In this PR, the SME mode-switching code needs to insert a stack-probe
loop for an alloca. This patch allows the target to do that.
There are two parts to it: allowing loops for insertions in blocks,
and allowing them for insertions on edges. The former can be handled
entirely within mode-switching itself, by recording which blocks have
had new branches inserted. The latter requires an extension to
commit_one_edge_insertion.
I think the extension to commit_one_edge_insertion makes logical sense,
since it already explicitly allows internal loops during RTL expansion.
The single-block find_sub_basic_blocks is a relatively recent addition,
so wouldn't have been available when the code was originally written.
The patch also has a small and obvious fix to make the aarch64 emit
hook cope with labels.
I've added specific -fstack-clash-protection versions of all
aarch64-sme.exp tests that previously failed because of this bug.
I've also added -fno-stack-clash-protection to the original versions
of these tests if they contain scans that assume no protection.
Tested on aarch64-linux-gnu. OK to install?
Richard
gcc/
PR target/113220
* cfgrtl.cc (commit_one_edge_insertion): Handle sequences that
contain jumps even if called after initial RTL expansion.
* mode-switching.cc: Include cfgbuild.h.
(optimize_mode_switching): Allow the sequence returned by the
emit hook to contain internal jumps. Record which blocks
contain such jumps and split the blocks at the end.
* config/aarch64/aarch64.cc (aarch64_mode_emit): Check for
non-debug insns when scanning the sequence.
gcc/testsuite/
PR target/113220
* gcc.target/aarch64/sme/call_sm_switch_5.c: Add
-fno-stack-clash-protection.
* gcc.target/aarch64/sme/call_sm_switch_5_scp.c: New test.
* gcc.target/aarch64/sme/sibcall_6_scp.c: New test.
* gcc.target/aarch64/sme/za_state_4.c: Add
-fno-stack-clash-protection.
* gcc.target/aarch64/sme/za_state_4_scp.c: New test.
* gcc.target/aarch64/sme/za_state_5.c: Add
-fno-stack-clash-protection.
* gcc.target/aarch64/sme/za_state_5_scp.c: New test.
---
gcc/cfgrtl.cc | 27 ++-
gcc/config/aarch64/aarch64.cc | 2 ++
gcc/mode-switching.cc | 15 +++
.../gcc.target/aarch64/sme/call_sm_switch_5.c | 2 +-
.../aarch64/sme/call_sm_switch_5_scp.c| 3 +++
.../gcc.target/aarch64/sme/sibcall_6_scp.c| 3 +++
.../gcc.target/aarch64/sme/za_state_4.c | 2 +-
.../gcc.target/aarch64/sme/za_state_4_scp.c | 3 +++
.../gcc.target/aarch64/sme/za_state_5.c | 2 +-
.../gcc.target/aarch64/sme/za_state_5_scp.c | 3 +++
10 files changed, 53 insertions(+), 9 deletions(-)
create mode 100644 gcc/testsuite/gcc.target/aarch64/sme/call_sm_switch_5_scp.c
create mode 100644 gcc/testsuite/gcc.target/aarch64/sme/sibcall_6_scp.c
create mode 100644 gcc/testsuite/gcc.target/aarch64/sme/za_state_4_scp.c
create mode 100644 gcc/testsuite/gcc.target/aarch64/sme/za_state_5_scp.c
diff --git a/gcc/cfgrtl.cc b/gcc/cfgrtl.cc
index 15259c5e984..304c429c99b 100644
--- a/gcc/cfgrtl.cc
+++ b/gcc/cfgrtl.cc
@@ -2018,6 +2018,21 @@ commit_one_edge_insertion (edge e)
insns = e->insns.r;
e->insns.r = NULL;
+ /* Allow the sequence to contain internal jumps, such as a memcpy loop
+ or an allocation loop. If such a sequence is emitted during RTL
+ expansion, we'll create the appropriate basic blocks later,
+ at the end of the pass. But if such a sequence is emitted after
+ initial expansion, we'll need to find the subblocks ourselves. */
+ bool contains_jump = false;
+ if (!currently_expanding_to_rtl)
+for (rtx_insn *insn = insns; insn; insn = NEXT_INSN (insn))
+ if (JUMP_P (insn))
+ {
+ rebuild_jump_labels_chain (insns);
+ contains_jump = true;
+ break;
+ }
+
/* Figure out where to put these insns. If the destination has
one predecessor, insert there. Except for the exit block. */
if (single_pred_p (e->dest) && e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun))
@@ -2112,13 +2127,13 @@ commit_one_edge_insertion (edge e)
delete_insn (before);
}
else
-/* Some builtin expanders, such as those for memset and memcpy,
- may generate loops and conditionals, and those may get emitted
- into edges. That's ok while expanding to rtl, basic block
- boundaries will be identified and split afterwards. ??? Need
- we check whether the destination labels of any inserted jumps
- are also part of the inserted sequence? */
+/* Sequences inserted after RTL expansion are expected to be SESE,
+ with only internal branches allowed. If the sequence jumps outside
+ itself then we do not know how to add the associated edges here. */
gcc_assert (!JUMP_P (last) ||
