[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [RFC PATCH v2 38/44] target/loongarch: Implement vbitsel vset
|
From: |
gaosong |
|
Subject: |
Re: [RFC PATCH v2 38/44] target/loongarch: Implement vbitsel vset |
|
Date: |
Fri, 14 Apr 2023 11:22:59 +0800 |
|
User-agent: |
Mozilla/5.0 (X11; Linux loongarch64; rv:68.0) Gecko/20100101 Thunderbird/68.7.0 |
在 2023/4/13 下午6:06, Richard Henderson 写道:
On 4/13/23 04:53, gaosong wrote:
在 2023/4/12 下午2:53, Richard Henderson 写道:
+#define SETANYEQZ(NAME, BIT, E) \
+void HELPER(NAME)(CPULoongArchState *env, uint32_t cd, uint32_t
vj) \
+{
\
+ int i; \
+ bool ret =
false; \
+ VReg *Vj = &(env->fpr[vj].vreg); \
+
\
+ for (i = 0; i < LSX_LEN/BIT; i++)
{ \
+ ret |= (Vj->E(i) ==
0); \
+ } \
+ env->cf[cd & 0x7] =
ret; \
+}
+SETANYEQZ(vsetanyeqz_b, 8, B)
+SETANYEQZ(vsetanyeqz_h, 16, H)
+SETANYEQZ(vsetanyeqz_w, 32, W)
+SETANYEQZ(vsetanyeqz_d, 64, D)
These could be inlined, though slightly harder.
C.f. target/arm/sve_helper.c, do_match2 (your n == 0).
Do you mean an inline like trans_vseteqz_v or just an inline helper
function?
I meant inline tcg code generation, instead of a call to a helper.
But even if we keep this in a helper, see do_match2 for avoiding the
loop over bytes.
Ok,
e.g
#define SETANYEQZ(NAME, MO) \
void HELPER(NAME)(CPULoongArchState *env, uint32_t cd, uint32_t vj) \
{ \
int i; \
bool ret = false; \
VReg *Vj = &(env->fpr[vj].vreg); \
\
ret = do_match2(0, (uint64_t)Vj->D(0), (uint64_t)Vj->D(1),
MO); \
env->cf[cd & 0x7] = ret; \
}
SETANYEQZ(vsetanyeqz_b, MO_8)
SETANYEQZ(vsetanyeqz_h, MO_16)
SETANYEQZ(vsetanyeqz_w, MO_32)
SETANYEQZ(vsetanyeqz_d, MO_64)
and
vsetanyeqz.b $fcc5 $vr11
v11 : {edc0004d576eef5b, ec03ec0fec03ea47}
------------------
do_match2
bits is 8
m1 is ec03ec0fec03ea47
m0 is edc0004d576eef5b
ones is 1010101
sings is 80808080
cmp1 is 0
cmp0 is edc0004d576eef5b
cmp1 is ec03ec0fec03ea47
cmp0 is 10000
cmp1 is 3000100
ret is 0
but, the results is not correct for vsetanyeqz.b. :-)
Well, 'ones' as printed above is only 4 bytes instead of 8, similarly
'sings'. That would certainly explain why it did not detect a zero in
byte 5 of 'm0'.
Some problem with your conversion of that function?
I copied do_match2 from arm. and my host is x86 machine.
...
uint64_t ones = dup_const(esz, 1); // esz = MO_8
uint64_t signs = ones << ( bits -1 ); // bits = 8
...
the dup_const() return 0x101010101010101.
but set the 'ones' is 0x1010101.
Thread 1 "qemu-loongarch6" hit Breakpoint 1, helper_vsetanyeqz_b
(env=0x555555a50910, cd=6, vj=3) at ../target/loongarch/lsx_helper.c:2906
2906 SETANYEQZ(vsetanyeqz_b, MO_8, B)
(gdb) s
do_match2 (n=0, m0=14467753019624114359, m1=14467753019624114359, esz=0)
at ../target/loongarch/lsx_helper.c:2868
2868 uint64_t bits = 8 << esz;
(gdb) s
2869 uint64_t ones = dup_const(esz, 1);
(gdb) s
dup_const (vece=0, c=1) at ../tcg/tcg-op-gvec.c:374
374 switch (vece) {
(gdb) finish
Run till exit from #0 dup_const (vece=0, c=1) at ../tcg/tcg-op-gvec.c:374
do_match2 (n=0, m0=14467753019624114359, m1=14467753019624114359, esz=0)
at ../target/loongarch/lsx_helper.c:2869
2869 uint64_t ones = dup_const(esz, 1);
Value returned is $16 = 72340172838076673
(gdb) disassemble $pc
Dump of assembler code for function do_match2:
0x00005555555fffdf <+0>: push %rbp
0x00005555555fffe0 <+1>: mov %rsp,%rbp
0x00005555555fffe3 <+4>: sub $0x50,%rsp
0x00005555555fffe7 <+8>: mov %rdi,-0x38(%rbp)
0x00005555555fffeb <+12>: mov %rsi,-0x40(%rbp)
0x00005555555fffef <+16>: mov %rdx,-0x48(%rbp)
0x00005555555ffff3 <+20>: mov %ecx,-0x4c(%rbp)
0x00005555555ffff6 <+23>: mov -0x4c(%rbp),%eax
0x00005555555ffff9 <+26>: mov $0x8,%edx
0x00005555555ffffe <+31>: mov %eax,%ecx
0x0000555555600000 <+33>: shl %cl,%edx
0x0000555555600002 <+35>: mov %edx,%eax
0x0000555555600004 <+37>: cltq
0x0000555555600006 <+39>: mov %rax,-0x28(%rbp)
0x000055555560000a <+43>: mov -0x4c(%rbp),%eax
0x000055555560000d <+46>: mov $0x1,%esi
0x0000555555600012 <+51>: mov %eax,%edi
0x0000555555600014 <+53>: mov $0x0,%eax
0x0000555555600019 <+58>: callq 0x5555556342c3 <dup_const>
=> 0x000055555560001e <+63>: cltq
0x0000555555600020 <+65>: mov %rax,-0x20(%rbp)
0x0000555555600024 <+69>: mov -0x28(%rbp),%rax
0x0000555555600028 <+73>: sub $0x1,%eax
0x000055555560002b <+76>: mov -0x20(%rbp),%rdx
0x000055555560002f <+80>: mov %eax,%ecx
0x0000555555600031 <+82>: shl %cl,%rdx
0x0000555555600034 <+85>: mov %rdx,%rax
0x0000555555600037 <+88>: mov %rax,-0x18(%rbp)
0x000055555560003b <+92>: lea 0x129df7(%rip),%rdi #
0x555555729e39
0x0000555555600042 <+99>: callq 0x555555583af0 <puts@plt>
0x0000555555600047 <+104>: mov -0x4c(%rbp),%eax
--Type <RET> for more, q to quit, c to continue without paging--q
Quit
(gdb) p/x $rax
$17 = 0x101010101010101
(gdb) si
0x0000555555600020 2869 uint64_t ones = dup_const(esz, 1);
(gdb) p/x $rax
$18 = 0x1010101
(gdb) disassemble $pc
Dump of assembler code for function do_match2:
0x00005555555fffdf <+0>: push %rbp
0x00005555555fffe0 <+1>: mov %rsp,%rbp
0x00005555555fffe3 <+4>: sub $0x50,%rsp
0x00005555555fffe7 <+8>: mov %rdi,-0x38(%rbp)
0x00005555555fffeb <+12>: mov %rsi,-0x40(%rbp)
0x00005555555fffef <+16>: mov %rdx,-0x48(%rbp)
0x00005555555ffff3 <+20>: mov %ecx,-0x4c(%rbp)
0x00005555555ffff6 <+23>: mov -0x4c(%rbp),%eax
0x00005555555ffff9 <+26>: mov $0x8,%edx
0x00005555555ffffe <+31>: mov %eax,%ecx
0x0000555555600000 <+33>: shl %cl,%edx
0x0000555555600002 <+35>: mov %edx,%eax
0x0000555555600004 <+37>: cltq
0x0000555555600006 <+39>: mov %rax,-0x28(%rbp)
0x000055555560000a <+43>: mov -0x4c(%rbp),%eax
0x000055555560000d <+46>: mov $0x1,%esi
0x0000555555600012 <+51>: mov %eax,%edi
0x0000555555600014 <+53>: mov $0x0,%eax
0x0000555555600019 <+58>: callq 0x5555556342c3 <dup_const>
0x000055555560001e <+63>: cltq
=> 0x0000555555600020 <+65>: mov %rax,-0x20(%rbp)
0x0000555555600024 <+69>: mov -0x28(%rbp),%rax
0x0000555555600028 <+73>: sub $0x1,%eax
0x000055555560002b <+76>: mov -0x20(%rbp),%rdx
0x000055555560002f <+80>: mov %eax,%ecx
0x0000555555600031 <+82>: shl %cl,%rdx
0x0000555555600034 <+85>: mov %rdx,%rax
0x0000555555600037 <+88>: mov %rax,-0x18(%rbp)
0x000055555560003b <+92>: lea 0x129df7(%rip),%rdi #
0x555555729e39
0x0000555555600042 <+99>: callq 0x555555583af0 <puts@plt>
0x0000555555600047 <+104>: mov -0x4c(%rbp),%eax
--Type <RET> for more, q to quit, c to continue without paging--q
Quit
(gdb) p/x ones
$19 = 0x7fffffffc850
(gdb) si
2871 uint64_t signs = ones << (bits - 1);
(gdb) p/x $rax
$20 = 0x1010101
(gdb) p/x ones
$21 = 0x1010101
After exec insn 'cltq' , the 'ones' is not we want.
Thanks.
Song Gao