T0b1/wasmtime
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Fix a corner case in fcvt_to_sint.i32.f64 legalization.

Browse Source 
An f64 can represent multiple values in the range INT_MIN-1 < x <=
INT_MIN which all truncate to INT_MIN, so comparing the input value
against INT_MIN is not good enough.

Instead, detect overflow on x <= INT_MIN-1 when INT_MIN-1 is an exact
floating point value.
This commit is contained in:
Jakob Stoklund Olesen
2017-09-28 14:24:39 -07:00
parent 8abcdac5a1
commit 2888ff5bf3
2 changed files with 19 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

24
lib/cretonne/src/isa/intel/enc_tables.rs
View File

@@ -222,15 +222,25 @@ fn expand_fcvt_to_sint(inst: ir::Inst, func: &mut ir::Function, cfg: &mut Contro
);
// Check for case 1: INT_MIN is the correct result.
// We use a `ueq` condition here because that can be translated into a single branch, and we
// already know that we don't have a NaN.
let fintmin = match xty {
ir::types::F32 => pos.ins().f32const(Ieee32::pow2(ty.lane_bits() - 1).neg()),
ir::types::F64 => pos.ins().f64const(Ieee64::pow2(ty.lane_bits() - 1).neg()),
// Determine the smallest floating point number that would convert to INT_MIN.
let mut overflow_cc = FloatCC::LessThan;
let output_bits = ty.lane_bits();
let flimit = match xty {
ir::types::F32 => pos.ins().f32const(Ieee32::pow2(output_bits - 1).neg()),
ir::types::F64 => {
// An f64 can represent `i32::min_value() - 1` exactly with precision to spare, so
// there are values less than -2^(N-1) that convert correctly to INT_MIN.
pos.ins().f64const(if output_bits < 64 {
overflow_cc = FloatCC::LessThanOrEqual;
Ieee64::with_float(-((1u64 << (output_bits - 1)) as f64) - 1.0)
} else {
Ieee64::pow2(output_bits - 1).neg()
})
}
_ => panic!("Can't convert {}", xty),
};
let in_range = pos.ins().fcmp(FloatCC::UnorderedOrEqual, x, fintmin);
pos.ins().trapz(in_range, ir::TrapCode::IntegerOverflow);
let overflow = pos.ins().fcmp(overflow_cc, x, flimit);
pos.ins().trapnz(overflow, ir::TrapCode::IntegerOverflow);
pos.ins().jump(done, &[]);
pos.insert_ebb(done);