<p>if you look at what LLVM actually spends its time doing (and I assume probably also GCC?), it's pointer chasing. basic blocks are doubly linked lists, etc. for a while I've had the idea that a from-scratch compiler re-implementation for GPGPUs might work out pretty well. and if not, it would at least be a couple years of super fun hacking. anyhow somehow I only now ran across this:</p><p>"Compilation On The GPU? A Feasibility Study"</p><p><a href="https://dl.acm.org/doi/pdf/10.1145/3528416.3530249" target="_blank" rel="nofollow noopener noreferrer" translate="no"><span class="invisible">https://</span><span class="ellipsis">dl.acm.org/doi/pdf/10.1145/352</span><span class="invisible">8416.3530249</span></a></p>
