https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114122
Bug ID: 114122
Summary: RISC-V: poor code generation in calling convention
with vlen > 4096
Product: gcc
Version: 14.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: ewlu at rivosinc dot com
Target Milestone: ---
godbolt: https://godbolt.org/z/9bKPWYn65
For vector sizes with vlen > 4096, we generate more than 2x insns than
necessary for the function.
ex: (from godbolt link)
v256si_RET1_ARG3:
li a5,128
vsetvli zero,a5,e32,m1,ta,ma
vle32.v v3,0(a2)
vle32.v v1,0(a1)
vle32.v v2,0(a3)
addi a1,a1,512 <-- all these addis are unnecessary
addi a2,a2,512
addi a3,a3,512
addi a4,a0,512
vadd.vv v1,v1,v3
vadd.vv v1,v1,v2
vse32.v v1,0(a0) <-- Return value set here, can return now
vle32.v v1,0(a1)
vle32.v v3,0(a2)
vle32.v v2,0(a3)
vadd.vv v1,v1,v3
vadd.vv v1,v1,v2
vse32.v v1,0(a4)
ret
Printing the gimple in the veclower2 (tree) pass, the optab_handler cannot find
an op which supports vlen 8192, which in turn generates vector constructors
that do not get completely optimized out.
<constructor 0x7efc492554f8
type <vector_type 0x7efc491b7a80 type <integer_type 0x7efc494435e8 int>
sizes-gimplified asm_written BLK
size <integer_cst 0x7efc4958bf60 constant 8192>
unit-size <integer_cst 0x7efc4958bfc0 constant 1024>
align:128 warn_if_not_align:0 symtab:1226528480 alias-set 1
canonical-type 0x7efc491b7a80 nunits:256
pointer_to_this <pointer_type 0x7efc4921fa80>>
length:2
val <ssa_name 0x7efc492521f8 type <vector_type 0x7efc491b7690>
def_stmt _10 = _8 + _9;
version:10>
val <ssa_name 0x7efc492522d0 type <vector_type 0x7efc491b7690>
def_stmt _13 = _11 + _12;
version:13>>
