On Fri, 10 Sept 2021 at 16:40, Richard Henderson < richard.hender...@linaro.org> wrote:
> On 9/10/21 3:36 PM, Philipp Tomsich wrote: > > Richard, > > > > Did you have a chance to consider what to do with clzw? > > I would prefer to avoid the extra extension instructions and change the > implementation > > (and would update the commit message to provide more context), but if > you insist on > > setting 'ctx->w' I'll just have the extra extensions emitted than delay > this series further… > > I don't mind not setting ctx->w, but bear in mind that UXL is going to > automatically set > this flag when executing RV32 on RV64. That's why I have written a tcg > patch set to > eliminate unnecessary sign-extensions. > Ok, thanks! Updated patches follow, once all test workloads have run… Just wondering regarding the UXL-comment: the clzw instruction will be an illegal encoding for RV32 (the w-form instructions are present on RV64 only), so it should never be encountered in a RV32 instruction stream. Did you mean that clz (the instruction operating on xlen-registers) would have ctx->w set for RV32 executing on RV64 ... or am I missing something fundamental? Philipp.
