https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83358
Bug ID: 83358
Summary: [8 Regression] division not converted with Intel
tuning since r253934
Product: gcc
Version: 8.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: trippels at gcc dot gnu.org
CC: hubicka at ucw dot cz
Target Milestone: ---
The following code snipped came up on the realworldtech forum recently:
#include <stdint.h>
__attribute__((noinline, noclone))
void bin2ascii(uint64_t val, char *dst) {
const int64_t POW10_10 = ((int64_t)10) * 1000 * 1000 * 1000;
int64_t hix = val / POW10_10;
int64_t lox = val % POW10_10;
int32_t v0 = hix / 100000;
int32_t v1 = hix % 100000;
int32_t v2 = lox / 100000;
int32_t v3 = lox % 100000;
for (int i = 4; i != 0; --i) {
dst[i + 0 * 5] = v0 % 10 + '0';
v0 /= 10;
dst[i + 1 * 5] = v1 % 10 + '0';
v1 /= 10;
dst[i + 2 * 5] = v2 % 10 + '0';
v2 /= 10;
dst[i + 3 * 5] = v3 % 10 + '0';
v3 /= 10;
}
dst[0 * 5] = v0 + '0';
dst[1 * 5] = v1 + '0';
dst[2 * 5] = v2 + '0';
dst[3 * 5] = v3 + '0';
dst[4 * 5] = 0;
}
int main() {
char ch[20];
for (int i = 0; i < 10000000; i++) {
bin2ascii(10010155021040540, ch);
}
}
Since r253934 gcc doesn't convert divisions anymore when tuning for Intel with
e.g. -mtune=haswell. The result is ~5x slower than before.
bin2ascii:
.LFB0:
.cfi_startproc
movq %rdi, %rax
pushq %rbx
.cfi_def_cfa_offset 16
.cfi_offset 3, -16
leaq 4(%rsi), %r9
movl $-858993459, %ebx
movabsq $-2601111570856684097, %rdx
movabsq $755578637259143235, %rcx
mulq %rdx
movabsq $10000000000, %rax
movq %rdx, %r8
shrq $38, %rdx
shrq $33, %r8
imulq %r8, %rax
subq %rax, %rdi
movq %rdx, %rax
mulq %rcx
shrq $7, %rdx
movl %edx, %r10d
imulq $100000, %rdx, %rdx
subl %edx, %r8d
movq %rdi, %rdx
shrq $5, %rdx
movq %rdx, %rax
mulq %rcx
movl $10, %ecx
shrq $7, %rdx
movl %edx, %r11d
imulq $100000, %rdx, %rdx
subl %edx, %edi
.L2:
movl %r10d, %eax
decq %r9
cltd
idivl %ecx
movl %r10d, %eax
addl $48, %edx
movb %dl, 1(%r9)
mull %ebx
movl %r8d, %eax
movl %edx, %r10d
cltd
idivl %ecx
movl %r8d, %eax
shrl $3, %r10d
addl $48, %edx
movb %dl, 6(%r9)
mull %ebx
movl %r11d, %eax
movl %edx, %r8d
cltd
idivl %ecx
movl %r11d, %eax
shrl $3, %r8d
addl $48, %edx
movb %dl, 11(%r9)
mull %ebx
movl %edi, %eax
movl %edx, %r11d
cltd
idivl %ecx
movl %edi, %eax
shrl $3, %r11d
addl $48, %edx
movb %dl, 16(%r9)
mull %ebx
movl %edx, %edi
shrl $3, %edi
cmpq %r9, %rsi
jne .L2
addl $48, %r10d
addl $48, %r8d
leal 48(%r11), %edx
addl $48, %edi
movb %r10b, (%rsi)
popq %rbx
.cfi_def_cfa_offset 8
movb %r8b, 5(%rsi)
movb %dl, 10(%rsi)
movb %dil, 15(%rsi)
movb $0, 20(%rsi)
ret
