https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120410
Bug ID: 120410
Summary: [15/16 Regression] [SH] XTRCT instruction is no longer
selected for shift after multiplication
Product: gcc
Version: 15.1.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: zack+gcc at buhman dot org
Target Milestone: ---
Target: sh*-*-*
Fixed point multiplication, as in:
int32_t fpmul_16_16(int32_t a, int32_t b)
{
int64_t tmp = (int64_t)a * (int64_t)b;
return tmp >> 16;
}
Is a common pattern on SH-2.
This function should be compiled as, and the GCC code generation in all
versions from versions 4 through 14 inclusive was:
dmuls.l r5,r4
sts macl,r0
sts mach,r1
rts
xtrct r1,r0
However, GCC >= 15.1 compiles the function as:
dmuls.l r5,r4
sts macl,r0
sts mach,r1
shll16 r1
shlr16 r0
rts
add r1,r0
This is a fairly impactful regression for the typical use-case for SH-2, where
large numbers of fixed-point multiplications are performed.
I confirmed this regression exists both in GCC 15.1 and the latest GCC 16
master.
https://godbolt.org/z/zaoPreK7b
