Results of llvm-tblgen are not reused in incremental compilation

[ZelinMa557]

I notice that every time I compile the clang & LLVM, the following things will be rebuilt whether or not relative code is changed:

[ 40%] Building ARMGenAsmMatcher.inc...
[ 40%] Building ARMGenAsmWriter.inc...
[ 40%] Building ARMGenCallingConv.inc...
[ 40%] Building ARMGenDisassemblerTables.inc...
[ 40%] Building ARMGenFastISel.inc...
[ 40%] Building ARMGenGlobalISel.inc...
[ 40%] Building ARMGenInstrInfo.inc...
[ 40%] Building ARMGenMCCodeEmitter.inc...
[ 40%] Building ARMGenMCPseudoLowering.inc...
[ 41%] Building ARMGenRegisterBank.inc...
[ 41%] Building ARMGenRegisterInfo.inc...
[ 41%] Building ARMGenSubtargetInfo.inc...
[ 41%] Building ARMGenSystemRegister.inc...

[ 43%] Building BPFGenAsmMatcher.inc...
[ 43%] Building BPFGenAsmWriter.inc...
[ 43%] Building BPFGenCallingConv.inc...
[ 43%] Building BPFGenDisassemblerTables.inc...
[ 43%] Building BPFGenGlobalISel.inc...
[ 43%] Building BPFGenInstrInfo.inc...
[ 43%] Building BPFGenMCCodeEmitter.inc...
[ 43%] Building BPFGenRegisterBank.inc...
[ 43%] Building BPFGenRegisterInfo.inc...
[ 43%] Building BPFGenSubtargetInfo.inc...

This costs a lot of time. Is there any chance to improve that?

[zero9178]

Are you using the Makefiles generator by any chance rather than ninja (-G Ninja with the cmake command line)? Last I recall make is not as capable in reading in the dep files produced by TableGen as well as pruning the build graph if the output didn't change. Ninja is generally much better than make

[ZelinMa557]

Yes, I was using Make to build clang/llvm. But I think we can do something to improve it since many people use Make to build llvm.