https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118687
Bug ID: 118687
Summary: RISC-V extensions for inline asm code (vs. llvm)
Product: gcc
Version: 15.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: vineetg at gcc dot gnu.org
CC: charlie at rivosinc dot com, jeffreyalaw at gmail dot com
Target Milestone: ---
I think I know the resolution but opening this bug for discussion anyways.
Charlie reported some kernel code has an inline asm whcih uses a LI + large
const, which can possibly be synthesized optimally with bitmanip extension.
long ret_from_fork()
{
long ret;
asm ("li %[ret],%[c1]\n"
: [ret] "=r" (ret)
: [c1] "i" (0x0123456789abcdefull));
return ret;
}
However gcc is ignoring the bitmanip flag and generating the "invalid asm"
ret_from_fork:
li a0,81985529216486895
ret
Ultimately the assembler synthesizes it using basic codegen (it is not expected
to re-implement the compiler anyways).
lui a0,0x92
addiw a0,a0,-1493
slli a0,a0,0xc
addi a0,a0,965
slli a0,a0,0xd
addi a0,a0,-1347
slli a0,a0,0xc
addi a0,a0,-529
ret
The question is - is this as designed or a bug. My guy says inline asm is meant
to keep the compiler out of the way so gcc is doing the right thing.
But then RISC-V llvm seems to be generating the optimal bitmanip enabled
sequence for the same test.
https://godbolt.org/z/W3G4e6qnv
