[libc-commits] [libc] [libc] Improve nearest_integer for non-AVX x86 targets. (PR #230529)

[Pavel Labath via libc-commits]

================
@@ -26,18 +26,36 @@
 namespace LIBC_NAMESPACE_DECL {
 namespace fputil {
 
+// Without AVX, `roundss` and `roundsd` are 2-operand instructions that preserve
----------------
labath wrote:

At least to me, this reads as if the behaviour of roundss changes depending on whether you have AVX, which doesn't sound right. What I think you mean is that without AVX, we only have this instruction with a particular quirk (i.e., we cannot use `vroundss`).

https://github.com/llvm/llvm-project/pull/230529