https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120479
Bug ID: 120479
Summary: missed opportunity to generate czero.nez
Product: gcc
Version: 16.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: bonzini at gnu dot org
Target Milestone: ---
Target: riscv
Given this code:
void f(unsigned *s, unsigned a0, unsigned a1, unsigned a2, unsigned a3)
{
unsigned r, c;
r = s[0] + a0;
s[0] = r;
c = r < a0; // (1)
r = s[1] + a1;
s[1] = r + c;
c = (r < a1) | (r == -1 ? c : 0); // (2)
r = s[2] + a2;
s[2] = r + c;
c = (r < a2) | (r == -1 ? c : 0);
s[3] += a3 + c;
}
with "-O2 -march=rv32g_zicond", GCC generates the following instructions to
compute the expression marked with (2):
; c from line (1) in a7; r+1 in t1
seqz t1,t1
and t3,a7,t1 ; t3 = (t1 == 0) ? 0 : a7
sltu a5,a2,a5 ; r<a1 in a5
or a5,a5,t3
The seqz+and sequence can be replaced by a single czero.nez instruction:
; c from line (1) in a7, r+1 in t1
czero.nez t3,a7,t1 ; t3 = (t1 != 0) ? a7 : 0
sltu a5,a2,a5 ; r<a1 in a5
or a5,a5,t3
