https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109537
Bug ID: 109537
Summary: Improve code generation for dynamic loop unrolling
Product: gcc
Version: 13.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: middle-end
Assignee: unassigned at gcc dot gnu.org
Reporter: helijia.i at foxmail dot com
Target Milestone: ---
For the current dynamic loop unrolling implementation, we will try to do the
followng transform:
'''
for (i = 0; i < n; i++)
body;
==> (LOOP->LPT_DECISION.TIMES == 3)
i = 0;
mod = n % 4;
switch (mod)
{
case 3:
body; i++;
case 2:
body; i++;
case 1:
body; i++;
case 0: ;
}
while (i < n)
{
body; i++;
body; i++;
body; i++;
body; i++;
}
'''
It would be better if we could carry out loop unrolling in the following way (R
== # unrolls)
'''
i=0; if (i > n-R-1) goto remain (not needed with loop bounds as shown)
for(; i< n-R-1; i+= R)
{
body;
body;
...
body; // R times
}
remain:
if (i < n)
for(; i < n; i++)
body
'''
For the following sample code:
'''
void matrix_add_const(int N1, int* A, int val)
{
int i, j;
for (j = 0; j < N1; j++)
A[j] += val;
}
'''
Gcc will generate more jump instructions compared to clang.
gcc's assemly
'''
$ cat unroll.gcc.s
.file "unroll.c"
.text
.p2align 4
.globl matrix_add_const
.type matrix_add_const, @function
matrix_add_const:
.LFB0:
.cfi_startproc
testl %edi, %edi
jle .L1
movslq %edi, %rdi
leaq -4(,%rdi,4), %rax
leaq (%rsi,%rdi,4), %rcx
shrq $2, %rax
addq $1, %rax
andl $7, %eax
je .L3
cmpq $1, %rax
je .L26
cmpq $2, %rax
je .L27
cmpq $3, %rax
je .L28
cmpq $4, %rax
je .L29
cmpq $5, %rax
je .L30
cmpq $6, %rax
jne .L41
.L31:
addl %edx, (%rsi)
addq $4, %rsi
.L30:
addl %edx, (%rsi)
addq $4, %rsi
.L29:
addl %edx, (%rsi)
addq $4, %rsi
.L28:
addl %edx, (%rsi)
addq $4, %rsi
.L27:
addl %edx, (%rsi)
addq $4, %rsi
.L26:
addl %edx, (%rsi)
addq $4, %rsi
cmpq %rcx, %rsi
je .L42
.L3:
addl %edx, (%rsi)
addl %edx, 4(%rsi)
addl %edx, 8(%rsi)
addl %edx, 12(%rsi)
addl %edx, 16(%rsi)
addl %edx, 20(%rsi)
addl %edx, 24(%rsi)
addl %edx, 28(%rsi)
addq $32, %rsi
cmpq %rcx, %rsi
jne .L3
.L1:
ret
.p2align 4,,10
.p2align 3
.L41:
addl %edx, (%rsi)
addq $4, %rsi
jmp .L31
.L42:
ret
.cfi_endproc
.LFE0:
.size matrix_add_const, .-matrix_add_const
.ident "GCC: (GNU) 13.0.0 20221022 (experimental)"
.section .note.GNU-stack,"",@progbits
'''
clang's assembly
'''
$ cat unroll.clang.s
.text
.file "unroll.c"
.globl matrix_add_const # -- Begin function matrix_add_const
.p2align 4, 0x90
.type matrix_add_const,@function
matrix_add_const: # @matrix_add_const
.cfi_startproc
# %bb.0:
testl %edi, %edi
jle .LBB0_11
# %bb.1:
movl %edi, %r9d
cmpl $8, %edi
jae .LBB0_3
# %bb.2:
xorl %ecx, %ecx
jmp .LBB0_10
.LBB0_3:
movl %r9d, %ecx
andl $-8, %ecx
movd %edx, %xmm0
pshufd $0, %xmm0, %xmm0 # xmm0 = xmm0[0,0,0,0]
leaq -8(%rcx), %rax
movq %rax, %rdi
shrq $3, %rdi
addq $1, %rdi
movl %edi, %r8d
andl $1, %r8d
testq %rax, %rax
je .LBB0_4
# %bb.5:
movq %r8, %rax
subq %rdi, %rax
xorl %edi, %edi
.p2align 4, 0x90
.LBB0_6: # =>This Inner Loop Header: Depth=1
movdqu (%rsi,%rdi,4), %xmm1
movdqu 16(%rsi,%rdi,4), %xmm2
movdqu 32(%rsi,%rdi,4), %xmm3
movdqu 48(%rsi,%rdi,4), %xmm4
paddd %xmm0, %xmm1
paddd %xmm0, %xmm2
movdqu %xmm1, (%rsi,%rdi,4)
movdqu %xmm2, 16(%rsi,%rdi,4)
paddd %xmm0, %xmm3
paddd %xmm0, %xmm4
movdqu %xmm3, 32(%rsi,%rdi,4)
movdqu %xmm4, 48(%rsi,%rdi,4)
addq $16, %rdi
addq $2, %rax
jne .LBB0_6
# %bb.7:
testq %r8, %r8
je .LBB0_9
.LBB0_8:
movdqu (%rsi,%rdi,4), %xmm1
movdqu 16(%rsi,%rdi,4), %xmm2
paddd %xmm0, %xmm1
paddd %xmm0, %xmm2
movdqu %xmm1, (%rsi,%rdi,4)
movdqu %xmm2, 16(%rsi,%rdi,4)
.LBB0_9:
cmpq %r9, %rcx
je .LBB0_11
.p2align 4, 0x90
.LBB0_10: # =>This Inner Loop Header: Depth=1
addl %edx, (%rsi,%rcx,4)
addq $1, %rcx
cmpq %rcx, %r9
jne .LBB0_10
.LBB0_11:
retq
.LBB0_4:
xorl %edi, %edi
testq %r8, %r8
jne .LBB0_8
jmp .LBB0_9
.Lfunc_end0:
.size matrix_add_const, .Lfunc_end0-matrix_add_const
.cfi_endproc
# -- End function
.ident "clang version 10.0.0-4ubuntu1 "
.section ".note.GNU-stack","",@progbits
.addrsig
'''
