This series, after some related and preparatory patches, contains
fixes for the problem of phi-sources being left in place after their
corresponding branches have been removed. This results in situations,
for example, where a phi-node has 3 arguments/inputs/sources but only
2 parent BBs and create all sort of problems, like:
* some dead code is not removed
* some optimizations involving phi-nodes are blocked (if-conversions)
* more importantly, these phi-nodes are invalid/meaningless and so
it's impossible to reason correctly about them.
Note: This series only fixes *some* of these problems. Other probably
still exists, especially when rewrite_branch() is used.
Luc Van Oostenryck (13):
Revert "simplify CBR-CBR on the same condition"
add testcases to check if phi-sources from removed targets are removed
too
remove insert_branch() redundant arg
simplify remove_parent()
fold remove_parent() into insert_branch()
let insert_branch() reuse the terminating instruction
move insert_branch() to flow.c
let insert_branch() return a status
rename insert_branch() to convert_to_jump()
add remove_phisources()
fix phisources during CBR-BR conversion
use convert_to_jump() when converting a CBR with same targets
fix phisources during SWITCH-BR conversion
flow.c | 205 +++++++++++++++-----------------
flow.h | 3 +
linearize.c | 36 ------
linearize.h | 1 -
simplify.c | 34 ++----
validation/optim/bad-phisrc1.c | 15 +++
validation/optim/bad-phisrc1a.c | 23 ++++
validation/optim/bad-phisrc2.c | 16 +++
validation/optim/bad-phisrc3.c | 20 ++++
9 files changed, 184 insertions(+), 169 deletions(-)
create mode 100644 validation/optim/bad-phisrc1.c
create mode 100644 validation/optim/bad-phisrc1a.c
create mode 100644 validation/optim/bad-phisrc2.c
create mode 100644 validation/optim/bad-phisrc3.c
--
2.31.0
