Support signed int divisor in onnx.Mod op

[negiyas]

This PR supports signed int divisor in onnx.Mod op as well as unsigned int divisor.　This PR includes the conversion pass, its lit and backend tests. The following comments explain difference between return values of math.rem and onnx.Mod op.

    // "math.rem" returns "minus" for minus dividend and "plus or zero" for plus dividend.
    // We call the math.rem's return value "mathRemaider".
    // However onnx.ModOp should return "minus" for minus divisor and "plus or zero" for plus divisor.
    // we call the value that onnx.Mod op should return "onnxMod". The following table shows mathRemainder,
    // onnxMod and their diference (=onnxMod-mathRemainder) for some inputs.
    //
    // dividend                |  7  |  7 | -7 | -7 |  6 |  6 | -6 | -6 |
    // divisor                 |  3  | -3 |  3 | -3 |  3 | -3 |  3 | -3 |
    // ------------------------+-----+----+----+----+----+----+----+----+
    // mathRemainder           |  1  |  1 | -1 | -1 |  0 |  0 |  0 |  0 |
    // onnxMod                 |  1  | -2 |  2 | -1 |  0 |  0 |  0 |  0 |
    // onnxMod - mathRemainder |  0  | -3 |  3 |  0 |  0 |  0 |  0 |  0 |
    //
    // The following code shows logic to get onnxMod from mathRemainder
    //
    // int dividend, divisor;
    // int mathRemainder = diviend % divisor;
    // int adjustedRemainder = mathRemainder + divisor;
    //
    // if ((mathRemainder != 0) && ((dividend < 0) ^ (divisor < 0))) # c.f. "^" shows "exclusive or".
    //   return adjustedRemainder;
    // return mathRemainder;

[chentong319]

Should we use llvm srem?

[negiyas]

Should we use llvm srem?
@chentong319 Thank you for useful comments.

I investigated definitions and behavior of llvm.srem, and arith.remsi (https://mlir.llvm.org/docs/Dialects/ArithOps/#arithremsi-arithremsiop), that is converted into llvm.srem. According to the investigation, these operations support signed integers, but they cannot generate right answers when the divisor is minus. If we simply use arith.remsi for onnx.Mod, the backend tests ("test_mod_mixed_sign_int64_cpu", "test_mod_mixed_sign_int32_cpu") fail as follows.

 Mismatched elements: 2 / 6 (33.3%)
Max absolute difference: 3
Max relative difference: 1.5
 x: array([ 0,  1,  5,  0, -1,  3], dtype=int32)
 y: array([ 0, -2,  5,  0,  2,  3], dtype=int32)

Accordingly, it seems that additional logic in this PR is necessary to generate right answers for the divisor-is-minus cases.
Your comments to simplify the code are very welcome! Thank you.

[chentong319]

LGTM!

[chentong319]

The code is little tricky. To understand it, I list out the case and result according to your code (NOT the comment):

1. dividend > 0 and divisor >0, rem unchanged
2. dividend < 0 and divisor <0, rem = (rem+divisor) + divisor = rem + 2*divisor
3. dividend < 0 and divisor > 0, rem = rem + divisor = rem + |divisor|
4. dividend > 0 and divisor < 0, rem = rem+divisor = rem - |divisor|

Is the second case correct? Please fix the difference of the comment and code.
Another question: do we need to check eq to zero again at line 1156? If the original rem is not zero, it will be zero after the added calculation.

[negiyas]

@chentong319 I updated the explanation and code for the part.
Thank you for useful comments. Can you kindly check the comments and code again?

I have questions in details

[chentong319]

LGTM!

[jenkins-droid]

Jenkins Linux amd64 Build #12613 [push] Support signed int divis... started at 01:29

[jenkins-droid]

Jenkins Linux s390x Build #12625 [push] Support signed int divis... started at 02:29

[jenkins-droid]

Jenkins Linux ppc64le Build #11618 [push] Support signed int divis... started at 02:38

[jenkins-droid]

Jenkins Linux amd64 Build #12613 [push] Support signed int divis... passed after 1 hr 0 min

[jenkins-droid]

Jenkins Linux s390x Build #12625 [push] Support signed int divis... passed after 1 hr 24 min

[jenkins-droid]

Jenkins Linux ppc64le Build #11618 [push] Support signed int divis... passed after 1 hr 46 min