| Message ID | 1455287246-3540549-1-git-send-email-arnd@arndb.de (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
Em Fri, 12 Feb 2016 15:27:18 +0100 Arnd Bergmann <arnd@arndb.de> escreveu: > I noticed a build error in some randconfig builds in the zl10353 driver: > > dvb-frontends/zl10353.c:138: undefined reference to `____ilog2_NaN' > dvb-frontends/zl10353.c:138: undefined reference to `__aeabi_uldivmod' > > The problem can be tracked down to the use of -fprofile-arcs (using > CONFIG_GCOV_PROFILE_ALL) in combination with CONFIG_PROFILE_ALL_BRANCHES > on gcc version 4.9 or higher, when it fails to reliably optimize > constant expressions. > > Using div_u64() instead of do_div() makes the code slightly more > readable by both humans and by gcc, which gives the compiler enough > of a break to figure it all out. I'm not against this patch, but we have 94 occurrences of do_div() just at the media subsystem. If this is failing here, it would likely fail with other drivers. So, I guess we should either fix do_div() or convert all such occurrences to do_div64(). > > Signed-off-by: Arnd Bergmann <arnd@arndb.de> > --- > drivers/media/dvb-frontends/zl10353.c | 6 ++---- > 1 file changed, 2 insertions(+), 4 deletions(-) > > diff --git a/drivers/media/dvb-frontends/zl10353.c b/drivers/media/dvb-frontends/zl10353.c > index ef9764a02d4c..160c88710553 100644 > --- a/drivers/media/dvb-frontends/zl10353.c > +++ b/drivers/media/dvb-frontends/zl10353.c > @@ -135,8 +135,7 @@ static void zl10353_calc_nominal_rate(struct dvb_frontend *fe, > > value = (u64)10 * (1 << 23) / 7 * 125; > value = (bw * value) + adc_clock / 2; > - do_div(value, adc_clock); > - *nominal_rate = value; > + *nominal_rate = div_u64(value, adc_clock); > > dprintk("%s: bw %d, adc_clock %d => 0x%x\n", > __func__, bw, adc_clock, *nominal_rate); > @@ -163,8 +162,7 @@ static void zl10353_calc_input_freq(struct dvb_frontend *fe, > if (ife > adc_clock / 2) > ife = adc_clock - ife; > } > - value = (u64)65536 * ife + adc_clock / 2; > - do_div(value, adc_clock); > + value = div_u64((u64)65536 * ife + adc_clock / 2, adc_clock); > *input_freq = -value; > > dprintk("%s: if2 %d, ife %d, adc_clock %d => %d / 0x%x\n", -- To unsubscribe from this list: send the line "unsubscribe linux-media" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Friday 12 February 2016 14:32:20 Mauro Carvalho Chehab wrote: > Em Fri, 12 Feb 2016 15:27:18 +0100 > Arnd Bergmann <arnd@arndb.de> escreveu: > > > I noticed a build error in some randconfig builds in the zl10353 driver: > > > > dvb-frontends/zl10353.c:138: undefined reference to `____ilog2_NaN' > > dvb-frontends/zl10353.c:138: undefined reference to `__aeabi_uldivmod' > > > > The problem can be tracked down to the use of -fprofile-arcs (using > > CONFIG_GCOV_PROFILE_ALL) in combination with CONFIG_PROFILE_ALL_BRANCHES > > on gcc version 4.9 or higher, when it fails to reliably optimize > > constant expressions. > > > > Using div_u64() instead of do_div() makes the code slightly more > > readable by both humans and by gcc, which gives the compiler enough > > of a break to figure it all out. > > I'm not against this patch, but we have 94 occurrences of do_div() > just at the media subsystem. If this is failing here, it would likely > fail with other drivers. So, I guess we should either fix do_div() or > convert all such occurrences to do_div64(). I agree that it's possible that the same problem exists elsewhere, but this is the only one that I ever saw (in five ranconfig builds out of 8035 last week). I also tried changing do_div() to be an inline function with just a small macro wrapper around it for the odd calling conventions, which also made this error go away. I would assume that Nico had a good reason for doing do_div() the way he did. In some other files, I saw the object code grow by a few instructions, but the examples I looked at were otherwise identical. I can imagine that there might be cases where the constant-argument optimization of do_div fails when we go through an inline function in some combination of Kconfig options and compiler version, though I don't think that was the case here. Nico, any other thoughts on this? Arnd -- To unsubscribe from this list: send the line "unsubscribe linux-media" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Fri, 12 Feb 2016, Arnd Bergmann wrote: > On Friday 12 February 2016 14:32:20 Mauro Carvalho Chehab wrote: > > Em Fri, 12 Feb 2016 15:27:18 +0100 > > Arnd Bergmann <arnd@arndb.de> escreveu: > > > > > I noticed a build error in some randconfig builds in the zl10353 driver: > > > > > > dvb-frontends/zl10353.c:138: undefined reference to `____ilog2_NaN' > > > dvb-frontends/zl10353.c:138: undefined reference to `__aeabi_uldivmod' > > > > > > The problem can be tracked down to the use of -fprofile-arcs (using > > > CONFIG_GCOV_PROFILE_ALL) in combination with CONFIG_PROFILE_ALL_BRANCHES > > > on gcc version 4.9 or higher, when it fails to reliably optimize > > > constant expressions. > > > > > > Using div_u64() instead of do_div() makes the code slightly more > > > readable by both humans and by gcc, which gives the compiler enough > > > of a break to figure it all out. > > > > I'm not against this patch, but we have 94 occurrences of do_div() > > just at the media subsystem. If this is failing here, it would likely > > fail with other drivers. So, I guess we should either fix do_div() or > > convert all such occurrences to do_div64(). > > I agree that it's possible that the same problem exists elsewhere, but this is > the only one that I ever saw (in five ranconfig builds out of 8035 last week). > > I also tried changing do_div() to be an inline function with just a small > macro wrapper around it for the odd calling conventions, which also made this > error go away. I would assume that Nico had a good reason for doing do_div() > the way he did. The do_div() calling convention predates my work on it. I assume it was originally done this way to better map onto the x86 instruction. > In some other files, I saw the object code grow by a few > instructions, but the examples I looked at were otherwise identical. > > I can imagine that there might be cases where the constant-argument optimization > of do_div fails when we go through an inline function in some combination > of Kconfig options and compiler version, though I don't think that was > the case here. What could be tried is to turn __div64_const32() into a static inline and see if that makes a difference with those gcc versions we currently accept. > Nico, any other thoughts on this? This is all related to the gcc bug for which I produced a test case here: http://article.gmane.org/gmane.linux.kernel.cross-arch/29801 Do you know if this is fixed in recent gcc? Nicolas -- To unsubscribe from this list: send the line "unsubscribe linux-media" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
diff --git a/drivers/media/dvb-frontends/zl10353.c b/drivers/media/dvb-frontends/zl10353.c index ef9764a02d4c..160c88710553 100644 --- a/drivers/media/dvb-frontends/zl10353.c +++ b/drivers/media/dvb-frontends/zl10353.c @@ -135,8 +135,7 @@ static void zl10353_calc_nominal_rate(struct dvb_frontend *fe, value = (u64)10 * (1 << 23) / 7 * 125; value = (bw * value) + adc_clock / 2; - do_div(value, adc_clock); - *nominal_rate = value; + *nominal_rate = div_u64(value, adc_clock); dprintk("%s: bw %d, adc_clock %d => 0x%x\n", __func__, bw, adc_clock, *nominal_rate); @@ -163,8 +162,7 @@ static void zl10353_calc_input_freq(struct dvb_frontend *fe, if (ife > adc_clock / 2) ife = adc_clock - ife; } - value = (u64)65536 * ife + adc_clock / 2; - do_div(value, adc_clock); + value = div_u64((u64)65536 * ife + adc_clock / 2, adc_clock); *input_freq = -value; dprintk("%s: if2 %d, ife %d, adc_clock %d => %d / 0x%x\n",
I noticed a build error in some randconfig builds in the zl10353 driver: dvb-frontends/zl10353.c:138: undefined reference to `____ilog2_NaN' dvb-frontends/zl10353.c:138: undefined reference to `__aeabi_uldivmod' The problem can be tracked down to the use of -fprofile-arcs (using CONFIG_GCOV_PROFILE_ALL) in combination with CONFIG_PROFILE_ALL_BRANCHES on gcc version 4.9 or higher, when it fails to reliably optimize constant expressions. Using div_u64() instead of do_div() makes the code slightly more readable by both humans and by gcc, which gives the compiler enough of a break to figure it all out. Signed-off-by: Arnd Bergmann <arnd@arndb.de> --- drivers/media/dvb-frontends/zl10353.c | 6 ++---- 1 file changed, 2 insertions(+), 4 deletions(-)