[profiling] Review the existing DXPAIRS instrumentation

[ericsnowcurrently]

If "DYNAMIC_EXECUTION_PROFILE" and "DXPAIRS" are defined then in the eval loop we keep a count for each opcode pair encountered (i.e. a 256x256 array of int). This is exposed by sys.getdxp() and there's a module to help review the results (Tools/scripts/analyze_dxp.py).

This information may be useful, particularly relative to work on superinstructions (#16), if gathered against meaningful workloads.

[ericsnowcurrently]

Here's a patch I was using (via `python3 -m pyperformance compile --patch ...`):

diff --git a/Include/Python.h b/Include/Python.h
index 86dbbcf6bd..48e26491d4 100644
--- a/Include/Python.h
+++ b/Include/Python.h
@@ -2,6 +2,9 @@
 #define Py_PYTHON_H
 /* Since this is a "meta-include" file, no #ifdef __cplusplus / extern "C" { */

+#define DYNAMIC_EXECUTION_PROFILE 1
+#define DXPAIRS 1
+
 /* Include nearly all Python header files */

 #include "patchlevel.h"
diff --git a/Modules/main.c b/Modules/main.c
index 2684d23067..28137d7509 100644
--- a/Modules/main.c
+++ b/Modules/main.c
@@ -658,6 +658,27 @@ pymain_exit_error(PyStatus status)
 }


+static const char *dump_dxp_script = "\
+import sys\n\
+\n\
+dxp = sys.getdxp()\n\
+\n\
+import json\n\
+import os\n\
+import os.path\n\
+\n\
+outdir = os.environ.get('ARTIFACTS_DIR') or '.'\n\
+filename = os.path.join(outdir, 'dxp.json')\n\
+with open(filename, 'w') as outfile:\n\
+    json.dump(dxp, outfile)\n\
+";
+
+static void dump_dxp(void)
+{
+    PyRun_SimpleString(dump_dxp_script);
+    PyErr_Clear();
+}
+
 int
 Py_RunMain(void)
 {
@@ -665,6 +686,8 @@ Py_RunMain(void)

     pymain_run_python(&exitcode);

+    dump_dxp();
+
     if (Py_FinalizeEx() < 0) {
         /* Value unlikely to be confused with a non-error exit status or
            other special meaning */

[ericsnowcurrently]

I have a branch that updates the existing script: https://github.com/ericsnowcurrently/cpython/tree/update-dxp-script.

[ericsnowcurrently]

Here are results from a short run (only runtime startup?):

=============
== Summary ==
=============

- Usage -

total:                                             678,134

Top 10 pairs:
  LOAD_FAST            --> LOAD_CONST               70,467
  STORE_FAST           --> LOAD_FAST                69,743
  STORE_FAST           --> STORE_FAST               68,203
  COMPARE_OP           --> POP_JUMP_IF_FALSE        67,497
  UNPACK_SEQUENCE      --> STORE_FAST               67,291
  LOAD_CONST           --> COMPARE_OP               67,225
  FOR_ITER             --> UNPACK_SEQUENCE          67,123
  POP_JUMP_IF_FALSE    --> FOR_ITER                 64,755
  POP_JUMP_IF_FALSE    --> LOAD_FAST                 6,658
  LOAD_FAST            --> LOAD_ATTR                 5,416

Top 10 op1:
  STORE_FAST                                       143,169
  LOAD_FAST                                        100,143
  LOAD_CONST                                        79,106
  POP_JUMP_IF_FALSE                                 74,098
  FOR_ITER                                          69,028
  COMPARE_OP                                        67,918
  UNPACK_SEQUENCE                                   67,336
  LOAD_GLOBAL                                       10,387
  LOAD_ATTR                                          8,421
  CALL_FUNCTION                                      5,363

Top 10 op2:
  STORE_FAST                                       143,169
  LOAD_FAST                                        100,143
  LOAD_CONST                                        79,106
  POP_JUMP_IF_FALSE                                 74,098
  FOR_ITER                                          69,028
  COMPARE_OP                                        67,918
  UNPACK_SEQUENCE                                   67,336
  LOAD_GLOBAL                                       10,387
  LOAD_ATTR                                          8,422
  CALL_FUNCTION                                      5,361

- Pairs -

total:                 661 / 65,536
total op1:              84 / 256
total op2:              86 / 256

Top 10 op1:
  LOAD_FAST             52
  LOAD_CONST            43
  LOAD_ATTR             36
  CALL_METHOD           29
  CALL_FUNCTION         28
  LOAD_GLOBAL           21
  BUILD_TUPLE           19
  STORE_NAME            16
  LOAD_DEREF            16
  POP_JUMP_IF_FALSE     15

Top 10 op2:
  LOAD_CONST            49
  STORE_FAST            42
  LOAD_FAST             39
  LOAD_GLOBAL           29
  CALL_FUNCTION         26
  EXTENDED_ARG          24
  RETURN_VALUE          23
  LOAD_NAME             19
  STORE_NAME            17
  CALL_METHOD           17

[gvanrossum]

Curious how this would change after merging my add-opcodes branch.

[gvanrossum]

I have a branch that updates the existing script: https://github.com/ericsnowcurrently/cpython/tree/update-dxp-script.

What I would love to see (not sure if your version does this yet): allow running the script as a main followed by arguments that make it run another module or script. See example in prof.py.

[ericsnowcurrently]

I'm closing this in favor of #44.