Skip to contentSkip to conversation

FRV/ARM unaligned access question

DISCUSSION Harvey Harrison 3 participants 8 messages Oct 8, 2008 – Oct 9, 2008

lore
Share & tools
lore
#0 lore raw

David, Russell,

I noticed that frv/arm are the only two arches that currently use open-coded byteshifting routines for both the cpu endianness and the other endianness whereas just about all the other arches use a packed-struct version for the cpu-endian and then the byteshifting versions (lifted from arm) for the other endianness.

For frv, this was due to the consolidation of unaligned access helpers which removed the assembly version. For ARM, I think it was done this way as the kernel could be compiled for either endianness and so it was just easiest to always have the byteshifting versions around.

Quoted 1 line

From the comments in the ARM version, time was spent to make sure the byteshifting

helpers used a small amount or registers. Was there a known problem with the packed-struct implementation on ARM that made you choose to use byte-shifting or was it just done this way to more easily support compiling for different endianness?

If there isn't an issue I'm missing, could ARM/FRV move over to the packed-struct version?

Cheers,

Harvey

#1 lore raw
Quoted 5 lines · Harvey Harrison

I noticed that frv/arm are the only two arches that currently use open-coded byteshifting routines for both the cpu endianness and the other endianness whereas just about all the other arches use a packed-struct version for the cpu-endian and then the byteshifting versions (lifted from arm) for the other endianness.

I'm sorry, I think you're mistaken. I've looked at x86, m68k and parisc, and they all use assembly for their swab functions in asm/byteorder.h.

Signature · 3 lines
Russell King
 Linux kernel    2.6 ARM Linux   - http://www.arm.linux.org.uk/
 maintainer of:
#2 lore raw
Quoted 11 lines · Russell King
Quoted 5 lines · Harvey Harrison

I noticed that frv/arm are the only two arches that currently use open-coded byteshifting routines for both the cpu endianness and the other endianness whereas just about all the other arches use a packed-struct version for the cpu-endian and then the byteshifting versions (lifted from arm) for the other endianness.

I'm sorry, I think you're mistaken. I've looked at x86, m68k and parisc, and they all use assembly for their swab functions in asm/byteorder.h.

Sorry, not talking about byteorder at the moment, talking about unaligned.h.

Cheers,

Harvey

#3 lore raw
Quoted 8 lines · Harvey Harrison
Quoted 7 lines · Russell King
Quoted 5 lines · Harvey Harrison

I noticed that frv/arm are the only two arches that currently use open-coded byteshifting routines for both the cpu endianness and the other endianness whereas just about all the other arches use a packed-struct version for the cpu-endian and then the byteshifting versions (lifted from arm) for the other endianness.

Quoted 7 lines
Quoted 4 lines

I'm sorry, I think you're mistaken. I've looked at x86, m68k and parisc, and they all use assembly for their swab functions in asm/byteorder.h.

Sorry, not talking about byteorder at the moment, talking about unaligned.h.

At the moment, I've no idea what effect it'll have. I'd need to run some tests to discover what the effect will be. Not sure when I'll get around to that.

If someone else can be found to evaluate what the effect would be...

Signature · 3 lines
Russell King
 Linux kernel    2.6 ARM Linux   - http://www.arm.linux.org.uk/
 maintainer of:
#4 lore raw
Quoted 14 lines · Russell King
Quoted 13 lines · Harvey Harrison
Quoted 11 lines · Russell King
Quoted 5 lines · Harvey Harrison

I noticed that frv/arm are the only two arches that currently use open-coded byteshifting routines for both the cpu endianness and the other endianness whereas just about all the other arches use a packed-struct version for the cpu-endian and then the byteshifting versions (lifted from arm) for the other endianness.

I'm sorry, I think you're mistaken. I've looked at x86, m68k and parisc, and they all use assembly for their swab functions in asm/byteorder.h.

Quoted 9 lines
Quoted 2 lines

Sorry, not talking about byteorder at the moment, talking about unaligned.h.

At the moment, I've no idea what effect it'll have. I'd need to run some tests to discover what the effect will be. Not sure when I'll get around to that.

If someone else can be found to evaluate what the effect would be...

I don't have hardware to test with, but I'll do some cross-compiles to investigate a bit. I was just curious if there was any known issues on arm, or a specific arm compiler that made you choose the implementation you did.

Cheers,

Harvey

#5 lore raw
In reply to Harvey Harrison
Quoted 2 lines · Harvey Harrison

If there isn't an issue I'm missing, could ARM/FRV move over to the packed-struct version?

Using this source:

	typedef unsigned char u8;
	typedef unsigned int u32;

	struct __una_u32 { u32 x __attribute__((packed)); };

	#if 0 // packed struct
	static inline u32 __get_unaligned_cpu32(const void *p)
	{
		const struct __una_u32 *ptr = (const struct __una_u32 *)p;
		return ptr->x;
	}

	static inline u32 get_unaligned_be32(const void *p)
	{
		return __get_unaligned_cpu32((const u8 *)p);
	}

	#else // manual byteshift
	static inline u32 __get_unaligned_be32(const u8 *p)
	{
		return p[0] << 24 | p[1] << 16 | p[2] << 8 | p[3];
	}

	static inline u32 get_unaligned_be32(const void *p)
	{
		return __get_unaligned_be32((const u8 *)p);
	}
	#endif

	u32 jump(u32 *p)
	{
		return get_unaligned_be32(p);
	}

I see the packed struct version compile (with -O2) to:

	jump:
		ldub @(gr8,gr0),gr6
		ldubi @(gr8,1),gr4
		ldubi @(gr8,2),gr5
		ldubi @(gr8,3),gr8
		slli gr6,#24,gr6
		slli gr4,#16,gr4
		or.p gr4, gr6, gr4
		slli gr5,#8,gr5
		or gr5, gr4, gr5
		or.p gr8, gr5, gr8
		ret

and the byteshift version compile to:

	jump:
		ldubi.p @(gr8,1),gr7
		mov gr8, gr5
		ldubi @(gr5,2),gr6
		ldub @(gr8,gr0),gr8
		ldubi @(gr5,3),gr9
		slli gr7,#16,gr7
		slli gr6,#8,gr6
		slli.p gr8,#24,gr8
		or gr6, gr7, gr6
		or gr8, gr9, gr8
		or.p gr8, gr6, gr8
		ret

so they're more or less equivalent, give or take the compiler using an extra instruction unnecessarily.

Switching to the packed struct algorithms also reduces the kernel size very slightly. Before:

	warthog>size vmlinux
	   text    data     bss     dec     hex filename
	2207836   66588  150189 2424613  24ff25 vmlinux

After:

	warthog>size vmlinux
	   text    data     bss     dec     hex filename
	2207804   66588  150189 2424581  24ff05 vmlinux

The attached patch boots okay over NFS.

David
---
[PATCH] FRV: Use packed-struct unalignment rather than manual-shift

From: David Howells <[email protected]>

Use the packed-struct unalignment algorithms rather than the manual-shift unalignment algorithms.

This makes the kernel very slightly smaller.

---

 include/asm-frv/unaligned.h |    4 ++--
 1 files changed, 2 insertions(+), 2 deletions(-)
include/asm-frv/unaligned.h +2 -2
index 839a2fb..d06b9bc 100644
@@ -12,8 +12,8 @@
 #ifndef _ASM_UNALIGNED_H
 #define _ASM_UNALIGNED_H
 
-#include <linux/unaligned/le_byteshift.h>
-#include <linux/unaligned/be_byteshift.h>
+#include <linux/unaligned/le_struct.h>
+#include <linux/unaligned/be_struct.h>
 #include <linux/unaligned/generic.h>
 
 #define get_unaligned	__get_unaligned_be
#6 lore raw
Quoted 9 lines · David Howells

Harvey Harrison <[email protected]> wrote:

Switching to the packed struct algorithms also reduces the kernel size very slightly. Before:

	warthog>size vmlinux
	   text    data     bss     dec     hex filename
	2207836   66588  150189 2424613  24ff25 vmlinux
Quoted 5 lines

After:

	warthog>size vmlinux
	   text    data     bss     dec     hex filename
	2207804   66588  150189 2424581  24ff05 vmlinux

Probably smaller as now the le values aren't being byteswapped anymore, only the native endianess can use the struct version. Care to look at the kernel size with the following instead?

From: Harvey Harrison <[email protected]>
Subject: [PATCH] frv: switch unaligned access to the packed-struct implementation

---
 include/asm-frv/unaligned.h |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)
include/asm-frv/unaligned.h +1 -1
index 839a2fb..6c61c05 100644
@@ -13,7 +13,7 @@
 #define _ASM_UNALIGNED_H
 
 #include <linux/unaligned/le_byteshift.h>
-#include <linux/unaligned/be_byteshift.h>
+#include <linux/unaligned/be_struct.h>
 #include <linux/unaligned/generic.h>
 
 #define get_unaligned	__get_unaligned_be
Signature · 1 line
1.6.0.2.471.g47a76
#7 lore raw
Quoted 3 lines · Harvey Harrison

Probably smaller as now the le values aren't being byteswapped anymore, only the native endianess can use the struct version. Care to look at the kernel size with the following instead?

For some unknown reason, the data segment shrank by four bytes. The text segment stayed the same though:

	warthog>size vmlinux
	   text    data     bss     dec     hex filename
	2207804   66584  150189 2424577  24ff01 vmlinux

I would guess that in that kernel almost nothing uses LE unalignment.

The kernel works, so feel free to add my Acked-by to the patch.

David