GH-115457: Support splitting and replication of micro ops. #115558
Conversation
|
Also, I like how clean this is in the DSL! |
| else if (oparg < _PyUop_Replication[opcode]) { | ||
| buffer[pc].opcode = opcode + oparg + 1; | ||
| } | ||
| else if (opcode == _JUMP_TO_TOP || opcode == _EXIT_TRACE) { |
There was a problem hiding this comment.
It's a static function in another file.
There was a problem hiding this comment.
We'll leave it up to Ken Jin to turn it into a macro or static inline in a header. (Because he's adding another opcode that could end the list of opcodes, a new JUMP variant.)
| @@ -483,6 +505,49 @@ def make_uop(name: str, op: parser.InstDef, inputs: list[parser.InputEffect]) -> | |||
| body=op.block.tokens, | |||
| properties=compute_properties(op), | |||
| ) | |||
| if effect_depends_on_oparg_1(op) and "split" in op.annotations: | |||
| result.properties.oparg_and_1 = True | |||
There was a problem hiding this comment.
For code consistency, shouldn't this be compute_properties?
There was a problem hiding this comment.
Shouldn't what be compute_properties?
There was a problem hiding this comment.
Woops I meant shouldn't this be computed in compute_properties ?
There was a problem hiding this comment.
The oparg_and_1 property only applies to the base uop, not the replicas as their behavior does not depend on oparg & 1. compute_properties computes the properties from the definition only, so we would need to modify oparg_and_1 anyway.
| else if (oparg < _PyUop_Replication[opcode]) { | ||
| buffer[pc].opcode = opcode + oparg + 1; | ||
| } | ||
| else if (opcode == _JUMP_TO_TOP || opcode == _EXIT_TRACE) { |
There was a problem hiding this comment.
We'll leave it up to Ken Jin to turn it into a macro or static inline in a header. (Because he's adding another opcode that could end the list of opcodes, a new JUMP variant.)
Adds the
splitandreplicate(N)annotationssplitsplits uops into two depending on the low bit of theoparg. This removes a few jumps from uops like_LOAD_ATTR_INSTANCE_VALUE.replicate(N)replicates the original uop for each oparg inrange(N). This is particularly valuable for uops that loop overoparg, but it is also useful to inline the oparg at build time rather than when patching.