Committed, thanks Kito.
On 2023/8/28 17:55, Kito Cheng via Gcc-patches wrote:
LGTM, that's much clearer than v1 to me :)
On Mon, Aug 28, 2023 at 5:54 PM Juzhe-Zhong <juzhe.zh...@rivai.ai> wrote:
This patch is fixing these bunch of ICE in "vect" testsuite:
FAIL: gcc.dg/vect/no-scevccp-outer-2.c (internal compiler error: in
anticipatable_occurrence_p, at config/riscv/riscv-vsetvl.cc:314)
FAIL: gcc.dg/vect/no-scevccp-outer-2.c (test for excess errors)
FAIL: gcc.dg/vect/pr109025.c (internal compiler error: in
anticipatable_occurrence_p, at config/riscv/riscv-vsetvl.cc:314)
FAIL: gcc.dg/vect/pr109025.c (test for excess errors)
FAIL: gcc.dg/vect/pr109025.c -flto -ffat-lto-objects (internal compiler error:
in anticipatable_occurrence_p, at config/riscv/riscv-vsetvl.cc:314)
FAIL: gcc.dg/vect/pr109025.c -flto -ffat-lto-objects (test for excess errors)
FAIL: gcc.dg/vect/pr42604.c (internal compiler error: in
anticipatable_occurrence_p, at config/riscv/riscv-vsetvl.cc:314)
FAIL: gcc.dg/vect/pr42604.c (test for excess errors)
FAIL: gcc.dg/vect/pr42604.c -flto -ffat-lto-objects (internal compiler error:
in anticipatable_occurrence_p, at config/riscv/riscv-vsetvl.cc:314)
FAIL: gcc.dg/vect/pr42604.c -flto -ffat-lto-objects (test for excess errors)
FAIL: gcc.dg/vect/vect-double-reduc-3.c (internal compiler error: in
anticipatable_occurrence_p, at config/riscv/riscv-vsetvl.cc:314)
FAIL: gcc.dg/vect/vect-double-reduc-3.c (test for excess errors)
FAIL: gcc.dg/vect/vect-double-reduc-3.c -flto -ffat-lto-objects (internal
compiler error: in anticipatable_occurrence_p, at
config/riscv/riscv-vsetvl.cc:314)
FAIL: gcc.dg/vect/vect-double-reduc-3.c -flto -ffat-lto-objects (test for
excess errors)
FAIL: gcc.dg/vect/vect-double-reduc-7.c (internal compiler error: in
anticipatable_occurrence_p, at config/riscv/riscv-vsetvl.cc:314)
FAIL: gcc.dg/vect/vect-double-reduc-7.c (test for excess errors)
FAIL: gcc.dg/vect/vect-double-reduc-7.c -flto -ffat-lto-objects (internal
compiler error: in anticipatable_occurrence_p, at
config/riscv/riscv-vsetvl.cc:314)
FAIL: gcc.dg/vect/vect-double-reduc-7.c -flto -ffat-lto-objects (test for
excess errors)
gcc/ChangeLog:
* config/riscv/riscv-vsetvl.cc (pass_vsetvl::earliest_fusion): Fix bug.
---
gcc/config/riscv/riscv-vsetvl.cc | 38 ++++++++++++++++++++++++++++++--
1 file changed, 36 insertions(+), 2 deletions(-)
diff --git a/gcc/config/riscv/riscv-vsetvl.cc b/gcc/config/riscv/riscv-vsetvl.cc
index 682f795c8e1..48e89fe2c03 100644
--- a/gcc/config/riscv/riscv-vsetvl.cc
+++ b/gcc/config/riscv/riscv-vsetvl.cc
@@ -3285,12 +3285,46 @@ pass_vsetvl::earliest_fusion (void)
gcc_assert (!(eg->flags & EDGE_ABNORMAL));
vector_insn_info new_info = vector_insn_info ();
profile_probability prob = src_block_info.probability;
+ /* We don't fuse user vsetvl into EMPTY or
+ DIRTY (EMPTY but polluted) block for these
+ following reasons:
+
+ - The user vsetvl instruction is configured as
+ no side effects that the previous passes
+ (GSCE, Loop-invariant, ..., etc)
+ should be able to do a good job on optimization
+ of user explicit vsetvls so we don't need to
+ PRE optimization (The user vsetvls should be
+ on the optimal local already before this pass)
+ again for user vsetvls in VSETVL PASS here
+ (Phase 3 && Phase 4).
+
+ - Allowing user vsetvls be optimized in PRE
+ optimization here (Phase 3 && Phase 4) will
+ complicate the codes so much so we prefer user
+ vsetvls be optimized in post-optimization
+ (Phase 5 && Phase 6). */
+ if (vsetvl_insn_p (expr.get_insn ()->rtl ()))
+ {
+ if (src_block_info.reaching_out.empty_p ())
+ continue;
+ else if (src_block_info.reaching_out.dirty_p ()
+ && !src_block_info.reaching_out.compatible_p (expr))
+ {
+ new_info.set_empty ();
+ /* Update probability as uninitialized status so that
+ we won't try to fuse any demand info into such EMPTY
+ block any more. */
+ prob = profile_probability::uninitialized ();
+ update_block_info (eg->src->index, prob, new_info);
+ continue;
+ }
+ }
if (src_block_info.reaching_out.empty_p ())
{
if (src_block_info.probability
- == profile_probability::uninitialized ()
- || vsetvl_insn_p (expr.get_insn ()->rtl ()))
+ == profile_probability::uninitialized ())
continue;
new_info = expr.global_merge (expr, eg->src->index);
new_info.set_dirty ();
--
2.36.3
--
Best,
Lehua
