SoylentNews is people
SoylentNews
from the please-let-the-vapors-condense dept.
From the lowRISC project's blog:
A high quality, upstream RISC-V backend for LLVM is perhaps the most frequently requested missing piece of the RISC-V software ecosystem… As always, you can track status here and find the code here.
RV32
100% of the GCC torture suite passes for RV32I at -O0, -O1, -O2, -O3, and -Os (after masking gcc-only tests). MC-layer (assembler) support for RV32IMAFD has now been implemented, as well as code generation for RV32IM.
RV64
This is the biggest change versus my last update. LLVM recently gained support for parameterising backends by register size, which allows code duplication to be massively reduced for architectures like RISC-V. As planned, I've gone ahead and implemented RV64I MC-layer and code generation support making use of this feature. I'm happy to report that 100% of the GCC torture suite passes for RV64I at O1, O2, O3 and Os (and there's a single compilation failure at O0). I'm very grateful for Krzysztof Parzyszek's (QUIC) work on variable-sized register classes, which has made it possible to parameterise the backend on XLEN in this way. That LLVM feature was actually motivated by requirements of the Hexagon architecture - I think this is a great example of how we can all benefit by contributing upstream to projects, even across different ISAs.
[...] Community members Luís Marques and David Craven have been experimenting with D and Rust support respectively.
[...] Approach and philosophy
As enthusiastic supporters of RISC-V, I think we all want to see a huge range of RISC-V core implementations, making different trade-offs or targeting different classes of applications. But we don't want to see that variety in the RISC-V ecosystem result in dozens of different vendor-specific compiler toolchains and a fractured software ecosystem. Unfortunately most work on LLVM for RISC-V has been invested in private/proprietary code bases or short-term prototypes. The work described in this post has been performed out in the open from the start, with a strong focus on code quality, testing, and on moving development upstream as quickly as possible - i.e. a solution for the long term.
(Score: 3, Informative) by TheRaven on Friday October 06 2017, @11:49AM (1 child)
You might want to look at Sun's MAJC architecture, which had a single register file that could be used for different types (integer, floating point, vector) depending on the instructions. More practically, architectural registers are a fiction on modern CPUs. The hardware has a lot more physical registers than it has architectural registers. These are mapped to physical registers on demand by the register rename unit (which is one of the most complex parts of a modern CPU). Typically, you have different banks of fixed-size registers, because it complicates rename logic to split them and you have redundant data in wires (i.e. heat) if you use larger rename registers than required, but there's nothing stopping you from using the same 128-bit rename registers for vectors and pointers (except that 128 bits is pretty small for a vector register these days).
To the other part of your post, sharing ALUs between 16-bit vectors and 128-bit integers, there's a huge difference in the circuitry between 8 16-bit adders and one 128-bit adder. If you don't have a carry in, then your entire structure is different. You could create a 128-bit adder and put an extra and gate on 7 of the carry lines, connected to an is-this-a-128-bit-integer control signal (with it becoming an 8-element 16-bit vector if the bit is not set), but it would be a staggeringly inefficient vector adder. You'd also be optimising for the wrong thing. Since the end of Dennard Scaling (about a decade ago), transistors are cheap, transistors that you are actually using are expensive. Having two separate pipelines, one for adding integers, one for adding vectors, and only using one at a time is only marginally more expensive than having just one, and having a combined one that does both and is less efficient than either is more expensive than having both.
sudo mod me up
(Score: 2) by maxwell demon on Friday October 06 2017, @03:35PM
Thanks; learned something new today.
The Tao of math: The numbers you can count are not the real numbers.