https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106910
Bug ID: 106910
Summary: roundss not vectorized
Product: gcc
Version: 13.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: pilarlatiesa at gmail dot com
Target Milestone: ---
GCC 7 and newer optimize `std::floor(float)` into `vroundss` with -O2 and
-march=skylake, which is great.
However, I can see in Compiler Explorer that the following example:
```
#include <cmath>
struct TVec { float x, y; };
struct TKey { int i, j; };
class TDom
{
private:
static int Floor(float const x)
{ return static_cast<int>(std::floor(x)); }
public:
TKey CalcKey(TVec const &) const;
};
TKey TDom::CalcKey(TVec const &r) const
{ return {Floor(r.x), Floor(r.y)}; }
```
produces:
```
vxorps %xmm1, %xmm1, %xmm1
vroundss $9, (%rsi), %xmm1, %xmm0
vroundss $9, 4(%rsi), %xmm1, %xmm1
vunpcklps %xmm1, %xmm0, %xmm0
vcvttps2dq %xmm0, %xmm2
vmovq %xmm2, %rax
ret
```
Couldn’t the pair of `vroundss` have been merged into a single `vroundps`?
