diff options
Diffstat (limited to 'js/src/jit-test/tests/jaeger/loops/multiply-by-int32min.js')
-rw-r--r-- | js/src/jit-test/tests/jaeger/loops/multiply-by-int32min.js | 59 |
1 files changed, 59 insertions, 0 deletions
diff --git a/js/src/jit-test/tests/jaeger/loops/multiply-by-int32min.js b/js/src/jit-test/tests/jaeger/loops/multiply-by-int32min.js new file mode 100644 index 000000000..bc692f148 --- /dev/null +++ b/js/src/jit-test/tests/jaeger/loops/multiply-by-int32min.js @@ -0,0 +1,59 @@ +function foo() +{ + // Range analysis incorrectly computes a range for the multiplication. Once + // that incorrect range is computed, the goal is to compute a new value whose + // range analysis *thinks* is in int32_t range, but which goes past it using + // JS semantics. + // + // On the final iteration, in JS semantics, the multiplication produces 0, and + // the next addition 0x7fffffff. Adding any positive integer to that goes + // past int32_t range: here, (0x7fffffff + 5) or 2147483652. + // + // Range analysis instead thinks the multiplication produces a value in the + // range [INT32_MIN, INT32_MIN], and the next addition a value in the range + // [-1, -1]. Adding any positive value to that doesn't overflow int32_t range + // but *does* overflow the actual range in JS semantics. Thus omitting + // overflow checks produces the value 0x80000004, which interpreting as signed + // is (INT32_MIN + 4) or -2147483644. + // + // For this test to trigger the bug it was supposed to trigger: + // + // * 0x7fffffff must be the LHS, not RHS, of the addition in the loop, and + // * i must not be incremented using ++ + // + // The first is required because JM LoopState doesn't treat *both* V + mul and + // mul + V as not overflowing, when V is known to be int32_t -- only V + mul. + // (JM pessimally assumes V's type might change before it's evaluated. This + // obviously can't happen if V is a constant, but JM's puny little mind + // doesn't detect this possibility now.) + // + // The second is required because JM LoopState only ignores integer overflow + // on multiplications if the enclosing loop is a "constrainedLoop" (the name + // of the relevant field). Loops become unconstrained when unhandled ops are + // found in the loop. Increment operators generate a DUP op, which is not + // presently a handled op, causing the loop to become unconstrained. + for (var i = 0; i < 15; i = i + 1) { + var y = (0x7fffffff + ((i & 1) * -2147483648)) + 5; + } + return y; +} +assertEq(foo(), (0x7fffffff + ((14 & 1) * -2147483648)) + 5); + +function bar() +{ + // Variation on the theme of the above test with -1 as the other half of the + // INT32_MIN multiplication, which *should* result in -INT32_MIN on multiply + // (exceeding int32_t range). + // + // Here, range analysis again thinks the range of the multiplication is + // INT32_MIN. We'd overflow-check except that adding zero (on the LHS, see + // above) prevents overflow checking, so range analysis thinks the range is + // [INT32_MIN, INT32_MIN] when -INT32_MIN is actually possible. This direct + // result of the multiplication is already out of int32_t range, so no need to + // add anything to bias it outside int32_t range to get a wrong result. + for (var i = 0; i < 17; i = i + 1) { + var y = (0 + ((-1 + (i & 1)) * -2147483648)); + } + return y; +} +assertEq(bar(), (0 + ((-1 + (16 & 1)) * -2147483648))); |