[Bug c/87944] Wrong code with LRA pushing stack local variable
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87944 Stefan "Bebbo" Franke changed: What|Removed |Added CC| |stefan at franke dot ms --- Comment #4 from Stefan "Bebbo" Franke --- the problem resides in lra-eliminations.c. It starts in void eliminate_regs_in_insn where a simple plus bypasses the stack offset handling: /* If we have a nonzero offset, and the source is already a simple REG, the following transformation would increase the cost of the insn by replacing a simple REG with (plus (reg sp) CST). So try only when we already had a PLUS before. */ if (known_eq (offset, 0) || plus_src) because lra_get_insn_recog_data (insn)->sp_offset is not considered. With the m68k architecture plus using lra you get this problem plus some similar ones, all related to arg / frame / stack pointer conversion with incorrect offsets. I fixed it locally with few more changes in lra-elimintaions.c
[Bug rtl-optimization/70961] Regrename ignores preferred_rename_class
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70961
Stefan "Bebbo" Franke changed:
What|Removed |Added
CC| |stefan at franke dot ms
--- Comment #6 from Stefan "Bebbo" Franke ---
Simply remove the outer loop and iterate only once:
has_preferred_class = (preferred_class != NO_REGS);
for (new_reg = 0; new_reg < FIRST_PSEUDO_REGISTER; new_reg++)
{
if (has_preferred_class
&& !TEST_HARD_REG_BIT(reg_class_contents[preferred_class],
new_reg))
continue;
if (!check_new_reg_p (old_reg, new_reg, this_head, *unavailable))
continue;
if (!best_rename)
return new_reg;
/* In the first pass, we force the renaming of registers that
don't belong to PREFERRED_CLASS to registers that do, even
though the latters were used not very long ago.
Also use a register if no best_new_reg was found till now */
if ((tick[best_new_reg] > tick[new_reg]
|| (old_reg == best_new_reg && new_reg < old_reg)))
best_new_reg = new_reg;
}
The last "if" is my experimental version. I am also testing (with m68k) an
extended preferred_rename_class2 method, which uses the register to return a
maybe better class:
preferred_class = (enum reg_class) targetm.preferred_rename_class2 (
super_class, old_reg);
And yes, the code is smaller and faster.
[Bug rtl-optimization/80126] New: df_ref_change_reg_with_loc does not update hard_regs_live_count
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80126
Bug ID: 80126
Summary: df_ref_change_reg_with_loc does not update
hard_regs_live_count
Product: gcc
Version: 6.3.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: rtl-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: stefan at franke dot ms
Target Milestone: ---
The method
static void
df_ref_change_reg_with_loc_1 (struct df_reg_info *old_df,
struct df_reg_info *new_df,
unsigned int new_regno, rtx loc)
does not update the hard_regs_live_count.
Old:
...
struct df_insn_info *insn_info = DF_REF_INSN_INFO (the_ref);
DF_REF_REGNO (the_ref) = new_regno;
...
Fixed:
...
struct df_insn_info *insn_info = DF_REF_INSN_INFO (the_ref);
if (DF_REF_FLAGS_IS_SET(the_ref, DF_HARD_REG_LIVE))
{
--df->hard_regs_live_count[DF_REF_REGNO(the_ref)];
++df->hard_regs_live_count[new_regno];
}
DF_REF_REGNO (the_ref) = new_regno;
...
[Bug rtl-optimization/80127] New: regrename_do_replace does not update the df refs
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80127
Bug ID: 80127
Summary: regrename_do_replace does not update the df refs
Product: gcc
Version: 6.3.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: rtl-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: stefan at franke dot ms
Target Milestone: ---
The method regrename_do_replace does neither update the df refs nor reset the
ever_live flag. Also the passes for stack variables are scheduled to early.
This leads to eliminated variables which are still pushed/popped to/from stack
but are never used.
Old code:
bool
regrename_do_replace (struct du_head *head, int regno)
{
...
mode = GET_MODE(*head->first->loc);
head->renamed = 1;
head->regno = regno;
head->nregs = hard_regno_nregs[regno][mode];
return true;
}
My code:
bool
regrename_do_replace (struct du_head *head, int regno)
{
...
mode = GET_MODE(*head->first->loc);
head->renamed = 1;
head->regno = regno;
head->nregs = hard_regno_nregs[regno][mode];
/* SBF: also update the current df info, move from base_regno -> regno. */
if (base_regno < FIRST_PSEUDO_REGISTER && regno < FIRST_PSEUDO_REGISTER)
for (chain = head->first; chain; chain = chain->next_use)
{
if (DEBUG_INSN_P (chain->insn))
continue;
/* undo regno patch - will be patched again */
if (REGNO (*chain->loc) == regno)
SET_REGNO(*chain->loc, base_regno);
df_ref_change_reg_with_loc (*chain->loc, regno);
SET_REGNO(*chain->loc, regno);
}
/* SBF: Mark the old regno as no longer used. */
if (!df->hard_regs_live_count[base_regno])
df_set_regs_ever_live (base_regno, false);
return true;
}
I also moved the stack creation passed after regrename:
passes.def:
...
NEXT_PASS (pass_regrename);
NEXT_PASS (pass_cprop_hardreg);
NEXT_PASS (pass_thread_prologue_and_epilogue);
NEXT_PASS (pass_rtl_dse2);
NEXT_PASS (pass_stack_adjustments);
...
=> No more push/pop for unused variables due to regrename.
[Bug tree-optimization/99364] New: Regression: loops not optimized to count to 0/-1
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99364
Bug ID: 99364
Summary: Regression: loops not optimized to count to 0/-1
Product: gcc
Version: tree-ssa
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: stefan at franke dot ms
Target Milestone: ---
consider this simple loop:
```
void ArrayFill2(int *array, int len) {
int i = 0;
for(; i < len; ++i)
array[i] = 50;
}
```
gcc 2.93 was able to convert this into a loop counting to zero.
gcc 6/7/8/9/10/11 can't do that.
if you change the code above:
```
void ArrayFill2(int *array) {
int i = 0;
int len = 42;
for(; i < len; ++i)
array[i] = 50;
}
```
then it works and you get this after ivcanon:
```
int i;
unsigned int i.0_4;
unsigned int _5;
int * _7;
unsigned int ivtmp_15;
unsigned int ivtmp_16;
:
:
# i_13 = PHI
# ivtmp_16 = PHI
i.0_4 = (unsigned int) i_13;
_5 = i.0_4 * 4;
_7 = array_6(D) + _5;
*_7 = 50;
i_9 = i_13 + 1;
ivtmp_15 = ivtmp_16 - 1;
if (ivtmp_15 != 0)
goto ;
else
goto ;
:
goto ;
:
return;
```
so the expected result for the version above, with len as parameter would be
something like this where the constant loop count is replaced by len-1:
```
int i;
unsigned int i.0_4;
unsigned int _5;
int * _7;
int ivtmp_14
unsigned int ivtmp_15;
unsigned int ivtmp_16;
:
ivtmp_14 = len_4(D) - 1;
:
# i_13 = PHI
# ivtmp_16 = PHI
i.0_4 = (unsigned int) i_13;
_5 = i.0_4 * 4;
_7 = array_6(D) + _5;
*_7 = 50;
i_9 = i_13 + 1;
ivtmp_15 = ivtmp_16 - 1;
if (ivtmp_15 != 0)
goto ;
else
goto ;
:
goto ;
:
return;
```
To handle this tree-ssa-loop-ivcanon must be able to handle no only const_int
loop limits but also a const loop limit provided in a variable.
