| Message ID | 1378376448-29036-5-git-send-email-famz@redhat.com |
|---|---|
| State | New |
| Headers | show |
On 09/05/2013 03:20 AM, Fam Zheng wrote: > 1) %.o in $(common-obj-m) is compiled to %.o, with > "QEMU_CFLAGS += -shared -fPIC". Then "linked" to %.mo, which is an > incremental object with "ln -r". This step is for consistency with > %.mod case and has no effect. As a general rule, you should avoid ld -r unless you know exactly what that implies for the given machine. This is the sort of thing that's highly likely to work on x86 while failing elsewhere. E.g. ARM and PPC ld would like to add trampolines for direct calls to reach the PLT entry. If you create one object that's too large then that's no longer possible. E.g. MIPS and Alpha ld require that any one input object file reference less than 64K worth of got entries. Every separate object file can address its own 64K segment of the got. If you combine too many objectsyou could overflow the got subsegments. If our modules are small enough, we may never see any of these, but... If you really require a single input file to ld for pattern matching purposes, I highly recommend an ar file, and then use --whole-archive to force the entire contents of the .a to be included. This preserves the original translation-unit boundaries for ld. r~
Il 05/09/2013 21:24, Richard Henderson ha scritto: > On 09/05/2013 03:20 AM, Fam Zheng wrote: >> 1) %.o in $(common-obj-m) is compiled to %.o, with >> "QEMU_CFLAGS += -shared -fPIC". Then "linked" to %.mo, which is an >> incremental object with "ln -r". This step is for consistency with >> %.mod case and has no effect. > > As a general rule, you should avoid ld -r unless you know exactly > what that implies for the given machine. This is the sort of thing > that's highly likely to work on x86 while failing elsewhere. > > E.g. ARM and PPC ld would like to add trampolines for direct calls > to reach the PLT entry. If you create one object that's too large > then that's no longer possible. > > E.g. MIPS and Alpha ld require that any one input object file > reference less than 64K worth of got entries. Every separate > object file can address its own 64K segment of the got. If you > combine too many objectsyou could overflow the got subsegments. > > If our modules are small enough, we may never see any of these, but... > > If you really require a single input file to ld for pattern matching > purposes, I highly recommend an ar file, and then use --whole-archive > to force the entire contents of the .a to be included. This preserves > the original translation-unit boundaries for ld. Yes, this is the reason why I suggested using libtool for this. It increases the cost of compilation by essentially doing it twice, on the other hand you can simply use .la convenience libraries and libtool will be able to link it transparently into either an executable or a shared library. Libtool used to be really bad, but most performance problems have been solved. Paolo
On 5 September 2013 20:41, Paolo Bonzini <pbonzini@redhat.com> wrote: > Libtool used to be really bad, but most performance problems have been > solved. How about the "it silently creates things in dot-directories, recompiles things at random times when it really shouldn't, and is an enormous insane shell script" problems? The positives would have to be really really strong before we let libtool into the codebase... -- PMM
On Thu, 09/05 12:24, Richard Henderson wrote: > On 09/05/2013 03:20 AM, Fam Zheng wrote: > > 1) %.o in $(common-obj-m) is compiled to %.o, with > > "QEMU_CFLAGS += -shared -fPIC". Then "linked" to %.mo, which is an > > incremental object with "ln -r". This step is for consistency with > > %.mod case and has no effect. > > As a general rule, you should avoid ld -r unless you know exactly > what that implies for the given machine. This is the sort of thing > that's highly likely to work on x86 while failing elsewhere. > > E.g. ARM and PPC ld would like to add trampolines for direct calls > to reach the PLT entry. If you create one object that's too large > then that's no longer possible. > > E.g. MIPS and Alpha ld require that any one input object file > reference less than 64K worth of got entries. Every separate > object file can address its own 64K segment of the got. If you > combine too many objectsyou could overflow the got subsegments. > > If our modules are small enough, we may never see any of these, but... > > If you really require a single input file to ld for pattern matching > purposes, I highly recommend an ar file, and then use --whole-archive > to force the entire contents of the .a to be included. This preserves > the original translation-unit boundaries for ld. > OK. Thanks for pointing out. The problem with ar file is that --whole-archive needs to be done specificly to modules .a, other libs shouldn't be linked with this flag. I'll drop this step by changing my pattern matching approach in Makefile, and just compile and link without such intermediate object file. Fam
On 09/05/2013 10:53 PM, Fam Zheng wrote: > OK. Thanks for pointing out. The problem with ar file is that --whole-archive > needs to be done specificly to modules .a, other libs shouldn't be linked with > this flag. See also: --whole-archive foo.a --no-whole-archive r~
On Fri, 09/06 00:20, Richard Henderson wrote: > On 09/05/2013 10:53 PM, Fam Zheng wrote: > > OK. Thanks for pointing out. The problem with ar file is that --whole-archive > > needs to be done specificly to modules .a, other libs shouldn't be linked with > > this flag. > > See also: > > --whole-archive foo.a --no-whole-archive Yes, this is a way to do it, but will complicates the link command, and may be more error prone. I expanded foo.mo to its contained .o for static linking in the next revision that's coming up. Thanks, Fam
Il 05/09/2013 23:45, Peter Maydell ha scritto: > On 5 September 2013 20:41, Paolo Bonzini <pbonzini@redhat.com> wrote: >> Libtool used to be really bad, but most performance problems have been >> solved. > > How about the "it silently creates things in dot-directories, > recompiles things at random times when it really shouldn't, > and is an enormous insane shell script" problems? The positives > would have to be really really strong before we let libtool > into the codebase... How about "we'd break compilation on your favorite OS"? :) Darwin ld doesn't support --whole-archive, it needs a different incantation. Relatively recent Solaris now has it, but Apple still holds the bastions. Paolo
diff --git a/Makefile b/Makefile index 806946e..130a497 100644 --- a/Makefile +++ b/Makefile @@ -56,7 +56,7 @@ Makefile: ; configure: ; .PHONY: all clean cscope distclean dvi html info install install-doc \ - pdf recurse-all speed test dist + pdf recurse-all speed test dist modules $(call set-vpath, $(SRC_PATH)) @@ -121,7 +121,22 @@ ifeq ($(CONFIG_SMARTCARD_NSS),y) include $(SRC_PATH)/libcacard/Makefile endif -all: $(DOCS) $(TOOLS) $(HELPERS-y) recurse-all +all: $(DOCS) $(TOOLS) $(HELPERS-y) recurse-all modules + +mod-obj-m = $(patsubst %.o,%$(DSOSUF),$(filter %.o,$(common-obj-m))) \ + $(patsubst %.mo,%$(DSOSUF),$(filter %.mo,$(common-obj-m))) +mod-obj-y = $(patsubst %.mo,%.mo.o,$(filter %.mo,$(common-obj-y))) + +# Generate rules for single file modules (%.mo: %.o). +# For multi file modules, dependencies should be listed explicitly in +# Makefile.objs +$(foreach o,$(mod-obj-m) $(mod-obj-y),$(eval \ + $(patsubst %$(DSOSUF),%.mo,$o): $(patsubst %.c,%.o,$(wildcard $(patsubst %$(DSOSUF),%.c,$o))) \ + )) + +modules: $(mod-obj-m) +modules: BUILD_DYNAMIC = 1 +modules: QEMU_CFLAGS += -shared -fPIC config-host.h: config-host.h-timestamp config-host.h-timestamp: config-host.mak @@ -155,7 +170,7 @@ subdir-dtc:dtc/libfdt dtc/tests dtc/%: mkdir -p $@ -$(SUBDIR_RULES): libqemuutil.a libqemustub.a $(common-obj-y) +$(SUBDIR_RULES): libqemuutil.a libqemustub.a $(common-obj-y) $(common-obj-m) ROMSUBDIR_RULES=$(patsubst %,romsubdir-%, $(ROMS)) romsubdir-%: @@ -235,6 +250,9 @@ clean: rm -f qemu-options.def find . -name '*.[oda]' -type f -exec rm -f {} + find . -name '*.l[oa]' -type f -exec rm -f {} + + find . -name '*'$(DSOSUF) -type f -exec rm -f {} + + find . -name '*.mo' -type f -exec rm -f {} + + rm -f $(TOOLS) $(HELPERS-y) qemu-ga TAGS cscope.* *.pod *~ */*~ rm -Rf .libs rm -f qemu-img-cmds.h diff --git a/Makefile.objs b/Makefile.objs index f46a4cd..2f178af 100644 --- a/Makefile.objs +++ b/Makefile.objs @@ -19,6 +19,8 @@ block-obj-y += qemu-coroutine.o qemu-coroutine-lock.o qemu-coroutine-io.o block-obj-y += qemu-coroutine-sleep.o block-obj-y += coroutine-$(CONFIG_COROUTINE_BACKEND).o +block-obj-m = block/ + ifeq ($(CONFIG_VIRTIO)$(CONFIG_VIRTFS)$(CONFIG_PCI),yyy) # Lots of the fsdev/9pcode is pulled in by vl.c via qemu_fsdev_add. # only pull in the actual virtio-9p device if we also enabled virtio. @@ -83,6 +85,9 @@ common-obj-$(CONFIG_SMARTCARD_NSS) += $(libcacard-y) common-obj-y += qmp-marshal.o common-obj-y += qmp.o hmp.o + +common-obj-m = $(block-obj-m) + endif ###################################################################### @@ -121,5 +126,8 @@ nested-vars += \ util-obj-y \ qga-obj-y \ block-obj-y \ - common-obj-y + block-obj-m \ + common-obj-y \ + common-obj-m + dummy := $(call unnest-vars) diff --git a/configure b/configure index 48e14b8..8f0a882 100755 --- a/configure +++ b/configure @@ -190,6 +190,7 @@ mingw32="no" gcov="no" gcov_tool="gcov" EXESUF="" +DSOSUF=".so" prefix="/usr/local" mandir="\${prefix}/share/man" datadir="\${prefix}/share" @@ -580,6 +581,7 @@ fi if test "$mingw32" = "yes" ; then EXESUF=".exe" + DSOSUF=".dll" QEMU_CFLAGS="-DWIN32_LEAN_AND_MEAN -DWINVER=0x501 $QEMU_CFLAGS" # enable C99/POSIX format strings (needs mingw32-runtime 3.15 or later) QEMU_CFLAGS="-D__USE_MINGW_ANSI_STDIO=1 $QEMU_CFLAGS" @@ -4165,6 +4167,7 @@ echo "LIBTOOLFLAGS=$LIBTOOLFLAGS" >> $config_host_mak echo "LIBS+=$LIBS" >> $config_host_mak echo "LIBS_TOOLS+=$libs_tools" >> $config_host_mak echo "EXESUF=$EXESUF" >> $config_host_mak +echo "DSOSUF=$DSOSUF" >> $config_host_mak echo "LIBS_QGA+=$libs_qga" >> $config_host_mak echo "POD2MAN=$POD2MAN" >> $config_host_mak echo "TRANSLATE_OPT_CFLAGS=$TRANSLATE_OPT_CFLAGS" >> $config_host_mak diff --git a/rules.mak b/rules.mak index 8276421..090eee1 100644 --- a/rules.mak +++ b/rules.mak @@ -58,6 +58,12 @@ endif %.o: %.dtrace $(call quiet-command,dtrace -o $@ -G -s $<, " GEN $(TARGET_DIR)$@") +%$(DSOSUF) %.mo.o: %.mo + $(call quiet-command,$(LD) $< -o $@ -shared," LD[M] $(TARGET_DIR)$@") + +%.mo: + $(call quiet-command,$(LD) $(sort $^) -r -o $@," LD $(TARGET_DIR)$@") + %$(EXESUF): %.o $(call LINK,$^)
Add necessary rules and flags for shared object generation. common-obj-m will include block-obj-m, as common-obj-y for block-obj-y. The rules introduced here are: QEMU_CFLAGS += -shared -fPIC, for all %.o of shared objects. 1) %.o in $(common-obj-m) is compiled to %.o, with "QEMU_CFLAGS += -shared -fPIC". Then "linked" to %.mo, which is an incremental object with "ln -r". This step is for consistency with %.mod case and has no effect. 2) %.mod in $(common-obj-m) is generated by "ln -r", with all its dependencies (multiple *.o) as input. Which means the list of dependency objects must be ruled out in each sub-Makefile.objs like: $(obj)/foo.mod: $(addprefix $(obj)/,bar.o baz.o qux.o) Notice that $(obj)/ is required for both target and dependency in the rule. Signed-off-by: Fam Zheng <famz@redhat.com> --- Makefile | 24 +++++++++++++++++++++--- Makefile.objs | 10 +++++++++- configure | 3 +++ rules.mak | 6 ++++++ 4 files changed, 39 insertions(+), 4 deletions(-)