[PATCH] net/crc: reduce usage of static arrays in net_crc_sse.c

[PATCH] net/crc: reduce usage of static arrays in net_crc_sse.c

[Shreesh Adiga]

Replace the clearing of lower 32 bits of XMM register with blend of
zero register.
Replace the clearing of upper 64 bits of XMM register with _mm_move_epi64.
Clang is able to optimize away the AND + memory operand with the
above sequence, however GCC is still emitting the code for AND with
memory operands which is being explicitly eliminated here.

Additionally replace the 48 byte crc_xmm_shift_tab with the contents of
shf_table which is 32 bytes, achieving the same functionality.

Signed-off-by: Shreesh Adiga <16567adigashreesh@gmail.com>
---
 lib/net/net_crc_sse.c | 30 +++++++-----------------------
 1 file changed, 7 insertions(+), 23 deletions(-)

diff --git a/lib/net/net_crc_sse.c b/lib/net/net_crc_sse.c
index 112dc94ac1..eec854e587 100644
--- a/lib/net/net_crc_sse.c
+++ b/lib/net/net_crc_sse.c
@@ -96,20 +96,13 @@ crcr32_reduce_128_to_64(__m128i data128, __m128i precomp)
 static __rte_always_inline uint32_t
 crcr32_reduce_64_to_32(__m128i data64, __m128i precomp)
 {
-	static const alignas(16) uint32_t mask1[4] = {
-		0xffffffff, 0xffffffff, 0x00000000, 0x00000000
-	};
-
-	static const alignas(16) uint32_t mask2[4] = {
-		0x00000000, 0xffffffff, 0xffffffff, 0xffffffff
-	};
 	__m128i tmp0, tmp1, tmp2;
 
-	tmp0 = _mm_and_si128(data64, _mm_load_si128((const __m128i *)mask2));
+	tmp0 = _mm_blend_epi16(_mm_setzero_si128(), data64, 252);
 
 	tmp1 = _mm_clmulepi64_si128(tmp0, precomp, 0x00);
 	tmp1 = _mm_xor_si128(tmp1, tmp0);
-	tmp1 = _mm_and_si128(tmp1, _mm_load_si128((const __m128i *)mask1));
+	tmp1 = _mm_move_epi64(tmp1);
 
 	tmp2 = _mm_clmulepi64_si128(tmp1, precomp, 0x10);
 	tmp2 = _mm_xor_si128(tmp2, tmp1);
@@ -118,13 +111,11 @@ crcr32_reduce_64_to_32(__m128i data64, __m128i precomp)
 	return _mm_extract_epi32(tmp2, 2);
 }
 
-static const alignas(16) uint8_t crc_xmm_shift_tab[48] = {
-	0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
-	0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+static const alignas(16) uint8_t crc_xmm_shift_tab[32] = {
+	0x00, 0x81, 0x82, 0x83, 0x84, 0x85, 0x86, 0x87,
+	0x88, 0x89, 0x8a, 0x8b, 0x8c, 0x8d, 0x8e, 0x8f,
 	0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,
-	0x08, 0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e, 0x0f,
-	0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
-	0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff
+	0x08, 0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e, 0x0f
 };
 
 /**
@@ -216,19 +207,12 @@ crc32_eth_calc_pclmulqdq(
 			0x80808080, 0x80808080, 0x80808080, 0x80808080
 		};
 
-		const alignas(16) uint8_t shf_table[32] = {
-			0x00, 0x81, 0x82, 0x83, 0x84, 0x85, 0x86, 0x87,
-			0x88, 0x89, 0x8a, 0x8b, 0x8c, 0x8d, 0x8e, 0x8f,
-			0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,
-			0x08, 0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e, 0x0f
-		};
-
 		__m128i last16, a, b;
 
 		last16 = _mm_loadu_si128((const __m128i *)&data[data_len - 16]);
 
 		temp = _mm_loadu_si128((const __m128i *)
-			&shf_table[data_len & 15]);
+			&crc_xmm_shift_tab[data_len & 15]);
 		a = _mm_shuffle_epi8(fold, temp);
 
 		temp = _mm_xor_si128(temp,
-- 
2.49.1

Re: [PATCH] net/crc: reduce usage of static arrays in net_crc_sse.c

[Thomas Monjalon]

Hello,

16/07/2025 12:34, Shreesh Adiga:
> Replace the clearing of lower 32 bits of XMM register with blend of
> zero register.
> Replace the clearing of upper 64 bits of XMM register with _mm_move_epi64.
> Clang is able to optimize away the AND + memory operand with the
> above sequence, however GCC is still emitting the code for AND with
> memory operands which is being explicitly eliminated here.
> 
> Additionally replace the 48 byte crc_xmm_shift_tab with the contents of
> shf_table which is 32 bytes, achieving the same functionality.
> 
> Signed-off-by: Shreesh Adiga <16567adigashreesh@gmail.com>

Sorry I'm not following.
Please could you start with defining the goal of this patch?
Is it a code simplification or a performance optimization?

Re: [PATCH] net/crc: reduce usage of static arrays in net_crc_sse.c

[Shreesh Adiga]

[-- Attachment #1: Type: text/plain, Size: 948 bytes --]

On Wed, Sep 24, 2025 at 8:28 PM Thomas Monjalon <thomas@monjalon.net> wrote:

> Hello,
>
> 16/07/2025 12:34, Shreesh Adiga:
> > Replace the clearing of lower 32 bits of XMM register with blend of
> > zero register.
> > Replace the clearing of upper 64 bits of XMM register with
> _mm_move_epi64.
> > Clang is able to optimize away the AND + memory operand with the
> > above sequence, however GCC is still emitting the code for AND with
> > memory operands which is being explicitly eliminated here.
> >
> > Additionally replace the 48 byte crc_xmm_shift_tab with the contents of
> > shf_table which is 32 bytes, achieving the same functionality.
> >
> > Signed-off-by: Shreesh Adiga <16567adigashreesh@gmail.com>
>
> Sorry I'm not following.
> Please could you start with defining the goal of this patch?
> Is it a code simplification or a performance optimization?

It is intended to be a minor performance optimization.

[-- Attachment #2: Type: text/html, Size: 1391 bytes --]

Re: [PATCH] net/crc: reduce usage of static arrays in net_crc_sse.c

[Thomas Monjalon]

29/09/2025 18:28, Shreesh Adiga:
> On Wed, Sep 24, 2025 at 8:28 PM Thomas Monjalon <thomas@monjalon.net> wrote:
> 
> > Hello,
> >
> > 16/07/2025 12:34, Shreesh Adiga:
> > > Replace the clearing of lower 32 bits of XMM register with blend of
> > > zero register.
> > > Replace the clearing of upper 64 bits of XMM register with
> > _mm_move_epi64.
> > > Clang is able to optimize away the AND + memory operand with the
> > > above sequence, however GCC is still emitting the code for AND with
> > > memory operands which is being explicitly eliminated here.
> > >
> > > Additionally replace the 48 byte crc_xmm_shift_tab with the contents of
> > > shf_table which is 32 bytes, achieving the same functionality.
> > >
> > > Signed-off-by: Shreesh Adiga <16567adigashreesh@gmail.com>
> >
> > Sorry I'm not following.
> > Please could you start with defining the goal of this patch?
> > Is it a code simplification or a performance optimization?
> 
> It is intended to be a minor performance optimization.

Please could you give some performance numbers in the commit log?

Re: [PATCH] net/crc: reduce usage of static arrays in net_crc_sse.c

[Shreesh Adiga]

[-- Attachment #1: Type: text/plain, Size: 1659 bytes --]

On Wed, Oct 1, 2025 at 1:25 PM Thomas Monjalon <thomas@monjalon.net> wrote:

> 29/09/2025 18:28, Shreesh Adiga:
> > On Wed, Sep 24, 2025 at 8:28 PM Thomas Monjalon <thomas@monjalon.net>
> wrote:
> >
> > > Hello,
> > >
> > > 16/07/2025 12:34, Shreesh Adiga:
> > > > Replace the clearing of lower 32 bits of XMM register with blend of
> > > > zero register.
> > > > Replace the clearing of upper 64 bits of XMM register with
> > > _mm_move_epi64.
> > > > Clang is able to optimize away the AND + memory operand with the
> > > > above sequence, however GCC is still emitting the code for AND with
> > > > memory operands which is being explicitly eliminated here.
> > > >
> > > > Additionally replace the 48 byte crc_xmm_shift_tab with the contents
> of
> > > > shf_table which is 32 bytes, achieving the same functionality.
> > > >
> > > > Signed-off-by: Shreesh Adiga <16567adigashreesh@gmail.com>
> > >
> > > Sorry I'm not following.
> > > Please could you start with defining the goal of this patch?
> > > Is it a code simplification or a performance optimization?
> >
> > It is intended to be a minor performance optimization.
>
> Please could you give some performance numbers in the commit log?
>
I don't think that this change can be reliably measured. The changes only
impact
the last stage crc 64 to 32 fold and the last 16 bytes computation. The
impact will only
be a couple of clock cycles at best. Reducing the static array usage also I
don't know
if it can be reliably measured especially since it is not affecting the
main loop.
This patch can be ignored if minor incremental changes are not desirable.

[-- Attachment #2: Type: text/html, Size: 2388 bytes --]

Re: [PATCH] net/crc: reduce usage of static arrays in net_crc_sse.c

[Thomas Monjalon]

01/10/2025 12:24, Shreesh Adiga:
> On Wed, Oct 1, 2025 at 1:25 PM Thomas Monjalon <thomas@monjalon.net> wrote:
> 
> > 29/09/2025 18:28, Shreesh Adiga:
> > > On Wed, Sep 24, 2025 at 8:28 PM Thomas Monjalon <thomas@monjalon.net>
> > wrote:
> > >
> > > > Hello,
> > > >
> > > > 16/07/2025 12:34, Shreesh Adiga:
> > > > > Replace the clearing of lower 32 bits of XMM register with blend of
> > > > > zero register.
> > > > > Replace the clearing of upper 64 bits of XMM register with
> > > > _mm_move_epi64.
> > > > > Clang is able to optimize away the AND + memory operand with the
> > > > > above sequence, however GCC is still emitting the code for AND with
> > > > > memory operands which is being explicitly eliminated here.
> > > > >
> > > > > Additionally replace the 48 byte crc_xmm_shift_tab with the contents
> > of
> > > > > shf_table which is 32 bytes, achieving the same functionality.
> > > > >
> > > > > Signed-off-by: Shreesh Adiga <16567adigashreesh@gmail.com>
> > > >
> > > > Sorry I'm not following.
> > > > Please could you start with defining the goal of this patch?
> > > > Is it a code simplification or a performance optimization?
> > >
> > > It is intended to be a minor performance optimization.
> >
> > Please could you give some performance numbers in the commit log?
> >
> I don't think that this change can be reliably measured. The changes only
> impact
> the last stage crc 64 to 32 fold and the last 16 bytes computation. The
> impact will only
> be a couple of clock cycles at best. Reducing the static array usage also I
> don't know
> if it can be reliably measured especially since it is not affecting the
> main loop.
> This patch can be ignored if minor incremental changes are not desirable.

Minor changes are desirable.
I'm just asking to understand the real impact of the change.
In general when doing an optimization we try to test it
and give some numbers.
Note we have a unit test for CRC in app/test/test_crc.c
but no performance test for it.

I'll wait for a review from an x86 maintainer.