* [Committed V2] RISC-V: Fix VSETVL PASS bug
@ 2023-12-06 14:26 Juzhe-Zhong
2023-12-06 20:08 ` Patrick O'Neill
0 siblings, 1 reply; 2+ messages in thread
From: Juzhe-Zhong @ 2023-12-06 14:26 UTC (permalink / raw)
To: gcc-patches; +Cc: Juzhe-Zhong
As PR112855 mentioned, the VSETVL PASS insert vsetvli in unexpected location.
Due to 2 reasons:
1. incorrect transparant computation LCM data. We need to check VL operand defs and uses.
2. incorrect fusion of unrelated edge which is the edge never reach the vsetvl expression.
PR target/112855
gcc/ChangeLog:
* config/riscv/riscv-vsetvl.cc (pre_vsetvl::compute_lcm_local_properties): Fix transparant LCM data.
(pre_vsetvl::earliest_fuse_vsetvl_info): Disable earliest fusion for unrelated edge.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/rvv/autovec/pr112855.c: New test.
---
gcc/config/riscv/riscv-vsetvl.cc | 66 ++++++++++++++++++-
.../gcc.target/riscv/rvv/autovec/pr112855.c | 26 ++++++++
2 files changed, 89 insertions(+), 3 deletions(-)
create mode 100644 gcc/testsuite/gcc.target/riscv/rvv/autovec/pr112855.c
diff --git a/gcc/config/riscv/riscv-vsetvl.cc b/gcc/config/riscv/riscv-vsetvl.cc
index 484a8b3a514..68f0be7e81d 100644
--- a/gcc/config/riscv/riscv-vsetvl.cc
+++ b/gcc/config/riscv/riscv-vsetvl.cc
@@ -2034,7 +2034,7 @@ private:
gcc_unreachable ();
}
- bool anticpatable_exp_p (const vsetvl_info &header_info)
+ bool anticipated_exp_p (const vsetvl_info &header_info)
{
if (!header_info.has_nonvlmax_reg_avl () && !header_info.has_vl ())
return true;
@@ -2645,7 +2645,7 @@ pre_vsetvl::compute_lcm_local_properties ()
}
}
- for (const insn_info *insn : bb->real_nondebug_insns ())
+ for (insn_info *insn : bb->real_nondebug_insns ())
{
if (info.has_nonvlmax_reg_avl ()
&& find_access (insn->defs (), REGNO (info.get_avl ())))
@@ -2653,6 +2653,59 @@ pre_vsetvl::compute_lcm_local_properties ()
bitmap_clear_bit (m_transp[bb_index], i);
break;
}
+
+ if (info.has_vl ()
+ && reg_mentioned_p (info.get_vl (), insn->rtl ()))
+ {
+ if (find_access (insn->defs (), REGNO (info.get_vl ())))
+ /* We can't fuse vsetvl into the blocks that modify the
+ VL operand since successors of such blocks will need
+ the value of those blocks are defining.
+
+ bb 4: def a5
+ / \
+ bb 5:use a5 bb 6:vsetvl a5, 5
+
+ The example above shows that we can't fuse vsetvl
+ from bb 6 into bb 4 since the successor bb 5 is using
+ the value defined in bb 4. */
+ ;
+ else
+ {
+ /* We can't fuse vsetvl into the blocks that use the
+ VL operand which has different value from the
+ vsetvl info.
+
+ bb 4: def a5
+ |
+ bb 5: use a5
+ |
+ bb 6: def a5
+ |
+ bb 7: use a5
+
+ The example above shows that we can't fuse vsetvl
+ from bb 6 into bb 5 since their value is different.
+ */
+ resource_info resource
+ = full_register (REGNO (info.get_vl ()));
+ def_lookup dl = crtl->ssa->find_def (resource, insn);
+ def_info *def
+ = dl.matching_set_or_last_def_of_prev_group ();
+ gcc_assert (def);
+ insn_info *def_insn = extract_single_source (
+ dyn_cast<set_info *> (def));
+ if (def_insn && vsetvl_insn_p (def_insn->rtl ()))
+ {
+ vsetvl_info def_info = vsetvl_info (def_insn);
+ if (m_dem.compatible_p (def_info, info))
+ continue;
+ }
+ }
+
+ bitmap_clear_bit (m_transp[bb_index], i);
+ break;
+ }
}
}
@@ -2663,7 +2716,7 @@ pre_vsetvl::compute_lcm_local_properties ()
vsetvl_info &footer_info = block_info.get_exit_info ();
if (header_info.valid_p ()
- && (anticpatable_exp_p (header_info) || block_info.full_available))
+ && (anticipated_exp_p (header_info) || block_info.full_available))
bitmap_set_bit (m_antloc[bb_index],
get_expr_index (m_exprs, header_info));
@@ -2920,6 +2973,13 @@ pre_vsetvl::earliest_fuse_vsetvl_info ()
|| eg->dest == EXIT_BLOCK_PTR_FOR_FN (cfun))
continue;
+ /* When multiple set bits in earliest edge, such edge may
+ have infinite loop in preds or succs or multiple conflict
+ vsetvl expression which make such edge is unrelated. We
+ don't perform fusion for such situation. */
+ if (bitmap_count_bits (e) != 1)
+ continue;
+
vsetvl_block_info &src_block_info = get_block_info (eg->src);
vsetvl_block_info &dest_block_info = get_block_info (eg->dest);
diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr112855.c b/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr112855.c
new file mode 100644
index 00000000000..f1fa6693d2d
--- /dev/null
+++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr112855.c
@@ -0,0 +1,26 @@
+/* { dg-do run } */
+/* { dg-options "-O3" } */
+/* { dg-require-effective-target rv64 } */
+/* { dg-require-effective-target riscv_v } */
+
+#include <assert.h>
+int a;
+int b = 100;
+int c[25];
+int d;
+int main() {
+ int e;
+ d = 0;
+ for (; d < 5; d++) {
+ e = 0;
+ for (; e < 5; e++)
+ c[d * 5 + e] = 0;
+ }
+ if (b)
+ if (a)
+ for (;;)
+ ;
+ b++;
+ int volatile f = *c;
+ assert(b == 101);
+}
--
2.36.3
^ permalink raw reply [flat|nested] 2+ messages in thread
* Re: [Committed V2] RISC-V: Fix VSETVL PASS bug
2023-12-06 14:26 [Committed V2] RISC-V: Fix VSETVL PASS bug Juzhe-Zhong
@ 2023-12-06 20:08 ` Patrick O'Neill
0 siblings, 0 replies; 2+ messages in thread
From: Patrick O'Neill @ 2023-12-06 20:08 UTC (permalink / raw)
To: Juzhe-Zhong, gcc-patches
Hi Juzhe,
An assert added in this patch is firing on a testcase on rv64gcv:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112888
Thanks,
Patrick
On 12/6/23 06:26, Juzhe-Zhong wrote:
> As PR112855 mentioned, the VSETVL PASS insert vsetvli in unexpected location.
>
> Due to 2 reasons:
> 1. incorrect transparant computation LCM data. We need to check VL operand defs and uses.
> 2. incorrect fusion of unrelated edge which is the edge never reach the vsetvl expression.
>
> PR target/112855
>
> gcc/ChangeLog:
>
> * config/riscv/riscv-vsetvl.cc (pre_vsetvl::compute_lcm_local_properties): Fix transparant LCM data.
> (pre_vsetvl::earliest_fuse_vsetvl_info): Disable earliest fusion for unrelated edge.
>
> gcc/testsuite/ChangeLog:
>
> * gcc.target/riscv/rvv/autovec/pr112855.c: New test.
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2023-12-06 20:08 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2023-12-06 14:26 [Committed V2] RISC-V: Fix VSETVL PASS bug Juzhe-Zhong
2023-12-06 20:08 ` Patrick O'Neill
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).