https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91201
Bug ID: 91201
Summary: [7~9 Regression] SIMD not generated for horizontal sum
of bytes in array
Product: gcc
Version: 9.1.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: bisqwit at iki dot fi
Target Milestone: ---
For this code —
typedef unsigned long long E;
const unsigned D = 2;
E bytes[D];
unsigned char sum()
{
E b[D]{};
//#pragma omp simd
for(unsigned n=0; n<D; ++n)
{
E temp = bytes[n];
temp += (temp >> 32);
temp += (temp >> 16);
temp += (temp >> 8);
b[n] = temp;
}
E result = 0;
//#pragma omp simd
for(unsigned n=0; n<D; ++n) result += b[n];
return result;
}
GCC 6.4 generates the following neat assembler code, but all versions since GCC
7 (including GCC 9.1) fail to utilize SIMD instructions at all.
vmovdqa xmm0, XMMWORD PTR bytes[rip]
vpsrlq xmm1, xmm0, 32
vpaddq xmm1, xmm1, xmm0
vpsrlq xmm0, xmm1, 16
vpaddq xmm1, xmm0, xmm1
vpsrlq xmm0, xmm1, 8
vpaddq xmm0, xmm0, xmm1
vpsrldq xmm1, xmm0, 8
vpaddq xmm0, xmm0, xmm1
vmovq rax, xmm0
ret
The code that GCC versions since 7.0, including and up to 9.1, generates, is:
mov rcx, QWORD PTR bytes[rip]
mov rdx, QWORD PTR bytes[rip+8]
mov rax, rcx
shr rax, 32
add rcx, rax
mov rax, rcx
shr rax, 16
add rcx, rax
mov rax, rdx
shr rax, 32
add rdx, rax
mov rax, rdx
shr rax, 16
add rdx, rax
mov rax, rcx
shr rax, 8
add rcx, rdx
add rcx, rax
shr rdx, 8
lea rax, [rcx+rdx]
ret
Tested using compiler options -Ofast -std=c++17 -pedantic -Wall -Wextra -W
-march=skylake. Tried also haswell, broadwell and znver1 for the -march option.
If I change the D constant to a larger one, such as 4 or 8, then SIMD
instructions will begin appearing. Interestingly with D=4, it uses stack as a
temporary, but with D=8, it manages without (on both AVX and non-AVX code).
If I uncomment the two OpenMP pragmas, then SIMD code will manifest, so it is
clear that the compiler _can_ generate the optimal code, but for some reason
chooses not to.
The testcase is a horizontal sum of all bytes in an array by the way.
Compiler Explorer link for quick testing: https://godbolt.org/z/azkXiL
