The compile time issue was discovered in SPEC 2017 wrf:
Use time and -ftime-report to analyze the profile data of SPEC 2017 wrf
compilation .
Before this patch (Lazy vsetvl):
scheduling : 121.89 ( 15%) 0.53 ( 11%) 122.72 ( 15%)
13M ( 1%)
machine dep reorg : 424.61 ( 53%) 1.84 ( 37%) 427.44 ( 53%)
5290k ( 0%)
real 13m27.074s
user 13m19.539s
sys 0m5.180s
Simple vsetvl:
machine dep reorg : 0.10 ( 0%) 0.00 ( 0%) 0.11 ( 0%)
4138k ( 0%)
real 6m5.780s
user 6m2.396s
sys 0m2.373s
The machine dep reorg is the compile time of VSETVL PASS (424 seconds) which
counts 53% of
the compilation time, spends much more time than scheduling.
After investigation, the critical patch of VSETVL pass is
compute_lcm_local_properties which
is called every iteration of phase 2 (earliest fusion) and phase 3 (global lcm).
This patch optimized the codes of compute_lcm_local_properties to reduce the
compilation time.
After this patch:
scheduling : 117.51 ( 27%) 0.21 ( 6%) 118.04 ( 27%)
13M ( 1%)
machine dep reorg : 80.13 ( 18%) 0.91 ( 26%) 81.26 ( 18%)
5290k ( 0%)
real 7m25.374s
user 7m20.116s
sys 0m3.795s
The optimization of this patch is very obvious, lazy VSETVL PASS: 424s (53%) ->
80s (18%) which
spend less time than scheduling.
Tested on both RV32 and RV64 no regression. Ok for trunk ?
PR target/113495
gcc/ChangeLog:
* config/riscv/riscv-vsetvl.cc (extract_single_source): Remove.
(pre_vsetvl::compute_vsetvl_def_data): Fix compile time issue.
(pre_vsetvl::compute_transparent): New function.
(pre_vsetvl::compute_lcm_local_properties): Fix compile time time issue.
---
gcc/config/riscv/riscv-vsetvl.cc | 184 ++++++++++---------------------
1 file changed, 60 insertions(+), 124 deletions(-)
diff --git a/gcc/config/riscv/riscv-vsetvl.cc b/gcc/config/riscv/riscv-vsetvl.cc
index d7b40a5c813..cec862329c5 100644
--- a/gcc/config/riscv/riscv-vsetvl.cc
+++ b/gcc/config/riscv/riscv-vsetvl.cc
@@ -599,14 +599,6 @@ extract_single_source (set_info *set)
return first_insn;
}
-static insn_info *
-extract_single_source (def_info *def)
-{
- if (!def)
- return nullptr;
- return extract_single_source (dyn_cast<set_info *> (def));
-}
-
static bool
same_equiv_note_p (set_info *set1, set_info *set2)
{
@@ -2374,6 +2366,7 @@ public:
}
void compute_vsetvl_def_data ();
+ void compute_transparent (const bb_info *);
void compute_lcm_local_properties ();
void fuse_local_vsetvl_info ();
@@ -2452,20 +2445,16 @@ pre_vsetvl::compute_vsetvl_def_data ()
{
for (unsigned i = 0; i < m_vsetvl_def_exprs.length (); i += 1)
{
- const vsetvl_info &info = *m_vsetvl_def_exprs[i];
- if (!info.has_nonvlmax_reg_avl ())
- continue;
- unsigned int regno;
- sbitmap_iterator sbi;
- EXECUTE_IF_SET_IN_BITMAP (m_reg_def_loc[bb->index ()], 0, regno,
- sbi)
- if (regno == REGNO (info.get_avl ()))
- {
- bitmap_set_bit (m_kill[bb->index ()], i);
- bitmap_set_bit (def_loc[bb->index ()],
- get_expr_index (m_vsetvl_def_exprs,
- m_unknow_info));
- }
+ auto *info = m_vsetvl_def_exprs[i];
+ if (info->has_nonvlmax_reg_avl ()
+ && bitmap_bit_p (m_reg_def_loc[bb->index ()],
+ REGNO (info->get_avl ())))
+ {
+ bitmap_set_bit (m_kill[bb->index ()], i);
+ bitmap_set_bit (def_loc[bb->index ()],
+ get_expr_index (m_vsetvl_def_exprs,
+ m_unknow_info));
+ }
}
continue;
}
@@ -2516,6 +2505,36 @@ pre_vsetvl::compute_vsetvl_def_data ()
sbitmap_vector_free (m_kill);
}
+/* Subroutine of compute_lcm_local_properties which Compute local transparent
+ BB. Note that the compile time is very sensitive to compute_transparent and
+ compute_lcm_local_properties, any change of these 2 functions should be
+ aware of the compile time changing of the program which has a large number
of
+ blocks, e.g SPEC 2017 wrf.
+
+ Current compile time profile of SPEC 2017 wrf:
+
+ 1. scheduling - 27%
+ 2. machine dep reorg (VSETVL PASS) - 18%
+
+ VSETVL pass should not spend more time than scheduling in compilation. */
+void
+pre_vsetvl::compute_transparent (const bb_info *bb)
+{
+ int num_exprs = m_exprs.length ();
+ unsigned bb_index = bb->index ();
+ for (int i = 0; i < num_exprs; i++)
+ {
+ auto *info = m_exprs[i];
+ if (info->has_nonvlmax_reg_avl ()
+ && bitmap_bit_p (m_reg_def_loc[bb_index], REGNO (info->get_avl ())))
+ bitmap_clear_bit (m_transp[bb_index], i);
+ else if (info->has_vl ()
+ && bitmap_bit_p (m_reg_def_loc[bb_index],
+ REGNO (info->get_vl ())))
+ bitmap_clear_bit (m_transp[bb_index], i);
+ }
+}
+
/* Compute the local properties of each recorded expression.
Local properties are those that are defined by the block, irrespective of
@@ -2572,7 +2591,7 @@ pre_vsetvl::compute_lcm_local_properties ()
bitmap_vector_clear (m_avloc, last_basic_block_for_fn (cfun));
bitmap_vector_clear (m_antloc, last_basic_block_for_fn (cfun));
- bitmap_vector_clear (m_transp, last_basic_block_for_fn (cfun));
+ bitmap_vector_ones (m_transp, last_basic_block_for_fn (cfun));
/* - If T is locally available at the end of a block, then T' must be
available at the end of the same block. Since some optimization has
@@ -2598,117 +2617,34 @@ pre_vsetvl::compute_lcm_local_properties ()
/* Compute m_transp */
if (block_info.empty_p ())
+ compute_transparent (bb);
+ else
{
- bitmap_ones (m_transp[bb_index]);
- for (int i = 0; i < num_exprs; i += 1)
- {
- const vsetvl_info &info = *m_exprs[i];
- if (!info.has_nonvlmax_reg_avl () && !info.has_vl ())
- continue;
-
- if (info.has_nonvlmax_reg_avl ())
- {
- unsigned int regno;
- sbitmap_iterator sbi;
- EXECUTE_IF_SET_IN_BITMAP (m_reg_def_loc[bb->index ()], 0,
- regno, sbi)
- {
- if (regno == REGNO (info.get_avl ()))
- bitmap_clear_bit (m_transp[bb->index ()], i);
- }
- }
-
- for (insn_info *insn : bb->real_nondebug_insns ())
- {
- if (info.has_nonvlmax_reg_avl ()
- && find_access (insn->defs (), REGNO (info.get_avl ())))
- {
- 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 ();
- insn_info *def_insn = extract_single_source (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 (m_transp[bb_index]);
+ vsetvl_info &header_info = block_info.get_entry_info ();
+ vsetvl_info &footer_info = block_info.get_exit_info ();
- bitmap_clear_bit (m_transp[bb_index], i);
- break;
- }
- }
- }
+ if (header_info.valid_p () && anticipated_exp_p (header_info))
+ bitmap_set_bit (m_antloc[bb_index],
+ get_expr_index (m_exprs, header_info));
- continue;
+ if (footer_info.valid_p ())
+ for (int i = 0; i < num_exprs; i += 1)
+ {
+ const vsetvl_info &info = *m_exprs[i];
+ if (!info.valid_p ())
+ continue;
+ if (available_exp_p (footer_info, info))
+ bitmap_set_bit (m_avloc[bb_index], i);
+ }
}
- vsetvl_info &header_info = block_info.get_entry_info ();
- vsetvl_info &footer_info = block_info.get_exit_info ();
-
- if (header_info.valid_p () && anticipated_exp_p (header_info))
- bitmap_set_bit (m_antloc[bb_index],
- get_expr_index (m_exprs, header_info));
-
- if (footer_info.valid_p ())
- for (int i = 0; i < num_exprs; i += 1)
- {
- const vsetvl_info &info = *m_exprs[i];
- if (!info.valid_p ())
- continue;
- if (available_exp_p (footer_info, info))
- bitmap_set_bit (m_avloc[bb_index], i);
- }
- }
-
- for (const bb_info *bb : crtl->ssa->bbs ())
- {
- unsigned bb_index = bb->index ();
if (invalid_opt_bb_p (bb->cfg_bb ()))
{
bitmap_clear (m_antloc[bb_index]);
bitmap_clear (m_transp[bb_index]);
}
+
/* Compute ae_kill for each basic block using:
~(TRANSP | COMP)
--
2.36.1
