This paper refutes your point and is inline with my understanding: https://carrv.github.io/2020/papers/CARRV2020_paper_12_Perot...
One counter point is the simpler the ISA the more work the compiler needs to do. So x86 is may be easier to write an optimize compiler for given it does some of the work for you (despite the bloat).
I wonder how much benchmarks on the current ARM and RISK-V chips will change over time due to compiler improvements. ARM probably already has alot of investment in some workloads...
EDIT: This presentation show code size comparisons with a RISC-V extension that adds multiply/divide and it get really close to ARM: https://riscv.org/wp-content/uploads/2019/12/12.10-12.50a-Co...
With compression and the multiply/divide extensions very similar to the ARM-M Thumb variants in code density.
https://riscv.org/announcements/2016/04/risc-v-offers-simple...
https://riscv.org/wp-content/uploads/2019/12/12.10-12.50a-Co...
Really fun time to be working on this part of the stack.
But for 64-bit, ARM has dropped support for Thumb, while RISC-V still supports compressed instructions. In that case, RISC-V is more dense than ARM. I can't find the numbers for that right now, but I read it just yesterday so I'm pretty sure I remember correctly.
Also keep in mind that Thumb is a mode that the ARM CPU has to switch to, while RISC-V compressed instructions can be mixed with normal instructions.