During some performance diagnostics I stumbled on this slightly wasteful
code in pcnet_cs.c which I made the patch included at the bottom for (two
minor comment fixes included).
Improvement:
instead of *always* calculating
lea 0x2c0(%edx),%ebx
and then additionally doing the
mov %edx,0xc0(%ebx)
addition *if we need it*,
we now do the *whole* calculation of
mov %edx,0x380(%ebx)
*only* if we need it.
This even manages to save us a whole 16-byte alignment buffer loss
in this compilation case.
Result: slightly improves IRQ handler performance in both shared and
non-shared IRQ case, which should make my rusty P3/700 a slight bit happier.
Thank you for your support,
Andreas Mohr
old asm result (using gcc 3.3.5):
000015a0 <ei_irq_wrapper>:
15a0: 55 push %ebp
15a1: 89 e5 mov %esp,%ebp
15a3: 53 push %ebx
15a4: 8d 9a c0 02 00 00 lea 0x2c0(%edx),%ebx
15aa: e8 fc ff ff ff call 15ab <ei_irq_wrapper+0xb>
15af: 83 f8 01 cmp $0x1,%eax
15b2: 74 03 je 15b7 <ei_irq_wrapper+0x17>
15b4: 5b pop %ebx
15b5: 5d pop %ebp
15b6: c3 ret
15b7: 31 d2 xor %edx,%edx
15b9: 89 93 c0 00 00 00 mov %edx,0xc0(%ebx)
15bf: eb f3 jmp 15b4 <ei_irq_wrapper+0x14>
15c1: eb 0d jmp 15d0 <ei_watchdog>
15c3: 90 nop
15c4: 90 nop
15c5: 90 nop
15c6: 90 nop
15c7: 90 nop
15c8: 90 nop
15c9: 90 nop
15ca: 90 nop
15cb: 90 nop
15cc: 90 nop
15cd: 90 nop
15ce: 90 nop
15cf: 90 nop
000015d0 <ei_watchdog>:
new asm result:
000015a0 <ei_irq_wrapper>:
15a0: 55 push %ebp
15a1: 89 e5 mov %esp,%ebp
15a3: 53 push %ebx
15a4: 89 d3 mov %edx,%ebx
15a6: e8 fc ff ff ff call 15a7 <ei_irq_wrapper+0x7>
15ab: 83 f8 01 cmp $0x1,%eax
15ae: 74 03 je 15b3 <ei_irq_wrapper+0x13>
15b0: 5b pop %ebx
15b1: 5d pop %ebp
15b2: c3 ret
15b3: 31 d2 xor %edx,%edx
15b5: 89 93 80 03 00 00 mov %edx,0x380(%ebx)
15bb: eb f3 jmp 15b0 <ei_irq_wrapper+0x10>
15bd: 8d 76 00 lea 0x0(%esi),%esi
000015c0 <ei_watchdog>:
Signed-off-by: Andrew Morton <akpm@osdl.org>
93ad4fb04f