Consider not running the verifier in non-debug builds

[pcwalton]

The LLVM verifier is the second slowest pass after LiveIRVariables (2.4%/2.085 s). It's unnecessary and rarely catches stuff that the assertions don't. We should only run it in debug rustc builds.

[metajack]

no-verify is already a debugging option, so I assume the compiler just needs some simple logic to add this to the driver in non-debug builds.

[bblum]

Related #3615. Nominating production-ready.

[graydon]

just a bug, removing milestone/nomination.

[huonw]

Triage: I assume that the verifier is Module Verifier in the output of -Z time-llvm-passes? If so, then the current timings are:

no-opt

   ---User Time---   --System Time--   --User+System--   ---Wall Time---  --- Name ---
  17.7440 ( 56.5%)   1.7320 ( 53.7%)  19.4760 ( 56.2%)  19.4380 ( 56.1%)  X86 DAG->DAG Instruction Selection
   3.0160 (  9.6%)   0.3800 ( 11.8%)   3.3960 (  9.8%)   3.4440 (  9.9%)  X86 Assembly / Object Emitter
   1.8000 (  5.7%)   0.1880 (  5.8%)   1.9880 (  5.7%)   1.9859 (  5.7%)  Fast Register Allocator
   1.6320 (  5.2%)   0.1480 (  4.6%)   1.7800 (  5.1%)   1.7558 (  5.1%)  Inliner for always_inline functions
   1.3880 (  4.4%)   0.1560 (  4.8%)   1.5440 (  4.5%)   1.5803 (  4.6%)  Module Verifier
   ... 2 passes ...
   0.6320 (  2.0%)   0.0600 (  1.9%)   0.6920 (  2.0%)   0.6389 (  1.8%)  Module Verifier
   ...

86s for the entire rustc invocation.

-O

   ---User Time---   --System Time--   --User+System--   ---Wall Time---  --- Name ---
  32.9560 ( 21.3%)   3.5840 ( 31.7%)  36.5400 ( 22.0%)  36.4753 ( 21.9%)  X86 DAG->DAG Instruction Selection
  11.2080 (  7.2%)   0.5440 (  4.8%)  11.7520 (  7.1%)  11.6393 (  7.0%)  Function Integration/Inlining
  10.2760 (  6.6%)   0.5080 (  4.5%)  10.7840 (  6.5%)  10.9164 (  6.6%)  Global Value Numbering
   7.4720 (  4.8%)   0.2560 (  2.3%)   7.7280 (  4.7%)   7.7136 (  4.6%)  Combine redundant instructions
   ... 35 passes ...
   0.6880 (  0.4%)   0.0960 (  0.8%)   0.7840 (  0.5%)   0.8407 (  0.5%)  Module Verifier
   ... 6 passes ...
   0.5800 (  0.4%)   0.0720 (  0.6%)   0.6520 (  0.4%)   0.6933 (  0.4%)  Module Verifier
   ... 6 passes ...
   0.5720 (  0.4%)   0.0400 (  0.4%)   0.6120 (  0.4%)   0.5771 (  0.3%)  Module Verifier
   ...

245s for the entire invocation.

[ahmedcharles]

Is this still relevant?

[thestinger]

Yes, it's still relevant. I think we do want to keep running this for now though.

[steveklabnik]

Triage:

-Z time-llvm-passes 2>&1 | grep "Module Verifier"

   0.1400 (  9.3%)   0.0360 ( 12.5%)   0.1760 (  9.8%)   0.1131 (  6.5%)  Module Verifier
   0.0640 (  4.2%)   0.0120 (  4.2%)   0.0760 (  4.2%)   0.0726 (  4.2%)  Module Verifier
   0.0680 (  4.5%)   0.0040 (  1.4%)   0.0720 (  4.0%)   0.0712 (  4.1%)  Module Verifier

-O -Z time-llvm-passes 2>&1| grep "Module Verifier"

   0.1240 (  1.1%)   0.0040 (  1.4%)   0.1280 (  1.2%)   0.1158 (  1.0%)  Module Verifier
   0.0560 (  0.5%)   0.0000 (  0.0%)   0.0560 (  0.5%)   0.0609 (  0.5%)  Module Verifier
   0.0440 (  0.4%)   0.0000 (  0.0%)   0.0440 (  0.4%)   0.0499 (  0.4%)  Module Verifier

tested with wtftw, takes 17s total on my machine

[brson]

It's a small improvement. Still doesn't seem worth turning it off to me. Are we ever going to want to turn off the LLVM module verifier? Can we close this issue?

cc @rust-lang/compiler

[brson]

You snooze you lose!