Hey Andy, On Tue, Oct 23, 2012 at 7:08 PM, Andy Walls <awalls@xxxxxxxxxxxxxxxx> wrote: > On Tue, 2012-10-23 at 16:57 -0300, Ezequiel Garcia wrote: >> This kind of memcpy() is error-prone. Its replacement with a struct >> assignment is prefered because it's type-safe and much easier to read. > > This one is a code maintenance win. :) > > See my comments at the end for the difference in assembled code on an > AMD x86_64 CPU using > $ gcc --version > gcc (GCC) 4.6.3 20120306 (Red Hat 4.6.3-2) > > >> Found by coccinelle. Hand patched and reviewed. >> Tested by compilation only. >> >> A simplified version of the semantic match that finds this problem is as >> follows: (http://coccinelle.lip6.fr/) >> >> // <smpl> >> @@ >> identifier struct_name; >> struct struct_name to; >> struct struct_name from; >> expression E; >> @@ >> -memcpy(&(to), &(from), E); >> +to = from; >> // </smpl> >> >> Cc: Andy Walls <awalls@xxxxxxxxxxxxxxxx> > > Signed-off-by: Andy Walls <awalls@xxxxxxxxxxxxxxxx> > > >> Signed-off-by: Peter Senna Tschudin <peter.senna@xxxxxxxxx> >> Signed-off-by: Ezequiel Garcia <elezegarcia@xxxxxxxxx> >> --- >> drivers/media/pci/ivtv/ivtv-i2c.c | 12 ++++-------- >> 1 files changed, 4 insertions(+), 8 deletions(-) >> >> diff --git a/drivers/media/pci/ivtv/ivtv-i2c.c b/drivers/media/pci/ivtv/ivtv-i2c.c >> index d47f41a..27a8466 100644 >> --- a/drivers/media/pci/ivtv/ivtv-i2c.c >> +++ b/drivers/media/pci/ivtv/ivtv-i2c.c >> @@ -719,13 +719,10 @@ int init_ivtv_i2c(struct ivtv *itv) >> return -ENODEV; >> } >> if (itv->options.newi2c > 0) { >> - memcpy(&itv->i2c_adap, &ivtv_i2c_adap_hw_template, >> - sizeof(struct i2c_adapter)); >> + itv->i2c_adap = ivtv_i2c_adap_hw_template; >> } else { >> - memcpy(&itv->i2c_adap, &ivtv_i2c_adap_template, >> - sizeof(struct i2c_adapter)); >> - memcpy(&itv->i2c_algo, &ivtv_i2c_algo_template, >> - sizeof(struct i2c_algo_bit_data)); >> + itv->i2c_adap = ivtv_i2c_adap_template; >> + itv->i2c_algo = ivtv_i2c_algo_template; >> } >> itv->i2c_algo.udelay = itv->options.i2c_clock_period / 2; >> itv->i2c_algo.data = itv; >> @@ -735,8 +732,7 @@ int init_ivtv_i2c(struct ivtv *itv) >> itv->instance); >> i2c_set_adapdata(&itv->i2c_adap, &itv->v4l2_dev); >> >> - memcpy(&itv->i2c_client, &ivtv_i2c_client_template, >> - sizeof(struct i2c_client)); >> + itv->i2c_client = ivtv_i2c_client_template; >> itv->i2c_client.adapter = &itv->i2c_adap; >> itv->i2c_adap.dev.parent = &itv->pdev->dev; >> > > I looked at the generated assembly with only this last change > implemented: > > $ objdump -h -r -d -l -s orig-ivtv-i2c.o.sav | less > [...] > 07e0 00000000 69767476 20696e74 65726e61 ....ivtv interna > 07f0 6c000000 00000000 00000000 00000000 l............... > 0800 00000000 00000000 00000000 00000000 ................ > 0810 00000000 00000000 00000000 00000000 ................ > 0820 00000000 00000000 00000000 00000000 ................ > 0830 00000000 00000000 00000000 00000000 ................ > [...] > init_ivtv_i2c(): > /home/andy/cx18dev/git/media_tree/drivers/media/video/ivtv/ivtv-i2c.c:738 > 13bb: 48 c7 c6 00 00 00 00 mov $0x0,%rsi > 13be: R_X86_64_32S .rodata+0x7e0 > 13c2: 48 8d bb 30 04 01 00 lea 0x10430(%rbx),%rdi > 13c9: b9 5a 00 00 00 mov $0x5a,%ecx > 13ce: f3 48 a5 rep movsq %ds:(%rsi),%es:(%rdi) > > > $ objdump -h -r -d -l -s orig-ivtv-i2c.o.sav | less > [...] > 07e0 00000000 69767476 20696e74 65726e61 ....ivtv interna > 07f0 6c000000 00000000 00000000 00000000 l............... > 0800 00000000 00000000 00000000 00000000 ................ > 0810 00000000 00000000 00000000 00000000 ................ > 0820 00000000 00000000 00000000 00000000 ................ > 0830 00000000 00000000 00000000 00000000 ................ > [...] > init_ivtv_i2c(): > /home/andy/cx18dev/git/media_tree/drivers/media/video/ivtv/ivtv-i2c.c:738 > 13bb: 48 8d bb 30 04 01 00 lea 0x10430(%rbx),%rdi > 13c2: 48 c7 c6 00 00 00 00 mov $0x0,%rsi > 13c5: R_X86_64_32S .rodata+0x7e0 > 13c9: b9 5a 00 00 00 mov $0x5a,%ecx > 13ce: f3 48 a5 rep movsq %ds:(%rsi),%es:(%rdi) > > > The generated code is reordered, but essentially identical. So I guess > in this instance, the preprocessor defines resolved such that an x86-64 > optimized memcpy() function was not used from the linux kernel source. > > Since all of these memcpy()'s are only called once for each board at > board initialization, performance here really doesn't matter here > anyway. (Unless one is insanely trying to shave microseconds off boot > time :P ) > > With other memcpy()/assignement_operator replacement patches, you may > wish to keep performance in mind, if you are patching a frequently > called function. > Thanks for your thorough review on generated assembly. It's certainly very helpful. However, IMHO, this kind of memcpy/assignment can't generate any performance difference, and we shouldn't worry about this unless it's a very-very-very hot path. On the other side, am I being too naive? I'd like to hear others opinion. Again: thanks! Ezequiel -- To unsubscribe from this list: send the line "unsubscribe linux-media" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
