Handle huge integers in OpenQASM 2 expression evaluator (#12140)

* Handle huge integers in OpenQASM 2 expression evaluator

This modifies the expression evaluator to directly parse the backing
string data of an integer token in a floating-point context, which lets
us handle numbers that would otherwise overflow a `usize`.  It's
possible for this to be totally valid, if, for example, the integer is a
multiple of some very large power of two that doesn't overflow a
double-precision float.

We already needed to immediately cast the integer to a float, so this is
just a more accurate way of doing the evaluation, and doesn't affect
when we use integers in other contexts.

* Clarify int/float split

crates/qasm2/src/expr.rs

@@ -501,8 +501,13 @@ impl<'a> ExprParser<'a> {
             | TokenType::Sin
             | TokenType::Sqrt
             | TokenType::Tan => Ok(Some(Atom::Function(token.ttype.into()))),
-            TokenType::Real => Ok(Some(Atom::Const(token.real(self.context)))),
-            TokenType::Integer => Ok(Some(Atom::Const(token.int(self.context) as f64))),
+            // This deliberately parses an _integer_ token as a float, since all OpenQASM 2.0
+            // integers can be interpreted as floats, and doing that allows us to gracefully handle
+            // cases where a huge float would overflow a `usize`.  Never mind that in such a case,
+            // there's almost certainly precision loss from the floating-point representating
+            // having insufficient mantissa digits to faithfully represent the angle mod 2pi;
+            // that's not our fault in the parser.
+            TokenType::Real | TokenType::Integer => Ok(Some(Atom::Const(token.real(self.context)))),
             TokenType::Pi => Ok(Some(Atom::Const(f64::consts::PI))),
             TokenType::Id => {
                 let id = token.text(self.context);
@@ -698,6 +703,11 @@ impl<'a> ExprParser<'a> {
 
     /// Parse a single expression completely. This is the only public entry point to the
     /// operator-precedence parser.
+    ///
+    /// .. note::
+    ///
+    ///     This evaluates in a floating-point context, including evaluating integer tokens, since
+    ///     the only places that expressions are valid in OpenQASM 2 is during gate applications.
     pub fn parse_expression(&mut self, cause: &Token) -> PyResult<Expr> {
         self.eval_expression(0, cause)
     }

crates/qasm2/src/lex.rs

@@ -262,9 +262,9 @@ impl Token {
     }
 
     /// If the token is a real number, this method can be called to evaluate its value.  Panics if
-    /// the token is not a real number.
+    /// the token is not a float or an integer.
     pub fn real(&self, context: &TokenContext) -> f64 {
-        if self.ttype != TokenType::Real {
+        if !(self.ttype == TokenType::Real || self.ttype == TokenType::Integer) {
             panic!()
         }
         context.text[self.index].parse().unwrap()

releasenotes/notes/qasm2-bigint-8eff42acb67903e6.yaml

@@ -0,0 +1,9 @@
+---
+fixes:
+  - |
+    The OpenQASM 2.0 parser (:func:`.qasm2.load` and :func:`.qasm2.loads`) can now evaluate
+    gate-angle expressions including integer operands that would overflow the system-size integer.
+    These will be evaluated in a double-precision floating-point context, just like the rest of the
+    expression always has been.  Beware: an arbitrarily large integer will not necessarily be
+    exactly representable in double-precision floating-point, so there is a chance that however the
+    circuit was generated, it had already lost all numerical precision modulo :math:`2\pi`.

test/python/qasm2/test_expression.py

@@ -123,6 +123,18 @@ def test_function_symbolic(self, function_str, function_py):
         actual = [float(x) for x in abstract_op.definition.data[0].operation.params]
         self.assertEqual(list(actual), expected)
 
+    def test_bigint(self):
+        """Test that an expression can be evaluated even if it contains an integer that will
+        overflow the integer handling."""
+        bigint = 1 << 200
+        # Sanity check that the number we're trying for is represented at full precision in floating
+        # point (which it should be - it's a power of two with fewer than 11 bits of exponent).
+        self.assertEqual(int(float(bigint)), bigint)
+        program = f"qreg q[1]; U({bigint}, -{bigint}, {bigint} * 2.0) q[0];"
+        parsed = qiskit.qasm2.loads(program)
+        parameters = list(parsed.data[0].operation.params)
+        self.assertEqual([bigint, -bigint, 2 * bigint], parameters)
+
 
 class TestPrecedenceAssociativity(QiskitTestCase):
     def test_precedence(self):