crc32-pclmul: remove useless relative addressing

Commit Message

Mikulas Patocka Sept. 7, 2017, 2:41 a.m. UTC

In 32-bit mode, the x86 architecture can hold full 32-bit pointers.
Therefore, the code that copies the current address to the %ecx register
and uses %ecx-relative addressing is useless, we could just use absolute
addressing.

The processors have a stack of return addresses for branch prediction. If 
we use a call instruction and pop the return address, it desynchronizes 
the return stack and causes branch prediction misses.

This patch also moves the data to the .rodata section.

Signed-off-by: Mikulas Patocka <[e-mail]>

---
 arch/x86/crypto/crc32-pclmul_asm.S |   17 ++++++-----------
 1 file changed, 6 insertions(+), 11 deletions(-)

Comments

Herbert Xu Oct. 7, 2017, 4:18 a.m. UTC | #1

On Wed, Sep 06, 2017 at 10:41:21PM -0400, Mikulas Patocka wrote:
> In 32-bit mode, the x86 architecture can hold full 32-bit pointers.
> Therefore, the code that copies the current address to the %ecx register
> and uses %ecx-relative addressing is useless, we could just use absolute
> addressing.
> 
> The processors have a stack of return addresses for branch prediction. If 
> we use a call instruction and pop the return address, it desynchronizes 
> the return stack and causes branch prediction misses.
> 
> This patch also moves the data to the .rodata section.
> 
> Signed-off-by: Mikulas Patocka <[e-mail]>

Patch applied.  Thanks.

Patch

Index: linux-2.6/arch/x86/crypto/crc32-pclmul_asm.S
===================================================================
--- linux-2.6.orig/arch/x86/crypto/crc32-pclmul_asm.S
+++ linux-2.6/arch/x86/crypto/crc32-pclmul_asm.S
@@ -41,6 +41,7 @@ 
 #include <asm/inst.h>
 
 
+.section .rodata
 .align 16
 /*
  * [x4*128+32 mod P(x) << 32)]'  << 1   = 0x154442bd4
@@ -111,19 +112,13 @@  ENTRY(crc32_pclmul_le_16) /* buffer and
 	pxor    CONSTANT, %xmm1
 	sub     $0x40, LEN
 	add     $0x40, BUF
-#ifndef __x86_64__
-	/* This is for position independent code(-fPIC) support for 32bit */
-	call    delta
-delta:
-	pop     %ecx
-#endif
 	cmp     $0x40, LEN
 	jb      less_64
 
 #ifdef __x86_64__
 	movdqa .Lconstant_R2R1(%rip), CONSTANT
 #else
-	movdqa .Lconstant_R2R1 - delta(%ecx), CONSTANT
+	movdqa .Lconstant_R2R1, CONSTANT
 #endif
 
 loop_64:/*  64 bytes Full cache line folding */
@@ -172,7 +167,7 @@  less_64:/*  Folding cache line into 128b
 #ifdef __x86_64__
 	movdqa  .Lconstant_R4R3(%rip), CONSTANT
 #else
-	movdqa  .Lconstant_R4R3 - delta(%ecx), CONSTANT
+	movdqa  .Lconstant_R4R3, CONSTANT
 #endif
 	prefetchnta     (BUF)
 
@@ -220,8 +215,8 @@  fold_64:
 	movdqa  .Lconstant_R5(%rip), CONSTANT
 	movdqa  .Lconstant_mask32(%rip), %xmm3
 #else
-	movdqa  .Lconstant_R5 - delta(%ecx), CONSTANT
-	movdqa  .Lconstant_mask32 - delta(%ecx), %xmm3
+	movdqa  .Lconstant_R5, CONSTANT
+	movdqa  .Lconstant_mask32, %xmm3
 #endif
 	psrldq  $0x04, %xmm2
 	pand    %xmm3, %xmm1
@@ -232,7 +227,7 @@  fold_64:
 #ifdef __x86_64__
 	movdqa  .Lconstant_RUpoly(%rip), CONSTANT
 #else
-	movdqa  .Lconstant_RUpoly - delta(%ecx), CONSTANT
+	movdqa  .Lconstant_RUpoly, CONSTANT
 #endif
 	movdqa  %xmm1, %xmm2
 	pand    %xmm3, %xmm1