I was wondering, if someone could point me into the right direction,
towards building just llc and opt (potentially 4.0 + patches) to distribute
them in the interim with compiler? The idea is that we’d like to pin our
llvm ir code generator to a specific set of llc and opt tools, with
potential patches applied, until we manage to get all those patches
upstreamed.
I’d like to find answers to the following question:
- How trivially can I build just llc and opt into (maybe static?)
distributable binaries?
- Would generating them for multiple architectures pose any specific
complications?
I was wondering, if someone could point me into the right direction,
towards building just llc and opt (potentially 4.0 + patches) to distribute
them in the interim with compiler? The idea is that we’d like to pin our
llvm ir code generator to a specific set of llc and opt tools, with
potential patches applied, until we manage to get all those patches
upstreamed.
I’d like to find answers to the following question:
- How trivially can I build just llc and opt into (maybe static?)
distributable binaries?
$ cmake path/to/llvm -DBUILD_SHARED_LIBS=OFF -DCMAKE_CXX_FLAGS=-static-libstdc++
$ make llc opt
- Would generating them for multiple architectures pose any specific
complications?
llc and opt are primarily developed as debugging tools to support LLVM testing. They are not intended to be shipped or used as code generators for frontends. They do often work in these settings but you may hit rough edges that nobody will care to fix and the behavior or commandline interface may change every release... A more stable alternative is usually to feed .ir/.bc files to clang or better link to the llvm libraries and generate the code from there.
I’ll heed your advice and try to replace our llc and opt invocations
with clang.
Thank you Jonathan, I’ll try to build clang in a static fashion as you
described and try to see how we can bundle that sensibly with ghc.
Hopefully, once our upstream patches made it into ghc, (e.g. D30770,
and D30812 as of right now), we’ll be able to drop shipping a custom
static clang with ghc, and can rely on the one provided by the system.
I’m trying to see if I can make do with the clang interface only.
How do I pass flags to opt or llc? Say I want to enable/disable tbaa,
which could be done with `--enable-tbaa=true/false` or `-mem2reg` or
`-globalopt`? Would the `-llvm` switch be, what I’m looking for or the
`-Xclang`, or something else entirely?
-mllvm xxx will pass options directly to the llvm libraries though for that the same caveats apply in that they are more of a developer/debug tool than a stable interface.
The stable interface to clang is the usual flags like -O0/-O3/-fXXX/-mXXX ...
If you find yourself in a position that you want to control the pass pipeline down to single passes (why would you want to enable/disable mem2reg really?) then you are relying on LLVM internals anyway. If you need/want that level of control then your best bet would be to link against the LLVM libraries or I guess shipping and relying on llc/opt is similarily good/bad. It will work but you will have to do more work to keep track with llvm releases, development of new passes, new behaviors, ...
what I’m observing right now is that replacing opt+llc with an clang invocation, and
subsequently fewer intermediate files, increases the consumed time with -O0 by 200%.
We used to always run opt with -mem2reg -globalopt, and I believe those are not part
of -O0 (is there an easy way to list all passes that -OX flags to clang imply for the
optimizer and code gen?).
Could the IR imply certain optimization passes to be required? Could llvm be taught
that say a certain calling convention needs a pass to run?
what I’m observing right now is that replacing opt+llc with an clang invocation, and
subsequently fewer intermediate files, increases the consumed time with -O0 by 200%.
We used to always run opt with -mem2reg -globalopt, and I believe those are not part
of -O0 (is there an easy way to list all passes that -OX flags to clang imply for the
optimizer and code gen?).
I believe most of the OptimizationLevel handling happens in lib/Passes/PassBuilder.cpp.
Could the IR imply certain optimization passes to be required? Could llvm be taught
that say a certain calling convention needs a pass to run?
Todays pass manager computes a static schedule before processing any functions so that wouldn't be possible. I don't know whether different pipelines per function would be possible with the new pass manager. Switching the optimization passes in the pipeline based on the calling convention seems odd to me; I would at least use a separate function attribute if I were to implement this.
what I’m observing right now is that replacing opt+llc with an clang invocation, and
subsequently fewer intermediate files, increases the consumed time with -O0 by 200%.
We used to always run opt with -mem2reg -globalopt, and I believe those are not part
of -O0 (is there an easy way to list all passes that -OX flags to clang imply for the
optimizer and code gen?).
No there isn’t.
Could the IR imply certain optimization passes to be required? Could llvm be taught
that say a certain calling convention needs a pass to run?
This could be done, but we’d need to here more about the motivation. I suspect the solution may end up different if we understand the exact use-case.
what I’m observing right now is that replacing opt+llc with an clang invocation, and
subsequently fewer intermediate files, increases the consumed time with -O0 by 200%.
We used to always run opt with -mem2reg -globalopt, and I believe those are not part
of -O0 (is there an easy way to list all passes that -OX flags to clang imply for the
optimizer and code gen?).
No there isn’t.
Could the IR imply certain optimization passes to be required? Could llvm be taught
that say a certain calling convention needs a pass to run?
This could be done, but we’d need to here more about the motivation. I suspect the solution may end up different if we understand the exact use-case.
what I’m observing right now is that replacing opt+llc with an clang invocation, and
subsequently fewer intermediate files, increases the consumed time with -O0 by 200%.
We used to always run opt with -mem2reg -globalopt, and I believe those are not part
of -O0 (is there an easy way to list all passes that -OX flags to clang imply for the
optimizer and code gen?).
No there isn’t.
Could the IR imply certain optimization passes to be required? Could llvm be taught
that say a certain calling convention needs a pass to run?
This could be done, but we’d need to here more about the motivation. I suspect the solution may end up different if we understand the exact use-case.
From that discussion I don’t see why you need mem2reg or globalopt. Sure those things are desirable to eliminate the extra copies and get performant code, but for that you wouldn’t use -O0 anyway…
what I’m observing right now is that replacing opt+llc with an clang invocation, and
subsequently fewer intermediate files, increases the consumed time with -O0 by 200%.
We used to always run opt with -mem2reg -globalopt, and I believe those are not part
of -O0 (is there an easy way to list all passes that -OX flags to clang imply for the
optimizer and code gen?).
No there isn’t.
Could the IR imply certain optimization passes to be required? Could llvm be taught
that say a certain calling convention needs a pass to run?
This could be done, but we’d need to here more about the motivation. I suspect the solution may end up different if we understand the exact use-case.
From that discussion I don’t see why you need mem2reg or globalopt. Sure those things are desirable to eliminate the extra copies and get performant code, but for that you wouldn’t use -O0 anyway…
I haven’t checked, but conceptually this looks very similar to the C++ “this parameter” which we probably load/store like crazy as well in -O0.
I’ll have to see what we do, that produces some .ll that fails to compile
at -O0. For now I’ve just bumped it to at least -O1. I’ll post agian,
once I have that broken ll file in front of me.