Don't use unaligned access for memcpy instructions

[nmoinvaz]

No description provided.

[phprus]

Compare without memcmp (see: eq_uint64_t_tmp):
https://godbolt.org/z/TavGsbTh9

GCC, Clang and MSVC does not call memcpy.
But I don't know if eq_uint64_t_tmp function will inlining.

On ARM and GCC < 9 there will be calls to memcmp.

[nmoinvaz]

@phprus the __attribute__((__may_alias__)) would be a good idea for a separate PR, I can see it definitely helps on old GCC.

[phprus]

@nmoinvaz
On GCC __may_alias__ and aligned(1) (to decrease alignment).

See link to stackoverflow in my comment:
#1307 (comment)
and
https://gcc.gnu.org/onlinedocs/gcc/Common-Type-Attributes.html

But Clang does not support the aligned attribute. I don't know how to force Clang to guarantee unaligned reads.

[codecov]

Codecov Report

Merging #1309 (52c3511) into develop (2174201) will decrease coverage by 0.02%.
The diff coverage is 91.66%.

@@             Coverage Diff             @@
##           develop    #1309      +/-   ##
===========================================
- Coverage    86.51%   86.48%   -0.03%     
===========================================
  Files          125      126       +1     
  Lines        10670    10671       +1     
  Branches      2664     2670       +6     
===========================================
- Hits          9231     9229       -2     
- Misses        1081     1083       +2     
- Partials       358      359       +1     

Flag	Coverage Δ
macos_clang	32.60% <ø> (ø)
macos_gcc	73.79% <36.66%> (-0.09%)	⬇️
ubuntu_clang	85.19% <73.33%> (+0.13%)	⬆️
ubuntu_clang_debug	84.84% <73.33%> (ø)
ubuntu_clang_inflate_allow_invalid_dist	84.91% <73.33%> (ø)
ubuntu_clang_inflate_strict	85.15% <73.33%> (ø)
ubuntu_clang_mmap	85.34% <73.33%> (ø)
ubuntu_clang_pigz	40.08% <50.00%> (ø)
ubuntu_clang_pigz_no_optim	41.49% <71.42%> (ø)
ubuntu_clang_pigz_no_threads	39.67% <50.00%> (ø)
ubuntu_clang_reduced_mem	85.58% <73.33%> (ø)
ubuntu_gcc	75.60% <63.33%> (-0.10%)	⬇️
ubuntu_gcc_aarch64	77.36% <66.66%> (-0.06%)	⬇️
ubuntu_gcc_aarch64_compat_no_opt	75.51% <14.28%> (-0.25%)	⬇️
ubuntu_gcc_aarch64_no_acle	76.08% <12.50%> (-0.19%)	⬇️
ubuntu_gcc_aarch64_no_neon	76.08% <14.28%> (-0.23%)	⬇️
ubuntu_gcc_armhf	77.31% <63.63%> (-0.07%)	⬇️
ubuntu_gcc_armhf_compat_no_opt	75.36% <15.78%> (-0.25%)	⬇️
ubuntu_gcc_armhf_no_acle	77.33% <63.63%> (-0.05%)	⬇️
ubuntu_gcc_armhf_no_neon	77.22% <94.73%> (-0.04%)	⬇️
ubuntu_gcc_armsf	77.13% <92.85%> (ø)
ubuntu_gcc_armsf_compat_no_opt	75.26% <0.00%> (-0.02%)	⬇️
ubuntu_gcc_benchmark	74.10% <75.00%> (-0.03%)	⬇️
ubuntu_gcc_compat_no_opt	76.91% <90.47%> (-0.03%)	⬇️
ubuntu_gcc_compat_sprefix	74.03% <75.00%> (-0.03%)	⬇️
ubuntu_gcc_m32	73.55% <75.00%> (-0.03%)	⬇️
ubuntu_gcc_mingw_i686	0.00% <0.00%> (ø)
ubuntu_gcc_mingw_x86_64	0.00% <0.00%> (ø)
ubuntu_gcc_no_avx2	74.74% <11.11%> (-0.19%)	⬇️
ubuntu_gcc_no_ctz	74.89% <87.50%> (-0.07%)	⬇️
ubuntu_gcc_no_ctzll	74.91% <87.50%> (-0.07%)	⬇️
ubuntu_gcc_no_pclmulqdq	73.92% <10.00%> (-0.23%)	⬇️
ubuntu_gcc_no_sse2	74.89% <12.50%> (-0.19%)	⬇️
ubuntu_gcc_no_sse4	74.55% <10.00%> (-0.22%)	⬇️
ubuntu_gcc_o1	74.40% <0.00%> (-0.05%)	⬇️
ubuntu_gcc_osb	∅ <ø> (∅)
ubuntu_gcc_pigz	37.95% <50.00%> (-0.07%)	⬇️
ubuntu_gcc_pigz_aarch64	38.97% <50.00%> (-0.24%)	⬇️
ubuntu_gcc_ppc	73.65% <83.33%> (ø)
ubuntu_gcc_ppc64	74.52% <66.66%> (ø)
ubuntu_gcc_ppc64le	74.46% <75.00%> (-0.05%)	⬇️
ubuntu_gcc_ppc_no_power8	74.58% <83.33%> (ø)
ubuntu_gcc_s390x	74.91% <83.33%> (ø)
ubuntu_gcc_s390x_dfltcc	72.15% <66.66%> (ø)
ubuntu_gcc_s390x_dfltcc_compat	73.69% <0.00%> (ø)
ubuntu_gcc_s390x_no_crc32	74.70% <83.33%> (ø)
ubuntu_gcc_sparc64	74.81% <83.33%> (ø)
ubuntu_gcc_sprefix	73.69% <75.00%> (-0.03%)	⬇️
win64_gcc	74.06% <69.56%> (+0.03%)	⬆️
win64_gcc_compat_no_opt	74.75% <100.00%> (?)

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
arch/power/chunkset_power8.c	100.00% <ø> (ø)
arch/x86/chunkset_sse2.c	28.57% <0.00%> (ø)
insert_string_tpl.h	100.00% <ø> (ø)
arch/arm/chunkset_neon.c	100.00% <100.00%> (ø)
arch/x86/chunkset_avx.c	100.00% <100.00%> (ø)
arch/x86/chunkset_sse41.c	100.00% <100.00%> (ø)
chunkset.c	100.00% <100.00%> (ø)
compare256.c	100.00% <100.00%> (ø)
deflate.h	100.00% <100.00%> (ø)
inffast.c	90.28% <100.00%> (ø)
... and 11 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 2174201...52c3511. Read the comment docs.

[nmoinvaz]

Rebased.

[shuffle2]

Compare without memcmp (see: eq_uint64_t_tmp): https://godbolt.org/z/TavGsbTh9

GCC, Clang and MSVC does not call memcpy. But I don't know if eq_uint64_t_tmp function will inlining.

On ARM and GCC < 9 there will be calls to memcmp.

Hi, I just want to point out that you've typo'd the msvc arguments there. You meant /O2 and put /02. Also, /O2 implies /Og /Oi /Ot /Oy /Ob2 /GF /Gy. So /Oi is not needed. /Ob3 would add more aggressive inlining but isn't needed in this case.

So in the end, msvc is generating code nearly same as gcc/clang here: https://godbolt.org/z/9MaYYsTbK
msvc/arm64 also emits the memcmp

[Dead2]

LGTM