[PATCH] x86: replace a few do_div() uses

Jan Beulich posted 1 patch 1 week, 5 days ago
Test gitlab-ci failed
Patches applied successfully (tree, apply log)
git fetch https://gitlab.com/xen-project/patchew/xen tags/patchew/01baee92-9d7f-5a2c-d63f-1de390bc10e2@suse.com

[PATCH] x86: replace a few do_div() uses

Posted by Jan Beulich 1 week, 5 days ago
When the macro's "return value" is not used, the macro use can be
replaced by a simply division, avoiding some obfuscation.

According to my observations, no change to generated code.

Signed-off-by: Jan Beulich <jbeulich@suse.com>
---
Arguably the ULL suffix (in write_watchdog_counter()) or the cast to
uint64_t (in div_sc()) aren't really needed in code which gets built for
64-bit only.

--- a/xen/arch/x86/cpu/intel.c
+++ b/xen/arch/x86/cpu/intel.c
@@ -392,9 +392,8 @@ static void intel_log_freq(const struct
             unsigned long long val = ecx;
 
             val *= ebx;
-            do_div(val, eax);
             printk("CPU%u: TSC: %u Hz * %u / %u = %Lu Hz\n",
-                   smp_processor_id(), ecx, ebx, eax, val);
+                   smp_processor_id(), ecx, ebx, eax, val / eax);
         }
         else if ( ecx | eax | ebx )
         {
--- a/xen/arch/x86/hpet.c
+++ b/xen/arch/x86/hpet.c
@@ -105,10 +105,7 @@ custom_param("hpet", parse_hpet_param);
 static inline unsigned long div_sc(unsigned long ticks, unsigned long nsec,
                                    int shift)
 {
-    uint64_t tmp = ((uint64_t)ticks) << shift;
-
-    do_div(tmp, nsec);
-    return (unsigned long) tmp;
+    return ((uint64_t)ticks << shift) / nsec;
 }
 
 /*
--- a/xen/arch/x86/nmi.c
+++ b/xen/arch/x86/nmi.c
@@ -292,10 +292,9 @@ static void clear_msr_range(unsigned int
 
 static inline void write_watchdog_counter(const char *descr)
 {
-    u64 count = (u64)cpu_khz * 1000;
+    uint64_t count = cpu_khz * 1000ULL / nmi_hz;
 
-    do_div(count, nmi_hz);
-    if(descr)
+    if ( descr )
         Dprintk("setting %s to -%#"PRIx64"\n", descr, count);
     wrmsrl(nmi_perfctr_msr, 0 - count);
 }
--- a/xen/arch/x86/time.c
+++ b/xen/arch/x86/time.c
@@ -610,8 +610,7 @@ static uint64_t xen_timer_cpu_frequency(
     struct vcpu_time_info *info = &this_cpu(vcpu_info)->time;
     uint64_t freq;
 
-    freq = 1000000000ULL << 32;
-    do_div(freq, info->tsc_to_system_mul);
+    freq = (1000000000ULL << 32) / info->tsc_to_system_mul;
     if ( info->tsc_shift < 0 )
         freq <<= -info->tsc_shift;
     else
@@ -2173,8 +2172,7 @@ void __init early_time_init(void)
     set_time_scale(&t->tsc_scale, tmp);
     t->stamp.local_tsc = boot_tsc_stamp;
 
-    do_div(tmp, 1000);
-    cpu_khz = (unsigned long)tmp;
+    cpu_khz = tmp / 1000;
     printk("Detected %lu.%03lu MHz processor.\n", 
            cpu_khz / 1000, cpu_khz % 1000);
 


Re: [PATCH] x86: replace a few do_div() uses

Posted by Andrew Cooper 1 week, 5 days ago
On 12/01/2022 09:00, Jan Beulich wrote:
> When the macro's "return value" is not used, the macro use can be
> replaced by a simply division, avoiding some obfuscation.
>
> According to my observations, no change to generated code.
>
> Signed-off-by: Jan Beulich <jbeulich@suse.com>

I like this change in principle, but see below.

do_div() needs to be deleted, because it's far too easy screw up.  At a
bare minimum, it should be replaced with a static inline that takes it's
first parameter by pointer, because then at least every callsite reads
correctly in terms of the C language.

> --- a/xen/arch/x86/time.c
> +++ b/xen/arch/x86/time.c
> @@ -610,8 +610,7 @@ static uint64_t xen_timer_cpu_frequency(
>      struct vcpu_time_info *info = &this_cpu(vcpu_info)->time;
>      uint64_t freq;
>  
> -    freq = 1000000000ULL << 32;
> -    do_div(freq, info->tsc_to_system_mul);
> +    freq = (1000000000ULL << 32) / info->tsc_to_system_mul;
>      if ( info->tsc_shift < 0 )
>          freq <<= -info->tsc_shift;

do_div()'s output is consumed here.  I don't think this hunk is safe to
convert.

~Andrew

Re: [PATCH] x86: replace a few do_div() uses

Posted by Jan Beulich 1 week, 5 days ago
On 12.01.2022 10:22, Andrew Cooper wrote:
> On 12/01/2022 09:00, Jan Beulich wrote:
>> When the macro's "return value" is not used, the macro use can be
>> replaced by a simply division, avoiding some obfuscation.
>>
>> According to my observations, no change to generated code.
>>
>> Signed-off-by: Jan Beulich <jbeulich@suse.com>
> 
> I like this change in principle, but see below.
> 
> do_div() needs to be deleted, because it's far too easy screw up.  At a
> bare minimum, it should be replaced with a static inline that takes it's
> first parameter by pointer, because then at least every callsite reads
> correctly in terms of the C language.

That ought to be a 2nd step, requiring agreement with Arm folks (and
adjustments to their code).

>> --- a/xen/arch/x86/time.c
>> +++ b/xen/arch/x86/time.c
>> @@ -610,8 +610,7 @@ static uint64_t xen_timer_cpu_frequency(
>>      struct vcpu_time_info *info = &this_cpu(vcpu_info)->time;
>>      uint64_t freq;
>>  
>> -    freq = 1000000000ULL << 32;
>> -    do_div(freq, info->tsc_to_system_mul);
>> +    freq = (1000000000ULL << 32) / info->tsc_to_system_mul;
>>      if ( info->tsc_shift < 0 )
>>          freq <<= -info->tsc_shift;
> 
> do_div()'s output is consumed here.  I don't think this hunk is safe to
> convert.

If by "output" you mean its "return value", then it clearly isn't
consumed. And I continue to think that I did express correctly the
effect do_div() did have on "freq".

Jan