Hi,
I'm still working on a new gcc-4.5.2 backend for a private processor.
I encountered a strange behavior and I'm unable to find what causes this
behavior.
As an overview, it seems that dse2 pass removes insn where it should not (optim
-O2, -O3)
Here is the code giving me headachs which returns 0 when it should return
0x3F800000 (hex representation of 1.0f) :
void f1(int *ret2) {
*ret2 = 2;
}
float f2(float par1) {
return par1;
}
void (*ff)() = f1;
int main() {
int x;
float af;
ff(&x);
af = f2(1.0f);
return *((int *)(&af));
}
When I try to simplify this sample code further, the problem disappear even if
apparenlty there is no relation between the f2 call and the ff call...
I watched the rtl dump and here is an extract of the interesting part dealing
with the *((int *)(&af)) with f2 previously inlined (-O2 -O3) :
* in expand dump :
LC0->$72
The 1.0f constant (LC0) is in data section.
Its a normal behavior because I have to insn to load a constant in $FP regs.
Hence I have to load the address of LC0 first.
(insn 9 8 10 3 (set:SI (reg:SI 72)
(symbol_ref/u:SI ("*.LC0") [flags 0x2])) -1 (nil))
M($72)->$73
Memory pointing LC0 is loaded into reg 73
(insn 10 9 11 3 (set:SF (reg:SF 73)
(mem:SF (reg:SI 72) [0 S4 A32])) -1 (nil))
$73->M($62+4)
reg 73 moved into mem(stack+4).
(There are no memory to memory moves so using the temp reg 73 to move M($72) to
M($62+4) is normal)
(insn 11 10 12 3 (set:SF (mem/c/i:SF (plus:SI (reg/f:SI 62 virtual-stack-vars)
(const_int 4 [0x4])) [3 af+0 S4
A32])
(reg:SF 73)) -1 (nil))
$62+4->$75
mem(stack+4) moved to reg 75
(insn 12 11 13 3 (set (reg:SI 75)
(plus:SI (reg/f:SI 62 virtual-stack-vars)
(const_int 4 [0x4]))))
M($75)->76
reg 75 to reg 76
(insn 13 12 14 3 (set (reg:SI 76)
(mem:SI (reg:SI 75) [4 S4 A32])) -1 (nil))
$76->$69
reg 76 to reg 69 which is the fixed return register
(I don't know why insn 12 don't move directly to reg 69... nevertheless it's
correct)
(insn 14 13 15 3 (set (reg:SI 69 [ <retval> ])
(reg:SI 76)) -1 (nil))
* in pro_and_epilogue dump (just before dse2) :
$C0+8->$C3
(insn 24 12 27 2 (set (reg/f:SI 3 $C3 [77])
(plus:SI (reg/f:SI 0 $C0)
(const_int 8 [0x8]))))
[...]
LC0->$C2
(insn 9 8 29 2 (set:SI (reg/f:SI 2 $C2 [72])
(symbol_ref/u:SI ("*.LC0") [flags 0x2])) 1
{movsi_internal} (expr_list:REG_EQUIV (symbol_ref/u:SI ("*.LC0") [flags 0x2])
(nil)))
M($C2)->$FP1
(insn 29 9 11 2 (set (reg:SF 41 $FP1)
(mem:SF (reg/f:SI 2 $C2 [72]) [0 S4 A32])) 10
{movsf_internal} (nil))
$FP1->M($C0+12)
(insn 11 29 13 2 (set:SF (mem/c/i:SF (plus:SI (reg/f:SI 0 $C0)
(const_int 12 [0xc])) [3 af+0 S4 A32])
(reg:SF 41 $FP1)) 10 {movsf_internal} (nil))
M($C3+4)->$R0
(insn 13 11 21 2 (set (reg/i:SI 16 $R0)
(mem:SI (plus:SI (reg/f:SI 3 $C3 [77])
(const_int 4 [0x4])) [4 S4 A32])) 1
{movsi_internal} (nil))
Every thing seems normal except the fact that $FP1 is moved in M($C0+12) and
then the stored value is read with M($C3+4) where $C3=$C0+8.
This is due to optimisation (factorization) done during previous passes with
the code before inline f2 call.
* in pro_ dse2_epilogue dump (we die here) :
$C0+8->$C3
(insn 24 12 27 2 (set (reg/f:SI 3 $C3 [77])
(plus:SI (reg/f:SI 0 $C0)
(const_int 8 [0x8]))))
M($C3+4)->$R0
(insn 13 8 21 2 (set (reg/i:SI 16 $R0)
(mem:SI (plus:SI (reg/f:SI 3 $C3 [77])
(const_int 4 [0x4])) [4 S4 A32])) 1
{movsi_internal} (expr_list:REG_DEAD (reg/f:SI 3 $C3 [77])
(nil)))
insn 9, 29 and 11 have been deleted ! :)
Here is what the dump says concerning insn 9, 29, 11 and 13
**scanning insn=9
mems_found = 0, cannot_delete = true
**scanning insn=29
mem: (reg/f:SI 2 $C2 [72])
after canon_rtx address: (reg/f:SI 2 $C2 [72])
after cselib_expand address: (symbol_ref/u:SI ("*.LC0") [flags 0x2])
after canon_rtx address: (symbol_ref/u:SI ("*.LC0") [flags 0x2])
gid=1 offset=0
processing const load gid=1[0..4)
mems_found = 0, cannot_delete = true
**scanning insn=11
mem: (plus:SI (reg/f:SI 0 $C0)
(const_int 12 [0xc]))
after canon_rtx address: (plus:SI (reg/f:SI 0 $C0)
(const_int 12 [0xc]))
gid=2 offset=12
processing const base store gid=2[12..16)
mems_found = 1, cannot_delete = false
**scanning insn=13
mem: (plus:SI (reg/f:SI 3 $C3 [77])
(const_int 4 [0x4]))
after canon_rtx address: (plus:SI (reg/f:SI 3 $C3 [77])
(const_int 4 [0x4]))
after cselib_expand address: (plus:SI (reg/f:SI 3 $C3 [77])
(const_int 4 [0x4]))
after canon_rtx address: (plus:SI (reg/f:SI 3 $C3 [77])
(const_int 4 [0x4]))
varying cselib base=10:7900 offset = 4
processing cselib load mem:(mem:SI (plus:SI (reg/f:SI 3 $C3 [77])
(const_int 4 [0x4])) [4 S4 A32])
processing cselib load against insn 11
mems_found = 0, cannot_delete = true
Locally deleting insn 11
deferring deletion of insn with uid = 11.
starting the processing of deferred insns
deleting insn with uid = 11.
ending the processing of deferred insns
DCE: Deleting insn 29
deleting insn with uid = 29.
DCE: Deleting insn 9
deleting insn with uid = 9.
DCE: Deleting insn 36
deleting insn with uid = 36
I tried to search if there were pending bugs concerning dse2 on gcc 4.5.2 but I
did not find anything.
I don't know if I forgot a macro or what I have done wrong. I don't know if the
problem comes from dse2 pass or is a consequence of an error done before.
Can someone help me to find out why my lovely insn are removed ?
Regards,
Selim
