* [PATCH] MIPS: crypto: Clean up useless assignment operations
@ 2024-06-28 8:41 WangYuli
2024-06-28 10:12 ` Herbert Xu
2024-06-28 15:51 ` WangYuli
0 siblings, 2 replies; 5+ messages in thread
From: WangYuli @ 2024-06-28 8:41 UTC (permalink / raw)
To: herbert, davem, tsbogend
Cc: linux-crypto, linux-mips, linux-kernel, WangYuli, Guan Wentao
When entering the "len & sizeof(u32)" branch, len must be less than 8.
So after one operation, len must be less than 4.
At this time, "len -= sizeof(u32)" is not necessary for 64-bit CPUs.
A similar issue has been solved at Loongarch.
Link: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?h=v6.10-rc5&id=fea1c949f6ca5059e12de00d0483645debc5b206
Signed-off-by: Guan Wentao <guanwentao@uniontech.com>
Signed-off-by: WangYuli <wangyuli@uniontech.com>
---
arch/mips/crypto/crc32-mips.c | 4 ++++
1 file changed, 4 insertions(+)
diff --git a/arch/mips/crypto/crc32-mips.c b/arch/mips/crypto/crc32-mips.c
index ec6d58008f8e..505d2d897849 100644
--- a/arch/mips/crypto/crc32-mips.c
+++ b/arch/mips/crypto/crc32-mips.c
@@ -94,7 +94,9 @@ static u32 crc32_mips_le_hw(u32 crc_, const u8 *p, unsigned int len)
CRC32(crc, value, w);
p += sizeof(u32);
+#ifndef CONFIG_64BIT
len -= sizeof(u32);
+#endif
}
if (len & sizeof(u16)) {
@@ -134,7 +136,9 @@ static u32 crc32c_mips_le_hw(u32 crc_, const u8 *p, unsigned int len)
CRC32C(crc, value, w);
p += sizeof(u32);
+#ifndef CONFIG_64BIT
len -= sizeof(u32);
+#endif
}
if (len & sizeof(u16)) {
--
2.43.4
^ permalink raw reply related [flat|nested] 5+ messages in thread
* Re: [PATCH] MIPS: crypto: Clean up useless assignment operations
2024-06-28 8:41 [PATCH] MIPS: crypto: Clean up useless assignment operations WangYuli
@ 2024-06-28 10:12 ` Herbert Xu
2024-06-28 15:51 ` WangYuli
1 sibling, 0 replies; 5+ messages in thread
From: Herbert Xu @ 2024-06-28 10:12 UTC (permalink / raw)
To: WangYuli
Cc: davem, tsbogend, linux-crypto, linux-mips, linux-kernel,
Guan Wentao
On Fri, Jun 28, 2024 at 04:41:17PM +0800, WangYuli wrote:
> When entering the "len & sizeof(u32)" branch, len must be less than 8.
> So after one operation, len must be less than 4.
> At this time, "len -= sizeof(u32)" is not necessary for 64-bit CPUs.
>
> A similar issue has been solved at Loongarch.
>
> Link: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?h=v6.10-rc5&id=fea1c949f6ca5059e12de00d0483645debc5b206
> Signed-off-by: Guan Wentao <guanwentao@uniontech.com>
> Signed-off-by: WangYuli <wangyuli@uniontech.com>
> ---
> arch/mips/crypto/crc32-mips.c | 4 ++++
> 1 file changed, 4 insertions(+)
>
> diff --git a/arch/mips/crypto/crc32-mips.c b/arch/mips/crypto/crc32-mips.c
> index ec6d58008f8e..505d2d897849 100644
> --- a/arch/mips/crypto/crc32-mips.c
> +++ b/arch/mips/crypto/crc32-mips.c
> @@ -94,7 +94,9 @@ static u32 crc32_mips_le_hw(u32 crc_, const u8 *p, unsigned int len)
>
> CRC32(crc, value, w);
> p += sizeof(u32);
> +#ifndef CONFIG_64BIT
> len -= sizeof(u32);
> +#endif
First of all, did you verify that this actually makes a difference?
Please post the actual assembly output with and without this patch.
If it does make a difference, you should avoid doing ifdefs as they
are more likely to cause build failures. Instead do something like
if (!IS_ENABLED(CONFIG_64BIT))
len -= sizeof(u32);
Cheers,
--
Email: Herbert Xu <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt
^ permalink raw reply [flat|nested] 5+ messages in thread
* Re: [PATCH] MIPS: crypto: Clean up useless assignment operations
2024-06-28 8:41 [PATCH] MIPS: crypto: Clean up useless assignment operations WangYuli
2024-06-28 10:12 ` Herbert Xu
@ 2024-06-28 15:51 ` WangYuli
2024-06-28 15:59 ` Wentao Guan
2024-06-28 16:42 ` Maciej W. Rozycki
1 sibling, 2 replies; 5+ messages in thread
From: WangYuli @ 2024-06-28 15:51 UTC (permalink / raw)
To: herbert
Cc: linux-crypto, linux-mips, linux-kernel, Guan Wentao, davem,
tsbogend
On 2024/6/28 18:12, Herbert Xu wrote:
> On Fri, Jun 28, 2024 at 04:41:17PM +0800, WangYuli wrote:
>> When entering the "len & sizeof(u32)" branch, len must be less than 8.
>> So after one operation, len must be less than 4.
>> At this time, "len -= sizeof(u32)" is not necessary for 64-bit CPUs.
>>
>> A similar issue has been solved at Loongarch.
>>
>> Link:
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?h=v6.10-rc5&id=fea1c949f6ca5059e12de00d0483645debc5b206
>> Signed-off-by: Guan Wentao <guanwentao@uniontech.com>
>> Signed-off-by: WangYuli <wangyuli@uniontech.com>
>> ---
>> arch/mips/crypto/crc32-mips.c | 4 ++++
>> 1 file changed, 4 insertions(+)
>>
>> diff --git a/arch/mips/crypto/crc32-mips.c
b/arch/mips/crypto/crc32-mips.c
>> index ec6d58008f8e..505d2d897849 100644
>> --- a/arch/mips/crypto/crc32-mips.c
>> +++ b/arch/mips/crypto/crc32-mips.c
>> @@ -94,7 +94,9 @@ static u32 crc32_mips_le_hw(u32 crc_, const u8 *p,
unsigned int len)
>>
>> CRC32(crc, value, w);
>> p += sizeof(u32);
>> +#ifndef CONFIG_64BIT
>> len -= sizeof(u32);
>> +#endif
>
> First of all, did you verify that this actually makes a difference?
> Please post the actual assembly output with and without this patch.
Sure.
The left shows the assembly after applying this patch, while the right
shows the origin. ( Generated by Clang 17.0.6 )
0000000000000018 <chksum_update>: 0000000000000018 <chksum_update>:
; ctx->crc = crc32_mips_le_hw(ctx->crc, data, length); ;
ctx->crc = crc32_mips_le_hw(ctx->crc, data, length);
18: 08 00 82 8c lw $2, 0x8($4) 18: 08 00 82 8c lw
$2, 0x8($4)
; while (len >= sizeof(u64)) { ; while (len >= sizeof(u64)) {
1c: 08 00 c1 2c sltiu $1, $6, 0x8 <chksum_init+0x8> 1c:
08 00 c1 2c sltiu $1, $6, 0x8 <chksum_init+0x8>
20: 06 00 20 f8 bnezc $1, 0x3c <chksum_update+0x24> 20:
06 00 20 f8 bnezc $1, 0x3c <chksum_update+0x24>
; return le64_to_cpu(__get_unaligned_t(__le64, p)); ; return
le64_to_cpu(__get_unaligned_t(__le64, p));
24: 00 00 a3 dc ld $3, 0x0($5) 24: 00 00 a3 dc ld
$3, 0x0($5)
; CRC32(crc, value, d); ; CRC32(crc, value, d);
28: cf 00 62 7c <unknown> crc32d v0,v1,v0 28: cf 00 62
7c <unknown> crc32d v0,v1,v0
; len -= sizeof(u64); ; len -= sizeof(u64);
2c: f8 ff c6 24 addiu $6, $6, -0x8
<chksumc_digest+0xfffffffffffffce0> 2c: f8 ff c6 24 addiu $6,
$6, -0x8 <chksumc_digest+0xfffffffffffffd48>
; while (len >= sizeof(u64)) { ; while (len >= sizeof(u64)) {
30: 08 00 c1 2c sltiu $1, $6, 0x8 <chksum_init+0x8> 30:
08 00 c1 2c sltiu $1, $6, 0x8 <chksum_init+0x8>
34: fb ff 20 10 beqz $1, 0x24 <chksum_update+0xc> 34: fb
ff 20 10 beqz $1, 0x24 <chksum_update+0xc>
38: 08 00 a5 64 daddiu $5, $5, 0x8 <chksum_init+0x8> 38:
08 00 a5 64 daddiu $5, $5, 0x8 <chksum_init+0x8>
; if (len & sizeof(u32)) { ; if (len & sizeof(u32)) {
3c: 04 00 c1 2c sltiu $1, $6, 0x4 <chksum_init+0x4> 3c:
04 00 c1 2c sltiu $1, $6, 0x4 <chksum_init+0x4>
40: 0a 00 20 10 beqz $1, 0x6c <chksum_update+0x54> 40: 04
00 20 f8 bnezc $1, 0x54 <chksum_update+0x3c>
44: 03 f8 c3 7c dext $3, $6, 0x0, 0x20
<chksum_update+0x8> ; return le32_to_cpu(__get_unaligned_t(__le32, p));
; if (len & sizeof(u16)) { 44: 00 00 a3 8c lw $3, 0x0($5)
48: 02 00 61 30 andi $1, $3, 0x2 <chksum_init+0x2>
; CRC32(crc, value, w);
4c: 0c 00 20 f8 bnezc $1, 0x80 <chksum_update+0x68>
; if (len & sizeof(u8)) { 48: 8f 00 62 7c <unknown> crc32w
v0,v1,v0
50: 01 00 61 30 andi $1, $3, 0x1 <chksum_init+0x1>
54: 02 00 20 d8 beqzc $1, 0x60 <chksum_update+0x48>
; len -= sizeof(u32);
; CRC32(crc, value, b); 4c: fc ff c6 24 addiu $6, $6,
-0x4 <chksumc_digest+0xfffffffffffffd4c>
58: 00 00 a3 90 lbu $3, 0x0($5) ; p += sizeof(u32);
50: 04 00 a5 64 daddiu $5, $5, 0x4 <chksum_init+0x4>
5c: 0f 00 62 7c <unknown> crc32b v0,v1,v0 ; if (len &
sizeof(u16)) {
54: 03 f8 c3 7c dext $3, $6, 0x0, 0x20 <chksum_update+0x8>
; ctx->crc = crc32_mips_le_hw(ctx->crc, data, length); 58: 02 00 61
30 andi $1, $3, 0x2 <chksum_init+0x2>
60: 08 00 82 ac sw $2, 0x8($4) 5c: 03 00 20 d8
beqzc $1, 0x6c <chksum_update+0x54>
; return 0; ; CRC32(crc, value, h);
64: 09 00 e0 03 jr $ra 60: 00 00 a6 94 lhu $6, 0x0($5)
68: 00 00 02 64 daddiu $2, $zero, 0x0 <chksum_init>
; return le32_to_cpu(__get_unaligned_t(__le32, p)); 64: 4f 00 c2
7c <unknown> crc32h v0,a2,v0
6c: 00 00 a6 8c lw $6, 0x0($5)
; CRC32(crc, value, w); ; p += sizeof(u16);
70: 8f 00 c2 7c <unknown> crc32w v0,a2,v0 68: 02 00 a5
64 daddiu $5, $5, 0x2 <chksum_init+0x2>
; if (len & sizeof(u16)) { ; if (len & sizeof(u8)) {
74: 02 00 61 30 andi $1, $3, 0x2 <chksum_init+0x2> 6c: 01
00 61 30 andi $1, $3, 0x1 <chksum_init+0x1>
78: f5 ff 20 10 beqz $1, 0x50 <chksum_update+0x38> 70: 02
00 20 d8 beqzc $1, 0x7c <chksum_update+0x64>
7c: 04 00 a5 64 daddiu $5, $5, 0x4 <chksum_init+0x4>
; CRC32(crc, value, h); ; CRC32(crc, value, b);
80: 00 00 a6 94 lhu $6, 0x0($5) 74: 00 00 a3 90
lbu $3, 0x0($5)
84: 4f 00 c2 7c <unknown> crc32h v0,a2,v0 78: 0f 00 62
7c <unknown> crc32b v0,v1,v0
; if (len & sizeof(u8)) { ; ctx->crc =
crc32_mips_le_hw(ctx->crc, data, length);
88: 01 00 61 30 andi $1, $3, 0x1 <chksum_init+0x1> 7c: 08
00 82 ac sw $2, 0x8($4)
8c: f4 ff 20 10 beqz $1, 0x60 <chksum_update+0x48> ;
return 0;
90: 02 00 a5 64 daddiu $5, $5, 0x2 <chksum_init+0x2> 80:
09 00 e0 03 jr $ra
94: 00 00 00 08 j 0x0 <chksum_init> 84: 00 00 02 64
daddiu $2, $zero, 0x0 <chksum_init>
In our testing, this assignment operation affects Clang's code expansion
and instruction reordering.
This redundant assignment operation confuses Clang and prevents us from
obtaining optimized
assembly code.
I extracted the 'crc32_mips_le_hw()' function as a user-mode demo to
analyze the assembly code
generated for it on MIPS64.
Link: https://godbolt.org/z/r4dGbhTGf
As you can see, regardless of the Clang or GCC version, this redundant
operation affects the generated
assembly code.
>
> If it does make a difference, you should avoid doing ifdefs as they
> are more likely to cause build failures. Instead do something like
>
> if (!IS_ENABLED(CONFIG_64BIT))
> len -= sizeof(u32);
Okay, I'll send a Patch V2 to fix this and update the commit message
based on above.
>
>
> Cheers,
--
WangYuli <wangyuli@uniontech.com>
^ permalink raw reply [flat|nested] 5+ messages in thread
* Re: [PATCH] MIPS: crypto: Clean up useless assignment operations
2024-06-28 15:51 ` WangYuli
@ 2024-06-28 15:59 ` Wentao Guan
2024-06-28 16:42 ` Maciej W. Rozycki
1 sibling, 0 replies; 5+ messages in thread
From: Wentao Guan @ 2024-06-28 15:59 UTC (permalink / raw)
To: 王昱力, herbert
Cc: linux-crypto, linux-mips, linux-kernel, davem, tsbogend
Something looks dirty with folding, the origin diff is following:
0000000000000018 <chksum_update>: 0000000000000018 <chksum_update>:
; ctx->crc = crc32_mips_le_hw(ctx->crc, data, length); ; ctx->crc = crc32_mips_le_hw(ctx->crc, data, length);
18: 08 00 82 8c lw $2, 0x8($4) 18: 08 00 82 8c lw $2, 0x8($4)
; while (len >= sizeof(u64)) { ; while (len >= sizeof(u64)) {
1c: 08 00 c1 2c sltiu $1, $6, 0x8 <chksum_init+0x8> 1c: 08 00 c1 2c sltiu $1, $6, 0x8 <chksum_init+0x8>
20: 06 00 20 f8 bnezc $1, 0x3c <chksum_update+0x24> 20: 06 00 20 f8 bnezc $1, 0x3c <chksum_update+0x24>
; return le64_to_cpu(__get_unaligned_t(__le64, p)); ; return le64_to_cpu(__get_unaligned_t(__le64, p));
24: 00 00 a3 dc ld $3, 0x0($5) 24: 00 00 a3 dc ld $3, 0x0($5)
; CRC32(crc, value, d); ; CRC32(crc, value, d);
28: cf 00 62 7c <unknown> crc32d v0,v1,v0 28: cf 00 62 7c <unknown> crc32d v0,v1,v0
; len -= sizeof(u64); ; len -= sizeof(u64);
2c: f8 ff c6 24 addiu $6, $6, -0x8 <chksumc_digest+0xfffffffffffffce0> 2c: f8 ff c6 24 addiu $6, $6, -0x8 <chksumc_digest+0xfffffffffffffd48>
; while (len >= sizeof(u64)) { ; while (len >= sizeof(u64)) {
30: 08 00 c1 2c sltiu $1, $6, 0x8 <chksum_init+0x8> 30: 08 00 c1 2c sltiu $1, $6, 0x8 <chksum_init+0x8>
34: fb ff 20 10 beqz $1, 0x24 <chksum_update+0xc> 34: fb ff 20 10 beqz $1, 0x24 <chksum_update+0xc>
38: 08 00 a5 64 daddiu $5, $5, 0x8 <chksum_init+0x8> 38: 08 00 a5 64 daddiu $5, $5, 0x8 <chksum_init+0x8>
; if (len & sizeof(u32)) { ; if (len & sizeof(u32)) {
3c: 04 00 c1 2c sltiu $1, $6, 0x4 <chksum_init+0x4> 3c: 04 00 c1 2c sltiu $1, $6, 0x4 <chksum_init+0x4>
40: 0a 00 20 10 beqz $1, 0x6c <chksum_update+0x54> 40: 04 00 20 f8 bnezc $1, 0x54 <chksum_update+0x3c>
44: 03 f8 c3 7c dext $3, $6, 0x0, 0x20 <chksum_update+0x8> ; return le32_to_cpu(__get_unaligned_t(__le32, p));
; if (len & sizeof(u16)) { 44: 00 00 a3 8c lw $3, 0x0($5)
48: 02 00 61 30 andi $1, $3, 0x2 <chksum_init+0x2> ; CRC32(crc, value, w);
4c: 0c 00 20 f8 bnezc $1, 0x80 <chksum_update+0x68>
; if (len & sizeof(u8)) { 48: 8f 00 62 7c <unknown> crc32w v0,v1,v0
50: 01 00 61 30 andi $1, $3, 0x1 <chksum_init+0x1>
54: 02 00 20 d8 beqzc $1, 0x60 <chksum_update+0x48> ; len -= sizeof(u32);
; CRC32(crc, value, b); 4c: fc ff c6 24 addiu $6, $6, -0x4 <chksumc_digest+0xfffffffffffffd4c>
58: 00 00 a3 90 lbu $3, 0x0($5) ; p += sizeof(u32);
50: 04 00 a5 64 daddiu $5, $5, 0x4 <chksum_init+0x4>
5c: 0f 00 62 7c <unknown> crc32b v0,v1,v0 ; if (len & sizeof(u16)) {
54: 03 f8 c3 7c dext $3, $6, 0x0, 0x20 <chksum_update+0x8>
; ctx->crc = crc32_mips_le_hw(ctx->crc, data, length); 58: 02 00 61 30 andi $1, $3, 0x2 <chksum_init+0x2>
60: 08 00 82 ac sw $2, 0x8($4) 5c: 03 00 20 d8 beqzc $1, 0x6c <chksum_update+0x54>
; return 0; ; CRC32(crc, value, h);
64: 09 00 e0 03 jr $ra 60: 00 00 a6 94 lhu $6, 0x0($5)
68: 00 00 02 64 daddiu $2, $zero, 0x0 <chksum_init>
; return le32_to_cpu(__get_unaligned_t(__le32, p)); 64: 4f 00 c2 7c <unknown> crc32h v0,a2,v0
6c: 00 00 a6 8c lw $6, 0x0($5)
; CRC32(crc, value, w); ; p += sizeof(u16);
70: 8f 00 c2 7c <unknown> crc32w v0,a2,v0 68: 02 00 a5 64 daddiu $5, $5, 0x2 <chksum_init+0x2>
; if (len & sizeof(u16)) { ; if (len & sizeof(u8)) {
74: 02 00 61 30 andi $1, $3, 0x2 <chksum_init+0x2> 6c: 01 00 61 30 andi $1, $3, 0x1 <chksum_init+0x1>
78: f5 ff 20 10 beqz $1, 0x50 <chksum_update+0x38> 70: 02 00 20 d8 beqzc $1, 0x7c <chksum_update+0x64>
7c: 04 00 a5 64 daddiu $5, $5, 0x4 <chksum_init+0x4>
; CRC32(crc, value, h); ; CRC32(crc, value, b);
80: 00 00 a6 94 lhu $6, 0x0($5) 74: 00 00 a3 90 lbu $3, 0x0($5)
84: 4f 00 c2 7c <unknown> crc32h v0,a2,v0 78: 0f 00 62 7c <unknown> crc32b v0,v1,v0
; if (len & sizeof(u8)) { ; ctx->crc = crc32_mips_le_hw(ctx->crc, data, length);
88: 01 00 61 30 andi $1, $3, 0x1 <chksum_init+0x1> 7c: 08 00 82 ac sw $2, 0x8($4)
8c: f4 ff 20 10 beqz $1, 0x60 <chksum_update+0x48> ; return 0;
90: 02 00 a5 64 daddiu $5, $5, 0x2 <chksum_init+0x2> 80: 09 00 e0 03 jr $ra
94: 00 00 00 08 j 0x0 <chksum_init> 84: 00 00 02 64 daddiu $2, $zero, 0x0 <chksum_init>
Guan Wentao
^ permalink raw reply [flat|nested] 5+ messages in thread
* Re: [PATCH] MIPS: crypto: Clean up useless assignment operations
2024-06-28 15:51 ` WangYuli
2024-06-28 15:59 ` Wentao Guan
@ 2024-06-28 16:42 ` Maciej W. Rozycki
1 sibling, 0 replies; 5+ messages in thread
From: Maciej W. Rozycki @ 2024-06-28 16:42 UTC (permalink / raw)
To: WangYuli
Cc: herbert, linux-crypto, linux-mips, linux-kernel, Guan Wentao,
David S. Miller, Thomas Bogendoerfer
On Fri, 28 Jun 2024, WangYuli wrote:
> As you can see, regardless of the Clang or GCC version, this redundant
> operation affects the generated
>
> assembly code.
Boy, this code is horribly structured!
Anyway, rather than making it yet worse with another #ifdef I'd suggest
replacing both `while' loops with equivalent `for' ones with `len' being
the worker variable, which will then make the code structure a little bit
better and as a side effect address the missed optimisation automagically.
You might also consider Herbert's suggestion to use IS_ENABLED, however
in the current shape of code I find it kind of pointless anyway. Instead
I think that it would make more sense to factor out the block bodies to
small static inline helpers and then restructure the call sites so that
IS_ENABLED controls the loops, the conditionals, and the choice between
them as applicable.
Maciej
^ permalink raw reply [flat|nested] 5+ messages in thread
end of thread, other threads:[~2024-06-28 16:42 UTC | newest]
Thread overview: 5+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2024-06-28 8:41 [PATCH] MIPS: crypto: Clean up useless assignment operations WangYuli
2024-06-28 10:12 ` Herbert Xu
2024-06-28 15:51 ` WangYuli
2024-06-28 15:59 ` Wentao Guan
2024-06-28 16:42 ` Maciej W. Rozycki
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox