https://gcc.gnu.org/bugzilla/show_bug.cgi?id=47764
--- Comment #6 from Carrot <carrot at google dot com> ---
Another example for ppc.
Following code is disassembled from sha1dgst.o in openssl which is compiled by
gcc
0000000000000000 <sha1_block_data_order>:
...
80: 82 5a 52 3f addis r26,r18,23170
84: 78 9a 4a 7e xor r10,r18,r19
88: 08 00 c4 8a lbz r22,8(r4)
8c: 88 00 1f ea ld r16,136(r31)
90: 0b 00 a4 8b lbz r29,11(r4)
94: 02 00 c4 8b lbz r30,2(r4)
98: 99 79 5a 3b addi r26,r26,31129
...
it uses two instructions to do (r18 + 23170 << 16 + 31129), this large constant
is used many times. In following command line sha1.gcc is disassembled from
sha1dgst.o.
$ grep 31129 sha1.gcc | wc
20 140 881
$ grep 23170 sha1.gcc | wc
20 140 886
If we load this large constant into a register, and use this register later, we
can save 18 instructions.
There are more such cases in the same functions:
$ grep 28378 sha1.gcc | wc
20 140 875
$ grep "\-5215" sha1.gcc | wc
20 140 867
$ grep "\-28900" sha1.gcc | wc
20 140 915
$ grep "\-17188" sha1.gcc | wc
20 140 916
$ grep "\-13725" sha1.gcc | wc
20 140 915
$ grep "\-15914" sha1.gcc | wc
20 140 914
More worse, these codes are inside a hot loop.
