Implement mul_u64_u64_div_u64_roundup()
v5 cover Implement mul_u64_u64_div_u64_roundup()
The pwm-stm32.c code wants a 'rounding up' version of mul_u64_u64_div_u64(). This can be done simply by adding 'divisor - 1' to the 128bit product. Implement mul_u64_add_u64_div_u64(a, b, c, d) = (a * b + c)/d based on the existing code.
Define mul_u64_u64_div_u64(a, b, d) as mul_u64_add_u64_div_u64(a, b, 0, d) and mul_u64_u64_div_u64_roundup(a, b, d) as mul_u64_add_u64_div_u64(a, b, d-1, d).
Only x86-64 has an optimsed (asm) version of the function. That is optimised to avoid the 'add c' when c is known to be zero. In all other cases the extra code will be noise compared to the software divide code.
The test module has been updated to test mul_u64_u64_div_u64_roundup() and also enhanced it to verify the C division code on x86-64 and the 32bit division code on 64bit.
Changes for v2:
- Rename the 'divisor' parameter from 'c' to 'd'.
- Add an extra patch to use BUG_ON() to trap zero divisors.
- Remove the last patch that ran the C code on x86-64 (I've a plan to do that differently).
Changes for v3:
- Replace the BUG_ON() (or panic in the original version) for zero divisors with a WARN_ONCE() and return zero.
- Remove the 'pre-multiply' check for small products. Completely non-trivial on 32bit systems.
- Use mul_u32_u32() and the new add_u64_u32() to stop gcc generating pretty much pessimal code for x86 with lots of register spills.
- Replace the 'bit at a time' divide with one that generates 16 bits per iteration on 32bit systems and 32 bits per iteration on 64bit. Massively faster, the tests run in under 1/3 the time.
Changes for v4:
No significant code changes.
- Rebase on 6.18-rc2
- Don't change the behaviour for overflow (return ~0ull) or divide by zero (execute ~0ul/0).
- Merge patches 8 and 9 to avoid bisection issues.
- Fix build of 32bit test cases on non-x86.
- Fix shell script that verifies test cases.
Changes for v5:
- Fix x86-64 asm code to correctly add non-constant 'add' values.
- Fix build on 32bit x86.
- Check for small dividends before divide overflow.
I've left the x86-64 faulting on both overflow and divide by zero. The patch to add an execption table entry to return ~0 for both doesn't seem to have been merged.
If merged it would make sense for the C version to return ~0 for both. Callers can check for a result of ~0 and then check the divisor if they care about overflow (etc).
(A valid quotent of ~0 is pretty unlikely and marginal changes to the input values are likely to generate a real overflow.)
The code that faulted on overflow was about to get invalid results because one of the 64bit inputs would itself wrap very soon.
David Laight (9):
lib: mul_u64_u64_div_u64() rename parameter 'c' to 'd'
lib: mul_u64_u64_div_u64() Combine overflow and divide by zero checks
lib: mul_u64_u64_div_u64() simplify check for a 64bit product
lib: Add mul_u64_add_u64_div_u64() and mul_u64_u64_div_u64_roundup()
lib: Add tests for mul_u64_u64_div_u64_roundup()
lib: test_mul_u64_u64_div_u64: Test both generic and arch versions
lib: mul_u64_u64_div_u64() optimise multiply on 32bit x86
lib: mul_u64_u64_div_u64() Optimise the divide code
lib: test_mul_u64_u64_div_u64: Test the 32bit code on 64bit
arch/x86/include/asm/div64.h | 39 ++++-- include/linux/math64.h | 59 ++++++++- lib/math/div64.c | 185 ++++++++++++++++++--------- lib/math/test_mul_u64_u64_div_u64.c | 191 ++++++++++++++++++++-------- 4 files changed, 354 insertions(+), 120 deletions(-)
Signature · 1 line
2.39.5
PATCH v5 1/9 lib: mul_u64_u64_div_u64() rename parameter 'c' to 'd'
Change to prototype from mul_u64_u64_div_u64(u64 a, u64 b, u64 c) to mul_u64_u64_div_u64(u64 a, u64 b, u64 d).
Using 'd' for 'divisor' makes more sense.
An upcoming change adds a 'c' parameter to calculate (a * b + c)/d.
--- lib/math/div64.c | 24 ++++++++++++------------ 1 file changed, 12 insertions(+), 12 deletions(-)
lib/math/div64.c +12 -12
@@ -184,10 +184,10 @@ u32 iter_div_u64_rem(u64 dividend, u32 divisor, u64 *remainder)EXPORT_SYMBOL (iter_div_u64_rem);#ifndef mul_u64_u64_div_u64 -u64 mul_u64_u64_div_u64 (u64 a,u64 b,u64 c) +u64 mul_u64_u64_div_u64 (u64 a,u64 b,u64 d) {if (ilog2 (a) +ilog2 (b) <=62 ) -return div64_u64 (a * b, c); +return div64_u64 (a * b, d);#if defined(__SIZEOF_INT128__)
@@ -212,37 +212,37 @@ u64 mul_u64_u64_div_u64(u64 a, u64 b, u64 c)#endif -/* make sure c is not zero, trigger runtime exception otherwise */ -if (unlikely (c ==0 )) { +/* make sure d is not zero, trigger runtime exception otherwise */ +if (unlikely (d ==0 )) {unsigned long zero =0 ;OPTIMIZER_HIDE_VAR (zero);return ~0UL /zero; } -int shift =__builtin_ctzll (c); +int shift =__builtin_ctzll (d);/* try reducing the fraction in case the dividend becomes <= 64 bits */ if ((n_hi >> shift) ==0 ) {u64 n = shift ? (n_lo >> shift) | (n_hi << (64 - shift)) : n_lo; -return div64_u64 (n, c >> shift); +return div64_u64 (n, d >> shift);/* * The remainder value if needed would be: -* res = div64_u64_rem(n, c >> shift, &rem); +* res = div64_u64_rem(n, d >> shift, &rem); * rem = (rem << shift) + (n_lo - (n << shift)); */ } -if (n_hi >= c) { +if (n_hi >= d) {/* overflow: result is unrepresentable in a u64 */ return -1 ; }/* Do the full 128 by 64 bits division */ - shift =__builtin_clzll (c); - c <<= shift; + shift =__builtin_clzll (d); + d <<= shift;int p =64 + shift;u64 res =0 ;
@@ -257,8 +257,8 @@ u64 mul_u64_u64_div_u64(u64 a, u64 b, u64 c) n_hi <<= shift; n_hi |= n_lo >> (64 - shift); n_lo <<= shift; -if (carry || (n_hi >= c)) { - n_hi -= c; +if (carry || (n_hi >= d)) { + n_hi -= d; res |=1ULL << p; } }while (n_hi);
Signature · 1 line
2.39.5
PATCH v5 2/9 lib: mul_u64_u64_div_u64() Combine overflow and divide by zero checks
In reply to David LaightSince the overflow check always triggers when the divisor is zero move the check for divide by zero inside the overflow check. This means there is only one test in the normal path.
---
V3 contained a different patch 2 that did different changes to the error paths.
lib/math/div64.c | 19 +++++++++---------- 1 file changed, 9 insertions(+), 10 deletions(-)
lib/math/div64.c +9 -10
@@ -212,12 +212,16 @@ u64 mul_u64_u64_div_u64(u64 a, u64 b, u64 d)#endif -/* make sure d is not zero, trigger runtime exception otherwise */ -if (unlikely (d ==0 )) { -unsigned long zero =0 ; +if (unlikely (n_hi >= d)) { +/* trigger runtime exception if divisor is zero */ +if (d ==0 ) { +unsigned long zero =0 ; -OPTIMIZER_HIDE_VAR (zero); -return ~0UL /zero; +OPTIMIZER_HIDE_VAR (zero); +return ~0UL /zero; + } +/* overflow: result is unrepresentable in a u64 */ +return ~0ULL ; }int shift =__builtin_ctzll (d);
@@ -234,11 +238,6 @@ u64 mul_u64_u64_div_u64(u64 a, u64 b, u64 d)*/ } -if (n_hi >= d) { -/* overflow: result is unrepresentable in a u64 */ -return -1 ; - } -/* Do the full 128 by 64 bits division */ shift =__builtin_clzll (d);
Signature · 1 line
2.39.5
PATCH v5 3/9 lib: mul_u64_u64_div_u64() simplify check for a 64bit product
In reply to David LaightIf the product is only 64bits div64_u64() can be used for the divide. Replace the pre-multiply check (ilog2(a) + ilog2(b) <= 62) with a simple post-multiply check that the high 64bits are zero.
This has the advantage of being simpler, more accurate and less code. It will always be faster when the product is larger than 64bits.
Most 64bit cpu have a native 64x64=128 bit multiply, this is needed (for the low 64bits) even when div64_u64() is called - so the early check gains nothing and is just extra code.
32bit cpu will need a compare (etc) to generate the 64bit ilog2() from two 32bit bit scans - so that is non-trivial.
(Never mind the mess of x86's 'bsr' and any oddball cpu without fast bit-scan instructions.)
Whereas the additional instructions for the 128bit multiply result are pretty much one multiply and two adds (typically the 'adc $0,%reg' can be run in parallel with the instruction that follows).
The only outliers are 64bit systems without 128bit mutiply and simple in order 32bit ones with fast bit scan but needing extra instructions to get the high bits of the multiply result. I doubt it makes much difference to either, the latter is definitely not mainstream.
If anyone is worried about the analysis they can look at the generated code for x86 (especially when cmov isn't used).
---
Split from patch 3 for v2.
Changes for v5:
- Test for small dividends before overflow.
lib/math/div64.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)
lib/math/div64.c +3 -3
@@ -186,9 +186,6 @@ EXPORT_SYMBOL(iter_div_u64_rem);#ifndef mul_u64_u64_div_u64u64 mul_u64_u64_div_u64 (u64 a,u64 b,u64 d) { -if (ilog2 (a) +ilog2 (b) <=62 ) -return div64_u64 (a * b, d); -#if defined(__SIZEOF_INT128__)/* native 64x64=128 bits multiplication */
@@ -212,6 +209,9 @@ u64 mul_u64_u64_div_u64(u64 a, u64 b, u64 d)#endif +if (!n_hi) +return div64_u64 (n_lo, d); +if (unlikely (n_hi >= d)) {/* trigger runtime exception if divisor is zero */ if (d ==0 ) {
Signature · 1 line
2.39.5
PATCH v5 4/9 lib: Add mul_u64_add_u64_div_u64() and mul_u64_u64_div_u64_roundup()
In reply to David LaightThe existing mul_u64_u64_div_u64() rounds down, a 'rounding up' variant needs 'divisor - 1' adding in between the multiply and divide so cannot easily be done by a caller.
Add mul_u64_add_u64_div_u64(a, b, c, d) that calculates (a * b + c)/d and implement the 'round down' and 'round up' using it.
Update the x86-64 asm to optimise for 'c' being a constant zero.
Add kerndoc definitions for all three functions.
---
Changes for v2 (formally patch 1/3):
- Reinstate the early call to div64_u64() on 32bit when 'c' is zero. Although I'm not convinced the path is common enough to be worth the two ilog2() calls.
Changes for v3 (formally patch 3/4):
- The early call to div64_u64() has been removed by patch 3. Pretty much guaranteed to be a pessimisation.
Changes for v4:
- For x86-64 split the multiply, add and divide into three asm blocks. (gcc makes a pigs breakfast of (u128)a * b + c)
- Change the kerndoc since divide by zero will (probably) fault.
Changes for v5:
- Fix test that excludes the add/adc asm block for constant zero 'add'.
arch/x86/include/asm/div64.h | 20 +++++++++------ include/linux/math64.h | 48 +++++++++++++++++++++++++++++++++++- lib/math/div64.c | 14 ++++++----- 3 files changed, 67 insertions(+), 15 deletions(-)
arch/x86/include/asm/div64.h +12 -8
@@ -84,21 +84,25 @@ static inline u64 mul_u32_u32(u32 a, u32 b)* Will generate an #DE when the result doesn't fit u64, could fix with an * __ex_table[] entry when it becomes an issue. */ -static inline u64 mul_u64_u64_div_u64 (u64 a,u64 mul,u64 div) +static inline u64 mul_u64_add_u64_div_u64 (u64 rax,u64 mul,u64 add,u64 div) { -u64 q; +u64 rdx; - asm ("mulq %2; divq %3" :"=a" (q) - :"a" (a),"rm" (mul),"rm" (div) - :"rdx" ); + asm ("mulq %[mul]" :"+a" (rax),"=d" (rdx) : [mul]"rm" (mul)); -return q; +if (!statically_true (!add)) + asm ("addq %[add], %[lo]; adcq $0, %[hi]" : + [lo]"+r" (rax), [hi]"+r" (rdx) : [add]"irm" (add)); + + asm ("divq %[div]" :"+a" (rax),"+d" (rdx) : [div]"rm" (div)); + +return rax; } -#define mul_u64_u64_div_u64 mul_u64_u64_div_u64 +#define mul_u64_add_u64_div_u64 mul_u64_add_u64_div_u64static inline u64 mul_u64_u32_div (u64 a,u32 mul,u32 div) { -return mul_u64_u64_div_u64 (a, mul, div); +return mul_u64_add_u64_div_u64 (a, mul,0 , div); }#define mul_u64_u32_div mul_u64_u32_div
include/linux/math64.h +47 -1
54 lines not on this page · raw message
lib/math/div64.c +8 -6
@@ -183,13 +183,13 @@ u32 iter_div_u64_rem(u64 dividend, u32 divisor, u64 *remainder) }EXPORT_SYMBOL (iter_div_u64_rem); -#ifndef mul_u64_u64_div_u64 -u64 mul_u64_u64_div_u64 (u64 a,u64 b,u64 d) +#ifndef mul_u64_add_u64_div_u64 +u64 mul_u64_add_u64_div_u64 (u64 a,u64 b,u64 c,u64 d) {#if defined(__SIZEOF_INT128__)/* native 64x64=128 bits multiplication */ -u128 prod = (u128 )a * b; +u128 prod = (u128 )a * b + c;u64 n_lo = prod, n_hi = prod >>64 ;#else
@@ -198,8 +198,10 @@ u64 mul_u64_u64_div_u64(u64 a, u64 b, u64 d)u32 a_lo = a, a_hi = a >>32 , b_lo = b, b_hi = b >>32 ;u64 x, y, z; - x = (u64 )a_lo * b_lo; - y = (u64 )a_lo * b_hi + (u32)(x >>32 ); +/* Since (x-1)(x-1) + 2(x-1) == x.x - 1 two u32 can be added to a u64 */ + x = (u64 )a_lo * b_lo + (u32 )c; + y = (u64 )a_lo * b_hi + (u32)(c >>32 ); + y += (u32)(x >>32 ); z = (u64 )a_hi * b_hi + (u32)(y >>32 ); y = (u64 )a_hi * b_lo + (u32 )y; z += (u32)(y >>32 );
@@ -265,5 +267,5 @@ u64 mul_u64_u64_div_u64(u64 a, u64 b, u64 d) return res; } -EXPORT_SYMBOL (mul_u64_u64_div_u64); +EXPORT_SYMBOL (mul_u64_add_u64_div_u64);#endif
Signature · 1 line
2.39.5
PATCH v5 5/9 lib: Add tests for mul_u64_u64_div_u64_roundup()
In reply to David LaightReplicate the existing mul_u64_u64_div_u64() test cases with round up. Update the shell script that verifies the table, remove the comment markers so that it can be directly pasted into a shell.
Rename the divisor from 'c' to 'd' to match mul_u64_add_u64_div_u64().
It any tests fail then fail the module load with -EINVAL.
---
Changes for v3:
- Rename 'c' to 'd' to match mul_u64_add_u64_div_u64()
Changes for v4:
- Fix shell script that verifies the table
lib/math/test_mul_u64_u64_div_u64.c | 122 +++++++++++++++++----------- 1 file changed, 73 insertions(+), 49 deletions(-)
lib/math/test_mul_u64_u64_div_u64.c +73 -49
@@ -10,61 +10,73 @@#include <linux/printk.h> #include <linux/math64.h> -typedef struct {u64 a;u64 b;u64 c;u64 result; }test_params ; +typedef struct {u64 a;u64 b;u64 d;u64 result;uint round_up;}test_params ;static test_params test_values[] = {/* this contains many edge values followed by a couple random values */ -{0xb ,0x7 ,0x3 ,0x19 }, -{0xffff0000 ,0xffff0000 ,0xf ,0x1110eeef00000000 }, -{0xffffffff ,0xffffffff ,0x1 ,0xfffffffe00000001 }, -{0xffffffff ,0xffffffff ,0x2 ,0x7fffffff00000000 }, -{0x1ffffffff ,0xffffffff ,0x2 ,0xfffffffe80000000 }, -{0x1ffffffff ,0xffffffff ,0x3 ,0xaaaaaaa9aaaaaaab }, -{0x1ffffffff ,0x1ffffffff ,0x4 ,0xffffffff00000000 }, -{0xffff000000000000 ,0xffff000000000000 ,0xffff000000000001 ,0xfffeffffffffffff }, -{0x3333333333333333 ,0x3333333333333333 ,0x5555555555555555 ,0x1eb851eb851eb851 }, -{0x7fffffffffffffff ,0x2 ,0x3 ,0x5555555555555554 }, -{0xffffffffffffffff ,0x2 ,0x8000000000000000 ,0x3 }, -{0xffffffffffffffff ,0x2 ,0xc000000000000000 ,0x2 }, -{0xffffffffffffffff ,0x4000000000000004 ,0x8000000000000000 ,0x8000000000000007 }, -{0xffffffffffffffff ,0x4000000000000001 ,0x8000000000000000 ,0x8000000000000001 }, -{0xffffffffffffffff ,0x8000000000000001 ,0xffffffffffffffff ,0x8000000000000001 }, -{0xfffffffffffffffe ,0x8000000000000001 ,0xffffffffffffffff ,0x8000000000000000 }, -{0xffffffffffffffff ,0x8000000000000001 ,0xfffffffffffffffe ,0x8000000000000001 }, -{0xffffffffffffffff ,0x8000000000000001 ,0xfffffffffffffffd ,0x8000000000000002 }, -{0x7fffffffffffffff ,0xffffffffffffffff ,0xc000000000000000 ,0xaaaaaaaaaaaaaaa8 }, -{0xffffffffffffffff ,0x7fffffffffffffff ,0xa000000000000000 ,0xccccccccccccccca }, -{0xffffffffffffffff ,0x7fffffffffffffff ,0x9000000000000000 ,0xe38e38e38e38e38b }, -{0x7fffffffffffffff ,0x7fffffffffffffff ,0x5000000000000000 ,0xccccccccccccccc9 }, -{0xffffffffffffffff ,0xfffffffffffffffe ,0xffffffffffffffff ,0xfffffffffffffffe }, -{0xe6102d256d7ea3ae ,0x70a77d0be4c31201 ,0xd63ec35ab3220357 ,0x78f8bf8cc86c6e18 }, -{0xf53bae05cb86c6e1 ,0x3847b32d2f8d32e0 ,0xcfd4f55a647f403c ,0x42687f79d8998d35 }, -{0x9951c5498f941092 ,0x1f8c8bfdf287a251 ,0xa3c8dc5f81ea3fe2 ,0x1d887cb25900091f }, -{0x374fee9daa1bb2bb ,0x0d0bfbff7b8ae3ef ,0xc169337bd42d5179 ,0x03bb2dbaffcbb961 }, -{0xeac0d03ac10eeaf0 ,0x89be05dfa162ed9b ,0x92bb1679a41f0e4b ,0xdc5f5cc9e270d216 }, +{0xb ,0x7 ,0x3 ,0x19 ,1 }, +{0xffff0000 ,0xffff0000 ,0xf ,0x1110eeef00000000 ,0 }, +{0xffffffff ,0xffffffff ,0x1 ,0xfffffffe00000001 ,0 }, +{0xffffffff ,0xffffffff ,0x2 ,0x7fffffff00000000 ,1 }, +{0x1ffffffff ,0xffffffff ,0x2 ,0xfffffffe80000000 ,1 }, +{0x1ffffffff ,0xffffffff ,0x3 ,0xaaaaaaa9aaaaaaab ,0 }, +{0x1ffffffff ,0x1ffffffff ,0x4 ,0xffffffff00000000 ,1 }, +{0xffff000000000000 ,0xffff000000000000 ,0xffff000000000001 ,0xfffeffffffffffff ,1 }, +{0x3333333333333333 ,0x3333333333333333 ,0x5555555555555555 ,0x1eb851eb851eb851 ,1 }, +{0x7fffffffffffffff ,0x2 ,0x3 ,0x5555555555555554 ,1 }, +{0xffffffffffffffff ,0x2 ,0x8000000000000000 ,0x3 ,1 }, +{0xffffffffffffffff ,0x2 ,0xc000000000000000 ,0x2 ,1 }, +{0xffffffffffffffff ,0x4000000000000004 ,0x8000000000000000 ,0x8000000000000007 ,1 }, +{0xffffffffffffffff ,0x4000000000000001 ,0x8000000000000000 ,0x8000000000000001 ,1 }, +{0xffffffffffffffff ,0x8000000000000001 ,0xffffffffffffffff ,0x8000000000000001 ,0 }, +{0xfffffffffffffffe ,0x8000000000000001 ,0xffffffffffffffff ,0x8000000000000000 ,1 }, +{0xffffffffffffffff ,0x8000000000000001 ,0xfffffffffffffffe ,0x8000000000000001 ,1 }, +{0xffffffffffffffff ,0x8000000000000001 ,0xfffffffffffffffd ,0x8000000000000002 ,1 }, +{0x7fffffffffffffff ,0xffffffffffffffff ,0xc000000000000000 ,0xaaaaaaaaaaaaaaa8 ,1 }, +{0xffffffffffffffff ,0x7fffffffffffffff ,0xa000000000000000 ,0xccccccccccccccca ,1 }, +{0xffffffffffffffff ,0x7fffffffffffffff ,0x9000000000000000 ,0xe38e38e38e38e38b ,1 }, +{0x7fffffffffffffff ,0x7fffffffffffffff ,0x5000000000000000 ,0xccccccccccccccc9 ,1 }, +{0xffffffffffffffff ,0xfffffffffffffffe ,0xffffffffffffffff ,0xfffffffffffffffe ,0 }, +{0xe6102d256d7ea3ae ,0x70a77d0be4c31201 ,0xd63ec35ab3220357 ,0x78f8bf8cc86c6e18 ,1 }, +{0xf53bae05cb86c6e1 ,0x3847b32d2f8d32e0 ,0xcfd4f55a647f403c ,0x42687f79d8998d35 ,1 }, +{0x9951c5498f941092 ,0x1f8c8bfdf287a251 ,0xa3c8dc5f81ea3fe2 ,0x1d887cb25900091f ,1 }, +{0x374fee9daa1bb2bb ,0x0d0bfbff7b8ae3ef ,0xc169337bd42d5179 ,0x03bb2dbaffcbb961 ,1 }, +{0xeac0d03ac10eeaf0 ,0x89be05dfa162ed9b ,0x92bb1679a41f0e4b ,0xdc5f5cc9e270d216 ,1 }, };/* * The above table can be verified with the following shell script: -* -* #!/bin/sh -* sed -ne 's/^{ \+\(.*\), \+\(.*\), \+\(.*\), \+\(.*\) },$/\1 \2 \3 \4/p' \ -* lib/math/test_mul_u64_u64_div_u64.c | -* while read a b c r; do -* expected=$( printf "obase=16; ibase=16; %X * %X / %X\n" $a $b $c | bc ) -* given=$( printf "%X\n" $r ) -* if [ "$expected" = "$given" ]; then -* echo "$a * $b / $c = $r OK" -* else -* echo "$a * $b / $c = $r is wrong" >&2 -* echo "should be equivalent to 0x$expected" >&2 -* exit 1 -* fi -* done + +#!/bin/sh +sed -ne 's/^{ \+\(.*\), \+\(.*\), \+\(.*\), \+\(.*\), \+\(.*\) },$/\1 \2 \3 \4 \5/p' \ +lib/math/test_mul_u64_u64_div_u64.c | +while read a b d r e; do +expected=$( printf "obase=16; ibase=16; %X * %X / %X\n" $a $b $d | bc ) +given=$( printf "%X\n" $r ) +if [ "$expected" = "$given" ]; then +echo "$a * $b / $d = $r OK" +else +echo "$a * $b / $d = $r is wrong" >&2 +echo "should be equivalent to 0x$expected" >&2 +exit 1 +fi +expected=$( printf "obase=16; ibase=16; (%X * %X + %X) / %X\n" $a $b $((d-1)) $d | bc ) +given=$( printf "%X\n" $((r + e)) ) +if [ "$expected" = "$given" ]; then +echo "$a * $b +/ $d = $(printf '%#x' $((r + e))) OK" +else +echo "$a * $b +/ $d = $(printf '%#x' $((r + e))) is wrong" >&2 +echo "should be equivalent to 0x$expected" >&2 +exit 1 +fi +done +*/ static int __init test_init (void ) { +int errors =0 ; +int tests =0 ;int i;pr_info ("Starting mul_u64_u64_div_u64() test\n" );
@@ -72,19 +84,31 @@ static int __init test_init(void)for (i =0 ; i <ARRAY_SIZE (test_values); i++) {u64 a = test_values[i].a;u64 b = test_values[i].b; -u64 c = test_values[i].c; +u64 d = test_values[i].d;u64 expected_result = test_values[i].result; -u64 result =mul_u64_u64_div_u64 (a, b, c); +u64 result =mul_u64_u64_div_u64 (a, b, d); +u64 result_up =mul_u64_u64_div_u64_roundup (a, b, d); + + tests +=2 ;if (result != expected_result) { -pr_err ("ERROR: 0x%016llx * 0x%016llx / 0x%016llx\n" , a, b, c); +pr_err ("ERROR: 0x%016llx * 0x%016llx / 0x%016llx\n" , a, b, d);pr_err ("ERROR: expected result: %016llx\n" , expected_result);pr_err ("ERROR: obtained result: %016llx\n" , result); + errors++; + } + expected_result += test_values[i].round_up; +if (result_up != expected_result) { +pr_err ("ERROR: 0x%016llx * 0x%016llx +/ 0x%016llx\n" , a, b, d); +pr_err ("ERROR: expected result: %016llx\n" , expected_result); +pr_err ("ERROR: obtained result: %016llx\n" , result_up); + errors++; } } -pr_info ("Completed mul_u64_u64_div_u64() test\n" ); -return 0 ; +pr_info ("Completed mul_u64_u64_div_u64() test, %d tests, %d errors\n" , + tests, errors); +return errors ? -EINVAL :0 ; } static void __exit test_exit(void)
Signature · 1 line
2.39.5
PATCH v5 6/9 lib: test_mul_u64_u64_div_u64: Test both generic and arch versions
In reply to David LaightChange the #if in div64.c so that test_mul_u64_u64_div_u64.c can compile and test the generic version (including the 'long multiply') on architectures (eg amd64) that define their own copy.
Test the kernel version and the locally compiled version on all arch. Output the time taken (in ns) on the 'test completed' trace.
For reference, on my zen 5, the optimised version takes ~220ns and the generic version ~3350ns.
Using the native multiply saves ~200ns and adding back the ilog2() 'optimisation' test adds ~50ms.
---
Changes for v4:
- Fix build on non x86 (eg arm32)
Changes for v5:
- Fix build on 32bit x86
lib/math/div64.c | 8 +++-- lib/math/test_mul_u64_u64_div_u64.c | 52 +++++++++++++++++++++++++---- 2 files changed, 51 insertions(+), 9 deletions(-)
lib/math/div64.c +6 -2
@@ -177,16 +177,18 @@ EXPORT_SYMBOL(div64_s64);* Iterative div/mod for use when dividend is not expected to be much * bigger than divisor. */ +#ifndef iter_div_u64_remu32 iter_div_u64_rem (u64 dividend,u32 divisor,u64 *remainder) {return __iter_div_u64_rem (dividend, divisor, remainder); }EXPORT_SYMBOL (iter_div_u64_rem); +#endif -#ifndef mul_u64_add_u64_div_u64 +#if !defined(mul_u64_add_u64_div_u64) || defined(test_mul_u64_add_u64_div_u64)u64 mul_u64_add_u64_div_u64 (u64 a,u64 b,u64 c,u64 d) { -#if defined(__SIZEOF_INT128__) +#if defined(__SIZEOF_INT128__) && !defined(test_mul_u64_add_u64_div_u64)/* native 64x64=128 bits multiplication */ u128 prod = (u128 )a * b + c;
@@ -267,5 +269,7 @@ u64 mul_u64_add_u64_div_u64(u64 a, u64 b, u64 c, u64 d) return res; } +#if !defined(test_mul_u64_add_u64_div_u64)EXPORT_SYMBOL (mul_u64_add_u64_div_u64);#endif +#endif
lib/math/test_mul_u64_u64_div_u64.c +45 -7
81 lines not on this page · raw message
Signature · 1 line
2.39.5
PATCH v5 7/9 lib: mul_u64_u64_div_u64() optimise multiply on 32bit x86
In reply to David Laightgcc generates horrid code for both ((u64)u32_a * u32_b) and (u64_a + u32_b). As well as the extra instructions it can generate a lot of spills to stack (including spills of constant zeros and even multiplies by constant zero).
mul_u32_u32() already exists to optimise the multiply. Add a similar add_u64_32() for the addition.
Disable both for clang - it generates better code without them.
Move the 64x64 => 128 multiply into a static inline helper function for code clarity.
No need for the a/b_hi/lo variables, the implicit casts on the function calls do the work for us.
Should have minimal effect on the generated code.
Use mul_u32_u32() and add_u64_u32() in the 64x64 => 128 multiply in mul_u64_add_u64_div_u64().
---
Changes for v4:
- merge in patch 8.
- Add comments about gcc being 'broken' for mixed 32/64 bit maths. clang doesn't have the same issues.
- Use a #define for define mul_add() to avoid 'defined but not used' errors.
arch/x86/include/asm/div64.h | 19 +++++++++++++++++ include/linux/math64.h | 11 ++++++++++ lib/math/div64.c | 40 +++++++++++++++++++++++------------- 3 files changed, 56 insertions(+), 14 deletions(-)
arch/x86/include/asm/div64.h +19 -0
@@ -60,6 +60,12 @@ static inline u64 div_u64_rem(u64 dividend, u32 divisor, u32 *remainder) }#define div_u64_rem div_u64_rem +/* +* gcc tends to zero extend 32bit values and do full 64bit maths. +* Define asm functions that avoid this. +* (clang generates better code for the C versions.) +*/ +#ifndef __clang__static inline u64 mul_u32_u32 (u32 a,u32 b) {u32 high, low;
@@ -71,6 +77,19 @@ static inline u64 mul_u32_u32(u32 a, u32 b) }#define mul_u32_u32 mul_u32_u32 +static inline u64 add_u64_u32 (u64 a,u32 b) +{ +u32 high = a >>32 , low = a; + + asm ("addl %[b], %[low]; adcl $0, %[high]" + : [low]"+r" (low), [high]"+r" (high) + : [b]"rm" (b) ); + +return low | (u64 )high <<32 ; +} +#define add_u64_u32 add_u64_u32 +#endif +/* * __div64_32() is never called on x86, so prevent the * generic definition from getting built.
include/linux/math64.h +11 -0
@@ -158,6 +158,17 @@ static inline u64 mul_u32_u32(u32 a, u32 b) }#endif +#ifndef add_u64_u32 +/* +* Many a GCC version also messes this up. +* Zero extending b and then spilling everything to stack. +*/ +static inline u64 add_u64_u32 (u64 a,u32 b) +{ +return a + b; +} +#endif +#if defined(CONFIG_ARCH_SUPPORTS_INT128 ) && defined(__SIZEOF_INT128__)#ifndef mul_u64_u32_shr
lib/math/div64.c +26 -14
@@ -186,33 +186,45 @@ EXPORT_SYMBOL(iter_div_u64_rem);#endif #if !defined(mul_u64_add_u64_div_u64) || defined(test_mul_u64_add_u64_div_u64) -u64 mul_u64_add_u64_div_u64 (u64 a,u64 b,u64 c,u64 d) -{ + +#define mul_add (a, b, c) add_u64_u32(mul_u32_u32(a, b), c) +#if defined(__SIZEOF_INT128__) && !defined(test_mul_u64_add_u64_div_u64) +static inline u64 mul_u64_u64_add_u64 (u64 *p_lo,u64 a,u64 b,u64 c) +{/* native 64x64=128 bits multiplication */ u128 prod = (u128 )a * b + c; -u64 n_lo = prod, n_hi = prod >>64 ; + + *p_lo = prod; +return prod >>64 ; +}#else -/* perform a 64x64=128 bits multiplication manually */ -u32 a_lo = a, a_hi = a >>32 , b_lo = b, b_hi = b >>32 ; +static inline u64 mul_u64_u64_add_u64 (u64 *p_lo,u64 a,u64 b,u64 c) +{ +/* perform a 64x64=128 bits multiplication in 32bit chunks */ u64 x, y, z;/* Since (x-1)(x-1) + 2(x-1) == x.x - 1 two u32 can be added to a u64 */ - x = (u64 )a_lo * b_lo + (u32 )c; - y = (u64 )a_lo * b_hi + (u32)(c >>32 ); - y += (u32)(x >>32 ); - z = (u64 )a_hi * b_hi + (u32)(y >>32 ); - y = (u64 )a_hi * b_lo + (u32 )y; - z += (u32)(y >>32 ); - x = (y <<32 ) + (u32 )x; - -u64 n_lo = x, n_hi = z; + x =mul_add (a, b, c); + y =mul_add (a, b >>32 , c >>32 ); + y =add_u64_u32 (y, x >>32 ); + z =mul_add (a >>32 , b >>32 , y >>32 ); + y =mul_add (a >>32 , b, y); + *p_lo = (y <<32 ) + (u32 )x; +return add_u64_u32 (z, y >>32 ); +}#endif +u64 mul_u64_add_u64_div_u64 (u64 a,u64 b,u64 c,u64 d) +{ +u64 n_lo, n_hi; + + n_hi =mul_u64_u64_add_u64 (&n_lo, a, b, c); +if (!n_hi)return div64_u64 (n_lo, d);
Signature · 1 line
2.39.5
PATCH v5 8/9 lib: mul_u64_u64_div_u64() Optimise the divide code
In reply to David LaightReplace the bit by bit algorithm with one that generates 16 bits per iteration on 32bit architectures and 32 bits on 64bit ones.
On my zen 5 this reduces the time for the tests (using the generic code) from ~3350ns to ~1000ns.
Running the 32bit algorithm on 64bit x86 takes ~1500ns. It'll be slightly slower on a real 32bit system, mostly due to register pressure.
The savings for 32bit x86 are much higher (tested in userspace). The worst case (lots of bits in the quotient) drops from ~900 clocks to ~130 (pretty much independant of the arguments).
Other 32bit architectures may see better savings.
It is possibly to optimise for divisors that span less than __LONG_WIDTH__/2 bits. However I suspect they don't happen that often and it doesn't remove any slow cpu divide instructions which dominate the result.
Typical improvements for 64bit random divides:
old new sandy bridge: 470 150 haswell: 400 144 piledriver: 960 467 I think rdpmc is very slow.
zen5: 244 80
(Timing is 'rdpmc; mul_div(); rdpmc' with the multiply depending on the first rdpmc and the second rdpmc depending on the quotient.)
Object code (64bit x86 test program): old 0x173 new 0x141.
---
Algorithm unchanged since v3.
lib/math/div64.c | 124 ++++++++++++++++++++++++++++++++--------------- 1 file changed, 85 insertions(+), 39 deletions(-)
lib/math/div64.c +85 -39
@@ -190,7 +190,6 @@ EXPORT_SYMBOL(iter_div_u64_rem);#define mul_add (a, b, c) add_u64_u32(mul_u32_u32(a, b), c)#if defined(__SIZEOF_INT128__) && !defined(test_mul_u64_add_u64_div_u64) -static inline u64 mul_u64_u64_add_u64 (u64 *p_lo,u64 a,u64 b,u64 c) {/* native 64x64=128 bits multiplication */
@@ -199,9 +198,7 @@ static inline u64 mul_u64_u64_add_u64(u64 *p_lo, u64 a, u64 b, u64 c) *p_lo = prod; return prod >> 64; } -#else -static inline u64 mul_u64_u64_add_u64 (u64 *p_lo,u64 a,u64 b,u64 c) {/* perform a 64x64=128 bits multiplication in 32bit chunks */
@@ -216,12 +213,37 @@ static inline u64 mul_u64_u64_add_u64(u64 *p_lo, u64 a, u64 b, u64 c) *p_lo = (y <<32 ) + (u32 )x;return add_u64_u32 (z, y >>32 ); } +#endif + +#ifndef BITS_PER_ITER +#define BITS_PER_ITER (__LONG_WIDTH__ >= 64 ? 32 : 16) +#endif + +#if BITS_PER_ITER ==32 +#define mul_u64_long_add_u64 (p_lo, a, b, c) mul_u64_u64_add_u64(p_lo, a, b, c) +#define add_u64_long (a, b) ((a) + (b)) +#else +#undef BITS_PER_ITER +#define BITS_PER_ITER 16 +static inline u32 mul_u64_long_add_u64 (u64 *p_lo,u64 a,u32 b,u64 c) +{ +u64 n_lo =mul_add (a, b, c); +u64 n_med =mul_add (a >>32 , b, c >>32 ); + + n_med =add_u64_u32 (n_med, n_lo >>32 ); + *p_lo = n_med <<32 | (u32 )n_lo; +return n_med >>32 ; +} +#define add_u64_long (a, b) add_u64_u32(a, b)#endif u64 mul_u64_add_u64_div_u64 (u64 a,u64 b,u64 c,u64 d) { -u64 n_lo, n_hi; +unsigned long d_msig, q_digit; +unsigned int reps, d_z_hi; +u64 quotient, n_lo, n_hi; +u32 overflow; n_hi =mul_u64_u64_add_u64 (&n_lo, a, b, c);
@@ -240,46 +262,70 @@ u64 mul_u64_add_u64_div_u64(u64 a, u64 b, u64 c, u64 d)return ~0ULL ; } -int shift =__builtin_ctzll (d); - -/* try reducing the fraction in case the dividend becomes <= 64 bits */ -if ((n_hi >> shift) ==0 ) { -u64 n = shift ? (n_lo >> shift) | (n_hi << (64 - shift)) : n_lo; - -return div64_u64 (n, d >> shift); -/* -* The remainder value if needed would be: -* res = div64_u64_rem(n, d >> shift, &rem); -* rem = (rem << shift) + (n_lo - (n << shift)); -*/ +/* Left align the divisor, shifting the dividend to match */ + d_z_hi =__builtin_clzll (d); +if (d_z_hi) { + d <<= d_z_hi; + n_hi = n_hi << d_z_hi | n_lo >> (64 - d_z_hi); + n_lo <<= d_z_hi; } -/* Do the full 128 by 64 bits division */ - - shift =__builtin_clzll (d); - d <<= shift; - -int p =64 + shift; -u64 res =0 ; -bool carry; + reps =64 /BITS_PER_ITER ; +/* Optimise loop count for small dividends */ +if (!(u32)(n_hi >>32 )) { + reps -=32 /BITS_PER_ITER ; + n_hi = n_hi <<32 | n_lo >>32 ; + n_lo <<=32 ; + } +#if BITS_PER_ITER ==16 +if (!(u32)(n_hi >>48 )) { + reps--; + n_hi =add_u64_u32 (n_hi <<16 , n_lo >>48 ); + n_lo <<=16 ; + } +#endif -do { - carry = n_hi >>63 ; - shift = carry ?1 :__builtin_clzll (n_hi); -if (p < shift) -break ; - p -= shift; - n_hi <<= shift; - n_hi |= n_lo >> (64 - shift); - n_lo <<= shift; -if (carry || (n_hi >= d)) { - n_hi -= d; - res |=1ULL << p; +/* Invert the dividend so we can use add instead of subtract. */ + n_lo = ~n_lo; + n_hi = ~n_hi; + +/* +* Get the most significant BITS_PER_ITER bits of the divisor. +* This is used to get a low 'guestimate' of the quotient digit. +*/ + d_msig = (d >> (64 -BITS_PER_ITER )) +1 ; + +/* +* Now do a 'long division' with BITS_PER_ITER bit 'digits'. +* The 'guess' quotient digit can be low and BITS_PER_ITER+1 bits. +* The worst case is dividing ~0 by 0x8000 which requires two subtracts. +*/ + quotient =0 ; +while (reps--) { + q_digit = (unsigned long )(~n_hi >> (64 -2 *BITS_PER_ITER )) / d_msig; +/* Shift 'n' left to align with the product q_digit * d */ + overflow = n_hi >> (64 -BITS_PER_ITER ); + n_hi =add_u64_u32 (n_hi <<BITS_PER_ITER , n_lo >> (64 -BITS_PER_ITER )); + n_lo <<=BITS_PER_ITER ; +/* Add product to negated divisor */ + overflow +=mul_u64_long_add_u64 (&n_hi, d, q_digit, n_hi); +/* Adjust for the q_digit 'guestimate' being low */ +while (overflow <0xffffffff >> (32 -BITS_PER_ITER )) { + q_digit++; + n_hi += d; + overflow += n_hi < d; } - }while (n_hi); -/* The remainder value if needed would be n_hi << p */ + quotient =add_u64_long (quotient <<BITS_PER_ITER , q_digit); + } -return res; +/* +* The above only ensures the remainder doesn't overflow, +* it can still be possible to add (aka subtract) another copy +* of the divisor. +*/ +if ((n_hi + d) > n_hi) + quotient++; +return quotient; }#if !defined(test_mul_u64_add_u64_div_u64)EXPORT_SYMBOL (mul_u64_add_u64_div_u64);
Signature · 1 line
2.39.5
PATCH v5 9/9 lib: test_mul_u64_u64_div_u64: Test the 32bit code on 64bit
In reply to David LaightThere are slight differences in the mul_u64_add_u64_div_u64() code between 32bit and 64bit systems.
Compile and test the 32bit version on 64bit hosts for better test coverage.
---
Changes for v4:
- Fix build on non x86-64
lib/math/test_mul_u64_u64_div_u64.c | 29 +++++++++++++++++++++++++++++ 1 file changed, 29 insertions(+)
lib/math/test_mul_u64_u64_div_u64.c +29 -0
@@ -74,6 +74,10 @@ done*/ static u64 test_mul_u64_add_u64_div_u64 (u64 a,u64 b,u64 c,u64 d); +#if __LONG_WIDTH__ >=64 +#define TEST_32BIT_DIV +static u64 test_mul_u64_add_u64_div_u64_32bit (u64 a,u64 b,u64 c,u64 d); +#endif static int __init test_run (unsigned int fn_no,const char *fn_name) {
@@ -100,6 +104,12 @@ static int __init test_run(unsigned int fn_no, const char *fn_name) result =test_mul_u64_add_u64_div_u64 (a, b,0 , d); result_up =test_mul_u64_add_u64_div_u64 (a, b, d -1 , d);break ; +#ifdef TEST_32BIT_DIV +case 2 : + result =test_mul_u64_add_u64_div_u64_32bit (a, b,0 , d); + result_up =test_mul_u64_add_u64_div_u64_32bit (a, b, d -1 , d); +break ; +#endif } tests +=2 ;
@@ -131,6 +141,10 @@ static int __init test_init(void)return -EINVAL ;if (test_run (1 ,"test_mul_u64_u64_div_u64" ))return -EINVAL ; +#ifdef TEST_32BIT_DIV +if (test_run (2 ,"test_mul_u64_u64_div_u64_32bit" )) +return -EINVAL ; +#endif return 0 ; }
@@ -153,6 +167,21 @@ static void __exit test_exit(void)#include "div64.c" +#ifdef TEST_32BIT_DIV +/* Recompile the generic code for 32bit long */ +#undef test_mul_u64_add_u64_div_u64 +#define test_mul_u64_add_u64_div_u64 test_mul_u64_add_u64_div_u64_32bit +#undef BITS_PER_ITER +#define BITS_PER_ITER 16 + +#define mul_u64_u64_add_u64 mul_u64_u64_add_u64_32bit +#undef mul_u64_long_add_u64 +#undef add_u64_long +#undef mul_add + +#include "div64.c" +#endif +module_init (test_init);module_exit (test_exit);
Signature · 1 line
2.39.5
Re: [PATCH v5 next 3/9] lib: mul_u64_u64_div_u64() simplify check for a 64bit product
In reply to David LaightQuoted 25 lines · David Laight
If the product is only 64bits
div64_u64()can be used for the divide. Replace the pre-multiply check (ilog2(a) + ilog2(b) <= 62) with a simple post-multiply check that the high 64bits are zero.This has the advantage of being simpler, more accurate and less code. It will always be faster when the product is larger than 64bits.
Most 64bit cpu have a native 64x64=128 bit multiply, this is needed (for the low 64bits) even when
div64_u64()is called - so the early check gains nothing and is just extra code.32bit cpu will need a compare (etc) to generate the 64bit
ilog2()from two 32bit bit scans - so that is non-trivial.
(Never mind the mess of x86's 'bsr' and any oddball cpu without fast bit-scan instructions.)
Whereas the additional instructions for the 128bit multiply result are pretty much one multiply and two adds (typically the 'adc $0,%reg' can be run in parallel with the instruction that follows).The only outliers are 64bit systems without 128bit mutiply and simple in order 32bit ones with fast bit scan but needing extra instructions to get the high bits of the multiply result. I doubt it makes much difference to either, the latter is definitely not mainstream.
Quoted 4 lines
If anyone is worried about the analysis they can look at the generated code for x86 (especially when cmov isn't used).
Signed-off-by David Laight <[email protected]>
Quoted 38 lines
---
Split from patch 3 for v2.
Changes for v5:
- Test for small dividends before overflow.lib/math/div64.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)lib/math/div64.c +3
-3@@ -186,9 +186,6 @@ EXPORT_SYMBOL(iter_div_u64_rem);#ifndef mul_u64_u64_div_u64u64 mul_u64_u64_div_u64 (u64 a,u64 b,u64 d) { -if (ilog2 (a) +ilog2 (b) <=62 ) -return div64_u64 (a * b, d); -#if defined(__SIZEOF_INT128__)/* native 64x64=128 bits multiplication */ @@ -212,6 +209,9 @@ u64 mul_u64_u64_div_u64(u64 a, u64 b, u64 d)#endif +if (!n_hi) +return div64_u64 (n_lo, d); +if (unlikely (n_hi >= d)) {/* trigger runtime exception if divisor is zero */ if (d ==0 ) {Signature · 1 line
2.39.5
H. Peter Anvin
reviewerRe: [PATCH v5 next 7/9] lib: mul_u64_u64_div_u64() optimise multiply on 32bit x86
In reply to David LaightQuoted 146 lines · David Laight
gcc generates horrid code for both ((u64)u32_a *
u32_b) and (u64_a+u32_b). As well as the extra instructions it can generate a lot of spills to stack (including spills of constant zeros and even multiplies by constant zero).
mul_u32_u32()already exists to optimise the multiply. Add a similaradd_u64_32()for the addition.
Disable both for clang - it generates better code without them.Move the 64x64 => 128 multiply into a static inline helper function for code clarity.
No need for the a/b_hi/lo variables, the implicit casts on the function calls do the work for us.
Should have minimal effect on the generated code.Use
mul_u32_u32()andadd_u64_u32()in the 64x64 => 128 multiply inmul_u64_add_u64_div_u64().Signed-off-by David Laight <[email protected]>Reviewed-by Nicolas Pitre <[email protected]>---
Changes for v4:
- merge in patch 8.
- Add comments about gcc being 'broken' for mixed 32/64 bit maths. clang doesn't have the same issues.
- Use a #define for definemul_add()to avoid 'defined but not used' errors.arch/x86/include/asm/div64.h | 19 +++++++++++++++++ include/linux/math64.h | 11 ++++++++++ lib/math/div64.c | 40 +++++++++++++++++++++++------------- 3 files changed, 56 insertions(+), 14 deletions(-)arch/x86/include/asm/div64.h +19
-0@@ -60,6 +60,12 @@ static inline u64 div_u64_rem(u64 dividend, u32 divisor, u32 *remainder) }#define div_u64_rem div_u64_rem +/* +* gcc tends to zero extend 32bit values and do full 64bit maths. +* Define asm functions that avoid this. +* (clang generates better code for the C versions.) +*/ +#ifndef __clang__static inline u64 mul_u32_u32 (u32 a,u32 b) {u32 high, low;@@ -71,6 +77,19 @@ static inline u64 mul_u32_u32(u32 a, u32 b) }#define mul_u32_u32 mul_u32_u32 +static inline u64 add_u64_u32 (u64 a,u32 b) +{ +u32 high = a >>32 , low = a; + + asm ("addl %[b], %[low]; adcl $0, %[high]" + : [low]"+r" (low), [high]"+r" (high) + : [b]"rm" (b) ); + +return low | (u64 )high <<32 ; +} +#define add_u64_u32 add_u64_u32 +#endif +/* * __div64_32() is never called on x86, so prevent the * generic definition from getting built. include/linux/math64.h +11
-0@@ -158,6 +158,17 @@ static inline u64 mul_u32_u32(u32 a, u32 b) }#endif +#ifndef add_u64_u32 +/* +* Many a GCC version also messes this up. +* Zero extending b and then spilling everything to stack. +*/ +static inline u64 add_u64_u32 (u64 a,u32 b) +{ +return a + b; +} +#endif +#if defined(CONFIG_ARCH_SUPPORTS_INT128 ) && defined(__SIZEOF_INT128__)#ifndef mul_u64_u32_shrlib/math/div64.c +20
-14@@ -186,33 +186,45 @@ EXPORT_SYMBOL(iter_div_u64_rem);#endif #if !defined(mul_u64_add_u64_div_u64) || defined(test_mul_u64_add_u64_div_u64) -u64 mul_u64_add_u64_div_u64 (u64 a,u64 b,u64 c,u64 d) -{ + +#define mul_add (a, b, c) add_u64_u32(mul_u32_u32(a, b), c) +#if defined(__SIZEOF_INT128__) && !defined(test_mul_u64_add_u64_div_u64) +static inline u64 mul_u64_u64_add_u64 (u64 *p_lo,u64 a,u64 b,u64 c) +{/* native 64x64=128 bits multiplication */ u128 prod = (u128 )a * b + c; -u64 n_lo = prod, n_hi = prod >>64 ; + + *p_lo = prod; +return prod >>64 ; +}#else -/* perform a 64x64=128 bits multiplication manually */ -u32 a_lo = a, a_hi = a >>32 , b_lo = b, b_hi = b >>32 ; +static inline u64 mul_u64_u64_add_u64 (u64 *p_lo,u64 a,u64 b,u64 c) +{ +/* perform a 64x64=128 bits multiplication in 32bit chunks */ u64 x, y, z;/* Since (x-1)(x-1) + 2(x-1) == x.x - 1 two u32 can be added to a u64 */ - x = (u64 )a_lo * b_lo + (u32 )c; - y = (u64 )a_lo * b_hi + (u32)(c >>32 ); - y += (u32)(x >>32 ); - z = (u64 )a_hi * b_hi + (u32)(y >>32 ); - y = (u64 )a_hi * b_lo + (u32 )y; - z += (u32)(y >>32 ); - x = (y <<32 ) + (u32 )x; - -u64 n_lo = x, n_hi = z; + x =mul_add (a, b, c); + y =mul_add (a, b >>32 , c >>32 ); + y =add_u64_u32 (y, x >>32 ); + z =mul_add (a >>32 , b >>32 , y >>32 ); + y =mul_add (a >>32 , b, y); + *p_lo = (y <<32 ) + (u32 )x; +return add_u64_u32 (z, y >>32 ); +}#endif
lib/math/div64.c EXPORT_SYMBOL(iter_div_u64_rem); · line 230
Quoted 10 lines
lib/math/div64.c +6
-0+u64 mul_u64_add_u64_div_u64 (u64 a,u64 b,u64 c,u64 d) +{ +u64 n_lo, n_hi; + + n_hi =mul_u64_u64_add_u64 (&n_lo, a, b, c); +if (!n_hi)return div64_u64 (n_lo, d);
By the way have you filed gcc bug reports for this?
H. Peter Anvin
reviewerRe: [PATCH v5 next 4/9] lib: Add mul_u64_add_u64_div_u64() and mul_u64_u64_div_u64_roundup()
In reply to David LaightQuoted 169 lines · David Laight
The existing
mul_u64_u64_div_u64()rounds down, a 'rounding up' variant needs 'divisor - 1' adding in between the multiply and divide so cannot easily be done by a caller.Add
mul_u64_add_u64_div_u64(a, b, c, d)that calculates (a * b + c)/d and implement the 'round down' and 'round up' using it.Update the x86-64 asm to optimise for 'c' being a constant zero.
Add kerndoc definitions for all three functions.
Signed-off-by David Laight <[email protected]>Reviewed-by Nicolas Pitre <[email protected]>---
Changes for v2 (formally patch 1/3):
- Reinstate the early call todiv64_u64()on 32bit when 'c' is zero. Although I'm not convinced the path is common enough to be worth the twoilog2()calls.Changes for v3 (formally patch 3/4):
- The early call todiv64_u64()has been removed by patch 3. Pretty much guaranteed to be a pessimisation.Changes for v4:
- For x86-64 split the multiply, add and divide into three asm blocks. (gcc makes a pigs breakfast of (u128)a * b + c)
- Change the kerndoc since divide by zero will (probably) fault.Changes for v5:
- Fix test that excludes the add/adc asm block for constant zero 'add'.arch/x86/include/asm/div64.h | 20 +++++++++------ include/linux/math64.h | 48 +++++++++++++++++++++++++++++++++++- lib/math/div64.c | 14 ++++++----- 3 files changed, 67 insertions(+), 15 deletions(-)arch/x86/include/asm/div64.h +12
-8@@ -84,21 +84,25 @@ static inline u64 mul_u32_u32(u32 a, u32 b)* Will generate an #DE when the result doesn't fit u64, could fix with an * __ex_table[] entry when it becomes an issue. */ -static inline u64 mul_u64_u64_div_u64 (u64 a,u64 mul,u64 div) +static inline u64 mul_u64_add_u64_div_u64 (u64 rax,u64 mul,u64 add,u64 div) { -u64 q; +u64 rdx; - asm ("mulq %2; divq %3" :"=a" (q) - :"a" (a),"rm" (mul),"rm" (div) - :"rdx" ); + asm ("mulq %[mul]" :"+a" (rax),"=d" (rdx) : [mul]"rm" (mul)); -return q; +if (!statically_true (!add)) + asm ("addq %[add], %[lo]; adcq $0, %[hi]" : + [lo]"+r" (rax), [hi]"+r" (rdx) : [add]"irm" (add)); + + asm ("divq %[div]" :"+a" (rax),"+d" (rdx) : [div]"rm" (div)); + +return rax; } -#define mul_u64_u64_div_u64 mul_u64_u64_div_u64 +#define mul_u64_add_u64_div_u64 mul_u64_add_u64_div_u64static inline u64 mul_u64_u32_div (u64 a,u32 mul,u32 div) { -return mul_u64_u64_div_u64 (a, mul, div); +return mul_u64_add_u64_div_u64 (a, mul,0 , div); }#define mul_u64_u32_div mul_u64_u32_divinclude/linux/math64.h +47
-1@@ -282,7 +282,53 @@ static inline u64 mul_u64_u32_div(u64 a, u32 mul, u32 divisor) }#endif /* mul_u64_u32_div */ -u64 mul_u64_u64_div_u64 (u64 a,u64 mul,u64 div); +/** +* mul_u64_add_u64_div_u64 - unsigned 64bit multiply, add, and divide +* @a: first unsigned 64bit multiplicand +* @b: second unsigned 64bit multiplicand +* @c: unsigned 64bit addend +* @d: unsigned 64bit divisor +* +* Multiply two 64bit values together to generate a 128bit product +* add a third value and then divide by a fourth. +* The Generic code divides by 0 if @d is zero and returns ~0 on overflow. +* Architecture specific code may trap on zero or overflow. +* +* Return: (@a * @b + @c) / @d +*/ +u64 mul_u64_add_u64_div_u64 (u64 a,u64 b,u64 c,u64 d); + +/** +* mul_u64_u64_div_u64 - unsigned 64bit multiply and divide +* @a: first unsigned 64bit multiplicand +* @b: second unsigned 64bit multiplicand +* @d: unsigned 64bit divisor +* +* Multiply two 64bit values together to generate a 128bit product +* and then divide by a third value. +* The Generic code divides by 0 if @d is zero and returns ~0 on overflow. +* Architecture specific code may trap on zero or overflow. +* +* Return: @a * @b / @d +*/ +#define mul_u64_u64_div_u64 (a, b, d) mul_u64_add_u64_div_u64(a, b, 0, d) + +/** +* mul_u64_u64_div_u64_roundup - unsigned 64bit multiply and divide rounded up +* @a: first unsigned 64bit multiplicand +* @b: second unsigned 64bit multiplicand +* @d: unsigned 64bit divisor +* +* Multiply two 64bit values together to generate a 128bit product +* and then divide and round up. +* The Generic code divides by 0 if @d is zero and returns ~0 on overflow. +* Architecture specific code may trap on zero or overflow. +* +* Return: (@a * @b + @d - 1) / @d +*/ +#define mul_u64_u64_div_u64_roundup (a, b, d) \ + ({ u64 _tmp = (d); mul_u64_add_u64_div_u64(a, b, _tmp - 1, _tmp); }) +/** * DIV64_U64_ROUND_UP - unsigned 64bit divide with 64bit divisor rounded up lib/math/div64.c +7
-5@@ -183,13 +183,13 @@ u32 iter_div_u64_rem(u64 dividend, u32 divisor, u64 *remainder) }EXPORT_SYMBOL (iter_div_u64_rem); -#ifndef mul_u64_u64_div_u64 -u64 mul_u64_u64_div_u64 (u64 a,u64 b,u64 d) +#ifndef mul_u64_add_u64_div_u64 +u64 mul_u64_add_u64_div_u64 (u64 a,u64 b,u64 c,u64 d) {#if defined(__SIZEOF_INT128__)/* native 64x64=128 bits multiplication */ -u128 prod = (u128 )a * b; +u128 prod = (u128 )a * b + c;u64 n_lo = prod, n_hi = prod >>64 ;#else @@ -198,8 +198,10 @@ u64 mul_u64_u64_div_u64(u64 a, u64 b, u64 d)u32 a_lo = a, a_hi = a >>32 , b_lo = b, b_hi = b >>32 ;u64 x, y, z; - x = (u64 )a_lo * b_lo; - y = (u64 )a_lo * b_hi + (u32)(x >>32 ); +/* Since (x-1)(x-1) + 2(x-1) == x.x - 1 two u32 can be added to a u64 */ + x = (u64 )a_lo * b_lo + (u32 )c; + y = (u64 )a_lo * b_hi + (u32)(c >>32 ); + y += (u32)(x >>32 ); z = (u64 )a_hi * b_hi + (u32)(y >>32 ); y = (u64 )a_hi * b_lo + (u32 )y; z += (u32)(y >>32 );
lib/math/div64.c u64 mul_u64_u64_div_u64(u64 a, u64 b, u64 d) · line 271
Quoted 6 lines
lib/math/div64.c +1
-1@@ -265,5 +267,5 @@ u64 mul_u64_u64_div_u64(u64 a, u64 b, u64 d) return res; } -EXPORT_SYMBOL (mul_u64_u64_div_u64); +EXPORT_SYMBOL (mul_u64_add_u64_div_u64);#endif
For the roundup case, I'm somewhat curious how this compares with doing:
cmp $1, %rdx
sbb $-1, %rax
... after the division. At least it means not needing to compute d - 1, saving an instruction as well as a register. Unfortunately using an lea instruction to compute %rax (which otherwise would incorporate both) isn't possible since it doesn't set the flags.
The cmp; sbb sequence should be no slower than add; adc – I'm saying "no slower" because %rdx is never written to, so I think this is provably a better sequence; whether or not it is measurable is another thing (but if we are tweaking this stuff...)
Re: [PATCH v5 next 7/9] lib: mul_u64_u64_div_u64() optimise multiply on 32bit x86
In reply to H. Peter AnvinQuoted 67 lines · On Wed, 05 Nov 2025 15:45:29 -0800 "H. Peter Anvin" <[email protected]> wrote:
Quoted 66 lines · David Laight
gcc generates horrid code for both ((u64)u32_a *
u32_b) and (u64_a+u32_b). As well as the extra instructions it can generate a lot of spills to stack (including spills of constant zeros and even multiplies by constant zero).
mul_u32_u32()already exists to optimise the multiply. Add a similaradd_u64_32()for the addition.
Disable both for clang - it generates better code without them.Move the 64x64 => 128 multiply into a static inline helper function for code clarity.
No need for the a/b_hi/lo variables, the implicit casts on the function calls do the work for us.
Should have minimal effect on the generated code.Use
mul_u32_u32()andadd_u64_u32()in the 64x64 => 128 multiply inmul_u64_add_u64_div_u64().Signed-off-by David Laight <[email protected]>Reviewed-by Nicolas Pitre <[email protected]>---
Changes for v4:
- merge in patch 8.
- Add comments about gcc being 'broken' for mixed 32/64 bit maths. clang doesn't have the same issues.
- Use a #define for definemul_add()to avoid 'defined but not used' errors.arch/x86/include/asm/div64.h | 19 +++++++++++++++++ include/linux/math64.h | 11 ++++++++++ lib/math/div64.c | 40 +++++++++++++++++++++++------------- 3 files changed, 56 insertions(+), 14 deletions(-)arch/x86/include/asm/div64.h +18
-0@@ -60,6 +60,12 @@ static inline u64 div_u64_rem(u64 dividend, u32 divisor, u32 *remainder) }#define div_u64_rem div_u64_rem +/* +* gcc tends to zero extend 32bit values and do full 64bit maths. +* Define asm functions that avoid this. +* (clang generates better code for the C versions.) +*/ +#ifndef __clang__static inline u64 mul_u32_u32 (u32 a,u32 b) {u32 high, low;@@ -71,6 +77,19 @@ static inline u64 mul_u32_u32(u32 a, u32 b) }#define mul_u32_u32 mul_u32_u32 +static inline u64 add_u64_u32 (u64 a,u32 b) +{ +u32 high = a >>32 , low = a; + + asm ("addl %[b], %[low]; adcl $0, %[high]" + : [low]"+r" (low), [high]"+r" (high) + : [b]"rm" (b) ); + +return low | (u64 )high <<32 ; +} +#define add_u64_u32 add_u64_u32 +#endif
...
Quoted 2 lines
By the way have you filed gcc bug reports for this?
As in the need for the asm() above?
No...
I doubt one was filed when the mul version was added either. ISTR that some very recent gcc versions were a bit better, but it depends on minor code changes and compiler options.
I suspect that internally gcc sometimes keeps a 64bit value as two 32bit ones, but at other times it is assigned to a 64bit internal register. If the latter happens it always promotes a 32bit value to 64 bits and assigns to another 64bit register.
At that point it won't split the 64bit registers - so a lot of spills to stack happen when it tries to assign real registers.
So breath on an 'A' (dx:ax) constraint and the generated code is horrid.
Even the lo | (u64)hi << 32 can generate 'or' instructions.
The same happens for int128 on 64bit.
David
Re: [PATCH v5 next 4/9] lib: Add mul_u64_add_u64_div_u64() and mul_u64_u64_div_u64_roundup()
In reply to H. Peter AnvinQuoted 68 lines · On Wed, 05 Nov 2025 16:26:05 -0800 "H. Peter Anvin" <[email protected]> wrote:
Quoted 67 lines · David Laight
The existing
mul_u64_u64_div_u64()rounds down, a 'rounding up' variant needs 'divisor - 1' adding in between the multiply and divide so cannot easily be done by a caller.Add
mul_u64_add_u64_div_u64(a, b, c, d)that calculates (a * b + c)/d and implement the 'round down' and 'round up' using it.Update the x86-64 asm to optimise for 'c' being a constant zero.
Add kerndoc definitions for all three functions.
Signed-off-by David Laight <[email protected]>Reviewed-by Nicolas Pitre <[email protected]>---
Changes for v2 (formally patch 1/3):
- Reinstate the early call todiv64_u64()on 32bit when 'c' is zero. Although I'm not convinced the path is common enough to be worth the twoilog2()calls.Changes for v3 (formally patch 3/4):
- The early call todiv64_u64()has been removed by patch 3. Pretty much guaranteed to be a pessimisation.Changes for v4:
- For x86-64 split the multiply, add and divide into three asm blocks. (gcc makes a pigs breakfast of (u128)a * b + c)
- Change the kerndoc since divide by zero will (probably) fault.Changes for v5:
- Fix test that excludes the add/adc asm block for constant zero 'add'.arch/x86/include/asm/div64.h | 20 +++++++++------ include/linux/math64.h | 48 +++++++++++++++++++++++++++++++++++- lib/math/div64.c | 14 ++++++----- 3 files changed, 67 insertions(+), 15 deletions(-)arch/x86/include/asm/div64.h +11
-7@@ -84,21 +84,25 @@ static inline u64 mul_u32_u32(u32 a, u32 b)* Will generate an #DE when the result doesn't fit u64, could fix with an * __ex_table[] entry when it becomes an issue. */ -static inline u64 mul_u64_u64_div_u64 (u64 a,u64 mul,u64 div) +static inline u64 mul_u64_add_u64_div_u64 (u64 rax,u64 mul,u64 add,u64 div) { -u64 q; +u64 rdx; - asm ("mulq %2; divq %3" :"=a" (q) - :"a" (a),"rm" (mul),"rm" (div) - :"rdx" ); + asm ("mulq %[mul]" :"+a" (rax),"=d" (rdx) : [mul]"rm" (mul)); -return q; +if (!statically_true (!add)) + asm ("addq %[add], %[lo]; adcq $0, %[hi]" : + [lo]"+r" (rax), [hi]"+r" (rdx) : [add]"irm" (add)); + + asm ("divq %[div]" :"+a" (rax),"+d" (rdx) : [div]"rm" (div)); + +return rax; } -#define mul_u64_u64_div_u64 mul_u64_u64_div_u64 +#define mul_u64_add_u64_div_u64 mul_u64_add_u64_div_u64
...
Quoted 2 lines
For the roundup case, I'm somewhat curious how this compares with doing:
I guess you are referring to the x86-64 asm version (left above).
Quoted 6 lines
cmp $1, %rdx sbb $-1, %rax... after the division. At least it means not needing to compute d - 1, saving an instruction as well as a register.
Quoted 7 lines
Unfortunately using an lea instruction to compute %rax (which otherwise
would incorporate both) isn't possible since it doesn't set the flags.The cmp; sbb sequence should be no slower than add;
adc – I'm saying "no slower" because %rdx is never written to, so I think this is provably a better sequence; whether or not it is measurable is another thing (but if we are tweaking this stuff...)
I wanted the same function as the non-x64-64 version and 'multiply and add' possibly has other uses.
The instruction to calculate 'd - 1' (if not a constant) will usually execute in parallel with an earlier instruction (eg the multiply) so will be pretty much 'zero cost'.
The add/adc pair are in the 'register dependency chain' - so add a clock each. The same is true for your cmp/sbb pair.
(Except on pre-broadwell Intel cpu where adc/sbb are two clocks. I've lost the full reference, the initial changes fixed 'adc $0,x' and generated the carry flag immediately and only delayed the result. The doc said 'adc $0,reg' not 'adc $const,reg' - so maybe the sbb $-1 was two clocks for longer.)
David