[loopsimplify] Many pointless phi nodes are created #559
Assignees
Comments
|
assigned to @lattner |
|
This should be easy to fix, I just didn't want to forget about it. -Chris |
|
This problem is fixed like so: Many fewer PHI nodes are created and inserted. I feel much better now. :) -Chris |
xlauko
pushed a commit
to trailofbits/instafix-llvm
that referenced
this issue
Mar 28, 2025
This commit extends the support for floating point attributes parsing by using the new `AsmParser::parseFloat(fltSemnatics, APFloat&)` interface. As a drive-by, this commit also harmonizes the cir.fp print/parse namespace usage, and adds the constraint of supporting only "CIRFPType"s for cir.fp in tablegen instead of verifying it manually in the parsing logic. --- This commit is based on top of a to-be-upstreamed commit which extends the upstream MLIR float type parsing. Upstream parsing of float type has full capability only through parsing the Builtin Dialect's `FloatAttr`. Thos commit exposes the same capabilities to downstream users. --- This PR should resolve (at least) `GCC-C-execute-ieee-fp-cmp-2` and `GCC-C-execute-ieee-fp-cmp-4`, paving the way to other `GCC-C-execute-ieee-*` tests passing from the SingleSource suite. It resolves llvm#559 .
This issue was closed.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Extended Description
In .general_ci_points_4 function of 177.mesa, for example, the following
preheader block is created by the loop simplify pass:
no_exit.0.preheader:
%iy.2.ph = phi int [ 0, %cond_true.0 ], [ 0, %cond_true.1 ], [ 0, %cond_false.1
]
%ix.4.ph = phi int [ 0, %cond_true.0 ], [ 0, %cond_true.1 ], [ 0, %cond_false.1
]
%y1.1.ph = phi int [ 0, %cond_true.0 ], [ 0, %cond_true.1 ], [ 0, %cond_false.1
]
%y0.1.ph = phi int [ 0, %cond_true.0 ], [ 0, %cond_true.1 ], [ 0, %cond_false.1
]
%x1.1.ph = phi int [ 0, %cond_true.0 ], [ 0, %cond_true.1 ], [ 0, %cond_false.1
]
%x0.1.ph = phi int [ 0, %cond_true.0 ], [ 0, %cond_true.1 ], [ 0, %cond_false.1
]
%z.0.ph = phi int [ 0, %cond_true.0 ], [ 0, %cond_true.1 ], [ 0, %cond_false.1 ]
%y.0.ph = phi int [ 0, %cond_true.0 ], [ 0, %cond_true.1 ], [ 0, %cond_false.1 ]
%x.0.ph = phi int [ 0, %cond_true.0 ], [ 0, %cond_true.1 ], [ 0, %cond_false.1 ]
This is a common occurance, so the loop simplify pass should try a bit harder to
clean things up for subsequent loop optimization passes.
-Chris
The text was updated successfully, but these errors were encountered: