search for: conv10

Thanks for the perf testing. I assume that DAG legalization is equipped to handle these cases fairly well, or someone would've complained by now... FWIW (and at least some of this can be blamed on me), instcombine already does the narrowing transforms without checking shouldChangeType() for binops like and/or/xor/udiv. The justification was that narrower ops are always better for

...i.i, align 2, !tbaa !5 %conv.i = zext i16 %3 to i32 %and.i256 = and i32 %conv.i, 32768 %tobool5.i = icmp eq i32 %and.i256, 0 br i1 %tobool5.i, label %if.end11.i, label %if.then6.i if.then6.i: ; preds = %if.end.i257 %and9.i = and i32 %conv.i, 32767 %conv10.i = trunc i32 %and9.i to i16 br label %if.end11.i if.end11.i: ; preds = %if.then6.i, %if.end.i257 %storemerge.i = phi i16 [ %conv10.i, %if.then6.i ], [ %3, %if.end.i257 ] %sign.0.i = phi i16 [ -1, %if.then6.i ], [ 0, %if.end.i257 ] store i16 %storemerg...