[pushed,RA] : Modify cost calculation for dealing with pseudo equivalences
Checks
Commit Message
This is the second attempt to improve RA cost calculation for pseudos
with equivalences. The patch explanation is in the log message.
The patch was successfully bootstrapped and tested on x86-64, aarch64,
and ppc64le. The patch was also benchmarked on x86-64 spec2017.
specfp2017 performance did not changed, specint2017 improved by 0.3%.
Comments
Hi Vladimir,
On Thu, 26 Oct 2023 at 16:00, Vladimir Makarov <vmakarov.gcc@gmail.com> wrote:
>
> This is the second attempt to improve RA cost calculation for pseudos
> with equivalences. The patch explanation is in the log message.
>
> The patch was successfully bootstrapped and tested on x86-64, aarch64,
> and ppc64le. The patch was also benchmarked on x86-64 spec2017.
> specfp2017 performance did not changed, specint2017 improved by 0.3%.
>
As reported by our CI, this patch causes a regression on arm:
FAIL: gcc.target/arm/eliminate.c scan-assembler-times r0,[\\t ]*sp 3
For this testcase, we used to generate:
str lr, [sp, #-4]!
sub sp, sp, #12
add r0, sp, #4
bl bar
add r0, sp, #4
bl bar
add r0, sp, #4
bl bar
add sp, sp, #12
ldr lr, [sp], #4
bx lr
After your patch, we generate:
push {r4, lr}
sub sp, sp, #8
add r4, sp, #4
mov r0, r4
bl bar
mov r0, r4
bl bar
mov r0, r4
bl bar
add sp, sp, #8
pop {r4, lr}
bx lr
which uses 1 more register and 1 more instruction.
Shall I file a bugzilla report for this?
Thanks,
Christophe
On 10/27/23 09:56, Christophe Lyon wrote:
> Hi Vladimir,
>
> On Thu, 26 Oct 2023 at 16:00, Vladimir Makarov <vmakarov.gcc@gmail.com> wrote:
>> This is the second attempt to improve RA cost calculation for pseudos
>> with equivalences. The patch explanation is in the log message.
>>
>> The patch was successfully bootstrapped and tested on x86-64, aarch64,
>> and ppc64le. The patch was also benchmarked on x86-64 spec2017.
>> specfp2017 performance did not changed, specint2017 improved by 0.3%.
>>
> As reported by our CI, this patch causes a regression on arm:
> FAIL: gcc.target/arm/eliminate.c scan-assembler-times r0,[\\t ]*sp 3
>
>
> For this testcase, we used to generate:
> str lr, [sp, #-4]!
> sub sp, sp, #12
> add r0, sp, #4
> bl bar
> add r0, sp, #4
> bl bar
> add r0, sp, #4
> bl bar
> add sp, sp, #12
> ldr lr, [sp], #4
> bx lr
>
> After your patch, we generate:
> push {r4, lr}
> sub sp, sp, #8
> add r4, sp, #4
> mov r0, r4
> bl bar
> mov r0, r4
> bl bar
> mov r0, r4
> bl bar
> add sp, sp, #8
> pop {r4, lr}
> bx lr
>
> which uses 1 more register and 1 more instruction.
>
> Shall I file a bugzilla report for this?
>
I started to work on this right after I got the message (yesterday). I
already have a patch and am going to commit it during an hour. So there
is no need to fill the PR.
On Fri, 27 Oct 2023 at 16:19, Vladimir Makarov <vmakarov@redhat.com> wrote:
>
>
> On 10/27/23 09:56, Christophe Lyon wrote:
> > Hi Vladimir,
> >
> > On Thu, 26 Oct 2023 at 16:00, Vladimir Makarov <vmakarov.gcc@gmail.com> wrote:
> >> This is the second attempt to improve RA cost calculation for pseudos
> >> with equivalences. The patch explanation is in the log message.
> >>
> >> The patch was successfully bootstrapped and tested on x86-64, aarch64,
> >> and ppc64le. The patch was also benchmarked on x86-64 spec2017.
> >> specfp2017 performance did not changed, specint2017 improved by 0.3%.
> >>
> > As reported by our CI, this patch causes a regression on arm:
> > FAIL: gcc.target/arm/eliminate.c scan-assembler-times r0,[\\t ]*sp 3
> >
> >
> > For this testcase, we used to generate:
> > str lr, [sp, #-4]!
> > sub sp, sp, #12
> > add r0, sp, #4
> > bl bar
> > add r0, sp, #4
> > bl bar
> > add r0, sp, #4
> > bl bar
> > add sp, sp, #12
> > ldr lr, [sp], #4
> > bx lr
> >
> > After your patch, we generate:
> > push {r4, lr}
> > sub sp, sp, #8
> > add r4, sp, #4
> > mov r0, r4
> > bl bar
> > mov r0, r4
> > bl bar
> > mov r0, r4
> > bl bar
> > add sp, sp, #8
> > pop {r4, lr}
> > bx lr
> >
> > which uses 1 more register and 1 more instruction.
> >
> > Shall I file a bugzilla report for this?
> >
> I started to work on this right after I got the message (yesterday). I
> already have a patch and am going to commit it during an hour. So there
> is no need to fill the PR.
>
Great, thanks for the quick fix!
Hey,
This is causing an ICE. Bug here: 112337 � arm: ICE in arm_effective_regno when compiling for MVE (gnu.org)<https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112337>
Regards,
Saurabh
commit f55cdce3f8dd8503e080e35be59c5f5390f6d95e
Author: Vladimir N. Makarov <vmakarov@redhat.com>
Date: Thu Oct 26 09:50:40 2023 -0400
[RA]: Modfify cost calculation for dealing with equivalences
RISCV target developers reported that pseudos with equivalence used in
a loop can be spilled. Simple changes of heuristics of cost
calculation of pseudos with equivalence or even ignoring equivalences
resulted in numerous testsuite failures on different targets or worse
spec2017 performance. This patch implements more sophisticated cost
calculations of pseudos with equivalences. The patch does not change
RA behaviour for targets still using the old reload pass instead of
LRA. The patch solves the reported problem and improves x86-64
specint2017 a bit (specfp2017 performance stays the same). The patch
takes into account how the equivalence will be used: will it be
integrated into the user insns or require an input reload insn. It
requires additional pass over insns. To compensate RA slow down, the
patch removes a pass over insns in the reload pass used by IRA before.
This also decouples IRA from reload more and will help to remove the
reload pass in the future if it ever happens.
gcc/ChangeLog:
* dwarf2out.cc (reg_loc_descriptor): Use lra_eliminate_regs when
LRA is used.
* ira-costs.cc: Include regset.h.
(equiv_can_be_consumed_p, get_equiv_regno, calculate_equiv_gains):
New functions.
(find_costs_and_classes): Call calculate_equiv_gains and redefine
mem_cost of pseudos with equivs when LRA is used.
* var-tracking.cc: Include ira.h and lra.h.
(vt_initialize): Use lra_eliminate_regs when LRA is used.
@@ -14311,7 +14311,9 @@ reg_loc_descriptor (rtx rtl, enum var_init_status initialized)
argument pointer and soft frame pointer rtx's.
Use DW_OP_fbreg offset DW_OP_stack_value in this case. */
if ((rtl == arg_pointer_rtx || rtl == frame_pointer_rtx)
- && eliminate_regs (rtl, VOIDmode, NULL_RTX) != rtl)
+ && (ira_use_lra_p
+ ? lra_eliminate_regs (rtl, VOIDmode, NULL_RTX)
+ : eliminate_regs (rtl, VOIDmode, NULL_RTX)) != rtl)
{
dw_loc_descr_ref result = NULL;
@@ -30,6 +30,7 @@ along with GCC; see the file COPYING3. If not see
#include "tm_p.h"
#include "insn-config.h"
#include "regs.h"
+#include "regset.h"
#include "ira.h"
#include "ira-int.h"
#include "addresses.h"
@@ -1757,6 +1758,145 @@ process_bb_node_for_costs (ira_loop_tree_node_t loop_tree_node)
process_bb_for_costs (bb);
}
+/* Check that reg REGNO can be changed by TO in INSN. Return true in case the
+ result insn would be valid one. */
+static bool
+equiv_can_be_consumed_p (int regno, rtx to, rtx_insn *insn)
+{
+ validate_replace_src_group (regno_reg_rtx[regno], to, insn);
+ bool res = verify_changes (0);
+ cancel_changes (0);
+ return res;
+}
+
+/* Return true if X contains a pseudo with equivalence. In this case also
+ return the pseudo through parameter REG. If the pseudo is a part of subreg,
+ return the subreg through parameter SUBREG. */
+
+static bool
+get_equiv_regno (rtx x, int ®no, rtx &subreg)
+{
+ subreg = NULL_RTX;
+ if (GET_CODE (x) == SUBREG)
+ {
+ subreg = x;
+ x = SUBREG_REG (x);
+ }
+ if (REG_P (x)
+ && (ira_reg_equiv[REGNO (x)].memory != NULL
+ || ira_reg_equiv[REGNO (x)].constant != NULL))
+ {
+ regno = REGNO (x);
+ return true;
+ }
+ RTX_CODE code = GET_CODE (x);
+ const char *fmt = GET_RTX_FORMAT (code);
+
+ for (int i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
+ if (fmt[i] == 'e')
+ {
+ if (get_equiv_regno (XEXP (x, i), regno, subreg))
+ return true;
+ }
+ else if (fmt[i] == 'E')
+ {
+ for (int j = 0; j < XVECLEN (x, i); j++)
+ if (get_equiv_regno (XVECEXP (x, i, j), regno, subreg))
+ return true;
+ }
+ return false;
+}
+
+/* A pass through the current function insns. Calculate costs of using
+ equivalences for pseudos and store them in regno_equiv_gains. */
+
+static void
+calculate_equiv_gains (void)
+{
+ basic_block bb;
+ int regno, freq, cost;
+ rtx subreg;
+ rtx_insn *insn;
+ machine_mode mode;
+ enum reg_class rclass;
+ bitmap_head equiv_pseudos;
+
+ ira_assert (allocno_p);
+ bitmap_initialize (&equiv_pseudos, ®_obstack);
+ for (regno = max_reg_num () - 1; regno >= FIRST_PSEUDO_REGISTER; regno--)
+ if (ira_reg_equiv[regno].init_insns != NULL
+ && (ira_reg_equiv[regno].memory != NULL
+ || (ira_reg_equiv[regno].constant != NULL
+ /* Ignore complicated constants which probably will be placed
+ in memory: */
+ && GET_CODE (ira_reg_equiv[regno].constant) != CONST_DOUBLE
+ && GET_CODE (ira_reg_equiv[regno].constant) != CONST_VECTOR
+ && GET_CODE (ira_reg_equiv[regno].constant) != LABEL_REF)))
+ {
+ rtx_insn_list *x;
+ for (x = ira_reg_equiv[regno].init_insns; x != NULL; x = x->next ())
+ {
+ insn = x->insn ();
+ rtx set = single_set (insn);
+
+ if (set == NULL_RTX || SET_DEST (set) != regno_reg_rtx[regno])
+ break;
+ bb = BLOCK_FOR_INSN (insn);
+ ira_curr_regno_allocno_map
+ = ira_bb_nodes[bb->index].parent->regno_allocno_map;
+ mode = PSEUDO_REGNO_MODE (regno);
+ rclass = pref[COST_INDEX (regno)];
+ ira_init_register_move_cost_if_necessary (mode);
+ if (ira_reg_equiv[regno].memory != NULL)
+ cost = ira_memory_move_cost[mode][rclass][1];
+ else
+ cost = ira_register_move_cost[mode][rclass][rclass];
+ freq = REG_FREQ_FROM_BB (bb);
+ regno_equiv_gains[regno] += cost * freq;
+ }
+ if (x != NULL)
+ /* We found complicated equiv or reverse equiv mem=reg. Ignore
+ them. */
+ regno_equiv_gains[regno] = 0;
+ else
+ bitmap_set_bit (&equiv_pseudos, regno);
+ }
+
+ FOR_EACH_BB_FN (bb, cfun)
+ {
+ freq = REG_FREQ_FROM_BB (bb);
+ ira_curr_regno_allocno_map
+ = ira_bb_nodes[bb->index].parent->regno_allocno_map;
+ FOR_BB_INSNS (bb, insn)
+ {
+ if (!INSN_P (insn) || !get_equiv_regno (PATTERN (insn), regno, subreg)
+ || !bitmap_bit_p (&equiv_pseudos, regno))
+ continue;
+ rtx subst = ira_reg_equiv[regno].memory;
+
+ if (subst == NULL)
+ subst = ira_reg_equiv[regno].constant;
+ ira_assert (subst != NULL);
+ mode = PSEUDO_REGNO_MODE (regno);
+ ira_init_register_move_cost_if_necessary (mode);
+ bool consumed_p = equiv_can_be_consumed_p (regno, subst, insn);
+
+ rclass = pref[COST_INDEX (regno)];
+ if (MEM_P (subst)
+ /* If it is a change of constant into double for example, the
+ result constant probably will be placed in memory. */
+ || (subreg != NULL_RTX && !INTEGRAL_MODE_P (GET_MODE (subreg))))
+ cost = ira_memory_move_cost[mode][rclass][1] + (consumed_p ? 0 : 1);
+ else if (consumed_p)
+ continue;
+ else
+ cost = ira_register_move_cost[mode][rclass][rclass];
+ regno_equiv_gains[regno] -= cost * freq;
+ }
+ }
+ bitmap_clear (&equiv_pseudos);
+}
+
/* Find costs of register classes and memory for allocnos or pseudos
and their best costs. Set up preferred, alternative and allocno
classes for pseudos. */
@@ -1855,6 +1995,12 @@ find_costs_and_classes (FILE *dump_file)
if (pass == 0)
pref = pref_buffer;
+ if (ira_use_lra_p && allocno_p && pass == 1)
+ /* It is a pass through all insns. So do it once and only for RA (not
+ for insn scheduler) when we already found preferable pseudo register
+ classes on the previous pass. */
+ calculate_equiv_gains ();
+
/* Now for each allocno look at how desirable each class is and
find which class is preferred. */
for (i = max_reg_num () - 1; i >= FIRST_PSEUDO_REGISTER; i--)
@@ -1947,6 +2093,17 @@ find_costs_and_classes (FILE *dump_file)
}
if (i >= first_moveable_pseudo && i < last_moveable_pseudo)
i_mem_cost = 0;
+ else if (ira_use_lra_p)
+ {
+ if (equiv_savings > 0)
+ {
+ i_mem_cost = 0;
+ if (ira_dump_file != NULL && internal_flag_ira_verbose > 5)
+ fprintf (ira_dump_file,
+ " Use MEM for r%d as the equiv savings is %d\n",
+ i, equiv_savings);
+ }
+ }
else if (equiv_savings < 0)
i_mem_cost = -equiv_savings;
else if (equiv_savings > 0)
@@ -2395,7 +2552,10 @@ ira_costs (void)
total_allocno_costs = (struct costs *) ira_allocate (max_struct_costs_size
* ira_allocnos_num);
initiate_regno_cost_classes ();
- calculate_elim_costs_all_insns ();
+ if (!ira_use_lra_p)
+ /* Process equivs in reload to update costs through hook
+ ira_adjust_equiv_reg_cost. */
+ calculate_elim_costs_all_insns ();
find_costs_and_classes (ira_dump_file);
setup_allocno_class_and_costs ();
finish_regno_cost_classes ();
@@ -2520,13 +2680,14 @@ ira_tune_allocno_costs (void)
}
}
-/* Add COST to the estimated gain for eliminating REGNO with its
- equivalence. If COST is zero, record that no such elimination is
- possible. */
+/* A hook from the reload pass. Add COST to the estimated gain for eliminating
+ REGNO with its equivalence. If COST is zero, record that no such
+ elimination is possible. */
void
ira_adjust_equiv_reg_cost (unsigned regno, int cost)
{
+ ira_assert (!ira_use_lra_p);
if (cost == 0)
regno_equiv_gains[regno] = 0;
else
@@ -107,6 +107,8 @@
#include "cfgrtl.h"
#include "cfganal.h"
#include "reload.h"
+#include "ira.h"
+#include "lra.h"
#include "calls.h"
#include "tree-dfa.h"
#include "tree-ssa.h"
@@ -10131,7 +10133,9 @@ vt_initialize (void)
#else
reg = arg_pointer_rtx;
#endif
- elim = eliminate_regs (reg, VOIDmode, NULL_RTX);
+ elim = (ira_use_lra_p
+ ? lra_eliminate_regs (reg, VOIDmode, NULL_RTX)
+ : eliminate_regs (reg, VOIDmode, NULL_RTX));
if (elim != reg)
{
if (GET_CODE (elim) == PLUS)
@@ -10151,7 +10155,9 @@ vt_initialize (void)
reg = arg_pointer_rtx;
fp_cfa_offset = ARG_POINTER_CFA_OFFSET (current_function_decl);
#endif
- elim = eliminate_regs (reg, VOIDmode, NULL_RTX);
+ elim = (ira_use_lra_p
+ ? lra_eliminate_regs (reg, VOIDmode, NULL_RTX)
+ : eliminate_regs (reg, VOIDmode, NULL_RTX));
if (elim != reg)
{
if (GET_CODE (elim) == PLUS)
@@ -10183,7 +10189,9 @@ vt_initialize (void)
#else
reg = arg_pointer_rtx;
#endif
- elim = eliminate_regs (reg, VOIDmode, NULL_RTX);
+ elim = (ira_use_lra_p
+ ? lra_eliminate_regs (reg, VOIDmode, NULL_RTX)
+ : eliminate_regs (reg, VOIDmode, NULL_RTX));
if (elim != reg)
{
if (GET_CODE (elim) == PLUS)