https://gcc.gnu.org/bugzilla/show_bug.cgi?id=117421
Bug ID: 117421
Summary: [RISCV] Use byte comparison instead of word comparison
Product: gcc
Version: 15.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: wojciech_mula at poczta dot onet.pl
Target Milestone: ---
Consider this simple function:
---
#include <string_view>
bool ext_is_gzip(std::string_view ext) {
return ext == "gzip";
}
---
For the x86 target, GCC nicely inlines compile-time constant, and produces the
code like that (it's from GCC 15, with `-O3 -march=icelake-server`):
---
ext_is_gzip(std::basic_string_view<char, std::char_traits<char> >):
xorl %eax, %eax
cmpq $4, %rdi
je .L5
ret
.L5:
cmpl $1885960807, (%rsi)
sete %al
ret
---
However, for the RISC-V target, GCC emits plain byte-by-byte comparison
(riscv64-unknown-linux-gnu-g++ (crosstool-NG UNKNOWN) 15.0.0 20241031
(experimental), with `-O3 -march=rv64gcv`):
---
ext_is_gzip(std::basic_string_view<char, std::char_traits<char> >):
addi sp,sp,-16
sd a0,0(sp)
sd a1,8(sp)
li a5,4
beq a0,a5,.L9
li a0,0
addi sp,sp,16
jr ra
.L9:
lbu a4,0(a1)
li a5,103
beq a4,a5,.L10
.L3:
li a0,1
.L4:
xori a0,a0,1
addi sp,sp,16
jr ra
.L10:
lbu a4,1(a1)
li a5,122
bne a4,a5,.L3
lbu a4,2(a1)
li a5,105
bne a4,a5,.L3
lbu a4,3(a1)
li a5,112
li a0,0
beq a4,a5,.L4
li a0,1
j .L4
---
My wild guess is that we have by default a high cost of placing huge
compile-time values in RISC-V. However, when I checked what is emitted for
"gzip" & "pizg" given as u32, then we have:
---
0: 677a7537 lui a0,0x677a7
4: 9705051b addiw a0,a0,-1680 # 677a6970
8: 70698537 lui a0,0x70698
c: a675051b addiw a0,a0,-1433 # 70697a67
---
A godbolt link for convenience: https://godbolt.org/z/e16bP369n.
