Skip to content

Commit 6d84bf9

Browse files
RalfJungAmanieu
authored andcommitted
Rust does not let you observe or mutate the floating-point register in well-defined ways
1 parent ca1b140 commit 6d84bf9

File tree

1 file changed

+19
-0
lines changed
  • crates/core_arch/src/x86

1 file changed

+19
-0
lines changed

crates/core_arch/src/x86/sse.rs

Lines changed: 19 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1365,6 +1365,15 @@ pub unsafe fn _mm_sfence() {
13651365

13661366
/// Gets the unsigned 32-bit value of the MXCSR control and status register.
13671367
///
1368+
/// Note that Rust makes no guarantees whatsoever about the contents of this regiser: Rust
1369+
/// floating-point operations may or may not result in this register getting updated with exception
1370+
/// state, and the register can change between two invocations of this function even when no
1371+
/// floating-point operations appear in the source code (since floating-point operations appearing
1372+
/// earlier or later can be reordered).
1373+
///
1374+
/// If you need to perform some floating-point operations and check whether they raised an
1375+
/// exception, use an inline assembly block for the entire sequence of operations.
1376+
///
13681377
/// For more info see [`_mm_setcsr`](fn._mm_setcsr.html)
13691378
///
13701379
/// [Intel's documentation](https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_getcsr)
@@ -1401,6 +1410,16 @@ pub unsafe fn _mm_getcsr() -> u32 {
14011410
/// * The *denormals-are-zero mode flag* turns all numbers which would be
14021411
/// denormalized (exponent bits are all zeros) into zeros.
14031412
///
1413+
/// Note that modfying the masking flags, rounding mode, or denormals-are-zero mode flags leads to
1414+
/// **immediate Undefined Behavior**: Rust assumes that these are always in their default state and
1415+
/// will optimize accordingly. This even applies when the register is altered and later reset to its
1416+
/// original value without any floating-point operations appearing in the source code between those
1417+
/// operations (since floating-point operations appearing earlier or later can be reordered).
1418+
///
1419+
/// If you need to perform some floating-point operations under a different masking flags, rounding
1420+
/// mode, or denormals-are-zero mode, use an inline assembly block and make sure to restore the
1421+
/// original MXCSR register state before the end of the block.
1422+
///
14041423
/// ## Exception Flags
14051424
///
14061425
/// * `_MM_EXCEPT_INVALID`: An invalid operation was performed (e.g., dividing

0 commit comments

Comments
 (0)