On 7/10/25 12:41 AM, Simon Richter wrote:
Hi,
I'd like to get GHDL working on riscv64 again, this builds an extra GCC
backend, "ghdl1", which works on most architectures, however here it
generates a few invalid assembler directives, so the build fails after
gcc has been built and installed, during the first invocation of the
newly built backend:
for i in std/v87/textio.vhdl std/v87/textio-body.vhdl; do \
cmd=" /build/ghdl-5.0.1+dfsg/builddir/gcc/build/gcc/ghdl -a --GHDL1=/
build/ghdl-5.0.1+dfsg/builddir/gcc/build/gcc/ghdl1 --std=87 --bootstrap
--work=std ../../src/$i"; \
echo $cmd; eval $cmd || exit 1; \
done
/build/ghdl-5.0.1+dfsg/builddir/gcc/build/gcc/ghdl -a --GHDL1=/build/
ghdl-5.0.1+dfsg/builddir/gcc/build/gcc/ghdl1 --std=87 --bootstrap --
work=std ../../src/std/v87/textio.vhdl
textio.s: Assembler messages:
textio.s:3: Error: the architecture string of -march and elf
architecture attributes cannot be empty
/build/ghdl-5.0.1+dfsg/builddir/gcc/build/gcc/ghdl:error: compilation error
make[3]: *** [../../libraries/Makefile.inc:187: lib/ghdl/gcc/vhdl/std/
v87/std-obj87.cf] Error 1
make[3]: Leaving directory '/build/ghdl-5.0.1+dfsg/builddir/gcc'
make[2]: *** [Makefile:334: libs.vhdl.gcc] Error 2
make[2]: Leaving directory '/build/ghdl-5.0.1+dfsg/builddir/gcc'
make[1]: *** [debian/rules:261: /build/ghdl-5.0.1+dfsg/builddir/stamps/
build-gcc] Error 2
make[1]: Leaving directory '/build/ghdl-5.0.1+dfsg'
make: *** [debian/rules:61: binary] Error 2
The generated source file begins with
.file "textio.vhdl"
.option pic
.attribute arch, ""
.attribute unaligned_access, 0
.attribute stack_align, 16
.text
.section .rodata
.align 2
so it is fairly obvious what is going wrong here: the RISC-V
architecture string is not initialized, and the backend defaults to 32
bit with no extensions. Manually specifying "-
march=rv64imafdc_zicsr_zifencei -mabi=lp64d" (i.e. the defaults of the
native compiler) on the ghdl1 command line causes it to correctly emit
.file "textio.vhdl"
.option pic
.attribute arch,
"rv64i2p1_m2p0_a2p1_f2p2_d2p2_c2p0_zicsr2p0_zifencei2p0"
.attribute unaligned_access, 0
.attribute stack_align, 16
.text
.section .rodata
.align 3
Is this a bug in the out-of-tree frontend initializing the backend
incorrectly, or should the compiler driver that invokes the frontend
always specify arch/abi on RISC-V?
Yes, this sounds like a problem in your ghdl bits. It should always be
emitting arch strings. Otherwise the rest of the tools have no way of
knowing the exact architecture you're targeting.
Jeff
