Need better regression tests for the optimizer

[wingo]

Sometimes we find optimizer bugs that have a precise test case we can run. It would be good to be able to drop input and expected output files into a directory so we can test; e.g.

regression-tests/foo

#!/bin/sh
pflua-optimize <<END
{ "if",
  { "!=",
    "len",
    0 },
  { ">",
    1,
    { "uint32",
      { "+",
        { "uint32",
          { "-",
            4294967295,
            { "uint32",
              { "/",
                1,
                "len" } } } },
        1 } } },
  { "fail" } }
END

regression-tests/foo.expected

Optimizing:
{ "if",
  { "!=",
    "len",
    0 },
  { ">",
    1,
    { "uint32",
      { "+",
        { "uint32",
          { "-",
            4294967295,
            { "uint32",
              { "/",
                1,
                "len" } } } },
        1 } } },
  { "fail" } }
Result:
{ "if",
  { "!=",
    "len",
    0 },
  { "<",
    { "uint32",
      { "+",
        { "-",
          4294967295,
          { "/",
            1,
            "len" } },
        1 } },
    1 },
  { "fail" } }

Then the harness would run the input file, as an executable, and cmp its output to the expected output. If the optimizer changes we might get spurious diffs but we should be seeing the changes, and a simple makefile rule can regenerate expected output. Dunno. Point being, we should let the property tester find the bugs, and the regression tests keep them from happening again.

[kbara]

Yes; we need unit tests for a handful of known bugs that we've fixed, whether or not we fit them into the regressions test framework.

[wingo]

@kbara would you mind hacking up something for this? that would be swell :)

[kbara]

I'll think it over, then hack something up, yes.

[kbara]

@andywingo Are you happy with the state of tests/ir-reg/* for this?