Skip to contentSkip to conversation

iio: frequency: adf4350: replace loop with order_base_2()

PATCH Superseded Neel Bullywon 5 participants 10 messages Mar 11 – Mar 23

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 iio: frequency: adf4350: replace loop with order_base_2()

Address the TODO in adf4350_set_freq() by replacing the iterative power-of-2 shift loop with the standard order_base_2() macro.

By utilizing DIV_ROUND_UP_ULL(), we
can calculate the required RF divider selection in a single step.

This ensures freq is properly shifted to meet or exceed the minimum VCO frequency.

---
 drivers/iio/frequency/adf4350.c | 16 ++++++----------
 1 file changed, 6 insertions(+), 10 deletions(-)
drivers/iio/frequency/adf4350.c +6 -10
index ed1741165f55..e07be6d5e67e 100644
@@ -17,6 +17,8 @@
 #include <linux/err.h>
 #include <linux/gcd.h>
 #include <linux/gpio/consumer.h>
+#include <linux/log2.h>
+#include <linux/math64.h>
 #include <asm/div64.h>
 #include <linux/clk.h>
 #include <linux/clk-provider.h>
@@ -149,18 +151,12 @@ static int adf4350_set_freq(struct adf4350_state *st, unsigned long long freq)
 	if (freq > ADF4350_MAX_OUT_FREQ || freq < st->min_out_freq)
 		return -EINVAL;
 
-	st->r4_rf_div_sel = 0;
-
 	/*
-	 * !\TODO: The below computation is making sure we get a power of 2
-	 * shift (st->r4_rf_div_sel) so that freq becomes higher or equal to
-	 * ADF4350_MIN_VCO_FREQ. This might be simplified with fls()/fls_long()
-	 * and friends.
+	 * Calculate the required RF divider selection (power of 2 shift)
+	 * to ensure the VCO frequency is >= ADF4350_MIN_VCO_FREQ.
 	 */
-	while (freq < ADF4350_MIN_VCO_FREQ) {
-		freq <<= 1;
-		st->r4_rf_div_sel++;
-	}
+	st->r4_rf_div_sel = order_base_2(DIV_ROUND_UP_ULL(ADF4350_MIN_VCO_FREQ, freq));
+	freq <<= st->r4_rf_div_sel;
 
 	if (freq > ADF4350_MAX_FREQ_45_PRESC) {
 		prescaler = ADF4350_REG1_PRESCALER;
Signature · 1 line
2.44.0
#1 lore raw
Quoted 8 lines · Neel Bullywon

Address the TODO in adf4350_set_freq() by replacing the iterative power-of-2 shift loop with the standard order_base_2() macro.

By utilizing DIV_ROUND_UP_ULL(), we
can calculate the required RF divider selection in a single step.

This ensures freq is properly shifted to meet or exceed the minimum VCO frequency.

Thanks! My comments below.

...

Quoted 5 lines
drivers/iio/frequency/adf4350.c +0 -4
-	st->r4_rf_div_sel = 0;
-
 	/*
-	 * !\TODO: The below computation is making sure we get a power of 2
-	 * shift (st->r4_rf_div_sel) so that freq becomes higher or equal to

drivers/iio/frequency/adf4350.c static int adf4350_set_freq(struct adf4350_state *st, unsigned long long freq) · line 158

Quoted 10 lines
drivers/iio/frequency/adf4350.c +3 -6
-	 * ADF4350_MIN_VCO_FREQ. This might be simplified with fls()/fls_long()
-	 * and friends.
+	 * Calculate the required RF divider selection (power of 2 shift)
+	 * to ensure the VCO frequency is >= ADF4350_MIN_VCO_FREQ.
 	 */
-	while (freq < ADF4350_MIN_VCO_FREQ) {
-		freq <<= 1;
-		st->r4_rf_div_sel++;
-	}
+	st->r4_rf_div_sel = order_base_2(DIV_ROUND_UP_ULL(ADF4350_MIN_VCO_FREQ, freq));

We don't need 64-bit division even with your approach as values are 32-bit for this case (yeah, I know that they have 64-bit types).

What you need is to do as suggested by comment: use fls_long() against the constant and freq to see the difference in value, if there is none the while will become just a check and a single shift.

Or consider it as the opposite problem: "How much the given constant is now bigger than freq?" With that it might be more obvious solution.

I.o.w. think about this a bit more.

drivers/iio/frequency/adf4350.c static int adf4350_set_freq(struct adf4350_state *st, unsigned long long freq) · line 159

Quoted 2 lines
drivers/iio/frequency/adf4350.c +1 -0
+	freq <<= st->r4_rf_div_sel;

I would move this blank line to be before freq shift, this will associate it with the check below.

Quoted 2 lines
drivers/iio/frequency/adf4350.c +0 -0
 	if (freq > ADF4350_MAX_FREQ_45_PRESC) {
 		prescaler = ADF4350_REG1_PRESCALER;
Signature · 2 lines
With Best Regards,
Andy Shevchenko
#2 lore raw

PATCH v2 1/1 iio: frequency: adf4350: replace loop with fls_long()

In reply to Neel Bullywon

Address the TODO in adf4350_set_freq() by replacing the iterative power-of-2 shift loop with a constant-time bitwise calculation.

By comparing the highest set bits of the target constant and freq using fls_long(), we can calculate the required RF divider selection in a single step without relying on expensive 64-bit division.

This ensures freq is properly shifted to meet or exceed the minimum VCO frequency.

---
Changes in v2:
- Use fls_long() instead of order_base_2(DIV_ROUND_UP_ULL()) to avoid unnecessary 64-bit division (Andy Shevchenko)
- Add correction check for mantissa edge case
- Adjust whitespace per review feedback

 drivers/iio/frequency/adf4350.c | 16 +++++++---------
 1 file changed, 7 insertions(+), 9 deletions(-)
drivers/iio/frequency/adf4350.c +7 -9
index ed1741165f55..2183deec179d 100644
@@ -17,6 +17,7 @@
 #include <linux/err.h>
 #include <linux/gcd.h>
 #include <linux/gpio/consumer.h>
+#include <linux/bitops.h>
 #include <asm/div64.h>
 #include <linux/clk.h>
 #include <linux/clk-provider.h>
@@ -151,17 +152,14 @@ static int adf4350_set_freq(struct adf4350_state *st, unsigned long long freq)
 
 	st->r4_rf_div_sel = 0;
 
-	/*
-	 * !\TODO: The below computation is making sure we get a power of 2
-	 * shift (st->r4_rf_div_sel) so that freq becomes higher or equal to
-	 * ADF4350_MIN_VCO_FREQ. This might be simplified with fls()/fls_long()
-	 * and friends.
-	 */
-	while (freq < ADF4350_MIN_VCO_FREQ) {
-		freq <<= 1;
-		st->r4_rf_div_sel++;
+	if (freq < ADF4350_MIN_VCO_FREQ) {
+		st->r4_rf_div_sel = fls_long(ADF4350_MIN_VCO_FREQ - 1) - fls_long(freq);
+		if ((freq << st->r4_rf_div_sel) < ADF4350_MIN_VCO_FREQ)
+			st->r4_rf_div_sel++;
 	}
 
+	freq <<= st->r4_rf_div_sel;
+
 	if (freq > ADF4350_MAX_FREQ_45_PRESC) {
 		prescaler = ADF4350_REG1_PRESCALER;
 		mdiv = 75;
Signature · 1 line
2.44.0
maintainer
#3 lore raw

Re: [PATCH v2] iio: frequency: adf4350: replace loop with fls_long()

Quoted 11 lines · On Sat, 14 Mar 2026 13:20:06 -0400 Neel Bullywon <[email protected]> wrote:

Address the TODO in adf4350_set_freq() by replacing the iterative power-of-2 shift loop with a constant-time bitwise calculation.

By comparing the highest set bits of the target constant and freq using fls_long(), we can calculate the required RF divider selection in a single step without relying on expensive 64-bit division.

This ensures freq is properly shifted to meet or exceed the minimum VCO frequency.

Hi Neel,

I'll wait for Andy to comment on the new implementation. In meantime a process comment. Don't send a new version in reply to old one. It tends to end up with deeply nested messy email threads. I'm not entirely sure where people doing this comes from. Maybe there is some part of the kernel where this style is requested? Or it's coming from other projects.

If you want to associate a new patch with an old one, put a link to lore after the change log.

No need to resend this time though!

Jonathan

Quoted 46 lines

---
Changes in v2:
- Use fls_long() instead of order_base_2(DIV_ROUND_UP_ULL()) to avoid unnecessary 64-bit division (Andy Shevchenko)
- Add correction check for mantissa edge case
- Adjust whitespace per review feedback

 drivers/iio/frequency/adf4350.c | 16 +++++++---------
 1 file changed, 7 insertions(+), 9 deletions(-)
drivers/iio/frequency/adf4350.c +7 -9
index ed1741165f55..2183deec179d 100644
@@ -17,6 +17,7 @@
 #include <linux/err.h>
 #include <linux/gcd.h>
 #include <linux/gpio/consumer.h>
+#include <linux/bitops.h>
 #include <asm/div64.h>
 #include <linux/clk.h>
 #include <linux/clk-provider.h>
@@ -151,17 +152,14 @@ static int adf4350_set_freq(struct adf4350_state *st, unsigned long long freq)
 
 	st->r4_rf_div_sel = 0;
 
-	/*
-	 * !\TODO: The below computation is making sure we get a power of 2
-	 * shift (st->r4_rf_div_sel) so that freq becomes higher or equal to
-	 * ADF4350_MIN_VCO_FREQ. This might be simplified with fls()/fls_long()
-	 * and friends.
-	 */
-	while (freq < ADF4350_MIN_VCO_FREQ) {
-		freq <<= 1;
-		st->r4_rf_div_sel++;
+	if (freq < ADF4350_MIN_VCO_FREQ) {
+		st->r4_rf_div_sel = fls_long(ADF4350_MIN_VCO_FREQ - 1) - fls_long(freq);
+		if ((freq << st->r4_rf_div_sel) < ADF4350_MIN_VCO_FREQ)
+			st->r4_rf_div_sel++;
 	}
 
+	freq <<= st->r4_rf_div_sel;
+
 	if (freq > ADF4350_MAX_FREQ_45_PRESC) {
 		prescaler = ADF4350_REG1_PRESCALER;
 		mdiv = 75;
#4 lore raw

Re: [PATCH v2] iio: frequency: adf4350: replace loop with fls_long()

In reply to Neel Bullywon
Quoted 9 lines · Neel Bullywon

Address the TODO in adf4350_set_freq() by replacing the iterative power-of-2 shift loop with a constant-time bitwise calculation.

By comparing the highest set bits of the target constant and freq using fls_long(), we can calculate the required RF divider selection in a single step without relying on expensive 64-bit division.

This ensures freq is properly shifted to meet or exceed the minimum VCO frequency.

...

drivers/iio/frequency/adf4350.c static int adf4350_set_freq(struct adf4350_state *st, unsigned long long freq) · line 161

Quoted 12 lines
drivers/iio/frequency/adf4350.c +5 -3
 	st->r4_rf_div_sel = 0;
 
-	while (freq < ADF4350_MIN_VCO_FREQ) {
-		freq <<= 1;
-		st->r4_rf_div_sel++;
+	if (freq < ADF4350_MIN_VCO_FREQ) {
+		st->r4_rf_div_sel = fls_long(ADF4350_MIN_VCO_FREQ - 1) - fls_long(freq);
+		if ((freq << st->r4_rf_div_sel) < ADF4350_MIN_VCO_FREQ)
+			st->r4_rf_div_sel++;
 	}
 
+	freq <<= st->r4_rf_div_sel;

Yeah, unfortunately it looks like too complicated in comparison with the original code. Have you checked binary output? My gut feelings that this gives also a lot of code in addition to.

Little optimisation can be like:

	if (freq < ADF4350_MIN_VCO_FREQ) {
		st->r4_rf_div_sel = fls_long(ADF4350_MIN_VCO_FREQ - 1) - fls_long(freq);
		freq <<= st->r4_rf_div_sel;
		if (freq < ADF4350_MIN_VCO_FREQ) {
			st->r4_rf_div_sel++;
			freq <<= 1;
		}
	}

but it is still too much.

...

Maybe we simply need to replace TODO with a NOTE explaining that if better algo is found we can replace.

Signature · 2 lines
With Best Regards,
Andy Shevchenko
maintainer
#5 lore raw

Re: [PATCH v2] iio: frequency: adf4350: replace loop with fls_long()

Quoted 41 lines · On Mon, 16 Mar 2026 14:40:39 +0200 Andy Shevchenko…
Quoted 9 lines · Neel Bullywon

Address the TODO in adf4350_set_freq() by replacing the iterative power-of-2 shift loop with a constant-time bitwise calculation.

By comparing the highest set bits of the target constant and freq using fls_long(), we can calculate the required RF divider selection in a single step without relying on expensive 64-bit division.

This ensures freq is properly shifted to meet or exceed the minimum VCO frequency.

...

drivers/iio/frequency/adf4350.c static int adf4350_set_freq(struct adf4350_state *st, unsigned long long freq) · line 161

Quoted 12 lines
drivers/iio/frequency/adf4350.c +5 -3
 	st->r4_rf_div_sel = 0;  
 
-	while (freq < ADF4350_MIN_VCO_FREQ) {
-		freq <<= 1;
-		st->r4_rf_div_sel++;
+	if (freq < ADF4350_MIN_VCO_FREQ) {
+		st->r4_rf_div_sel = fls_long(ADF4350_MIN_VCO_FREQ - 1) - fls_long(freq);
+		if ((freq << st->r4_rf_div_sel) < ADF4350_MIN_VCO_FREQ)
+			st->r4_rf_div_sel++;
 	}
 
+	freq <<= st->r4_rf_div_sel;  

Yeah, unfortunately it looks like too complicated in comparison with the original code. Have you checked binary output? My gut feelings that this gives also a lot of code in addition to.

Little optimisation can be like:

	if (freq < ADF4350_MIN_VCO_FREQ) {
		st->r4_rf_div_sel = fls_long(ADF4350_MIN_VCO_FREQ - 1) - fls_long(freq);
		freq <<= st->r4_rf_div_sel;
		if (freq < ADF4350_MIN_VCO_FREQ) {
			st->r4_rf_div_sel++;
			freq <<= 1;
		}
	}
Quoted 7 lines

but it is still too much.

...

Maybe we simply need to replace TODO with a NOTE explaining that if better algo is found we can replace.

This comment update makes sense to me as will make people look for previous attempts before sending a new one!

Neel, you ran into tricky challenge - thanks for giving it a go!

Jonathan

#6 lore raw

Re: [PATCH v2] iio: frequency: adf4350: replace loop with fls_long()

In reply to Neel Bullywon
Quoted 6 lines · On Sat, 14 Mar 2026 13:20:06 -0400 Neel Bullywon <[email protected]> wrote:

Address the TODO in adf4350_set_freq() by replacing the iterative power-of-2 shift loop with a constant-time bitwise calculation.

By comparing the highest set bits of the target constant and freq using fls_long(), we can calculate the required RF divider selection in a single step without relying on expensive 64-bit division.

Where is the 64bit division?
(apart from in v1)
Indeed where are the 64bit values at all.
If this code is used on 32bit it has to work with a 32bit long. Which makes be think that the 'freq' variable should be u32 (or possibly u64 if frequencies above 4GHz are likely - which I doubt).

In any case this looks like initialisation code and the existing loop has the advantage of being 'obviously correct' and small.

David

Quoted 51 lines

This ensures freq is properly shifted to meet or exceed the minimum VCO frequency.

---
Changes in v2:
- Use fls_long() instead of order_base_2(DIV_ROUND_UP_ULL()) to avoid unnecessary 64-bit division (Andy Shevchenko)
- Add correction check for mantissa edge case
- Adjust whitespace per review feedback

 drivers/iio/frequency/adf4350.c | 16 +++++++---------
 1 file changed, 7 insertions(+), 9 deletions(-)
drivers/iio/frequency/adf4350.c +7 -9
index ed1741165f55..2183deec179d 100644
@@ -17,6 +17,7 @@
 #include <linux/err.h>
 #include <linux/gcd.h>
 #include <linux/gpio/consumer.h>
+#include <linux/bitops.h>
 #include <asm/div64.h>
 #include <linux/clk.h>
 #include <linux/clk-provider.h>
@@ -151,17 +152,14 @@ static int adf4350_set_freq(struct adf4350_state *st, unsigned long long freq)
 
 	st->r4_rf_div_sel = 0;
 
-	/*
-	 * !\TODO: The below computation is making sure we get a power of 2
-	 * shift (st->r4_rf_div_sel) so that freq becomes higher or equal to
-	 * ADF4350_MIN_VCO_FREQ. This might be simplified with fls()/fls_long()
-	 * and friends.
-	 */
-	while (freq < ADF4350_MIN_VCO_FREQ) {
-		freq <<= 1;
-		st->r4_rf_div_sel++;
+	if (freq < ADF4350_MIN_VCO_FREQ) {
+		st->r4_rf_div_sel = fls_long(ADF4350_MIN_VCO_FREQ - 1) - fls_long(freq);
+		if ((freq << st->r4_rf_div_sel) < ADF4350_MIN_VCO_FREQ)
+			st->r4_rf_div_sel++;
 	}
 
+	freq <<= st->r4_rf_div_sel;
+
 	if (freq > ADF4350_MAX_FREQ_45_PRESC) {
 		prescaler = ADF4350_REG1_PRESCALER;
 		mdiv = 75;
#7 lore raw

Re: [PATCH v2] iio: frequency: adf4350: replace loop with fls_long()

Quoted 10 lines · David Laight
Quoted 6 lines · On Sat, 14 Mar 2026 13:20:06 -0400 Neel Bullywon <[email protected]> wrote:

Address the TODO in adf4350_set_freq() by replacing the iterative power-of-2 shift loop with a constant-time bitwise calculation.

By comparing the highest set bits of the target constant and freq using fls_long(), we can calculate the required RF divider selection in a single step without relying on expensive 64-bit division.

Quoted 6 lines

Where is the 64bit division?
(apart from in v1)
Indeed where are the 64bit values at all.
If this code is used on 32bit it has to work with a 32bit long. Which makes be think that the 'freq' variable should be u32 (or possibly u64 if frequencies above 4GHz are likely - which I doubt).

I don't know about _this_ device, but before looking into datasheet I wouldn't put a low probability on the frequencies higher than 4.3GHz. We have (or going to have) devices that work with up to 26GHz frequencies in this folder.

Quoted 2 lines

In any case this looks like initialisation code and the existing loop has the advantage of being 'obviously correct' and small.

Signature · 2 lines
With Best Regards,
Andy Shevchenko
#8 lore raw

Re: [PATCH v2] iio: frequency: adf4350: replace loop with fls_long()

Quoted 18 lines · Andy Shevchenko
Quoted 10 lines · David Laight
Quoted 6 lines · On Sat, 14 Mar 2026 13:20:06 -0400 Neel Bullywon <[email protected]> wrote:

Address the TODO in adf4350_set_freq() by replacing the iterative power-of-2 shift loop with a constant-time bitwise calculation.

By comparing the highest set bits of the target constant and freq using fls_long(), we can calculate the required RF divider selection in a single step without relying on expensive 64-bit division.

Quoted 6 lines

Where is the 64bit division?
(apart from in v1)
Indeed where are the 64bit values at all.
If this code is used on 32bit it has to work with a 32bit long. Which makes be think that the 'freq' variable should be u32 (or possibly u64 if frequencies above 4GHz are likely - which I doubt).

Quoted 3 lines

I don't know about _this_ device, but before looking into datasheet I wouldn't put a low probability on the frequencies higher than 4.3GHz. We have (or going to have) devices that work with up to 26GHz frequencies in this folder.

Yes, it goes up to 4.4GHz.

- Nuno Sá

Quoted 3 lines
Quoted 3 lines

In any case this looks like initialisation code and the existing loop has the advantage of being 'obviously correct' and small.

#9 lore raw

Re: [PATCH v2] iio: frequency: adf4350: replace loop with fls_long()

Quoted 19 lines · On Mon, 23 Mar 2026 11:08:15 +0000 Nuno Sá <[email protected]> wrote:
Quoted 18 lines · Andy Shevchenko
Quoted 16 lines · David Laight
Quoted 6 lines · On Sat, 14 Mar 2026 13:20:06 -0400 Neel Bullywon <[email protected]> wrote:

Address the TODO in adf4350_set_freq() by replacing the iterative power-of-2 shift loop with a constant-time bitwise calculation.

By comparing the highest set bits of the target constant and freq using fls_long(), we can calculate the required RF divider selection in a single step without relying on expensive 64-bit division.

Where is the 64bit division?
(apart from in v1)
Indeed where are the 64bit values at all.
If this code is used on 32bit it has to work with a 32bit long. Which makes be think that the 'freq' variable should be u32 (or possibly u64 if frequencies above 4GHz are likely - which I doubt).

Quoted 5 lines
Quoted 3 lines

I don't know about _this_ device, but before looking into datasheet I wouldn't put a low probability on the frequencies higher than 4.3GHz. We have (or going to have) devices that work with up to 26GHz frequencies in this folder.

Yes, it goes up to 4.4GHz.

In which case all the 'long' need to be u64.

David

Quoted 6 lines

- Nuno Sá

Quoted 3 lines
Quoted 3 lines

In any case this looks like initialisation code and the existing loop has the advantage of being 'obviously correct' and small.