https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86627
Bug ID: 86627
Summary: Inefficient division of 128-bit ints by small constant
integers
Product: gcc
Version: 9.0
Status: UNCONFIRMED
Severity: enhancement
Priority: P3
Component: rtl-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: tkoenig at gcc dot gnu.org
Target Milestone: ---
For division by small integers which fit into registers,
gcc uses quite a few tricks. For 128-bit integers, a function is
called instead, which takes much longer. Even a division by two
uses a library function.
Test case:
long half_64 (long a)
{
return a/2;
}
long onethird_64 (long a)
{
return a/3;
}
__int128_t half_128 (__int128_t a)
{
return a/2;
}
__int128_t onethird_128 (__int128_t a)
{
return a/3;
}
This gets translated on x86_64-pc-linux-gnu into
.file "halb.c"
.text
.p2align 4
.globl half_64
.type half_64, @function
half_64:
.LFB0:
.cfi_startproc
movq %rdi, %rax
shrq $63, %rax
addq %rdi, %rax
sarq %rax
ret
.cfi_endproc
.LFE0:
.size half_64, .-half_64
.p2align 4
.globl onethird_64
.type onethird_64, @function
onethird_64:
.LFB1:
.cfi_startproc
movabsq $6148914691236517206, %rdx
movq %rdi, %rax
sarq $63, %rdi
imulq %rdx
movq %rdx, %rax
subq %rdi, %rax
ret
.cfi_endproc
.LFE1:
.size onethird_64, .-onethird_64
.globl __divti3
.p2align 4
.globl half_128
.type half_128, @function
half_128:
.LFB2:
.cfi_startproc
subq $8, %rsp
.cfi_def_cfa_offset 16
movl $2, %edx
xorl %ecx, %ecx
call __divti3
addq $8, %rsp
.cfi_def_cfa_offset 8
ret
.cfi_endproc
.LFE2:
.size half_128, .-half_128
.p2align 4
.globl onethird_128
.type onethird_128, @function
onethird_128:
.LFB3:
.cfi_startproc
subq $8, %rsp
.cfi_def_cfa_offset 16
movl $3, %edx
xorl %ecx, %ecx
call __divti3
addq $8, %rsp
.cfi_def_cfa_offset 8
ret
.cfi_endproc
.LFE3:
.size onethird_128, .-onethird_128
.ident "GCC: (GNU) 9.0.0 20180711 (experimental)"
.section .note.GNU-stack,"",@progbits
