procyberian/gcc
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

[PR119285][IRA]: Use an additional way of reg equiv invariant substitution correctness

Browse source 
Patch for PR114991 resulted in 5% decrease of SPEC2017 lbm performance
on Zen2 and Zen4.  For one RTL insn of lbm, LRA with PR114991 patch
can not confirm that the equivalence insertion will create a valid RTL
insn.  This resulted in that the pseudo equiv was assumed costly and
pseudo was assigned to hard reg (caller saved as the pseudo lives
through calls) and some other pseudos did not get hard regs as it was
before PR114991 patch.  The insn in question is `pseudo1 = pseduo2 +
pseudo3` where pseudo2 has equiv `hard_reg + const`.  The old code
recognized the insn after equiv substitution as LEA.  The new code
failed.  The patch here makes to use two ways for equiv subsbtitution
correctness, the old one and new one (mostly for memory addresses
where the old code fails to find the substitution correctness).  So
given patch fixes lbm performance degradation and actually makes GCC
to generate the same code as it was before PR114991 patch.

gcc/ChangeLog:

	PR rtl-optimization/119285
	* ira-costs.cc (equiv_can_be_consumed_p): Use 2 ways for
	recognizing a valid insn after equiv insertion.
This commit is contained in:
Vladimir N. Makarov 2025-03-17 15:21:46 -04:00
parent 96698551b8
commit 8e0e17677a
1 changed files with 14 additions and 15 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

29
gcc/ira-costs.cc
View file

@ -1794,21 +1794,6 @@ validate_autoinc_and_mem_addr_p (rtx x)
static bool
equiv_can_be_consumed_p (int regno, rtx to, rtx_insn *insn, bool invariant_p)
{
if (invariant_p)
{
/* We use more expensive code for the invariant because we need to
simplify the result insn as the invariant can be arithmetic rtx
inserted into another arithmetic rtx. */
rtx pat = PATTERN (insn);
int code = INSN_CODE (insn);
PATTERN (insn) = copy_rtx (pat);
PATTERN (insn)
= simplify_replace_rtx (PATTERN (insn), regno_reg_rtx[regno], to);
bool res = !insn_invalid_p (insn, false);
PATTERN (insn) = pat;
INSN_CODE (insn) = code;
return res;
}
validate_replace_src_group (regno_reg_rtx[regno], to, insn);
/* We can change register to equivalent memory in autoinc rtl. Some code
including verify_changes assumes that autoinc contains only a register.
@ -1817,6 +1802,20 @@ equiv_can_be_consumed_p (int regno, rtx to, rtx_insn *insn, bool invariant_p)
if (res)
res = verify_changes (0);
cancel_changes (0);
if (!res && invariant_p)
{
/* Here we use more expensive code for the invariant because we need to
simplify the result insn as the invariant can be arithmetic rtx
inserted into another arithmetic rtx, e.g. into memory address. */
rtx pat = PATTERN (insn);
int code = INSN_CODE (insn);
PATTERN (insn) = copy_rtx (pat);
PATTERN (insn)
= simplify_replace_rtx (PATTERN (insn), regno_reg_rtx[regno], to);
res = !insn_invalid_p (insn, false);
PATTERN (insn) = pat;
INSN_CODE (insn) = code;
}
return res;
}