12
我編寫了一個小程序，用於測量循環中的時間（通過內聯Sparc彙編代碼片段）。在Sparc 32位上處理值> 2^32的整數
一切都是正確的，直到我設置迭代次數大約在4.0 + 9（2^32以上）以上。
下面的代碼片段：
#include <stdio.h>
#include <sys/time.h>
#include <unistd.h>
#include <math.h>
#include <stdint.h>
int main (int argc, char *argv[])
{
// For indices
int i;
// Set the number of executions
int nRunning = atoi(argv[1]);
// Set the sums
double avgSum = 0.0;
double stdSum = 0.0;
// Average of execution time
double averageRuntime = 0.0;
// Standard deviation of execution time
double deviationRuntime = 0.0;
// Init sum
unsigned long long int sum = 0;
// Number of iterations
unsigned long long int nLoop = 4000000000ULL;
//uint64_t nLoop = 4000000000;
// DEBUG
printf("sizeof(unsigned long long int) = %zu\n",sizeof(unsigned long long int));
printf("sizeof(unsigned long int) = %zu\n",sizeof(unsigned long int));
// Time intervals
struct timeval tv1, tv2;
double diff;
// Loop for multiple executions
for (i=0; i<nRunning; i++)
{
// Start time
gettimeofday (&tv1, NULL);
// Loop with Sparc assembly into C source
asm volatile ("clr %%g1\n\t"
"clr %%g2\n\t"
"mov %1, %%g1\n" // %1 = input parameter
"loop:\n\t"
"add %%g2, 1, %%g2\n\t"
"subcc %%g1, 1, %%g1\n\t"
"bne loop\n\t"
"nop\n\t"
"mov %%g2, %0\n" // %0 = output parameter
: "=r" (sum) // output
: "r" (nLoop) // input
: "g1", "g2"); // clobbers
// End time
gettimeofday (&tv2, NULL);
// Compute runtime for loop
diff = (tv2.tv_sec - tv1.tv_sec) * 1000000ULL + (tv2.tv_usec - tv1.tv_usec);
// Summing diff time
avgSum += diff;
stdSum += (diff*diff);
// DEBUG
printf("diff = %e\n", diff);
printf("avgSum = %e\n", avgSum);
}
// Compute final averageRuntime
averageRuntime = avgSum/nRunning;
// Compute standard deviation
deviationRuntime = sqrt(stdSum/nRunning-averageRuntime*averageRuntime);
// Print results
printf("(Average Elapsed time, Standard deviation) = %e usec %e usec\n", averageRuntime, deviationRuntime);
// Print sum from assembly loop
printf("Sum = %llu\n", sum);
例如，nLoop < 2^32，我得到正確的值diff，avgSum和stdSum。事實上，printf，與nLoop = 4.0e+9，得出：
sizeof(unsigned long long int) = 8
sizeof(unsigned long int) = 4
diff = 9.617167e+06
avgSum = 9.617167e+06
diff = 9.499878e+06
avgSum = 1.911704e+07
(Average Elapsed time, Standard deviation) = 9.558522e+06 usec 5.864450e+04 usec
Sum = 4000000000
的代碼被編譯在Debian Sparc 32 bits Etch與gcc 4.1.2。
不幸的是，如果我拿例如nLoop = 5.0e+9，我會得到測量時間小而不正確的值;這裏是在這種情況下，printf的輸出：
sizeof(unsigned long long int) = 8
sizeof(unsigned long int) = 4
diff = 5.800000e+01
avgSum = 5.800000e+01
diff = 4.000000e+00
avgSum = 6.200000e+01
(Average Elapsed time, Standard deviation) = 3.100000e+01 usec 2.700000e+01 usec
Sum = 5000000000
我不知道在哪裏的問題可能來自使用uint64_t但沒有成功，我做其他檢查。
也許問題是我用32位操作系統處理large integers (> 2^32)或者它可能是不支持8字節整數的程序集內聯代碼。
如果有人能夠給我一些線索來解決這個錯誤，
問候
更新1：
繼@Andrew Henle的建議，我採取了同樣的代碼，但不是行內的Sparc彙編代碼片段，我只是放了一個簡單的循環。
下面是用簡單的迴路，其已得到nLoop = 5.0e+9（見行「unsigned long long int nLoop = 5000000000ULL;」的節目，所以上面的limit 2^32-1：
#include <stdio.h>
#include <stdlib.h>
#include <sys/time.h>
#include <unistd.h>
#include <math.h>
#include <stdint.h>
int main (int argc, char *argv[])
{
// For indices of nRunning
int i;
// For indices of nRunning
unsigned long long int j;
// Set the number of executions
int nRunning = atoi(argv[1]);
// Set the sums
unsigned long long int avgSum = 0;
unsigned long long int stdSum = 0;
// Average of execution time
double averageRuntime = 0.0;
// Standard deviation of execution time
double deviationRuntime = 0.0;
// Init sum
unsigned long long int sum;
// Number of iterations
unsigned long long int nLoop = 5000000000ULL;
// DEBUG
printf("sizeof(unsigned long long int) = %zu\n",sizeof(unsigned long long int));
printf("sizeof(unsigned long int) = %zu\n",sizeof(unsigned long int));
// Time intervals
struct timeval tv1, tv2;
unsigned long long int diff;
// Loop for multiple executions
for (i=0; i<nRunning; i++)
{
// Reset sum
sum = 0;
// Start time
gettimeofday (&tv1, NULL);
// Loop with Sparc assembly into C source
/* asm volatile ("clr %%g1\n\t"
"clr %%g2\n\t"
"mov %1, %%g1\n" // %1 = input parameter
"loop:\n\t"
"add %%g2, 1, %%g2\n\t"
"subcc %%g1, 1, %%g1\n\t"
"bne loop\n\t"
"nop\n\t"
"mov %%g2, %0\n" // %0 = output parameter
: "=r" (sum) // output
: "r" (nLoop) // input
: "g1", "g2"); // clobbers
*/
// Classic loop
for (j=0; j<nLoop; j++)
sum ++;
// End time
gettimeofday (&tv2, NULL);
// Compute runtime for loop
diff = (unsigned long long int) ((tv2.tv_sec - tv1.tv_sec) * 1000000 + (tv2.tv_usec - tv1.tv_usec));
// Summing diff time
avgSum += diff;
stdSum += (diff*diff);
// DEBUG
printf("diff = %llu\n", diff);
printf("avgSum = %llu\n", avgSum);
printf("stdSum = %llu\n", stdSum);
// Print sum from assembly loop
printf("Sum = %llu\n", sum);
}
// Compute final averageRuntime
averageRuntime = avgSum/nRunning;
// Compute standard deviation
deviationRuntime = sqrt(stdSum/nRunning-averageRuntime*averageRuntime);
// Print results
printf("(Average Elapsed time, Standard deviation) = %e usec %e usec\n", averageRuntime, deviationRuntime);
return 0;
}
此代碼段工作正常，即，可變sum打印爲 （見 「printf("Sum = %llu\n", sum)」）：
Sum = 5000000000
所以，問題來自於與Sparc的會議樓的版本
。 我懷疑，在該彙編代碼，行"mov %1, %%g1\n" // %1 = input parameter要差些存儲nLoop成%g1 register（我認爲%g1是一個32位寄存器，因此不能存儲值以上2^32-1）。
然而，在該行的輸出中的參數（變量sum）：
"mov %%g2, %0\n" // %0 = output parameter
高於極限，因爲它是等於50億。
附上與大會環路版本之間並沒有它的Vimdiff可以：
在左邊，程序彙編，就沒事了，不大會（只是一個簡單的循環，而不是
我提醒你我的問題是，對於nLoop> 2^32-1並且使用匯編循環，我會在執行結束時獲得有效的sum參數，但無效（太短）average和standard deviation次（花費在循環中）;這裏是的輸出示例：
sizeof(unsigned long long int) = 8
sizeof(unsigned long int) = 4
diff = 17
avgSum = 17
stdSum = 289
Sum = 5000000000
diff = 4
avgSum = 21
stdSum = 305
Sum = 5000000000
(Average Elapsed time, Standard deviation) = 1.000000e+01 usec 7.211103e+00 usec
隨着服用nLoop = 4.0e+9，即nLoop = 4000000000ULL，是沒有問題的，時間值是有效的。
更新2：
我通過生成彙編代碼搜索更深入。與nLoop = 4000000000 (4.0e+9)版本下面是：
.file "loop-WITH-asm-inline-4-Billions.c"
.section ".rodata"
.align 8
.LLC1:
.asciz "sizeof(unsigned long long int) = %zu\n"
.align 8
.LLC2:
.asciz "sizeof(unsigned long int) = %zu\n"
.align 8
.LLC3:
.asciz "diff = %llu\n"
.align 8
.LLC4:
.asciz "avgSum = %llu\n"
.align 8
.LLC5:
.asciz "stdSum = %llu\n"
.align 8
.LLC6:
.asciz "Sum = %llu\n"
.global __udivdi3
.global __cmpdi2
.global __floatdidf
.align 8
.LLC7:
.asciz "(Average Elapsed time, Standard deviation) = %e usec %e usec\n"
.align 8
.LLC0:
.long 0
.long 0
.section ".text"
.align 4
.global main
.type main, #function
.proc 04
main:
save %sp, -248, %sp
st %i0, [%fp+68]
st %i1, [%fp+72]
ld [%fp+72], %g1
add %g1, 4, %g1
ld [%g1], %g1
mov %g1, %o0
call atoi, 0
nop
mov %o0, %g1
st %g1, [%fp-68]
st %g0, [%fp-64]
st %g0, [%fp-60]
st %g0, [%fp-56]
st %g0, [%fp-52]
sethi %hi(.LLC0), %g1
or %g1, %lo(.LLC0), %g1
ldd [%g1], %f8
std %f8, [%fp-48]
sethi %hi(.LLC0), %g1
or %g1, %lo(.LLC0), %g1
ldd [%g1], %f8
std %f8, [%fp-40]
mov 0, %g2
sethi %hi(4000000000), %g3
std %g2, [%fp-24]
sethi %hi(.LLC1), %g1
or %g1, %lo(.LLC1), %o0
mov 8, %o1
call printf, 0
nop
sethi %hi(.LLC2), %g1
or %g1, %lo(.LLC2), %o0
mov 4, %o1
call printf, 0
nop
st %g0, [%fp-84]
b .LL2
nop
.LL3:
st %g0, [%fp-32]
st %g0, [%fp-28]
add %fp, -92, %g1
mov %g1, %o0
mov 0, %o1
call gettimeofday, 0
nop
ldd [%fp-24], %o4
clr %g1
clr %g2
mov %o4, %g1
loop:
add %g2, 1, %g2
subcc %g1, 1, %g1
bne loop
nop
mov %g2, %o4
std %o4, [%fp-32]
add %fp, -100, %g1
mov %g1, %o0
mov 0, %o1
call gettimeofday, 0
nop
ld [%fp-100], %g2
ld [%fp-92], %g1
sub %g2, %g1, %g2
sethi %hi(999424), %g1
or %g1, 576, %g1
smul %g2, %g1, %g3
ld [%fp-96], %g2
ld [%fp-88], %g1
sub %g2, %g1, %g1
add %g3, %g1, %g1
st %g1, [%fp-12]
sra %g1, 31, %g1
st %g1, [%fp-16]
ldd [%fp-64], %o4
ldd [%fp-16], %g2
addcc %o5, %g3, %g3
addx %o4, %g2, %g2
std %g2, [%fp-64]
ld [%fp-16], %g2
ld [%fp-12], %g1
smul %g2, %g1, %g4
ld [%fp-16], %g2
ld [%fp-12], %g1
smul %g2, %g1, %g1
add %g4, %g1, %g4
ld [%fp-12], %g2
ld [%fp-12], %g1
umul %g2, %g1, %g3
rd %y, %g2
add %g4, %g2, %g4
mov %g4, %g2
ldd [%fp-56], %o4
addcc %o5, %g3, %g3
addx %o4, %g2, %g2
std %g2, [%fp-56]
sethi %hi(.LLC3), %g1
or %g1, %lo(.LLC3), %o0
ld [%fp-16], %o1
ld [%fp-12], %o2
call printf, 0
nop
sethi %hi(.LLC4), %g1
or %g1, %lo(.LLC4), %o0
ld [%fp-64], %o1
ld [%fp-60], %o2
call printf, 0
nop
sethi %hi(.LLC5), %g1
or %g1, %lo(.LLC5), %o0
ld [%fp-56], %o1
ld [%fp-52], %o2
call printf, 0
nop
sethi %hi(.LLC6), %g1
or %g1, %lo(.LLC6), %o0
ld [%fp-32], %o1
ld [%fp-28], %o2
call printf, 0
nop
ld [%fp-84], %g1
add %g1, 1, %g1
st %g1, [%fp-84]
.LL2:
ld [%fp-84], %g2
ld [%fp-68], %g1
cmp %g2, %g1
bl .LL3
nop
ld [%fp-68], %g1
sra %g1, 31, %g1
ld [%fp-68], %g3
mov %g1, %g2
ldd [%fp-64], %o0
mov %g2, %o2
mov %g3, %o3
call __udivdi3, 0
nop
mov %o0, %g2
mov %o1, %g3
std %g2, [%fp-136]
ldd [%fp-136], %o0
mov 0, %o2
mov 0, %o3
call __cmpdi2, 0
nop
mov %o0, %g1
cmp %g1, 1
bl .LL6
nop
ldd [%fp-136], %o0
call __floatdidf, 0
nop
std %f0, [%fp-144]
b .LL5
nop
.LL6:
ldd [%fp-136], %o4
and %o4, 0, %g2
and %o5, 1, %g3
ld [%fp-136], %o5
sll %o5, 31, %g1
ld [%fp-132], %g4
srl %g4, 1, %o5
or %o5, %g1, %o5
ld [%fp-136], %g1
srl %g1, 1, %o4
or %g2, %o4, %g2
or %g3, %o5, %g3
mov %g2, %o0
mov %g3, %o1
call __floatdidf, 0
nop
std %f0, [%fp-144]
ldd [%fp-144], %f8
ldd [%fp-144], %f10
faddd %f8, %f10, %f8
std %f8, [%fp-144]
.LL5:
ldd [%fp-144], %f8
std %f8, [%fp-48]
ld [%fp-68], %g1
sra %g1, 31, %g1
ld [%fp-68], %g3
mov %g1, %g2
ldd [%fp-56], %o0
mov %g2, %o2
mov %g3, %o3
call __udivdi3, 0
nop
mov %o0, %g2
mov %o1, %g3
std %g2, [%fp-128]
ldd [%fp-128], %o0
mov 0, %o2
mov 0, %o3
call __cmpdi2, 0
nop
mov %o0, %g1
cmp %g1, 1
bl .LL8
nop
ldd [%fp-128], %o0
call __floatdidf, 0
nop
std %f0, [%fp-120]
b .LL7
nop
.LL8:
ldd [%fp-128], %o4
and %o4, 0, %g2
and %o5, 1, %g3
ld [%fp-128], %o5
sll %o5, 31, %g1
ld [%fp-124], %g4
srl %g4, 1, %o5
or %o5, %g1, %o5
ld [%fp-128], %g1
srl %g1, 1, %o4
or %g2, %o4, %g2
or %g3, %o5, %g3
mov %g2, %o0
mov %g3, %o1
call __floatdidf, 0
nop
std %f0, [%fp-120]
ldd [%fp-120], %f8
ldd [%fp-120], %f10
faddd %f8, %f10, %f8
std %f8, [%fp-120]
.LL7:
ldd [%fp-48], %f8
ldd [%fp-48], %f10
fmuld %f8, %f10, %f8
ldd [%fp-120], %f10
fsubd %f10, %f8, %f8
std %f8, [%fp-112]
ldd [%fp-112], %f8
fsqrtd %f8, %f8
std %f8, [%fp-152]
ldd [%fp-152], %f10
ldd [%fp-152], %f8
fcmpd %f10, %f8
nop
fbe .LL9
nop
ldd [%fp-112], %o0
call sqrt, 0
nop
std %f0, [%fp-152]
.LL9:
ldd [%fp-152], %f8
std %f8, [%fp-40]
sethi %hi(.LLC7), %g1
or %g1, %lo(.LLC7), %o0
ld [%fp-48], %o1
ld [%fp-44], %o2
ld [%fp-40], %o3
ld [%fp-36], %o4
call printf, 0
nop
mov 0, %g1
mov %g1, %i0
restore
jmp %o7+8
nop
.size main, .-main
.ident "GCC: (GNU) 4.1.2 20061115 (prerelease) (Debian 4.1.1-21)"
.section ".note.GNU-stack"
當我生成彙編語言代碼版本nLoop = 5000000000 (5.0e+9)，差異是在如下圖所示（與vimdiff）：
的「塊4十億「版本：
mov 0, %g2
sethi %hi(4000000000), %g3
被替換爲」5十億「版本由：
mov 1, %g2
sethi %hi(705032192), %g3
or %g3, 512, %g3
我可以看到，5.0+e9不能在32位進行編碼，由於指令
sethi %hi(705032192), %g3
矛盾的是，當我編譯版本「5個十億」彙編代碼，所述輸出中參數sum計算得很好，即等於5 Billions，我無法解釋它。
歡迎任何幫助或評論，謝謝。
您似乎在彙編代碼中訪問'sum'，它是'unsigned long long'。當然，你必須調整你的asm代碼來匹配參數的大小和類型。您是否嘗試使用C代碼並讓編譯器工作？如果編譯器支持8個字節的整數值，它可以創建代碼來操作它們。 – Gerhardh
@ Gerhardh-如果你看看printf的輸出結果，你可以看到'sum'的計算結果很好（第一個例子爲4.0e + 9，第二個爲5.0e + 9）。在這兩種情況下，'sum'被聲明爲'unsigned long long int'。我不明白爲什麼在裝配輸入參數中使用'nLoop> 2^32'的情況並非如此？ – youpilat13
您是否使用匯編代碼進行64位計算？您可能不需要 –