https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119493
Jakub Jelinek <jakub at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |jamborm at gcc dot gnu.org
--- Comment #12 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
For musttail, perhaps SRA could avoid changing the path from musttail call
return to the return stmt.
I've tried
--- gcc/tree-sra.cc.jj 2025-01-02 11:23:27.527363147 +0100
+++ gcc/tree-sra.cc 2025-03-31 09:28:36.535135797 +0200
@@ -1668,9 +1668,11 @@ scan_function (void)
t = gimple_call_lhs (stmt);
if (t && !disqualify_if_bad_bb_terminating_stmt (stmt, t, NULL))
{
+ if (gimple_call_must_tail_p (as_a <gcall *> (stmt)))
+ disqualify_base_of_expr (t, "result of musttail call.");
/* If the STMT is a call to DEFERRED_INIT, avoid setting
cannot_scalarize_away_bitmap. */
- if (gimple_call_internal_p (stmt, IFN_DEFERRED_INIT))
+ else if (gimple_call_internal_p (stmt, IFN_DEFERRED_INIT))
ret |= !!build_access_from_expr_1 (t, stmt, true);
else
ret |= build_access_from_expr (t, stmt, true);
but that is insufficient for #c0 with -O2,
instead of
D.2926 = foo (D.2927); [must tail call]
SR.5_16 = D.2926.n_recurses;
<bb 6> [local count: 116930483]:
# SR.5_22 = PHI <0(4), SR.5_16(5)>
<bb 7> [local count: 354334800]:
# SR.4_24 = PHI <0(3), SR.5_22(6)>
<bb 8> [local count: 1073741824]:
# _9 = PHI <0(2), SR.4_24(7)>
D.2905.n_recurses = _9;
return D.2905;
we get
D.2926 = foo (D.2927); [must tail call]
pretmp_9 = D.2926.n_recurses;
<bb 6> [local count: 116930483]:
# prephitmp_1 = PHI <0(4), pretmp_9(5)>
<bb 7> [local count: 354334800]:
# SR.4_24 = PHI <0(3), prephitmp_1(6)>
<bb 8> [local count: 1073741824]:
# _4 = PHI <0(2), SR.4_24(7)>
D.2905.n_recurses = _4;
return D.2905;
which is pretty much the same thing.
