Skip to content

arm: memcpy: add NEON paths for aligned copies #18855

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
xiaoxiang781216 merged 1 commit into from
May 9, 2026
+165 −9
Merged

arm: memcpy: add NEON paths for aligned copies #18855

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
58 changes: 55 additions & 3 deletions libs/libc/machine/arm/armv7-a/arch_memcpy.S
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -260,7 +260,7 @@ def_fn ARCH_LIBCFUN(memcpy) p2align=6
cmp tmp1, tmp2
bne .Lcpy_notaligned

#ifdef USE_VFP
#if defined(USE_VFP) && !defined(USE_NEON)
/* Magic dust alert! Force VFP on Cortex-A9. Experiments show
that the FP pipeline is much better at streaming loads and
stores. This is outside the critical loop. */
Expand Down Expand Up @@ -290,7 +290,40 @@ def_fn ARCH_LIBCFUN(memcpy) p2align=6
bge .Lcpy_body_long

.Lcpy_body_medium: /* Count in tmp2. */
#ifdef USE_VFP
#ifdef USE_NEON
/* Use NEON multi-register transfers with destination alignment
hints for aligned copies. */
1:
vld1.8 {d0-d3}, [src]!
vld1.8 {d4-d7}, [src]!
pld [src, #(prefetch_lines * 64)]
subs tmp2, tmp2, #64
vst1.8 {d0-d3}, [ALIGN(dst, 64)]!
vst1.8 {d4-d7}, [ALIGN(dst, 64)]!
bge 1b
tst tmp2, #0x3f
beq .Ldone

.Ltail63aligned: /* Count in tmp2. */
/* Use NEON 8-byte vld1/vst1 for the tail. */
and tmp1, tmp2, #0x38
rsb tmp1, tmp1, #(56 - PC_OFFSET + INSN_SIZE)
add pc, pc, tmp1
vld1.8 {d0}, [src]! /* 14 words to go. */
vst1.8 {d0}, [dst]!
vld1.8 {d0}, [src]! /* 12 words to go. */
vst1.8 {d0}, [dst]!
vld1.8 {d0}, [src]! /* 10 words to go. */
vst1.8 {d0}, [dst]!
vld1.8 {d0}, [src]! /* 8 words to go. */
vst1.8 {d0}, [dst]!
vld1.8 {d0}, [src]! /* 6 words to go. */
vst1.8 {d0}, [dst]!
vld1.8 {d0}, [src]! /* 4 words to go. */
vst1.8 {d0}, [dst]!
vld1.8 {d0}, [src]! /* 2 words to go. */
vst1.8 {d0}, [dst]!
#elif defined(USE_VFP)
1:
vldr d0, [src, #0]
subs tmp2, tmp2, #64
Expand Down Expand Up @@ -411,7 +444,26 @@ def_fn ARCH_LIBCFUN(memcpy) p2align=6

/* Long copy. We know that there's at least (prefetch_lines * 64)
bytes to go. */
#ifdef USE_VFP
#ifdef USE_NEON
/* Use NEON multi-register transfers with prefetching for long
copies. */
pld [src, #0]
pld [src, #64]
pld [src, #128]
pld [src, #192]
pld [src, #256]
1:
vld1.8 {d0-d3}, [src]!
vld1.8 {d4-d7}, [src]!
pld [src, #(prefetch_lines * 64)]
subs tmp2, tmp2, #64
vst1.8 {d0-d3}, [ALIGN(dst, 64)]!
vst1.8 {d4-d7}, [ALIGN(dst, 64)]!
bge 1b
tst tmp2, #0x3f
beq .Ldone
b .Ltail63aligned
#elif defined(USE_VFP)
/* Don't use PLD. Instead, read some data in advance of the current
copy position into a register. This should act like a PLD
operation but we won't have to repeat the transfer. */
Expand Down
58 changes: 55 additions & 3 deletions libs/libc/machine/arm/armv7-r/arch_memcpy.S
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -258,7 +258,7 @@ def_fn ARCH_LIBCFUN(memcpy) p2align=6
cmp tmp1, tmp2
bne .Lcpy_notaligned

#ifdef USE_VFP
#if defined(USE_VFP) && !defined(USE_NEON)
/* Magic dust alert! Force VFP on Cortex-A9. Experiments show
that the FP pipeline is much better at streaming loads and
stores. This is outside the critical loop. */
Expand Down Expand Up @@ -288,7 +288,40 @@ def_fn ARCH_LIBCFUN(memcpy) p2align=6
bge .Lcpy_body_long

.Lcpy_body_medium: /* Count in tmp2. */
#ifdef USE_VFP
#ifdef USE_NEON
/* Use NEON multi-register transfers with destination alignment
hints for aligned copies. */
1:
vld1.8 {d0-d3}, [src]!
vld1.8 {d4-d7}, [src]!
pld [src, #(prefetch_lines * 64)]
subs tmp2, tmp2, #64
vst1.8 {d0-d3}, [ALIGN(dst, 64)]!
vst1.8 {d4-d7}, [ALIGN(dst, 64)]!
bge 1b
tst tmp2, #0x3f
beq .Ldone

.Ltail63aligned: /* Count in tmp2. */
/* Use NEON 8-byte vld1/vst1 for the tail. */
and tmp1, tmp2, #0x38
rsb tmp1, tmp1, #(56 - PC_OFFSET + INSN_SIZE)
add pc, pc, tmp1
vld1.8 {d0}, [src]! /* 14 words to go. */
vst1.8 {d0}, [dst]!
vld1.8 {d0}, [src]! /* 12 words to go. */
vst1.8 {d0}, [dst]!
vld1.8 {d0}, [src]! /* 10 words to go. */
vst1.8 {d0}, [dst]!
vld1.8 {d0}, [src]! /* 8 words to go. */
vst1.8 {d0}, [dst]!
vld1.8 {d0}, [src]! /* 6 words to go. */
vst1.8 {d0}, [dst]!
vld1.8 {d0}, [src]! /* 4 words to go. */
vst1.8 {d0}, [dst]!
vld1.8 {d0}, [src]! /* 2 words to go. */
vst1.8 {d0}, [dst]!
#elif defined(USE_VFP)
1:
vldr d0, [src, #0]
subs tmp2, tmp2, #64
Expand Down Expand Up @@ -409,7 +442,26 @@ def_fn ARCH_LIBCFUN(memcpy) p2align=6

/* Long copy. We know that there's at least (prefetch_lines * 64)
bytes to go. */
#ifdef USE_VFP
#ifdef USE_NEON
/* Use NEON multi-register transfers with prefetching for long
copies. */
pld [src, #0]
pld [src, #64]
pld [src, #128]
pld [src, #192]
pld [src, #256]
1:
vld1.8 {d0-d3}, [src]!
vld1.8 {d4-d7}, [src]!
pld [src, #(prefetch_lines * 64)]
subs tmp2, tmp2, #64
vst1.8 {d0-d3}, [ALIGN(dst, 64)]!
vst1.8 {d4-d7}, [ALIGN(dst, 64)]!
bge 1b
tst tmp2, #0x3f
beq .Ldone
b .Ltail63aligned
#elif defined(USE_VFP)
/* Don't use PLD. Instead, read some data in advance of the current
copy position into a register. This should act like a PLD
operation but we won't have to repeat the transfer. */
Expand Down
58 changes: 55 additions & 3 deletions libs/libc/machine/arm/armv8-r/arch_memcpy.S
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -258,7 +258,7 @@ def_fn memcpy p2align=6
cmp tmp1, tmp2
bne .Lcpy_notaligned

#ifdef USE_VFP
#if defined(USE_VFP) && !defined(USE_NEON)
/* Magic dust alert! Force VFP on Cortex-A9. Experiments show
that the FP pipeline is much better at streaming loads and
stores. This is outside the critical loop. */
Expand Down Expand Up @@ -288,7 +288,40 @@ def_fn memcpy p2align=6
bge .Lcpy_body_long

.Lcpy_body_medium: /* Count in tmp2. */
#ifdef USE_VFP
#ifdef USE_NEON
/* Use NEON multi-register transfers with destination alignment
hints for aligned copies. */
1:
vld1.8 {d0-d3}, [src]!
vld1.8 {d4-d7}, [src]!
pld [src, #(prefetch_lines * 64)]
subs tmp2, tmp2, #64
vst1.8 {d0-d3}, [ALIGN(dst, 64)]!
vst1.8 {d4-d7}, [ALIGN(dst, 64)]!
bge 1b
tst tmp2, #0x3f
beq .Ldone

.Ltail63aligned: /* Count in tmp2. */
/* Use NEON 8-byte vld1/vst1 for the tail. */
and tmp1, tmp2, #0x38
rsb tmp1, tmp1, #(56 - PC_OFFSET + INSN_SIZE)
add pc, pc, tmp1
vld1.8 {d0}, [src]! /* 14 words to go. */
vst1.8 {d0}, [dst]!
vld1.8 {d0}, [src]! /* 12 words to go. */
vst1.8 {d0}, [dst]!
vld1.8 {d0}, [src]! /* 10 words to go. */
vst1.8 {d0}, [dst]!
vld1.8 {d0}, [src]! /* 8 words to go. */
vst1.8 {d0}, [dst]!
vld1.8 {d0}, [src]! /* 6 words to go. */
vst1.8 {d0}, [dst]!
vld1.8 {d0}, [src]! /* 4 words to go. */
vst1.8 {d0}, [dst]!
vld1.8 {d0}, [src]! /* 2 words to go. */
vst1.8 {d0}, [dst]!
#elif defined(USE_VFP)
1:
vldr d0, [src, #0]
subs tmp2, tmp2, #64
Expand Down Expand Up @@ -409,7 +442,26 @@ def_fn memcpy p2align=6

/* Long copy. We know that there's at least (prefetch_lines * 64)
bytes to go. */
#ifdef USE_VFP
#ifdef USE_NEON
/* Use NEON multi-register transfers with prefetching for long
copies. */
pld [src, #0]
pld [src, #64]
pld [src, #128]
pld [src, #192]
pld [src, #256]
1:
vld1.8 {d0-d3}, [src]!
vld1.8 {d4-d7}, [src]!
pld [src, #(prefetch_lines * 64)]
subs tmp2, tmp2, #64
vst1.8 {d0-d3}, [ALIGN(dst, 64)]!
vst1.8 {d4-d7}, [ALIGN(dst, 64)]!
bge 1b
tst tmp2, #0x3f
beq .Ldone
b .Ltail63aligned
#elif defined(USE_VFP)
/* Don't use PLD. Instead, read some data in advance of the current
copy position into a register. This should act like a PLD
operation but we won't have to repeat the transfer. */
Expand Down
Loading