[llvm-dev] Musings on the TableGen -emit-dag-isel backend

[Paul C. Anagnostopoulos via llvm-dev]

A rather notorious aspect of TableGen is the time required to run the
-emit-dag-isel backend on some targets, including AMDGPU and X86. I added a
timing feature to TableGen and timed the AMDGPU run.

===-------------------------------------------------------------------------===
                             TableGen Phase Timing
===-------------------------------------------------------------------------===
  Total Execution Time: 733.6103 seconds (733.8740 wall clock)

   ---User Time---   --System Time--   --User+System--   ---Wall Time---  --- Name ---
  645.0017 ( 87.9%)   0.2340 (100.0%)  645.2357 ( 88.0%)  645.2709 ( 87.9%)  Emit matcher table
  70.4501 (  9.6%)   0.0000 (  0.0%)  70.4501 (  9.6%)  70.5510 (  9.6%)  Convert to matchers
  14.6329 (  2.0%)   0.0000 (  0.0%)  14.6329 (  2.0%)  14.7638 (  2.0%)  Parse, build records
   2.1996 (  0.3%)   0.0000 (  0.0%)   2.1996 (  0.3%)   2.1871 (  0.3%)  Sort patterns
   1.0920 (  0.1%)   0.0000 (  0.0%)   1.0920 (  0.1%)   1.0961 (  0.1%)  Optimize matchers
   0.0000 (  0.0%)   0.0000 (  0.0%)   0.0000 (  0.0%)   0.0050 (  0.0%)  Write output
  733.3763 (100.0%)   0.2340 (100.0%)  733.6103 (100.0%)  733.8740 (100.0%)  Total

As you can see, most of the time is spent emitting the C++ code. A simple
step toward reducing the time is to use the --omit-comments option. However,
I am informed that trying to debug the pattern matching table without
comments is a hopeless task.

===-------------------------------------------------------------------------===
                             TableGen Phase Timing
===-------------------------------------------------------------------------===
  Total Execution Time: 162.9274 seconds (162.9173 wall clock)

   ---User Time---   --System Time--   --User+System--   ---Wall Time---  --- Name ---
  75.0833 ( 46.1%)   0.0468 ( 42.9%)  75.1301 ( 46.1%)  75.1313 ( 46.1%)  Emit matcher table
  69.7948 ( 42.9%)   0.0000 (  0.0%)  69.7948 ( 42.8%)  69.8050 ( 42.8%)  Convert to matchers
  14.6173 (  9.0%)   0.0468 ( 42.9%)  14.6641 (  9.0%)  14.6668 (  9.0%)  Parse, build records
   2.2308 (  1.4%)   0.0000 (  0.0%)   2.2308 (  1.4%)   2.2191 (  1.4%)  Sort patterns
   1.0920 (  0.7%)   0.0000 (  0.0%)   1.0920 (  0.7%)   1.0921 (  0.7%)  Optimize matchers
   0.0000 (  0.0%)   0.0156 ( 14.3%)   0.0156 (  0.0%)   0.0030 (  0.0%)  Write output
  162.8182 (100.0%)   0.1092 (100.0%)  162.9274 (100.0%)  162.9173 (100.0%)  Total

Emitting the C++ code for most pattern operators is straightforward.
However, three operators are more time-consuming: Matcher::Scope,
SwitchOpcode, and SwitchType. These operators take a list of child patterns,
each of which is emitted as a 1- to 3-byte size followed by the child's
pattern bytes. The size is coded as a variable-length sequence of bytes,
as is every integer in the matcher table. (Just for interest, the average
number of children per Scope operator is about 2.7.)

In order to minimize the length of the size, the backend performs a sort of
relaxation algorithm, where it first tries a 1-byte size. If that fails, it
tries the actual required number of bytes. Trying involves calculating the
offset in the table for the child and then recursively generating the entire
child, passing it a string buffer as its output stream. When the size is
finally determined, that string buffer is appended to the actual buffer
passed to the generation function.

Why does the number of bytes in the size matter to the child? Because the
offset of the child is determined by that size, and the offset is passed to
the child and then included in many comments emitted by the child. If the
size is wrong, then the offset is wrong, and then the comments are wrong.

So it's clear that repetitive code generating is done because of the
relaxation algorithm. How bad is it? It turns out that the algorithm is
about O(1.7^(n-1)), where n is the depth of the pattern matching tree. The
depth of the AMDGPU tree is 13. Here are some interesting statistics:

Number of pattern operators at depth 11:         35,000
Number of times those operators are regenerated: 20 million

I think we have found the problem. So what can be done? I tried a quick and
dirty experiment. I forced all the child sizes to occupy a 3-byte length, so
that the relaxation algorithm was no longer necesseary. The results are
shown here.

===-------------------------------------------------------------------------===
                             TableGen Phase Timing
===-------------------------------------------------------------------------===
  Total Execution Time: 90.7302 seconds (90.8242 wall clock)

   ---User Time---   --System Time--   --User+System--   ---Wall Time---  --- Name ---
  69.7324 ( 76.9%)   0.0000 (  0.0%)  69.7324 ( 76.9%)  69.7320 ( 76.8%)  Convert to matchers
  14.5705 ( 16.1%)   0.0312 ( 33.3%)  14.6017 ( 16.1%)  14.6958 ( 16.2%)  Parse, build records
   3.0576 (  3.4%)   0.0624 ( 66.7%)   3.1200 (  3.4%)   3.1192 (  3.4%)  Emit matcher table
   2.1840 (  2.4%)   0.0000 (  0.0%)   2.1840 (  2.4%)   2.1891 (  2.4%)  Sort patterns
   1.0920 (  1.2%)   0.0000 (  0.0%)   1.0920 (  1.2%)   1.0831 (  1.2%)  Optimize matchers
   0.0000 (  0.0%)   0.0000 (  0.0%)   0.0000 (  0.0%)   0.0050 (  0.0%)  Write output
  90.6366 (100.0%)   0.0936 (100.0%)  90.7302 (100.0%)  90.8242 (100.0%)  Total

Now the matcher emitter phase is insignificant! Unfortunately, the size of
the matchter table increases from 451,430 bytes to 469,612 bytes, an
increase of 18,182 bytes or 4%.

So one solution is not to care about the 4% increase and always use 3-byte
child sizes. A second solution is to add an option that specifies the
starting number of child size bytes and retains the relaxation algorithm.
When building the system, we would specify --child-size-bytes=1. When building
for debugging, you could specify --child-size-bytes=3.

Comments and suggestion gratefully accepted.