Support LLVM raw profile versions 8, 9, and 10 #3
Conversation
Signed-off-by: Saman Dehghan <samaan.dehghan@gmail.com>
| #endif | ||
| .names_size = END_NAMES - START_NAMES, | ||
| .counters_delta = (uintptr_t)START_COUNTERS, | ||
| .names_delta = (uintptr_t)START_NAMES, |
There was a problem hiding this comment.
Why (uintptr_t) is removed? On my side this does not compile with either GCC or Clang
There was a problem hiding this comment.
Let's focus on #2 , I just put this here as a record for a version before adding clang-11 to clang-13 support. ignore this PR.
|
|
My steps of x86 testing, starting with a fresh Debian bookworm VM
## Dependencyies
sudo apt install build-essential libncurses-dev flex bison \
kconfig-frontends-nox python3-dev acpica-tools uuid-dev pkgconf libglib2.0-dev \
libpixman-1-dev libyajl-dev python3-venv ninja-build \
lsb-release wget software-properties-common gnupg
wget https://apt.llvm.org/llvm.sh -O /tmp/llvm.sh
chmod +x /tmp/llvm.sh
sudo /tmp/llvm.sh 20
rm /tmp/llvm.sh
## Build Xen
# git clone git@github.com:enum-class/xen.git --branch V3 --depth=10 xen-source
git clone https://github.com/whentojump/xen.git --branch V3-wentao --depth=10 xen-source
cd xen-source
LLVM_DIR=/lib/llvm-20/bin
# ❓ I don't actually know whether we need them all or just CC
MAKE_FLAGS=(
"CC=$LLVM_DIR/clang"
# "LD=$LLVM_DIR/ld.lld"
# "AR=$LLVM_DIR/llvm-ar"
# "NM=$LLVM_DIR/llvm-nm"
# "STRIP=$LLVM_DIR/llvm-strip"
# "OBJCOPY=$LLVM_DIR/llvm-objcopy"
# "OBJDUMP=$LLVM_DIR/llvm-objdump"
# "READELF=$LLVM_DIR/llvm-readelf"
# "HOSTCC=$LLVM_DIR/clang"
# "HOSTCXX=$LLVM_DIR/clang++"
# "HOSTAR=$LLVM_DIR/llvm-ar"
# "HOSTLD=$LLVM_DIR/ld.lld"
)
# "--libdir=/lib/x86_64-linux-gnu/" makes life easier rather than the
# default /usr/local/lib and LD_LIBRARY_PATH hassle
./configure --disable-xen --disable-docs --libdir=/lib/x86_64-linux-gnu/
# Build "tools" only with GNU
# ❓ This step builds a xen anyway under "tools/firmware/xen-dir/xen-root/xen". What is it?
make -j$(nproc) dist-tools
sudo make -j$(nproc) install-tools
# Build "xen" only with Clang/LLVM
make ${MAKE_FLAGS[@]} -C xen defconfig
pushd xen
kconfig-tweak -d LIVEPATCH
kconfig-tweak -e COVERAGE
popd
make ${MAKE_FLAGS[@]} -C xen olddefconfig
make ${MAKE_FLAGS[@]} -j$(nproc) efi-y= dist-xen
sudo make ${MAKE_FLAGS[@]} -j$(nproc) efi-y= install-xen
## Prepare for the next boot
# I don't have easy access to my VM's GRUB screen so this
# is the only way for me to select Xen as boot target.
# It's somewhat good as it's noninteractive.
# But proceed with caution and carefully consider what string
# should be put here in your setup. E.g., run
# "sudo grep menuentry /boot/grub/grub.cfg | less"
#
# sudo sed -i 's|GRUB_DEFAULT=0|GRUB_DEFAULT="Debian GNU/Linux, with Xen hypervisor"|' /etc/default/grub
#
sudo update-grub
sudo reboot
## After reboot
# ❓ I don't know why only I need this (not you or Erkai :))). Otherwise I have
# "xencall: error: Could not obtain handle on privileged command interface: No such file or directory
# xencov: opening xc interface: No such file or directory"
sudo systemctl start xencommons.service
sudo systemctl enable xencommons.service
sudo xencov read | tee >/dev/null default.profraw
file default.profraw
/lib/llvm-20/bin/llvm-profdata merge default.profraw -o default.profdata
file default.profdata
# Replace it with your path
XEN_SOURCE_DIR=/vagrant/xen-source/
TMP_REPORT_DIR=`mktemp -d`
/lib/llvm-20/bin/llvm-cov show \
-instr-profile default.profdata \
-output-dir $TMP_REPORT_DIR \
-show-directory-coverage \
-show-branches=count \
-use-color=false \
$XEN_SOURCE_DIR/xen/xen-syms
less $TMP_REPORT_DIR/index.txt
less $TMP_REPORT_DIR/coverage/$XEN_SOURCE_DIR/xen/common/coverage/llvm.c.txt
rm -r $TMP_REPORT_DIR |
|
One more note on linking, related to some of my previous comments/questions: If LD is GNU linker (both your and my current way), there is a sensible delay around On the other hand, if I additionally use LLVM linker as well MAKE_FLAGS=(
"CC=$LLVM_DIR/clang"
"LD=$LLVM_DIR/ld.lld"
)
|
|
Allow me to cc @erkaii as well to see if he has some insights. But no obligation. Thanks! |
I also see this warnings |
I did not test without round-up on x86. I will test and let you know the result 👍 |
I didn't use llvm linker. My cpu is core-i5 4300U. CPUs are not comparable. And since linking is single-threaded most probably that's the reason of difference in time. I will try cloudlab and let you know the result |
I see. It's possible that warning is specific to Arm, which I never ran. I can take a brief look |
Does it mean our code has some issue that with LLVM linker it doesn't work ? |
In my own test, it looks like so. Please do verify.
I personally don't want to do that... I believe a lot of alignment in our code is assuming 64-bit. Let's focus on x86_64 and aarch64. For more, let's wait for opinions from the mailing list / Boeing. |
|
I just realize my "fast" builds all happened in an NFS directory; outside it, mine also becomes slow. The full reasoning behind it: Expand/Collapse AI ChatQI am debugging a flaky slow linking problem. For this or that reason, when I build in an NFS, it's fast; otherwise it's slow. I suspect some decisions in Makefiles are different. The script is @build-both.sh . The relevant log is @chat.txt . This is the fast version. In the slow version, some steps around quoted Makefile seems the bottleneck. Analyze why the quoted Makefile part is not run in the fast run. # xen/arch/x86/Makefile
extra-$(XEN_BUILD_PE) += efi.lds
ifeq ($(XEN_BUILD_PE),y)
$(TARGET).efi: $(objtree)/prelink.o $(note_file) $(obj)/efi.lds $(obj)/efi/relocs-dummy.o $(obj)/efi/mkreloc
ifeq ($(CONFIG_DEBUG_INFO),y)
$(if $(filter --strip-debug,$(EFI_LDFLAGS)),echo,:) "Will strip debug info from $(@F)"
endif
$(foreach base, $(VIRT_BASE) $(ALT_BASE), \
$(LD) $(call EFI_LDFLAGS,$(base)) -T $(obj)/efi.lds $< $(relocs-dummy) \
$(objtree)/common/symbols-dummy.o $(note_file_option) \
-o $(dot-target).$(base).0 &&) :
$(MKRELOC) $(foreach base,$(VIRT_BASE) $(ALT_BASE),$(dot-target).$(base).0) \
> $(dot-target).0r.S
$(NM) -pa --format=sysv $(dot-target).$(VIRT_BASE).0 \AI'll search the Xen x86 Makefiles to see how [3 tools called] I'll run a few quick searches in parallel to see how [4 tools called] I'll open the two relevant Makefiles to cite the exact lines that gate [2 tools called] I checked the x86 build rules. The heavy EFI link recipe only runs when Key bits: EFI_LDFLAGS := $(patsubst -m%,-mi386pep,$(LDFLAGS)) --subsystem=10 --enable-long-section-names
LD_PE_check_cmd = $(call ld-option,$(EFI_LDFLAGS) --image-base=0x100000000 -o $(efi-check).efi $(efi-check).o)
XEN_BUILD_PE := $(LD_PE_check_cmd)
...
$(info FINDME FINDME XEN_BUILD_PE=$(XEN_BUILD_PE))What’s happening:
FINDME FINDME XEN_BUILD_PE=y
UPD include/xen/compile.h
if [ -r xen.efi -a -n '/usr/lib64/efi' ]; then \
... install xen.efi ...Status: I inspected the Makefiles and your log. The EFI link block isn’t run in the fast (NFS) build because
Q2What are the downside of not rebuilding EFI? In a non NFS directory, can I also skip it for faster testing? A2
Example (fits your script that runs dist-xen): make CC=/lib/llvm-20/bin/clang -C xen efi-y= dist-xenIf you want to ensure no stale rm -f xen/xen.efi && make CC=/lib/llvm-20/bin/clang -C xen efi-y= dist-xenOptional: you can also set
Q3If I'm going to boot Xen via GRUB, can I skip it? A3
Quick check: look at your GRUB stanza. If it says To skip in non-NFS builds for faster loops: rm -f xen/xen.efi && make CC=/lib/llvm-20/bin/clang -C xen efi-y= dist-xen
Let me see if there is a general way to speed up all x86 builds. make ${MAKE_FLAGS[@]} -j$(nproc) efi-y= dist-xen
sudo make ${MAKE_FLAGS[@]} -j$(nproc) efi-y= install-xenThe current measurement of
|
No description provided.