[0035] Revise createMatrix to create a unique reference #750
Conversation
After losing all its type information to annotateMatrix, createMatrix had no purpose and would have resulted in all calls being merged as redundant, however, some means of uniquely identifying a matrix is needed that annotateMatrix can't do because it needs to be generated for parameters as well as locals and globals (assuming we allow globals). Previous handles for resources have had global variables that are able to uniquely identify them and workgraph handles have had parameters to the entry functions. Lacking either of these, it was impossible to distinguish allocas created for the initial declaration of a matrix and one that was created for a parameter that should get merged when inlining that function. By generating a unique createMatrix call where the local variable is declared or in the entry block for global variables. Addresses some issues in microsoft#696 though not in the way I initially proposed there
pow2clk
changed the title
|
@pow2clk - I think that things have moved on enough that we should probably close this PR? |
|
While I acknowledge the spec has moved in some direction since this began its languishing period, I don't think another solution to the problem it solves has been proposed. With no parameters, the create function needs to be marked read/write to prevent invalid deduplication. This prevents removal of unused create calls. Even if they are mangled to match the mangled handle proposed, read/write is still needed to prevent merging creation of matrices with identical types. |
|
Ok - please resolve the conflicts with this PR so we can get it reviewed and merged. |
|
The alternate approach to this is in #769, which removes the |
|
I hadn't looked it over much. It definitely does. Closing. |
3 participants
After losing all its type information to annotateMatrix, createMatrix had no purpose and would have resulted in all calls being merged as redundant, however, some means of uniquely identifying a matrix is needed that annotateMatrix can't do because it needs to be generated for parameters as well as locals and globals (assuming we allow globals).
Previous handles for resources have had global variables that are able to uniquely identify them and workgraph handles have had parameters to the entry functions. Lacking either of these, it was impossible to distinguish allocas created for the initial declaration of a matrix and one that was created for a parameter that should get merged when inlining that function. By generating a unique createMatrix call where the local variable is declared or in the entry block for global variables.
Addresses some issues in #696 though not in the way I initially proposed there