自定义LLVM opt管道的替代方法有哪些？

Question

Question

There is no way from command line to safely disable a specific pass in the O1/O2/O3 opt pipeline. What are the alternate ways to customize or tailor the opt pipeline of LLVM passes?

Background and Problem

I am writing a pass to outline specific regions of code and I need to reduce the number of inputs & outputs to the outlined function. In some cases, the opt tool raises some (invariant) GEP instructions (for example, from a loop body). This happens with O1/O2/O3 flags. I do not want this since it creates more number of inputs/outputs to my outlined function. I prefer to have the GEP done in the outlined function.

Failed attempts

• The opt O1 pass arguments are shown below. I used "print-after" command in opt and looked at the IR to find that the "reassociate" transformation creates this change. I went over the disable options in opt --help-hidden and I do not find a way to disable this specific pass or LICM.
• I tried looking at the LLVM source code Reassociate.cpp. I couldn't identify any ways to disable this specific transformation from the code either.
• I have also tried to manually enter specific pass names to the opt command. However some passes need the O1/2/3 flag to be set. If I do not set the flag, these passes do not make the necessary transformation. Example, I need the loop to be unrolled but it does not unroll without O1/2/3 flag, even if i explicitly enter the flag in the opt command.

Pass arguments in O1 stage

 Pass Arguments:  -targetlibinfo -tti -targetpassconfig -tbaa -scoped-noalias 
    -assumption-cache-tracker -profile-summary-info -forceattrs -inferattrs 
    -ipsccp -called-value-propagation -globalopt -domtree -mem2reg -deadargelim 
    -domtree -basicaa -aa -loops -lazy-branch-prob -lazy-block-freq 
    -opt-remark-emitter -instcombine -simplifycfg -basiccg -globals-aa -prune-eh 
    -always-inline -functionattrs -domtree -sroa -basicaa -aa -memoryssa -early-cse-memssa 
    -speculative-execution -basicaa -aa -lazy-value-info -jump-threading -correlated-propagation 
    -simplifycfg -domtree -basicaa -aa -loops -lazy-branch-prob -lazy-block-freq 
    -opt-remark-emitter -instcombine -libcalls-shrinkwrap -loops -branch-prob -block-freq 
    -lazy-branch-prob -lazy-block-freq -opt-remark-emitter -pgo-memop-opt -basicaa -aa 
    -loops -lazy-branch-prob -lazy-block-freq -opt-remark-emitter -tailcallelim -simplifycfg 
    -reassociate -domtree -loops -loop-simplify -lcssa-verification -lcssa -basicaa -aa 
    -scalar-evolution -loop-rotate -licm -loop-unswitch -simplifycfg -domtree -basicaa -aa -loops
    -lazy-branch-prob -lazy-block-freq -opt-remark-emitter -instcombine -loop-simplify -lcssa-verification -lcssa -scalar-evolution -indvars
    -loop-idiom -loop-deletion -loop-unroll -phi-values -memdep -memcpyopt -sccp -demanded-bits -bdce -basicaa -aa -lazy-branch-prob
    -lazy-block-freq -opt-remark-emitter -instcombine -lazy-value-info -jump-threading -correlated-propagation -basicaa -aa -phi-values
    -memdep -dse -loops -loop-simplify -lcssa-verification -lcssa -basicaa -aa -scalar-evolution -licm -postdomtree -adce -simplifycfg 
    -domtree -basicaa -aa -loops -lazy-branch-prob -lazy-block-freq -opt-remark-emitter -instcombine -barrier -basiccg -rpo-functionattrs 
    -globalopt -globaldce -basiccg -globals-aa -float2int -domtree -loops -loop-simplify -lcssa-verification -lcssa -basicaa -aa -scalar-evolution
    -loop-rotate -loop-accesses -lazy-branch-prob -lazy-block-freq -opt-remark-emitter -loop-distribute -branch-prob -block-freq -scalar-evolution 
    -basicaa -aa -loop-accesses -demanded-bits -lazy-branch-prob -lazy-block-freq -opt-remark-emitter -loop-vectorize -loop-simplify -scalar-evolution 
    -aa -loop-accesses -loop-load-elim -basicaa -aa -lazy-branch-prob -lazy-block-freq -opt-remark-emitter -instcombine -simplifycfg -domtree
    -basicaa -aa -loops -lazy-branch-prob -lazy-block-freq -opt-remark-emitter -instcombine -loop-simplify -lcssa-verification -lcssa 
    -scalar-evolution -loop-unroll -lazy-branch-prob -lazy-block-freq -opt-remark-emitter -instcombine -loop-simplify -lcssa-verification 
    -lcssa -scalar-evolution -licm -alignment-from-assumptions -strip-dead-prototypes -domtree -loops -branch-prob -block-freq -loop-simplify
    -lcssa-verification -lcssa -basicaa -aa -scalar-evolution -branch-prob -block-freq -loop-sink -lazy-branch-prob -lazy-block-freq
    -opt-remark-emitter -instsimplify -div-rem-pairs -simplifycfg -domtree -sroa -verify -print-module

I think what I need is

• to move the definition close to its use (Any comments on how I can do this at IR stage?) or
• find a way to safely disable a pass in the opt pipeline (licm, reassociate). I believe this is not possible from command line. If I have to modify the LLVM source code locally (for hacking the opt pipeline to accomplish what I need), what would be the right place to do this safely? Any pointers will be helpful; or
• to manually enter the passes I need to run in the opt command without any of the O1/O2/O3 flags But I am not able to find a way to do any of these that works. I am using LLVM release 7.00 .

Possible solution?

Can I add only the required passes (subset of the O1 pipeline) as AddRequired in getAnalysisUsage() in my pass definition and make opt run my pass with O0 flag enabled? Will this work? I can try this out.

Answer 1

According to this thread from summer 2018, there is no general way to disable a specific pass from clang:

Some passes have supported options to disable them, eg -fno-vectorize and -fno-unroll-loops, but there is no general option. Since it's not useful in general to disable arbitrary options, some handywork is required.

The original poster in the thread seemed to be working on a patch to enable this behavior, but as far as I can tell it never landed.