<div dir="ltr"><div class="gmail_extra"><div class="gmail_quote">On Mon, Apr 20, 2015 at 6:57 AM, Bradley Smith <span dir="ltr"><<a href="mailto:bradley.smith@arm.com" target="_blank">bradley.smith@arm.com</a>></span> wrote:<blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
However, when this IR goes down to LLVM there is no context of the bitfields<br>
left so LLVM just sees an i64 load from an i40. InstCombine sees this and<br>
blindly (but correctly I believe, based on the information it has) converts<br>
this load to the ABI alignment of i64, which is now incorrect.<br></blockquote><div><br></div><div>This seems suspicious to me. I always thought that 'load i32..., align 1' was a supported way to ask the backend to do the load slicing if needed, and this instcombine breaks that.</div></div></div></div>