Fix constant folding of booleans casted to integers

[annagrin]

Description

Fix constant folding of casts of integers.

Using the MLIR below

%0 = arith.constant <c> : i8
%1 = cc.cast unsigned %0: i32

The constant folding creates the following code without changing <c>:

%1 = arith.constant <c>: i32

This leads to incorrect values in constants when constant folding is invoked, for example,
(unsigned)(true) is currently converted to a -1 integer constant.

See added tests in targettests/execution/cast.cpp for more examples.

This PR updates constant folding to cast the source of the cast instruction to its original type before folding into a constant of a new type.

Example:

struct testConstBool {
  auto operator()() __qpu__ {
    cudaq::qubit q;
    unsigned i0 = (unsigned)(true);
    if (i0 == 1) {
      x(q);
    }
  }
};

int main() {
  auto counts = cudaq::sample(testConstBool{});
  counts.dump();
}

Quake before the fix:

func.func @__nvqpp__mlirgen__testConstBool () attributes {"cudaq-entrypoint", "cudaq-kernel"} {
    %c-1_i32 = arith.constant -1 : i32
    %c1_i32 = arith.constant 1 : i32
    %0 = quake.alloca !quake.ref
    %1 = cc.alloca i32
    cc.store %c-1_i32, %1 : !cc.ptr<i32>
    %2 = cc.load %1 : !cc.ptr<i32>
    %3 = arith.cmpi eq, %2, %c1_i32 : i32
    cc.if(%3) {
      quake.x %0 : (!quake.ref) -> ()
    }
    return
  }

Quake after the fix:

func.func @__nvqpp__mlirgen__testConstBool () attributes {"cudaq-entrypoint", "cudaq-kernel"} {
    %c1_i32 = arith.constant 1 : i32
    %0 = quake.alloca !quake.ref
    %1 = cc.alloca i32
    cc.store %c1_i32, %1 : !cc.ptr<i32>
    %2 = cc.load %1 : !cc.ptr<i32>
    %3 = arith.cmpi eq, %2, %c1_i32 : i32
    cc.if(%3) {
      quake.x %0 : (!quake.ref) -> ()
    }
    return
  }

Result before the fix:
{ 0:0000 }

Result after the fix:
{ 1:0000 }

Closes: #2804

[github-actions]

CUDA Quantum Docs Bot: A preview of the documentation can be found here.

[nvidia-dobri]

Change looks good, might be useful to know the context of when this is important / what the downstream breakage we saw was.

I'm a bit confused by the statement:

(unsigned)(true) is currently converted to a -1 constant in MLIR

What does -1 in unsigned mean? U..._MAX?

[annagrin]

Change looks good, might be useful to know the context of when this is important / what the downstream breakage we saw was.

I included an example in the description and the reference to the bug it fixes

I'm a bit confused by the statement:

(unsigned)(true) is currently converted to a -1 constant in MLIR

What does -1 in unsigned mean? U..._MAX?

Constant folding folds the cast, so in this particular example we end up with an integer constant with value of -1, of type i32 (see example in the description). It would behave the same as uint32 max, yes, which is incorrect in this case, we should const fold it to arith.constant 1 instead.

[github-actions]

CUDA Quantum Docs Bot: A preview of the documentation can be found here.

[github-actions]

CUDA Quantum Docs Bot: A preview of the documentation can be found here.

[github-actions]

CUDA Quantum Docs Bot: A preview of the documentation can be found here.

[github-actions]

CUDA Quantum Docs Bot: A preview of the documentation can be found here.

[github-actions]

CUDA Quantum Docs Bot: A preview of the documentation can be found here.

[github-actions]

CUDA Quantum Docs Bot: A preview of the documentation can be found here.

[github-actions]

CUDA Quantum Docs Bot: A preview of the documentation can be found here.

[github-actions]

CUDA Quantum Docs Bot: A preview of the documentation can be found here.

[schweitzpgi]

Thanks for all the tests!

[github-actions]

CUDA Quantum Docs Bot: A preview of the documentation can be found here.

[github-actions]

CUDA Quantum Docs Bot: A preview of the documentation can be found here.

[github-actions]

CUDA Quantum Docs Bot: A preview of the documentation can be found here.