Souper crash in codegen from valid transformation about souper

[regehr]

we're never supposed to synthesize a phi, but that transformation seems to contain a new phi -- so the crash is correct but the optimization looks weird. can you isolate the code that leads to this transformation so we can understand what's going on here?

[regehr]

%11 looks like it was supposed to be %4, so we need to understand how that happened. if %11 was %4 then we'd not have a new phi and we'd not get the crash. that looks like the bug to me.

[manasij7479]

The 'block' should be same too.
Somewhere we are not reusing the block from LHS during synthesis.

This is how to synthesize this with souper-check without involving sclang
./souper-check -reinfer-rhs -souper-enumerative-synthesis-max-instructions=1 -souper-dataflow-pruning -souper-double-check /tmp/phi.opt
Where phi.opt contains the transformation in the issue.
I'll investigate what''s causing this to be synthesized.

[regehr]

run under valgrind and such -- there could be an uninitialized bug that is causing things to get garbled.

yes, the key question is why that block wasn't getting reused -- if that happened, it wouldn't look like we synthesized a phi on the RHS, which of course we did not.

[regehr]

@manasij7479 reminder to fix this