https://bugs.llvm.org/show_bug.cgi?id=43495
Bug ID: 43495
Summary: [AARCH64] Clang should use xtn/shrn for shuffles
Product: libraries
Version: 9.0
Hardware: Other
OS: Linux
Status: NEW
Severity: enhancement
Priority: P
Component: Backend: AArch64
Assignee: [email protected]
Reporter: [email protected]
CC: [email protected],
[email protected], [email protected],
[email protected]
On aarch64, shuffles suck.
Not only do you need to split up the instructions and use ext more, but they
are so much slower than on ARMv7a.
uint32x2_t get_even_lanes(uint32x4_t x)
{
return __builtin_shufflevector(x, x, 0, 2);
}
uint32x2_t get_odd_lanes(uint32x4_t x)
{
return __builtin_shufflevector(x, x, 1, 3);
}
uint32x2x2_t vzip_pairwise(uint32x4_t x)
{
return vzip_u32(vget_low_u32(x), vget_high_u32(x));
}
Clang-9 emits this:
get_even_lanes:
ext v1.16b, v0.16b, v0.16b, #8
zip1 v0.2s, v0.2s, v1.2s
ret
get_odd_lanes:
ext v1.16b, v0.16b, v0.16b, #8
zip1 v0.2s, v0.2s, v1.2s
ret
vzip_pairwise:
ext v1.16b, v0.16b, v0.16b, #8
zip1 v2.2s, v0.2s, v1.2s
zip2 v1.2s, v0.2s, v1.2s
mov v0.16b, v2.16b
ret
This is garbage.
It significantly better to do this instead:
get_even_lanes:
xtn v0.2s, v0.2d
ret
get_odd_lanes:
shrn v0.2s, v0.2d, #32
ret
vzip_pairwise:
shrn v1.2s, v0.2d, #32
xtn v0.2s, v0.2d
ret
Side note: on 32-bit, if we kill the source, vshrn+vmovn and vrevNq+vuzpq (and
only using onne result) both lose a vzip in place, and is the best 0213 shuffle
aside from vld2.
--
You are receiving this mail because:
You are on the CC list for the bug._______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
