Missing library routines for double-float conversion when compiling with march=rv32imf?

[tcal-x]

In CFU playground, when using the conda-provided RISC-V compiler, there are link-time errors (missing __extendsfdf2 and __truncdfsf2) when targeting a rv32imf CPU. When using the SiFive toolchain, the binary links fine. This is only circumstantial evidence, I admit, but could we be missing some library routines?

The errors like like this:

/home/tim/google/CFU-Playground/env/conda/envs/cfu-common/bin/../lib/gcc/riscv32-elf/10.1.0/../../../../riscv32-elf/bin/ld: src/tensorflow/lite/micro/micro_string.o: in function `.L0 ':
/home/tim/google/CFU-Playground/proj/proj_template_v/build/src/tensorflow/lite/micro/micro_string.cc:280: undefined reference to `__truncdfsf2'
/home/tim/google/CFU-Playground/env/conda/envs/cfu-common/bin/../lib/gcc/riscv32-elf/10.1.0/../../../../riscv32-elf/bin/ld: src/tensorflow/lite/micro/micro_utils.o: in function `tflite::SignedSymmetricPerChannelQuantize(float const*, TfLiteIntArray*, int, signed char*, float*)':
/home/tim/google/CFU-Playground/proj/proj_template_v/build/src/tensorflow/lite/micro/micro_utils.cc:77: undefined reference to `__extendsfdf2'

To reproduce in CFU Playground, with current main branch, do the following after completing setup:

$ make env
$ make enter
$ cd proj/proj_template_v
$ make EXTRA_LITEX_ARGS="--cpu-variant=fpu+cfu" software