spatel added a comment.
In https://reviews.llvm.org/D49606#1170278, @craig.topper wrote:
> Here are the IR patterns for this that work. Not sure if we can do this
> directly in C, we need a 128 bit type, but maybe we can emit it from
> CGBuiltin.cpp?
>
> define i64 @__shiftleft128(i64 %x, i64 %y, i8 %amt) {
> %a = zext i64 %x to i128
> %b = zext i64 %y to i128
> %c = shl i128 %b, 64
> %d = or i128 %a, %c
> %amtmask = and i8 %amt, 63
> %e = zext i8 %amtmask to i128
> %f = shl i128 %d, %e
> %g = lshr i128 %f, 64
> %h = trunc i128 %g to i64
> ret i64 %h
> }
>
> define i64 @__shiftright128(i64 %x, i64 %y, i8 %amt) {
> %a = zext i64 %x to i128
> %b = zext i64 %y to i128
> %c = shl i128 %b, 64
> %d = or i128 %a, %c
> %amtmask = and i8 %amt, 63
> %e = zext i8 %amtmask to i128
> %f = lshr i128 %d, %e
> %g = trunc i128 %f to i64
> ret i64 %g
> }
>
I’m not at my dev machine, but this is exactly the definition of funnel shift,
no? Unless that got reverted, adding/modifying clang builtins was the next step
in the plan for those intrinsics. We probably need some backend work to match
the variable shift version, but shift-by constant should already work.
https://reviews.llvm.org/D49606
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
