Add muldiv_c and muxadd peepopts
#4740
Conversation
| for (auto bit : add_y) | ||
| if (nusers(bit) != 2) | ||
| reject; |
There was a problem hiding this comment.
| for (auto bit : add_y) | |
| if (nusers(bit) != 2) | |
| reject; | |
| if (nusers(add_y) != 2) | |
| reject; |
is all we need and more idiomatic.
We know nusers(add_y) >= 2 by this being an output from add and mux being connected to it too. If any bit has an extra fanout then nusers(add_y) > 2
| std::swap(add_a, add_b); | ||
| endcode | ||
|
|
||
| match mux |
There was a problem hiding this comment.
We can also match on the case of swapped mux inputs
| match mux | ||
| // Select mux of form s ? (a + b) : a, allow leading 0s when A_WIDTH != Y_WIDTH | ||
| select mux->type == $mux | ||
| index <SigSpec> port(mux, \A) === SigSpec({Const(State::S0, GetSize(add_y)-GetSize(add_a)), add_a}) |
There was a problem hiding this comment.
This should be zero-padded if !param(add, \A_SIGNED).bool() and sign-extended otherwise (provided add_a is the A input)
There was a problem hiding this comment.
Also I am not sure this doesn't crash on add_y < add_a which I don't think we rule out anywhere
|
|
||
| match add | ||
| // Select adder | ||
| select add->type == $add |
There was a problem hiding this comment.
| select add->type == $add | |
| select add->type == $add | |
| choice <IdString> A {\A, \B} | |
| define <IdString> B (A == \A ? \B : \A) | |
| set add_y port(add, \Y) | |
| set add_a port(add, A) | |
| set add_b port(add, B) | |
| set add_a_signed param(add, (A == \A) ? \A_SIGNED : \B_SIGNED)) | |
| set add_a_id A |
and replace the branch-and-swap pattern further below
Requires new state
state <bool> add_a_signed
state <IdString> add_a_id
| mux->setPort(\A, Const(State::S0, GetSize(add_b))); | ||
| mux->setPort(\B, add_b); | ||
| mux->setPort(\Y, mid); | ||
| add->setPort(\B, mid); |
There was a problem hiding this comment.
Wrong side of the adder if the swap above was hit (we need to use add_a_id instead)
|
Thanks for the feedback on |
| std::swap(a, b_const); | ||
| endcode | ||
|
|
||
| match div |
There was a problem hiding this comment.
No check that the divider and multiplier are connected
What are the reasons/motivation for this change?
This PR adds two peep-hole optimizations that improve netlist quality:
muldiv_c:y = (a * b_const) / c_const ===> a * eval(b_const / c_const)ifb_const % c_const == 0. This implements basic constant propagation for multipliers and dividers with divisible constants.muxadd:y = s ? (a + b) : a ===> y = a + (s ? b : 0). This provides useful restructuring for improving logic optimization.s ? b : 0can be optimized to a bitwise AND in subsequent optimizations.Explain how this is achieved.
passes/pmgen/peepopt_muldiv_c.pmg:muldiv_cpeepoptpasses/pmgen/peepopt_muxadd.pmg:muxaddpeepoptpasses/pmgen/peepopt.cc: Add to docs and run peepopts duringpeepoptpasspasses/pmgen/Makefile.inc: Add the new sourcesIf applicable, please suggest to reviewers how they can test the change.
equiv_optor FOSSeqy