Skip to contentSkip to conversation

ARM: lib: delay-loop: Add align directive to fix BogoMIPS calculation

RFC In review Fabio Estevam 2 participants 4 messages Nov 22, 2013 – Nov 30, 2013 Arm (32-bit)

lore
Share & tools
lore
#0 lore raw

RFC 1/1 ARM: lib: delay-loop: Add align directive to fix BogoMIPS calculation

From: Fabio Estevam <[email protected]>

Currently mx53 (CortexA8) running at 1GHz reports:
Calibrating delay loop... 663.55 BogoMIPS (lpj=3317760)

Tom Evans verified that alignments of 0x0 and 0x8 run the two instructions of __loop_delay in one clock cycle (1 clock/loop), while alignments of 0x4 and 0xc take 3 clocks to run the loop twice. (1.5 clock/loop)

The original object code looks like this:

00000010 <__loop_const_udelay>:
10: e3e01000 mvn r1, #0
14: e51f201c ldr r2, [pc, #-28] ; 0 <__loop_udelay-0x8>
18: e5922000 ldr r2, [r2]
1c: e0800921 add r0, r0, r1, lsr #18
20: e1a00720 lsr r0, r0, #14
24: e0822b21 add r2, r2, r1, lsr #22
28: e1a02522 lsr r2, r2, #10
2c: e0000092 mul r0, r2, r0
30: e0800d21 add r0, r0, r1, lsr #26
34: e1b00320 lsrs r0, r0, #6
38: 01a0f00e moveq pc, lr

0000003c <__loop_delay>:
3c: e2500001 subs r0, r0, #1
40: 8afffffe bhi 3c <__loop_delay>
44: e1a0f00e mov pc, lr

After adding the 'align 3' directive to __loop_delay (align to 8 bytes):

00000010 <__loop_const_udelay>:
10: e3e01000 mvn r1, #0
14: e51f201c ldr r2, [pc, #-28] ; 0 <__loop_udelay-0x8>
18: e5922000 ldr r2, [r2]
1c: e0800921 add r0, r0, r1, lsr #18
20: e1a00720 lsr r0, r0, #14
24: e0822b21 add r2, r2, r1, lsr #22
28: e1a02522 lsr r2, r2, #10
2c: e0000092 mul r0, r2, r0
30: e0800d21 add r0, r0, r1, lsr #26
34: e1b00320 lsrs r0, r0, #6
38: 01a0f00e moveq pc, lr
3c: e320f000 nop {0}

00000040 <__loop_delay>:
40: e2500001 subs r0, r0, #1
44: 8afffffe bhi 40 <__loop_delay>
48: e1a0f00e mov pc, lr
4c: e320f000 nop {0}

, which now reports:
Calibrating delay loop... 996.14 BogoMIPS (lpj=4980736)

Some more test results:

On mx31 (ARM1136) running at 532 MHz, before the patch:
Calibrating delay loop... 351.43 BogoMIPS (lpj=1757184)

On mx31 (ARM1136) running at 532 MHz after the patch:
Calibrating delay loop... 528.79 BogoMIPS (lpj=2643968)

Also tested on mx6 (CortexA9) and on mx27 (ARM926), which shows the same BogoMIPS value before and after this patch.

---
 arch/arm/lib/delay-loop.S | 1 +
 1 file changed, 1 insertion(+)
arch/arm/lib/delay-loop.S +1 -0
index 36b668d..5e5673b8 100644
@@ -40,6 +40,7 @@ ENTRY(__loop_const_udelay)			@ 0 <= r0 <= 0x7fffff06
 /*
  * loops = r0 * HZ * loops_per_jiffy / 1000000
  */
+		.align 3
 
 @ Delay routine
 ENTRY(__loop_delay)
Signature · 1 line
1.8.1.2
#1 lore raw

Hi Russell,

Quoted 62 lines · Fabio Estevam

From: Fabio Estevam <[email protected]>

Currently mx53 (CortexA8) running at 1GHz reports:
Calibrating delay loop... 663.55 BogoMIPS (lpj=3317760)

Tom Evans verified that alignments of 0x0 and 0x8 run the two instructions of __loop_delay in one clock cycle (1 clock/loop), while alignments of 0x4 and 0xc take 3 clocks to run the loop twice. (1.5 clock/loop)

The original object code looks like this:

00000010 <__loop_const_udelay>:

  10:   e3e01000        mvn     r1, #0
  14:   e51f201c        ldr     r2, [pc, #-28]  ; 0 <__loop_udelay-0x8>
  18:   e5922000        ldr     r2, [r2]
  1c:   e0800921        add     r0, r0, r1, lsr #18
  20:   e1a00720        lsr     r0, r0, #14
  24:   e0822b21        add     r2, r2, r1, lsr #22
  28:   e1a02522        lsr     r2, r2, #10
  2c:   e0000092        mul     r0, r2, r0
  30:   e0800d21        add     r0, r0, r1, lsr #26
  34:   e1b00320        lsrs    r0, r0, #6
  38:   01a0f00e        moveq   pc, lr

0000003c <__loop_delay>:

  3c:   e2500001        subs    r0, r0, #1
  40:   8afffffe        bhi     3c <__loop_delay>
  44:   e1a0f00e        mov     pc, lr

After adding the 'align 3' directive to __loop_delay (align to 8 bytes):

00000010 <__loop_const_udelay>:

  10:   e3e01000        mvn     r1, #0
  14:   e51f201c        ldr     r2, [pc, #-28]  ; 0 <__loop_udelay-0x8>
  18:   e5922000        ldr     r2, [r2]
  1c:   e0800921        add     r0, r0, r1, lsr #18
  20:   e1a00720        lsr     r0, r0, #14
  24:   e0822b21        add     r2, r2, r1, lsr #22
  28:   e1a02522        lsr     r2, r2, #10
  2c:   e0000092        mul     r0, r2, r0
  30:   e0800d21        add     r0, r0, r1, lsr #26
  34:   e1b00320        lsrs    r0, r0, #6
  38:   01a0f00e        moveq   pc, lr
  3c:   e320f000        nop     {0}

00000040 <__loop_delay>:

  40:   e2500001        subs    r0, r0, #1
  44:   8afffffe        bhi     40 <__loop_delay>
  48:   e1a0f00e        mov     pc, lr
  4c:   e320f000        nop     {0}

, which now reports:
Calibrating delay loop... 996.14 BogoMIPS (lpj=4980736)

Some more test results:

On mx31 (ARM1136) running at 532 MHz, before the patch:
Calibrating delay loop... 351.43 BogoMIPS (lpj=1757184)

On mx31 (ARM1136) running at 532 MHz after the patch:
Calibrating delay loop... 528.79 BogoMIPS (lpj=2643968)

Quoted 6 lines

Also tested on mx6 (CortexA9) and on mx27 (ARM926), which shows the same BogoMIPS value before and after this patch.

Any comments on this, please?

Regards,

Fabio Estevam

#2 lore raw
Quoted 65 lines · Fabio Estevam

Hi Russell,

Quoted 62 lines · Fabio Estevam

From: Fabio Estevam <[email protected]>

Currently mx53 (CortexA8) running at 1GHz reports:
Calibrating delay loop... 663.55 BogoMIPS (lpj=3317760)

Tom Evans verified that alignments of 0x0 and 0x8 run the two instructions of __loop_delay in one clock cycle (1 clock/loop), while alignments of 0x4 and 0xc take 3 clocks to run the loop twice. (1.5 clock/loop)

The original object code looks like this:

00000010 <__loop_const_udelay>:

  10:   e3e01000        mvn     r1, #0
  14:   e51f201c        ldr     r2, [pc, #-28]  ; 0 <__loop_udelay-0x8>
  18:   e5922000        ldr     r2, [r2]
  1c:   e0800921        add     r0, r0, r1, lsr #18
  20:   e1a00720        lsr     r0, r0, #14
  24:   e0822b21        add     r2, r2, r1, lsr #22
  28:   e1a02522        lsr     r2, r2, #10
  2c:   e0000092        mul     r0, r2, r0
  30:   e0800d21        add     r0, r0, r1, lsr #26
  34:   e1b00320        lsrs    r0, r0, #6
  38:   01a0f00e        moveq   pc, lr

0000003c <__loop_delay>:

  3c:   e2500001        subs    r0, r0, #1
  40:   8afffffe        bhi     3c <__loop_delay>
  44:   e1a0f00e        mov     pc, lr

After adding the 'align 3' directive to __loop_delay (align to 8 bytes):

00000010 <__loop_const_udelay>:

  10:   e3e01000        mvn     r1, #0
  14:   e51f201c        ldr     r2, [pc, #-28]  ; 0 <__loop_udelay-0x8>
  18:   e5922000        ldr     r2, [r2]
  1c:   e0800921        add     r0, r0, r1, lsr #18
  20:   e1a00720        lsr     r0, r0, #14
  24:   e0822b21        add     r2, r2, r1, lsr #22
  28:   e1a02522        lsr     r2, r2, #10
  2c:   e0000092        mul     r0, r2, r0
  30:   e0800d21        add     r0, r0, r1, lsr #26
  34:   e1b00320        lsrs    r0, r0, #6
  38:   01a0f00e        moveq   pc, lr
  3c:   e320f000        nop     {0}

00000040 <__loop_delay>:

  40:   e2500001        subs    r0, r0, #1
  44:   8afffffe        bhi     40 <__loop_delay>
  48:   e1a0f00e        mov     pc, lr
  4c:   e320f000        nop     {0}

, which now reports:
Calibrating delay loop... 996.14 BogoMIPS (lpj=4980736)

Some more test results:

On mx31 (ARM1136) running at 532 MHz, before the patch:
Calibrating delay loop... 351.43 BogoMIPS (lpj=1757184)

On mx31 (ARM1136) running at 532 MHz after the patch:
Calibrating delay loop... 528.79 BogoMIPS (lpj=2643968)

Quoted 8 lines
Quoted 6 lines

Also tested on mx6 (CortexA9) and on mx27 (ARM926), which shows the same BogoMIPS value before and after this patch.

Any comments on this, please?

Any chance that you could run hackbench, and build the kernel with -falign-functions=32, comparing the kernel without and with this option ?

If alignment has as much effect as the above suggests, the results may be interesting.

As far as this patch is concerned, I'm happy with it, please put it in the patch system, thanks.

Thanks.

#3 lore raw
Quoted 6 lines · Russell King - ARM Linux

Any chance that you could run hackbench, and build the kernel with -falign-functions=32, comparing the kernel without and with this option ?

If alignment has as much effect as the above suggests, the results may be interesting.

Yes, that would be interesting to know.

I will try to run this test next week, or maybe Tom Evans could try it sooner if he has a chance.

Quoted 3 lines

As far as this patch is concerned, I'm happy with it, please put it in the patch system, thanks.

Sent it as 7907/1.

Regards,

Fabio Estevam