[mlir] Use of MLIR Properties slows compilation

[psteinfeld]

Starting with patch https://reviews.llvm.org/D148298, compilation times for some Fortran benchmarks slowed significantly. Here's a small Fortran test case that illustrates the problem:

Program slow
  Integer, Parameter :: x(2,4,6,8,10,12) = &
    Reshape( [ (i,i=1,46080) ], [ 2,4,6,8,10,12 ] )
End Program

If I run the bbc tool on this program with the -emit-fir option, compilation stops after the initial creation of the FIR code. If I run this command on a compiler produced with and without the patch listed above, I get almost identical results. Here's the command line I used:

time bbc -emit-fir slow.f90

On the other hand, if I run the bbc tool without the emit-fir option, the bbc tool runs all of the optimization passes. In this case, the execution time is almost twice as slow. Here's the command line I used:

time bbc slow.f90

I also noted that the optimized and unoptimized MLIR produced is identical.

[llvmbot]

@llvm/issue-subscribers-mlir

[joker-eph]

I've been trying to build flang on my MacBook, but clang is quite unhappy with it right now. So many files (like flang/lib/Lower/ConvertExpr.cpp) are taking > 5min to compile (this a M2 MacBook Pro!), is there a known issue here?

[joker-eph]

This is fixing it: https://reviews.llvm.org/D150540 (and making it slightly faster than before actually)

[psteinfeld]

I've been trying to build flang on my MacBook, but clang is quite unhappy with it right now. So many files (like flang/lib/Lower/ConvertExpr.cpp) are taking > 5min to compile (this a M2 MacBook Pro!), is there a known issue here?

I have an M2 MacBook Pro with 64G of RAM. I can build all of flang from scratch in a little over 17 minutes.