https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110492
Bug ID: 110492
Summary: Attempted optimization of switch statement pessimizes
it instead
Product: gcc
Version: 13.1.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: qufanat at gmail dot com
Target Milestone: ---
This happens on my local GCC 11.3, but you can also see it on 13.1 at this
godbolt link https://godbolt.org/z/G3qecWxPr
I'm creating a switch statement of hashed strings, which compiles to a binary
search on the hashes, all well and good. However, with -O3 specified, GCC
peels back the last multiplication of the hash for some of the comparison
branches, but is unable to do it for others, resulting in longer assembly with
twice as many comparisons as is necessary. Here is lines 15..19 in
get_choice_1()
(end of the hash loop)
imul esi, eax, 16777619
test dl, dl
jne .L3
(start of the switch)
cmp eax, 1954414351
je .L8
cmp esi, 1901626525
ja .L4
eax is the hash without the last imul, and esi is the final hash. If we
prevent inlining of the hash function, the compiler can't make this
"optimization" and gives the assembly I expect. Here is lines 90..93 in
get_choice_2()
call hash32_noinline(char const*)
cmp eax, 1901626525
je .L33
jbe .L46
Now it only does one comparison per entry and uses it for both the == and <=
branches.
This isn't that important to my program but I thought you'd like to know.
