https://gcc.gnu.org/bugzilla/show_bug.cgi?id=77824
Bug ID: 77824
Summary: unreachable code in SLSR GIMPLE pass
Product: gcc
Version: 7.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: ebotcazou at gcc dot gnu.org
Target Milestone: ---
I played a little with the SLSR GIMPLE pass and, while it's clearly written and
very well documented, it contains an oversight which deactivates a small part
of the processing:
/* Given GS which is a copy of a scalar integer type, make at least one
appropriate entry in the candidate table.
This interface is included for completeness, but is unnecessary
if this pass immediately follows a pass that performs copy
propagation, such as DOM. */
static void
slsr_process_copy (gimple *gs, tree rhs1, bool speed)
slsr_process_copy is never invoked because it's called from:
case MODIFY_EXPR:
slsr_process_copy (gs, rhs1, speed);
break;
and there are no MODIFY_EXPRs left in GIMPLE. I think that all occurrences of
MODIFY_EXPR should be replaced with SSA_NAME in the file. And, at least in the
case I played with, the additional processing makes a difference because there
are unpropagated copies in the GIMPLE IR.
Another minor nit: in replace_mult_candidate
if (wi::fits_shwi_p (bump)
&& bump.to_shwi () != HOST_WIDE_INT_MIN
/* It is not useful to replace casts, copies, or adds of
an SSA name and a constant. */
&& cand_code != MODIFY_EXPR
&& !CONVERT_EXPR_CODE_P (cand_code)
&& cand_code != PLUS_EXPR
&& cand_code != POINTER_PLUS_EXPR
&& cand_code != MINUS_EXPR)
Why not just testing c->kind instead of cand_code?
