strange usage of pointers in sin/cos optimization

[davemilter]

With such code, clang (17.0.1) able to optimize (-fno-math-errno -O3)
sin/cos calls to one call of sincos:

#include <cmath>

struct Point3D {
    double x;
    double y;
    double z;
};

Point3D f(double lat, double lon) noexcept
{
   double lat_sin = sin(lat), lat_cos = cos(lat);
   double lon_sin = sin(lon), lon_cos = cos(lon);
   
   return {lat_cos * lon_cos , lat_cos * lon_sin, lat_sin};
}

but for some reason it can not use pointers properly:

        lea     rdi, [rsp + 40]
        lea     rsi, [rsp + 32]
        call    sincos@PLT
        movsd   xmm0, qword ptr [rsp + 40]      # xmm0 = mem[0],zero
        movsd   qword ptr [rsp + 8], xmm0       # 8-byte Spill
        lea     rdi, [rsp + 24]
        lea     rsi, [rsp + 16]
        movsd   xmm0, qword ptr [rsp]           # 8-byte Reload
        call    sincos@PLT

clang/llvm passes pointer to sincos via [rsp + 40] and [rsp + 32],
then move stored value to [rsp + 8] and [rsp].

If it can use [rsp + 8] and [rsp] directly,
this removes at least two extra save and loads.

godbolt.org link

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: None (davemilter)

With such code, clang (17.0.1) able to optimize (-fno-math-errno -O3) sin/cos calls to one call of sincos:

#include <cmath>

struct Point3D {
    double x;
    double y;
    double z;
};

Point3D f(double lat, double lon) noexcept
{
   double lat_sin = sin(lat), lat_cos = cos(lat);
   double lon_sin = sin(lon), lon_cos = cos(lon);
   
   return {lat_cos * lon_cos , lat_cos * lon_sin, lat_sin};
}

but for some reason it can not use pointers properly:

        lea     rdi, [rsp + 40]
        lea     rsi, [rsp + 32]
        call    sincos@<!-- -->PLT
        movsd   xmm0, qword ptr [rsp + 40]      # xmm0 = mem[0],zero
        movsd   qword ptr [rsp + 8], xmm0       # 8-byte Spill
        lea     rdi, [rsp + 24]
        lea     rsi, [rsp + 16]
        movsd   xmm0, qword ptr [rsp]           # 8-byte Reload
        call    sincos@<!-- -->PLT

clang/llvm passes pointer to sincos via [rsp + 40] and [rsp + 32],
then move stored value to [rsp + 8] and [rsp].

If it can use [rsp + 8] and [rsp] directly,
this removes at least two extra save and loads.

godbolt.org link

[davemilter]

@dtcxzyw , not sure about backend:x86,

for armv8-a direct usage of sincos also gives benefits.
clang generates only 22 instruction instead of 25,
and use less stack space:

armv8-a godbolt link