https://gcc.gnu.org/bugzilla/show_bug.cgi?id=22141
--- Comment #44 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
With the #c43 patch, the included store_merging_10.c improves on x86_64 from:
movzbl (%rdi), %eax
- andl $-19, %eax
+ andl $-32, %eax
orl $13, %eax
movb %al, (%rdi)
in foo and
- orb $1, (%rdi)
movl (%rdi), %eax
- andl $-131071, %eax
+ andl $2147352576, %eax
+ orl $1, %eax
movl %eax, (%rdi)
- shrl $24, %eax
- andl $127, %eax
- movb %al, 3(%rdi)
in bar. foo is something combine.c managed to optimize too, but bar it
couldn't.
In store_merging_11.c on x86_64, bar is the same and foo changed:
- movabsq $578437695752115969, %rax
- movl $0, 9(%rdi)
- movb $0, 15(%rdi)
- movq %rax, 1(%rdi)
- xorl %eax, %eax
- movw %ax, 13(%rdi)
+ movl $23, %eax
+ movb $1, 1(%rdi)
+ movl $117835012, 4(%rdi)
+ movw %ax, 2(%rdi)
+ movq $8, 8(%rdi)
which is not only shorter, but all the stores are aligned.
On ppc64le in store_merging_10.c the difference is:
- lwz 9,0(3)
+ lbz 9,0(3)
rlwinm 9,9,0,0,26
ori 9,9,0xd
- stw 9,0(3)
+ stb 9,0(3)
in foo and
lwz 9,0(3)
+ rlwinm 9,9,0,1,14
ori 9,9,0x1
- rlwinm 9,9,0,31,14
- rlwinm 9,9,0,1,31
stw 9,0(3)
in bar, and store_merging_11.c the difference is:
- lis 8,0x807
- li 9,0
- ori 8,8,0x605
- li 10,0
- sldi 8,8,32
- stw 9,9(3)
- sth 9,13(3)
- oris 8,8,0x400
- stb 10,15(3)
- ori 8,8,0x1701
- mtvsrd 0,8
- stfd 0,1(3)
+ lis 9,0x706
+ li 7,1
+ li 8,23
+ ori 9,9,0x504
+ li 10,8
+ stb 7,1(3)
+ sth 8,2(3)
+ stw 9,4(3)
+ std 10,8(3)
in foo and no changes in bar.
What the patch doesn't implement yet, but could be also possible for
allow_unaligned case is in store_merging_11.c when we are storing 15 bytes
store
8 bytes at offset 1 and 8 bytes at offset 8 (i.e. create two overlapping
stores, in this case one aligned and one unaligned).
