XLS: IR Lowering
As part of codegen, IR constructs are lowered to Verilog/SystemVerilog constructs. In some cases, this lowering is simple and direct, but some IR constructs don't map directly to Verilog constructs.
Some XLS types are flattened into simpler types when IR gets lowered to RTL. This flattening is performed here (also see the more-commented header file). The following summarizes how types are flattened.
Arrays are flattened such that the last element occupies the most significant
bits. Elements are concatenated via the SystemVerilog concatenation operation.
This matches the flattening of SystemVerilog packed arrays (e.g.
foo) to unpacked arrays declared like
logic bar[N:0]. For example, a
4-element array of 4-bit UInts would be flattened as
[0x3, 0x4, 0x5, 0x6] => 0x6543
Tuples are flattened by concatenating each leaf element of the tuple. If an element of a tuple is an array, this is equivalent to first flattening the array and treating the flattened array as a leaf element. Concatenation is performed via the SystemVerilog concatenation operation. The zero-th tuple element will end up occupying the most significant bits in the flattened output. For example, a 4-tuple of 4-bit UInts would be flattened as
(0x3, 0x4, 0x5, 0x6) => 0x3456
and a 2-tuple of length 2 arrays of 4-bit UInts would be flattened as
([0x3, 0x4], [0x5, 0x6]) => 0x4365
DSLX structs are lowered to tuples in the IR, so there's no separate handling of structs.
Tokens are not represented in RTL.
Entities with zero width are unrepresented in RTL. Where a zero-width value is used, a zero-valued literal can be substituted.
Asserts and covers are only represented when producing SystemVerilog, and are unrepresented when producing Verilog.