Make u32 shift and rotation instructions always checked #1135
Conversation
| | u32shl <br> - *(47 cycles)* <br> u32shl.*b* <br> - *(4 cycles)* | [b, a, ...] | [c, ...] | $c \leftarrow (a \cdot 2^b) \mod 2^{32}$ <br> Fails if $a \ge 2^{32}$ or $b > 31$ | | ||
| | u32shr <br> - *(47 cycles)*<br> u32shr.*b* <br> - *(4 cycles)* | [b, a, ...] | [c, ...] | $c \leftarrow \lfloor a/2^b \rfloor$ <br> Fails if $a \ge 2^{32}$ or $b > 31$ | | ||
| | u32rotl <br> - *(47 cycles)* <br> u32rotl.*b* <br> - *(4 cycles)* | [b, a, ...] | [c, ...] | Computes $c$ by rotating a 32-bit representation of $a$ to the left by $b$ bits. <br> Fails if $a \ge 2^{32}$ or $b > 31$ | | ||
| | u32rotr <br> - *(59 cycles)* <br> u32rotr.*b* <br> - *(6 cycles)* | [b, a, ...] | [c, ...] | Computes $c$ by rotating a 32-bit representation of $a$ to the right by $b$ bits. <br> Fails if $a \ge 2^{32}$ or $b > 31$ | |
There was a problem hiding this comment.
Were the previous counts incorrect? Or did this really increase cycle counts for non-immediate operations by 7 - 15 cycles?
Also, would be curious to know how these changes affected things like SHA256 and BLAKE3 computations (we have these in stdlib - maybe would be a good idea to put together simple examples with them, and then see before/after cycle counts).
There was a problem hiding this comment.
Indeed, these values are not correct. I'm not even sure why were these values like this before. I've updated the functions comments and documentation.
|
I tested 
|
|
Based on yesterday's discussion, I think we should probably leave the implementation as is (i.e., keep the unchecked implementation). A couple of things we could still do:
|
|
btw, if you have the code for running blake3 and sha256, would be great to add it to the examples. |
This small PR changes a little bit
u32shl,u32shr,u32rotlandu32totrinstructions to make them always checked — input values are always validated to be a u32 and a value less than 32.