Added sink_alloc

[skeqiqevian]

sink_alloc is roughly the inverse operation of lift_alloc. However, sink_alloc and lift_alloc are incorrect as it stands. The plan is to come back and fix the issue (described below) when we have rewritten analysis.

In writing this scheduling operation, I noticed an analysis error in lift_alloc. However, fixing this analysis requires data dependent analysis.

Consider the following:

def test_lift_and_sink_alloc_should_fail():
    @proc
    def foo1():
        for i in seq(0, 10):
            a: i8[11] @ DRAM
            a[i] += 1.0
            a[i + 1] = 1.0

    @proc
    def foo2():
        a: i8[11] @ DRAM
        for i in seq(0, 10):
            a[i] += 1.0
            a[i + 1] = 1.0

These programs are not equivalent because of the cross-iteration read/write dependencies. However, the current lift_alloc lacks analysis to reason about this and permits this transformation.

[codecov-commenter]

Codecov Report

Attention: 1 lines in your changes are missing coverage. Please review.

Comparison is base (4c6bfdb) 86.73% compared to head (a51d3c1) 86.75%.

❗ Your organization needs to install the Codecov GitHub app to enable full functionality.

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #501      +/-   ##
==========================================
+ Coverage   86.73%   86.75%   +0.02%     
==========================================
  Files          78       78              
  Lines       18341    18378      +37     
==========================================
+ Hits        15908    15944      +36     
- Misses       2433     2434       +1     

Files	Coverage Δ
src/exo/LoopIR_scheduling.py	91.07% <100.00%> (+0.07%)	⬆️
src/exo/internal_cursors.py	95.45% <ø> (ø)
src/exo/stdlib/scheduling.py	83.65% <ø> (ø)
tests/test_schedules.py	99.43% <100.00%> (+<0.01%)	⬆️
src/exo/API_scheduling.py	88.23% <85.71%> (-0.03%)	⬇️

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.