[Mesa-dev] [PATCH] radeon/llvm: Convert to Automake
Tom Stellard
tom at stellard.net
Tue Sep 4 11:28:00 PDT 2012
On Tue, Sep 04, 2012 at 11:10:52AM -0700, Matt Turner wrote:
> On Fri, Aug 31, 2012 at 1:41 PM, Tom Stellard <tom at stellard.net> wrote:
> > On Fri, Aug 31, 2012 at 01:26:33PM -0700, Matt Turner wrote:
> >> On Fri, Aug 31, 2012 at 6:04 AM, Tom Stellard <tom at stellard.net> wrote:
> >> > On Thu, Aug 30, 2012 at 02:11:22PM -0700, Matt Turner wrote:
> >> >> On Thu, Aug 30, 2012 at 1:19 PM, Tom Stellard <tom at stellard.net> wrote:
> >> >> > On Thu, Aug 30, 2012 at 10:47:55AM -0700, Matt Turner wrote:
> >> >> >> On Thu, Aug 30, 2012 at 9:55 AM, <tstellar at gmail.com> wrote:
> >> >> >> > From: Tom Stellard <thomas.stellard at amd.com>
> >> >> >> >
> >> >> >> > ---
> >> >> >> > Hi Matt,
> >> >> >> >
> >> >> >> > This patch applies to your automake-gallium branch.
> >> >> >>
> >> >> >> Thanks a bunch!
> >> >> >>
> >> >> >> > configure.ac | 5 ++
> >> >> >> > src/gallium/drivers/radeon/.gitignore | 1 +
> >> >> >> > src/gallium/drivers/radeon/Makefile | 76 -----------------------
> >> >> >> > src/gallium/drivers/radeon/Makefile.am | 87 +++++++++++++++++++++++++++
> >> >> >> > src/gallium/drivers/radeon/Makefile.sources | 32 ++++++++++-
> >> >> >> > 5 files changed, 124 insertions(+), 77 deletions(-)
> >> >> >> > delete mode 100644 src/gallium/drivers/radeon/Makefile
> >> >> >> > create mode 100644 src/gallium/drivers/radeon/Makefile.am
> >> >> >> >
> >> >> >> > + $(LLVM_CFLAGS)
> >> >> >> > +
> >> >> >> > +CXXFLAGS+= $(LLVM_CXXFLAGS)
> >> >> >>
> >> >> >> Do we really need this? We already have LLVM_CXXFLAGS in AM_CXXFLAGS.
> >> >> >>
> >> >> >
> >> >> > Probably not.
> >> >> >> > +
> >> >> >> > +tablegen = $(LLVM_BINDIR)/llvm-tblgen -I $(LLVM_INCLUDEDIR) $1 $2 -o $3
> >> >> >> > +
> >> >> >> > +libradeon_la_SOURCES = \
> >> >> >> > + $(GENERATED_SOURCES) \
> >> >> >> > + $(CXX_SOURCES) \
> >> >> >> > + $(C_SOURCES)
> >> >> >> > +
> >> >> >> > +CLEANFILES = $(GENERATED_SOURCES)
> >> >> >>
> >> >> >> We should probably set BUILT_SOURCES here.
> >> >> >>
> >> >> >
> >> >> > OK
> >> >> >
> >> >> >> > +
> >> >> >> > +SIRegisterInfo.td: SIGenRegisterInfo.pl
> >> >> >> > + $(PERL) $^ > $@
> >> >> >> > +
> >> >> >> > +SIRegisterGetHWRegNum.inc: SIGenRegisterInfo.pl
> >> >> >> > + $(PERL) $^ $@ > /dev/null
> >> >> >> > +
> >> >> >> > +R600Intrinsics.td: R600IntrinsicsNoOpenCL.td R600IntrinsicsOpenCL.td
> >> >> >> > +if HAVE_R600_LLVM_INTRINSICS
> >> >> >> > + cp R600IntrinsicsNoOpenCL.td R600Intrinsics.td
> >> >> >> > +else
> >> >> >> > + cp R600IntrinsicsOpenCL.td R600Intrinsics.td
> >> >> >> > +endif
> >> >> >> > +
> >> >> >> > +R600RegisterInfo.td: R600GenRegisterInfo.pl
> >> >> >> > + $(PERL) $^ > $@
> >> >> >>
> >> >> >> It's nice to put $(AM_V_GEN) before generator calls.
> >> >> >>
> >> >> >
> >> >> > You mean like:
> >> >> >
> >> >> > $(AM_V_GEN) $(PERL) $^ > $@
> >> >>
> >> >> Yeah, exactly.
> >> >>
> >> >> >> > +
> >> >> >> > +AMDGPUGenRegisterInfo.inc: $(TD_FILES)
> >> >> >> > + $(call tablegen, -gen-register-info, AMDGPU.td, $@)
> >> >> >> > +
> >> >> >> > +AMDGPUGenInstrInfo.inc: $(TD_FILES)
> >> >> >> > + $(call tablegen, -gen-instr-info, AMDGPU.td, $@)
> >> >> >> > +
> >> >> >> > +AMDGPUGenAsmWriter.inc: $(TD_FILES)
> >> >> >> > + $(call tablegen, -gen-asm-writer, AMDGPU.td, $@)
> >> >> >> > +
> >> >> >> > +AMDGPUGenDAGISel.inc: $(TD_FILES)
> >> >> >> > + $(call tablegen, -gen-dag-isel, AMDGPU.td, $@)
> >> >> >> > +
> >> >> >> > +AMDGPUGenCallingConv.inc: $(TD_FILES)
> >> >> >> > + $(call tablegen, -gen-callingconv, AMDGPU.td, $@)
> >> >> >> > +
> >> >> >> > +AMDGPUGenSubtargetInfo.inc: $(TD_FILES)
> >> >> >> > + $(call tablegen, -gen-subtarget, AMDGPU.td, $@)
> >> >> >> > +
> >> >> >> > +AMDGPUGenEDInfo.inc: $(TD_FILES)
> >> >> >> > + $(call tablegen, -gen-enhanced-disassembly-info, AMDGPU.td, $@)
> >> >> >> > +
> >> >> >> > +AMDGPUGenIntrinsics.inc: $(TD_FILES)
> >> >> >> > + $(call tablegen, -gen-tgt-intrinsic, AMDGPU.td, $@)
> >> >> >> > +
> >> >> >> > +AMDGPUGenCodeEmitter.inc: $(TD_FILES)
> >> >> >> > + $(call tablegen, -gen-emitter, AMDGPU.td, $@)
> >> >> >> > +
> >> >> >> > +AMDGPUGenMCCodeEmitter.inc: $(TD_FILES)
> >> >> >> > + $(call tablegen, -mc-emitter -gen-emitter, AMDGPU.td, $@)
> >> >> >> > +
> >> >> >> > +AMDGPUGenDFAPacketizer.inc: $(TD_FILES)
> >> >> >> > + $(call tablegen, -gen-dfa-packetizer, AMDGPU.td, $@)
> >> >> >> > +
> >> >> >> > +noinst_PROGRAMS = loader
> >> >> >> > +
> >> >> >> > +loader_SOURCES = loader.cpp
> >> >> >> > +
> >> >> >> > +loader_LDADD = libradeon.la $(LLVM_LIBS)
> >> >> >> > +loader_LDFLAGS = \
> >> >> >> > + $(LLVM_LDFLAGS) \
> >> >> >> > + $(LD_FLAGS)
> >> >> >> > +
> >> >> >> > +#XXX: Delete this when all targets that rely on radeon are converted to
> >> >> >> > +# automake.
> >> >> >> > +all-local: libradeon.la
> >> >> >> > + ln -f $(builddir)/.libs/libradeon.a $(builddir)/libradeon.a
> >> >> >> > +
> >> >> >> > +CLEANFILES += libradeon.a
> >> >> >> > diff --git a/src/gallium/drivers/radeon/Makefile.sources b/src/gallium/drivers/radeon/Makefile.sources
> >> >> >> > index 687acb3..d8df295 100644
> >> >> >> > --- a/src/gallium/drivers/radeon/Makefile.sources
> >> >> >> > +++ b/src/gallium/drivers/radeon/Makefile.sources
> >> >> >> > @@ -1,4 +1,34 @@
> >> >> >> >
> >> >> >> > +TD_FILES := \
> >> >> >> > + AMDGPU.td \
> >> >> >> > + AMDGPUInstrInfo.td \
> >> >> >> > + AMDGPUInstructions.td \
> >> >> >> > + AMDGPUIntrinsics.td \
> >> >> >> > + AMDGPURegisterInfo.td \
> >> >> >> > + AMDILBase.td \
> >> >> >> > + AMDILInstrInfo.td \
> >> >> >> > + AMDILIntrinsics.td \
> >> >> >> > + AMDILRegisterInfo.td \
> >> >> >> > + Processors.td \
> >> >> >> > + R600InstrInfo.td \
> >> >> >> > + R600Instructions.td \
> >> >> >> > + R600Intrinsics.td \
> >> >> >> > + R600Intrinsics.td \
> >> >> >> > + R600IntrinsicsNoOpenCL.td \
> >> >> >> > + R600IntrinsicsOpenCL.td \
> >> >> >> > + R600OpenCLIntrinsics.td \
> >> >> >> > + R600RegisterInfo.td \
> >> >> >> > + R600RegisterInfo.td \
> >> >> >> > + R600Schedule.td \
> >> >> >> > + SIInstrFormats.td \
> >> >> >> > + SIInstrInfo.td \
> >> >> >> > + SIInstructions.td \
> >> >> >> > + SIIntrinsics.td \
> >> >> >> > + SIRegisterInfo.td \
> >> >> >> > + SIRegisterInfo.td \
> >> >> >> > + SISchedule.td
> >> >> >> > +
> >> >> >> > +
> >> >> >> > GENERATED_SOURCES := \
> >> >> >> > R600Intrinsics.td \
> >> >> >> > R600RegisterInfo.td \
> >> >> >> > @@ -16,7 +46,7 @@ GENERATED_SOURCES := \
> >> >> >> > AMDGPUGenMCCodeEmitter.inc \
> >> >> >> > AMDGPUGenDFAPacketizer.inc
> >> >> >> >
> >> >> >> > -CPP_SOURCES := \
> >> >> >> > +CXX_SOURCES := \
> >> >> >> > AMDIL7XXDevice.cpp \
> >> >> >> > AMDILCFGStructurizer.cpp \
> >> >> >> > AMDILDevice.cpp \
> >> >> >> > --
> >> >> >> > 1.7.7.6
> >> >> >>
> >> >> >> When I build with this, I get:
> >> >> >>
> >> >> >> CXX AMDIL7XXDevice.lo
> >> >> >> In file included from AMDIL7XXDevice.cpp:10:0:
> >> >> >> AMDGPUSubtarget.h:22:38: fatal error: AMDGPUGenSubtargetInfo.inc: No
> >> >> >> such file or directory
> >> >> >> compilation terminated.
> >> >> >> make: *** [AMDIL7XXDevice.lo] Error 1
> >> >> >>
> >> >> >> and seeing AMDGPUGenSubtargetInfo.inc in GENERATED_SOURCES, if I try to make it:
> >> >> >>
> >> >> >> make: *** No rule to make target `R600InstrInfo.td', needed by
> >> >> >> `AMDGPUGenSubtargetInfo.inc'. Stop.
> >> >> >>
> >> >> >> After removing R600InstrInfo.td, R600IntrinsicsOpenCL.td, and
> >> >> >> R600OpenCLIntrinsics.td from TD_FILES in Makefile.sources,
> >> >> >> AMDIL7XXDevice.lo still fails to build, but I can now build
> >> >> >> AMDGPUGenSubtargetInfo.inc.
> >> >> >>
> >> >> >> So, we either need to do more finely grained dependencies on *.inc
> >> >> >> files, or put BUILT_SOURCES = $(GENERATED_SOURCES) to make sure
> >> >> >> everything is generated before we start compiling. The first would
> >> >> >> allow slightly better parallelism in the build, but the second is much
> >> >> >> easier.
> >> >> >>
> >> >> >
> >> >> > Let's do BUILT_SOURCES = $(GENERATED_SOURCES) for now. Do you want me
> >> >> > to send you a new patch, or have you already made these changes locally?
> >> >>
> >> >> Works for me. Even with the changes I made, I couldn't get loader to
> >> >> link, so you might want to look at that. If the fix is simple, just
> >> >> tell me what it is and I'll fix these things locally.
> >> >
> >> > The loader program is only for debugging and is not built by default in
> >> > Mesa master. If it is the only blocker you have you can remove it
> >> > from the default build, and actually since it can take a while to link
> >> > when using debug builds, it probably shouldn't be built by default
> >> > anyway.
> >> >
> >> > I can still take a look at fixing the linking problems, but I need
> >> > to know:
> >> > 1. What are the errors?
> >> > 2. What version of LLVM are you using?
> >> > 3. What prefix do you have LLVM installed in?
> >> >
> >> > -Tom
> >>
> >> Of course. Sorry -- I thought you would be able to reproduce locally.
> >>
> >> The error is>
> >> gmake[1]: Entering directory
> >> `/home/mattst88/projects/mesa/src/gallium/drivers/radeon'
> >> /bin/sh ../../../../libtool --tag=CXX --mode=link g++
> >> -I/usr/include -DNDEBUG -D_GNU_SOURCE -D__STDC_CONSTANT_MACROS
> >> -D__STDC_FORMAT_MACROS -D__STDC_LIMIT_MACROS
> >> -fvisibility-inlines-hidden -fno-exceptions -fPIC -Woverloaded-virtual
> >> -Wcast-qual -O0 -pipe -Wall -fno-strict-aliasing -fno-builtin-memcmp
> >> -L/usr/lib64/llvm -lpthread -lffi -ldl -lm -o loader loader.o
> >> libradeon.la -lLLVMMCJIT -lLLVMBitWriter -lLLVMX86CodeGen
> >> -lLLVMSelectionDAG -lLLVMAsmPrinter -lLLVMX86AsmParser -lLLVMMCParser
> >> -lLLVMX86Disassembler -lLLVMX86Desc -lLLVMX86Info -lLLVMX86AsmPrinter
> >> -lLLVMX86Utils -lLLVMJIT -lLLVMRuntimeDyld -lLLVMExecutionEngine
> >> -lLLVMCodeGen -lLLVMScalarOpts -lLLVMInstCombine -lLLVMTransformUtils
> >> -lLLVMipa -lLLVMAnalysis -lLLVMTarget -lLLVMMC -lLLVMObject -lLLVMCore
> >> -lLLVMSupport
> >> libtool: link: g++ -I/usr/include -DNDEBUG -D_GNU_SOURCE
> >> -D__STDC_CONSTANT_MACROS -D__STDC_FORMAT_MACROS -D__STDC_LIMIT_MACROS
> >> -fvisibility-inlines-hidden -fno-exceptions -fPIC -Woverloaded-virtual
> >> -Wcast-qual -O0 -pipe -Wall -fno-strict-aliasing -fno-builtin-memcmp
> >> -o loader loader.o -L/usr/lib64/llvm -lpthread -lffi -ldl -lm
> >> ./.libs/libradeon.a -lLLVMMCJIT -lLLVMBitWriter -lLLVMX86CodeGen
> >> -lLLVMSelectionDAG -lLLVMAsmPrinter -lLLVMX86AsmParser -lLLVMMCParser
> >> -lLLVMX86Disassembler -lLLVMX86Desc -lLLVMX86Info -lLLVMX86AsmPrinter
> >> -lLLVMX86Utils -lLLVMJIT -lLLVMRuntimeDyld -lLLVMExecutionEngine
> >> -lLLVMCodeGen -lLLVMScalarOpts -lLLVMInstCombine -lLLVMTransformUtils
> >> -lLLVMipa -lLLVMAnalysis -lLLVMTarget -lLLVMMC -lLLVMObject -lLLVMCore
> >> -lLLVMSupport
> >> loader.o: In function `llvm::ParseIR(llvm::MemoryBuffer*,
> >> llvm::SMDiagnostic&, llvm::LLVMContext&)':
> >> loader.cpp:(.text._ZN4llvm7ParseIREPNS_12MemoryBufferERNS_12SMDiagnosticERNS_11LLVMContextE[llvm::ParseIR(llvm::MemoryBuffer*,
> >> llvm::SMDiagnostic&, llvm::LLVMContext&)]+0x80): undefined reference
> >> to `llvm::ParseBitcodeFile(llvm::MemoryBuffer*, llvm::LLVMContext&,
> >> std::basic_string<char, std::char_traits<char>, std::allocator<char>
> >> >*)'
> >> loader.cpp:(.text._ZN4llvm7ParseIREPNS_12MemoryBufferERNS_12SMDiagnosticERNS_11LLVMContextE[llvm::ParseIR(llvm::MemoryBuffer*,
> >> llvm::SMDiagnostic&, llvm::LLVMContext&)]+0x185): undefined reference
> >> to `llvm::ParseAssembly(llvm::MemoryBuffer*, llvm::Module*,
> >> llvm::SMDiagnostic&, llvm::LLVMContext&)'
> >> collect2: ld returned 1 exit status
> >>
> >
> > Ok, I think I see the problem. I was building with --enable-opencl,
> > which was adding the libs required by loader to $(LLVM_LIBS). In the
> > current Makefile, we are getting the libraries required for the
> > loader like this:
> >
> > LOADER_LIBS=$(shell llvm-config --libs bitreader asmparser)
> >
> > We then appending the LOADER_LIBS to the linker arguments after the
> > $(LLVM_LIBS). Actually, we should be using $(LLVM_BINDIR)/llvm-config
> > here. So, I think doing something similar in the automake file should work.
> >
> >> I'm using llvm-3.1 from my distribution, which means the prefix is /usr.
> >>
> >> If it's useful, I'm able to build working LLVMpipe using my automake patches.
> >>
> >> To confirm, should I remove R600InstrInfo.td, R600IntrinsicsOpenCL.td,
> >> and R600OpenCLIntrinsics.td from TD_FILES?
> >
> > I think I see the problem here too. Those files need to be in TD_FILES,
> > except that R600OpenCLIntrinsics.td does not exist, it should be replaced
> > with R600InstrinsicsNoOpenCL.td.
>
> Typo that confused me. In_s_trinsics. R600IntrinsicsNoOpenCL.td is
> already in TD_FILES, along with a pair of other duplicates.
>
Oops, I overlooked that, so just remove R600OpenCLIntrinsics.td along
with the duplicated files and it should be fine.
> And I don't understand why R600InstrInfo.td should be in TD_FILES. If
> I understand correctly, these files generate .h/.cpp files, and
> R600InstrInfo.{h,cpp} already exist. I'm going to assume it's not
> needed.
All files generated by tablegen have the .inc suffix. The main tablegen
file AMDGPU.td includes all the other .td files and is passed to
llvm-tablegen for each .inc file that needs to be generated.
R600InstrInfo.{h,cpp} are not generated files.
-Tom
More information about the mesa-dev
mailing list