https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126809
Bug ID: 126809
Summary: fold a division of a value clamped at the divisor
Product: gcc
Version: 17.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: enhancement
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: ktkachov at gcc dot gnu.org
Target Milestone: ---
A value clamped at C reaches C exactly when the unclamped one does, so the
quotient by C is one there and zero below, which is a comparison, and the
remainder is the mirror:
MIN (X, C) / C -> X >= C
MIN (X, C) % C -> X < C ? X : 0
unsigned f (unsigned a) { unsigned t = a < 100 ? a : 100; return t / 100; }
at -O2 on aarch64 generates:
f:
cmp w0, 100
mov w1, 100
csel w0, w0, w1, ls
mov w1, 34079
movk w1, 0x51eb, lsl 16
umull x0, w0, w1
lsr x0, x0, 37
ret
but Clang can do:
f:
cmp w0, #99
cset w0, hi
ret
