normal-server-release THIRD_PARTY_HEAP=$PWD/../mmtk-openjdk/openjdk images
+$ rm -rf /tmp/$USER/pgo-data/*
+```
+We clear the `/tmp/$USER/pgo-data` directory as during compilation, the JVM we
+have created is used in a bootstrap process, resulting in profile data being
+emitted.
+
+We then run `fop` in order to get some profiling data. Note that your location
+for the DaCapo benchmarks may be different:
+
+```bash
+MMTK_PLAN=GenImmix MMTK_STRESS_FACTOR=4194304 MMTK_PRECISE_STRESS=false ./build/linux-x86_64-normal-server-release/images/jdk/bin/java -XX:MetaspaceSize=500M -XX:+DisableExplicitGC -XX:-TieredCompilation -Xcomp -XX:+UseThirdPartyHeap -Xms60M -Xmx60M -jar /usr/share/benchmarks/dacapo/dacapo-evaluation-git-6e411f33.jar -n 5 fop
+```
+
+We have to merge the profiling data into something we can feed into the Rust
+compiler using `llvm-profdata`:
+
+```console
+$ /opt/rust/toolchains/1.66.1-x86_64-unknown-linux-gnu/lib/rustlib/x86_64-unknown-linux-gnu/bin/llvm-profdata merge -o /tmp/$USER/pgo-data/merged.profdata /tmp/$USER/pgo-data
+```
+
+The location of your version of `llvm-profdata` may be different to what we
+have above. *Make sure to only use a version of `llvm-profdata` that matches
+your Rust version.*
+
+Finally, we build a new image using the profiling data as an input:
+
+```console
+$ RUSTFLAGS="-Cprofile-use=/tmp/$USER/pgo-data/merged.profdata -Cllvm-args=-pgo-warn-missing-function" make CONF=linux-x86_64-normal-server-release THIRD_PARTY_HEAP=$PWD/../mmtk-openjdk/openjdk images
+```
+
+We now have an OpenJDK build under
+`./build/linux-x86_64-normal-server-release/images/jdk` with MMTk that has been
+optimized using PGO.
+
+For ease of use, we have provided an example script which does the above in
+`.github/scripts/pgo-build.sh` that you may adapt for your purposes. Note that
+you may have to change the location of `llvm-profdata`.
+
 ### Location of Mark-bit
 The location of the mark-bit can be specified by the environment variable
 `MARK_IN_HEADER`. By default, the mark-bit is located on the side (in a side

mmtk/src/abi.rs

@@ -83,17 +83,14 @@ impl Klass {
         &*(self as *const _ as usize as *const T)
     }
     /// Force slow-path for instance size calculation?
-    #[inline(always)]
     const fn layout_helper_needs_slow_path(lh: i32) -> bool {
         (lh & Self::LH_INSTANCE_SLOW_PATH_BIT) != 0
     }
     /// Get log2 array element size
-    #[inline(always)]
     const fn layout_helper_log2_element_size(lh: i32) -> i32 {
         (lh >> Self::LH_LOG2_ELEMENT_SIZE_SHIFT) & Self::LH_LOG2_ELEMENT_SIZE_MASK
     }
     /// Get array header size
-    #[inline(always)]
     const fn layout_helper_header_size(lh: i32) -> i32 {
         (lh >> Self::LH_HEADER_SIZE_SHIFT) & Self::LH_HEADER_SIZE_MASK
     }
@@ -281,7 +278,6 @@ pub struct OopDesc {
 }
 
 impl OopDesc {
-    #[inline(always)]
     pub fn start(&self) -> Address {
         unsafe { mem::transmute(self) }
     }
@@ -300,15 +296,13 @@ pub type Oop = &'static OopDesc;
 
 /// Convert ObjectReference to Oop
 impl From<ObjectReference> for &OopDesc {
-    #[inline(always)]
     fn from(o: ObjectReference) -> Self {
         unsafe { mem::transmute(o) }
     }
 }
 
 /// Convert Oop to ObjectReference
 impl From<&OopDesc> for ObjectReference {
-    #[inline(always)]
     fn from(o: &OopDesc) -> Self {
         unsafe { mem::transmute(o) }
     }
@@ -324,13 +318,11 @@ impl OopDesc {
     }
 
     /// Slow-path for calculating object instance size
-    #[inline(always)]
     unsafe fn size_slow(&self) -> usize {
         ((*UPCALLS).get_object_size)(self.into())
     }
 
     /// Calculate object instance size
-    #[inline(always)]
     pub unsafe fn size(&self) -> usize {
         let klass = self.klass;
         let lh = klass.layout_helper;

mmtk/src/object_model.rs

@@ -21,7 +21,6 @@ impl ObjectModel<OpenJDK> for VMObjectModel {
     const UNIFIED_OBJECT_REFERENCE_ADDRESS: bool = true;
     const OBJECT_REF_OFFSET_LOWER_BOUND: isize = 0;
 
-    #[inline]
     fn copy(
         from: ObjectReference,
         copy: CopySemantics,
@@ -81,22 +80,18 @@ impl ObjectModel<OpenJDK> for VMObjectModel {
         unimplemented!()
     }
 
-    #[inline(always)]
     fn ref_to_object_start(object: ObjectReference) -> Address {
         object.to_raw_address()
     }
 
-    #[inline(always)]
     fn ref_to_address(object: ObjectReference) -> Address {
         object.to_raw_address()
     }
 
-    #[inline(always)]
     fn ref_to_header(object: ObjectReference) -> Address {
         object.to_raw_address()
     }
 
-    #[inline(always)]
     fn address_to_ref(address: Address) -> ObjectReference {
         ObjectReference::from_raw_address(address)
     }

mmtk/src/object_scanning.rs

@@ -13,7 +13,6 @@ trait OopIterate: Sized {
 }
 
 impl OopIterate for OopMapBlock {
-    #[inline]
     fn oop_iterate(&self, oop: Oop, closure: &mut impl EdgeVisitor<OpenJDKEdge>) {
         let start = oop.get_field_address(self.offset);
         for i in 0..self.count as usize {
@@ -24,7 +23,6 @@ impl OopIterate for OopMapBlock {
 }
 
 impl OopIterate for InstanceKlass {
-    #[inline]
     fn oop_iterate(&self, oop: Oop, closure: &mut impl EdgeVisitor<OpenJDKEdge>) {
         let oop_maps = self.nonstatic_oop_maps();
         for map in oop_maps {
@@ -34,7 +32,6 @@ impl OopIterate for InstanceKlass {
 }
 
 impl OopIterate for InstanceMirrorKlass {
-    #[inline]
     fn oop_iterate(&self, oop: Oop, closure: &mut impl EdgeVisitor<OpenJDKEdge>) {
         self.instance_klass.oop_iterate(oop, closure);
         // if (Devirtualizer::do_metadata(closure)) {
@@ -75,7 +72,6 @@ impl OopIterate for InstanceMirrorKlass {
 }
 
 impl OopIterate for InstanceClassLoaderKlass {
-    #[inline]
     fn oop_iterate(&self, oop: Oop, closure: &mut impl EdgeVisitor<OpenJDKEdge>) {
         self.instance_klass.oop_iterate(oop, closure);
         // if (Devirtualizer::do_metadata(closure)) {
@@ -89,7 +85,6 @@ impl OopIterate for InstanceClassLoaderKlass {
 }
 
 impl OopIterate for ObjArrayKlass {
-    #[inline]
     fn oop_iterate(&self, oop: Oop, closure: &mut impl EdgeVisitor<OpenJDKEdge>) {
         let array = unsafe { oop.as_array_oop() };
         for oop in unsafe { array.data::<Oop>(BasicType::T_OBJECT) } {
@@ -99,15 +94,13 @@ impl OopIterate for ObjArrayKlass {
 }
 
 impl OopIterate for TypeArrayKlass {
-    #[inline]
     fn oop_iterate(&self, _oop: Oop, _closure: &mut impl EdgeVisitor<OpenJDKEdge>) {
         // Performance tweak: We skip processing the klass pointer since all
         // TypeArrayKlasses are guaranteed processed via the null class loader.
     }
 }
 
 impl OopIterate for InstanceRefKlass {
-    #[inline]
     fn oop_iterate(&self, oop: Oop, closure: &mut impl EdgeVisitor<OpenJDKEdge>) {
         use crate::abi::*;
         use crate::api::{add_phantom_candidate, add_soft_candidate, add_weak_candidate};
@@ -135,11 +128,9 @@ impl OopIterate for InstanceRefKlass {
 }
 
 impl InstanceRefKlass {
-    #[inline]
     fn should_scan_weak_refs() -> bool {
         !*SINGLETON.get_options().no_reference_types
     }
-    #[inline]
     fn process_ref_as_strong(oop: Oop, closure: &mut impl EdgeVisitor<OpenJDKEdge>) {
         let referent_addr = Self::referent_address(oop);
         closure.visit_edge(referent_addr);
@@ -155,7 +146,6 @@ fn oop_iterate_slow(oop: Oop, closure: &mut impl EdgeVisitor<OpenJDKEdge>, tls:
     }
 }
 
-#[inline]
 fn oop_iterate(oop: Oop, closure: &mut impl EdgeVisitor<OpenJDKEdge>) {
     let klass_id = oop.klass.id;
     debug_assert!(
@@ -192,7 +182,6 @@ fn oop_iterate(oop: Oop, closure: &mut impl EdgeVisitor<OpenJDKEdge>) {
     }
 }
 
-#[inline]
 pub fn scan_object(
     object: ObjectReference,
     closure: &mut impl EdgeVisitor<OpenJDKEdge>,

openjdk/CompileThirdPartyHeap.gmk

@@ -10,6 +10,11 @@ MMTK_CPP_ROOT = $(THIRD_PARTY_HEAP)
 OPENJDK_VERSION=`cd $(MMTK_RUST_ROOT) ; cargo read-manifest --manifest-path=Cargo.toml | python3 -c 'import json,sys; print(json.load(sys.stdin)["metadata"]["openjdk"]["openjdk_version"])'`
 OPENJDK_LOCAL_VERSION=`git rev-parse HEAD`
 
+# Get the current host triple. This is used for the PGO build as mentioned in
+# the best practices for PGO here:
+# https://doc.rust-lang.org/rustc/profile-guided-optimization.html#a-complete-cargo-workflow
+HOST_TRIPLE=`rustc -vV | grep host | cut -d' ' -f2`
+
 ifdef MMTK_PLAN
   GC_FEATURES=--features $(MMTK_PLAN)
 endif
@@ -61,9 +66,10 @@ $(LIB_MMTK): FORCE
 		echo -e $(YELLOW)Local OpenJDK version $(OPENJDK_LOCAL_VERSION)$(NC); \
 		echo -e $(YELLOW)mmtk/Cargo.toml OpenJDK version $(OPENJDK_VERSION)$(NC); \
 	fi
-	echo "cd $(MMTK_RUST_ROOT) && cargo build $(CARGO_PROFILE_FLAG) $(GC_FEATURES)"
-	cd $(MMTK_RUST_ROOT) && cargo build $(CARGO_PROFILE_FLAG) $(GC_FEATURES)
-	cp $(MMTK_RUST_ROOT)/target/$(CARGO_PROFILE)/libmmtk_openjdk.so $(LIB_MMTK)
+	cargo --version
+	echo "cd $(MMTK_RUST_ROOT) && cargo build $(CARGO_PROFILE_FLAG) --target $(HOST_TRIPLE) $(GC_FEATURES)"
+	cd $(MMTK_RUST_ROOT) && cargo build $(CARGO_PROFILE_FLAG) --target $(HOST_TRIPLE) $(GC_FEATURES)
+	cp $(MMTK_RUST_ROOT)/target/$(HOST_TRIPLE)/$(CARGO_PROFILE)/libmmtk_openjdk.so $(LIB_MMTK)
 
 JVM_LIBS += -L$(JVM_LIB_OUTPUTDIR) -lmmtk_openjdk
 JVM_LDFLAGS += '-Wl,-rpath,$$ORIGIN'