https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112824
Bug ID: 112824
Summary: Stack spills and vector splitting with vector builtins
Product: gcc
Version: 14.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: elrodc at gmail dot com
Target Milestone: ---
I am not sure which component to place this under, but selected
tree-optimization as I suspect this is some sort of alias analysis failure
preventing the removal of stack allocations.
Godbolt link, reproduces on GCC trunk and 13.2:
https://godbolt.org/z/4TPx17Mbn
Clang has similar problems in my actual test case, but they don't show up in
this minimal example I made. Although Clang isn't perfect here either: it fails
to fuse fmadd + masked vmovapd, while GCC does succeed in fusing them.
For reference, code behind the godbolt link is:
#include <bit>
#include <concepts>
#include <cstddef>
#include <cstdint>
template <ptrdiff_t W, typename T>
using Vec [[gnu::vector_size(W * sizeof(T))]] = T;
// Omitted: 16 without AVX, 32 without AVX512F,
// or for forward compatibility some AVX10 may also mean 32-only
static constexpr ptrdiff_t VectorBytes = 64;
template<typename T>
static constexpr ptrdiff_t VecWidth = 64 <= sizeof(T) ? 1 : 64/sizeof(T);
template <typename T, ptrdiff_t N> struct Vector{
static constexpr ptrdiff_t L = N;
T data[L];
static constexpr auto size()->ptrdiff_t{return N;}
};
template <std::floating_point T, ptrdiff_t N> struct Vector<T,N>{
static constexpr ptrdiff_t W = N >= VecWidth<T> ? VecWidth<T> :
ptrdiff_t(std::bit_ceil(size_t(N)));
static constexpr ptrdiff_t L = (N/W) + ((N%W)!=0);
using V = Vec<W,T>;
V data[L];
static constexpr auto size()->ptrdiff_t{return N;}
};
/// should be trivially copyable
/// codegen is worse when passing by value, even though it seems like it should
make
/// aliasing simpler to analyze?
template<typename T, ptrdiff_t N>
[[gnu::always_inline]] constexpr auto operator+(Vector<T,N> x, Vector<T,N> y)
-> Vector<T,N> {
Vector<T,N> z;
for (ptrdiff_t n = 0; n < Vector<T,N>::L; ++n) z.data[n] = x.data[n] +
y.data[n];
return z;
}
template<typename T, ptrdiff_t N>
[[gnu::always_inline]] constexpr auto operator*(Vector<T,N> x, Vector<T,N> y)
-> Vector<T,N> {
Vector<T,N> z;
for (ptrdiff_t n = 0; n < Vector<T,N>::L; ++n) z.data[n] = x.data[n] *
y.data[n];
return z;
}
template<typename T, ptrdiff_t N>
[[gnu::always_inline]] constexpr auto operator+(T x, Vector<T,N> y) ->
Vector<T,N> {
Vector<T,N> z;
for (ptrdiff_t n = 0; n < Vector<T,N>::L; ++n) z.data[n] = x + y.data[n];
return z;
}
template<typename T, ptrdiff_t N>
[[gnu::always_inline]] constexpr auto operator*(T x, Vector<T,N> y) ->
Vector<T,N> {
Vector<T,N> z;
for (ptrdiff_t n = 0; n < Vector<T,N>::L; ++n) z.data[n] = x * y.data[n];
return z;
}
template <typename T, ptrdiff_t N> struct Dual {
T value;
Vector<T, N> partials;
};
// Here we have a specialization for non-power-of-2 `N`
template <typename T, ptrdiff_t N>
requires(std::floating_point<T> && (std::popcount(size_t(N))>1))
struct Dual<T,N> {
Vector<T, N+1> data;
};
template<ptrdiff_t W, typename T>
consteval auto firstoff(){
static_assert(std::same_as<T,double>, "type not implemented");
if constexpr (W==2) return Vec<2,int64_t>{0,1} != 0;
else if constexpr (W == 4) return Vec<4,int64_t>{0,1,2,3} != 0;
else if constexpr (W == 8) return Vec<8,int64_t>{0,1,2,3,4,5,6,7} != 0;
else static_assert(false, "vector width not implemented");
}
template <typename T, ptrdiff_t N>
[[gnu::always_inline]] constexpr auto operator+(Dual<T, N> a,
Dual<T, N> b)
-> Dual<T, N> {
if constexpr (std::floating_point<T> && (std::popcount(size_t(N))>1)){
Dual<T,N> c;
for (ptrdiff_t l = 0; l < Vector<T,N>::L; ++l)
c.data.data[l] = a.data.data[l] + b.data.data[l];
return c;
} else return {a.value + b.value, a.partials + b.partials};
}
template <typename T, ptrdiff_t N>
[[gnu::always_inline]] constexpr auto operator*(Dual<T, N> a,
Dual<T, N> b)
-> Dual<T, N> {
if constexpr (std::floating_point<T> && (std::popcount(size_t(N))>1)){
using V = typename Vector<T,N>::V;
V va = V{}+a.data.data[0][0], vb = V{}+b.data.data[0][0];
V x = va * b.data.data[0];
Dual<T,N> c;
c.data.data[0] = firstoff<Vector<T,N>::W,T>() ? x + vb*a.data.data[0] : x;
for (ptrdiff_t l = 1; l < Vector<T,N>::L; ++l)
c.data.data[l] = va*b.data.data[l] + vb*a.data.data[l];
return c;
} else return {a.value * b.value, a.value * b.partials + b.value *
a.partials};
}
void prod(Dual<Dual<double,7>,2> &c, const Dual<Dual<double,7>,2> &a, const
Dual<Dual<double,7>,2>&b){
c = a*b;
}
void prod(Dual<Dual<double,8>,2> &c, const Dual<Dual<double,8>,2> &a, const
Dual<Dual<double,8>,2>&b){
c = a*b;
}
GCC 13.2 asm, when compiling with
-std=gnu++23 -march=skylake-avx512 -mprefer-vector-width=512 -O3
prod(Dual<Dual<double, 7l>, 2l>&, Dual<Dual<double, 7l>, 2l> const&,
Dual<Dual<double, 7l>, 2l> const&):
push rbp
mov eax, -2
kmovb k1, eax
mov rbp, rsp
and rsp, -64
sub rsp, 264
vmovdqa ymm4, YMMWORD PTR [rsi+128]
vmovapd zmm8, ZMMWORD PTR [rsi]
vmovapd zmm9, ZMMWORD PTR [rdx]
vmovdqa ymm6, YMMWORD PTR [rsi+64]
vmovdqa YMMWORD PTR [rsp+8], ymm4
vmovdqa ymm4, YMMWORD PTR [rdx+96]
vbroadcastsd zmm0, xmm8
vmovdqa ymm7, YMMWORD PTR [rsi+96]
vbroadcastsd zmm1, xmm9
vmovdqa YMMWORD PTR [rsp-56], ymm6
vmovdqa ymm5, YMMWORD PTR [rdx+128]
vmovdqa ymm6, YMMWORD PTR [rsi+160]
vmovdqa YMMWORD PTR [rsp+168], ymm4
vxorpd xmm4, xmm4, xmm4
vaddpd zmm0, zmm0, zmm4
vaddpd zmm1, zmm1, zmm4
vmovdqa YMMWORD PTR [rsp-24], ymm7
vmovdqa ymm7, YMMWORD PTR [rdx+64]
vmovapd zmm3, ZMMWORD PTR [rsp-56]
vmovdqa YMMWORD PTR [rsp+40], ymm6
vmovdqa ymm6, YMMWORD PTR [rdx+160]
vmovdqa YMMWORD PTR [rsp+200], ymm5
vmulpd zmm2, zmm0, zmm9
vmovdqa YMMWORD PTR [rsp+136], ymm7
vmulpd zmm5, zmm1, zmm3
vbroadcastsd zmm3, xmm3
vmovdqa YMMWORD PTR [rsp+232], ymm6
vaddpd zmm3, zmm3, zmm4
vmovapd zmm7, zmm2
vmovapd zmm2, ZMMWORD PTR [rsp+8]
vfmadd231pd zmm7{k1}, zmm8, zmm1
vmovapd zmm6, zmm5
vmovapd zmm5, ZMMWORD PTR [rsp+136]
vmulpd zmm1, zmm1, zmm2
vfmadd231pd zmm6{k1}, zmm9, zmm3
vbroadcastsd zmm2, xmm2
vmovapd zmm3, ZMMWORD PTR [rsp+200]
vaddpd zmm2, zmm2, zmm4
vmovapd ZMMWORD PTR [rdi], zmm7
vfmadd231pd zmm1{k1}, zmm9, zmm2
vmulpd zmm2, zmm0, zmm5
vbroadcastsd zmm5, xmm5
vmulpd zmm0, zmm0, zmm3
vbroadcastsd zmm3, xmm3
vaddpd zmm5, zmm5, zmm4
vaddpd zmm3, zmm3, zmm4
vfmadd231pd zmm2{k1}, zmm8, zmm5
vfmadd231pd zmm0{k1}, zmm8, zmm3
vaddpd zmm2, zmm2, zmm6
vaddpd zmm0, zmm0, zmm1
vmovapd ZMMWORD PTR [rdi+64], zmm2
vmovapd ZMMWORD PTR [rdi+128], zmm0
vzeroupper
leave
ret
prod(Dual<Dual<double, 8l>, 2l>&, Dual<Dual<double, 8l>, 2l> const&,
Dual<Dual<double, 8l>, 2l> const&):
push rbp
mov rbp, rsp
and rsp, -64
sub rsp, 648
vmovdqa ymm5, YMMWORD PTR [rsi+224]
vmovdqa ymm3, YMMWORD PTR [rsi+352]
vmovapd zmm0, ZMMWORD PTR [rdx+64]
vmovdqa ymm2, YMMWORD PTR [rsi+320]
vmovdqa YMMWORD PTR [rsp+104], ymm5
vmovdqa ymm5, YMMWORD PTR [rdx+224]
vmovdqa ymm7, YMMWORD PTR [rsi+128]
vmovdqa YMMWORD PTR [rsp+232], ymm3
vmovsd xmm3, QWORD PTR [rsi]
vmovdqa ymm6, YMMWORD PTR [rsi+192]
vmovdqa YMMWORD PTR [rsp+488], ymm5
vmovdqa ymm4, YMMWORD PTR [rdx+192]
vmovapd zmm1, ZMMWORD PTR [rsi+64]
vbroadcastsd zmm5, xmm3
vmovdqa YMMWORD PTR [rsp+200], ymm2
vmovdqa ymm2, YMMWORD PTR [rdx+320]
vmulpd zmm8, zmm5, zmm0
vmovdqa YMMWORD PTR [rsp+8], ymm7
vmovdqa ymm7, YMMWORD PTR [rsi+256]
vmovdqa YMMWORD PTR [rsp+72], ymm6
vmovdqa ymm6, YMMWORD PTR [rdx+128]
vmovdqa YMMWORD PTR [rsp+584], ymm2
vmovsd xmm2, QWORD PTR [rdx]
vmovdqa YMMWORD PTR [rsp+136], ymm7
vmovdqa ymm7, YMMWORD PTR [rdx+256]
vmovdqa YMMWORD PTR [rsp+392], ymm6
vmovdqa ymm6, YMMWORD PTR [rdx+352]
vmulsd xmm10, xmm3, xmm2
vmovdqa YMMWORD PTR [rsp+456], ymm4
vbroadcastsd zmm4, xmm2
vfmadd231pd zmm8, zmm4, zmm1
vmovdqa YMMWORD PTR [rsp+520], ymm7
vmovdqa YMMWORD PTR [rsp+616], ymm6
vmulpd zmm9, zmm4, ZMMWORD PTR [rsp+72]
vmovsd xmm6, QWORD PTR [rsp+520]
vmulpd zmm4, zmm4, ZMMWORD PTR [rsp+200]
vmulpd zmm11, zmm5, ZMMWORD PTR [rsp+456]
vmovsd QWORD PTR [rdi], xmm10
vmulpd zmm5, zmm5, ZMMWORD PTR [rsp+584]
vmovapd ZMMWORD PTR [rdi+64], zmm8
vfmadd231pd zmm9, zmm0, QWORD PTR [rsp+8]{1to8}
vfmadd231pd zmm4, zmm0, QWORD PTR [rsp+136]{1to8}
vmovsd xmm0, QWORD PTR [rsp+392]
vmulsd xmm7, xmm3, xmm0
vbroadcastsd zmm0, xmm0
vmulsd xmm3, xmm3, xmm6
vfmadd132pd zmm0, zmm11, zmm1
vbroadcastsd zmm6, xmm6
vfmadd132pd zmm1, zmm5, zmm6
vfmadd231sd xmm7, xmm2, QWORD PTR [rsp+8]
vfmadd132sd xmm2, xmm3, QWORD PTR [rsp+136]
vaddpd zmm0, zmm0, zmm9
vaddpd zmm1, zmm1, zmm4
vmovapd ZMMWORD PTR [rdi+192], zmm0
vmovsd QWORD PTR [rdi+128], xmm7
vmovsd QWORD PTR [rdi+256], xmm2
vmovapd ZMMWORD PTR [rdi+320], zmm1
vzeroupper
leave
ret
Note all the stores to/loads from rsp, and the use of ymm registers.