Opened 13 months ago

Last modified 12 months ago

#15929 new task

Explore whether adding XRay attributes to LLVM IR is worthwhile

Reported by: mpickering Owned by:
Priority: normal Milestone:
Component: Compiler Version: 8.6.2
Keywords: newcomer Cc: adamse
Operating System: Unknown/Multiple Architecture: Unknown/Multiple
Type of failure: None/Unknown Test Case:
Blocked By: Blocking:
Related Tickets: Differential Rev(s):
Wiki Page:


LLVM has support for XRay which is a tracing framework developed by Google. It might be possible to add the relevant function attributes to the IR we generate to get access to detailed traces which can be analysed using existing tooling.

Change History (5)

comment:1 Changed 13 months ago by mpickering

I tried this out. The diff to add support was very simple.

I then tried to compile the following program and to enable the XRay trace.

module Main where                                                               
main :: IO ()                                                                   
main = print "abc" >> print (f 1000) >> print (g 1000) >> print (qux B)         
data T = A | B deriving Show                                                    
f 0 = 1                                                                         
f x = f (x -1) + 1                                                              
g 1 = 0                                                                         
g x = g (x - 1) + f (x - 1)                                                     
{-# NOINLINE qux #-}                                                            
qux A = A                                                                       
qux B = qux A 

The first thing to note was that it was necessary to use clang-7.0.0 which I achieved by using this diff to ghc.nix.

Then I compiled with the following sequence of commands:

inplace/bin/ghc-stage2 -fllvm -ddump-llvm llvm.hs -opta="-v" -pgma=clang -pgmc=clang -pgml=clang -optl="-fxray-instrument" -optl="-fxray-instruction-threshold=1" -O2 -fforce-recomp
XRAY_OPTIONS="patch_premain=true xray_mode=xray-basic verbosity=1" ./llvm

Then the log file was analysed using the following command

llvm-xray account  -sort=sum -sortorder=dsc -instr_map ./llvm  xray-log.llvm.XEyh0y 

Which produced the very unhelpful profiling output.

Functions with latencies: 7
   funcid      count [      min,       med,       90p,       99p,       max]       sum  function    
        7         18 [ 0.000006,  0.000006,  0.000021,  0.000040,  0.000040]  0.000175  <invalid>:0:0: @(422fd0)
        5         14 [ 0.000006,  0.000006,  0.000014,  0.000021,  0.000021]  0.000110  <invalid>:0:0: @(422e90)
       10          8 [ 0.000006,  0.000006,  0.000042,  0.000042,  0.000042]  0.000104  <invalid>:0:0: @(4231a8)
        8         10 [ 0.000006,  0.000006,  0.000040,  0.000040,  0.000040]  0.000096  <invalid>:0:0: @(423070)
        6         14 [ 0.000006,  0.000006,  0.000006,  0.000006,  0.000006]  0.000086  <invalid>:0:0: @(422f30)
       11         13 [ 0.000006,  0.000006,  0.000006,  0.000008,  0.000008]  0.000082  <invalid>:0:0: @(423260)
        9          7 [ 0.000006,  0.000006,  0.000014,  0.000014,  0.000014]  0.000055  <invalid>:0:0: @(423110)

Should be easy now for someone else more familiar with LLVM to investigate.

comment:2 Changed 13 months ago by mpickering

Keywords: newcomer added

comment:3 Changed 12 months ago by bgamari

Milestone: 8.6.3

Ticket retargeted after milestone closed

comment:4 Changed 12 months ago by mpickering

I investigated some more and get the basic support working. None of the call stack stuff is useful for GHC as every call is a tail call so the call hierarchy is completely flat.

Here is the branch:

See this thread on llvm-devs for some of the problems and suggestions of the LLVM devs:

Here's an example comment I was using to compile which correctly instrumented a binary.

_build/stage1/bin/ghc -fllvm -ddump-llvm llvm.hs -opta="-v" -pgma=clang -pgmc=clang -pgml=clang -optl="-fxray-instrument" -optl="-fxray-instruction-threshold=1" -optlo="-O0" -O2 -g3 -fforce-recomp nofib/real/fulsom/Main.hs  -inofib/real/fulsom/


  1. You have to pass -O0 to opt as otherwise the LLVM verifier complains about our debugging information. See the email list for more information about this.
  2. It didn't seem like every call was actually counted in some simple tests but I didn't look too closely.

comment:5 Changed 12 months ago by adamse

Cc: adamse added
Note: See TracTickets for help on using tickets.