OpenE2K
/
qemu-e2k
13
4
Fork 0
Code Issues 4 Pull Requests Projects 2 Releases Activity

e2k: add whole class of packed min & max

This commit is contained in:
Alibek Omarov 2021-01-15 05:37:37 +03:00 committed by Denis Drakhnia
parent 27defe956a
commit 3e98a47b0a
3 changed files with 62 additions and 75 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
target/e2k/helper.h
View File

@ -37,9 +37,17 @@ DEF_HELPER_4(dam_unlock_addr, int, env, i64, int, int)
/* Packed Min/Max */
DEF_HELPER_2(pminub, i64, i64, i64)
DEF_HELPER_2(pminsb, i64, i64, i64)
DEF_HELPER_2(pminuh, i64, i64, i64)
DEF_HELPER_2(pminsh, i64, i64, i64)
DEF_HELPER_2(pminuw, i64, i64, i64)
DEF_HELPER_2(pminsw, i64, i64, i64)
DEF_HELPER_2(pmaxub, i64, i64, i64)
DEF_HELPER_2(pmaxsb, i64, i64, i64)
DEF_HELPER_2(pmaxuh, i64, i64, i64)
DEF_HELPER_2(pmaxsh, i64, i64, i64)
DEF_HELPER_2(pmaxuw, i64, i64, i64)
DEF_HELPER_2(pmaxsw, i64, i64, i64)
/* Float 32/64 Ops */
#define DEF_HELPER_3_32_64(name) \

107
target/e2k/helper_vec.c
View File

@ -14,14 +14,24 @@ static uint8_t reverse_bits(uint8_t b)
return b;
}
#define vec64_ub 8
#define vec64_uh 4
#define vec64_uw 2
#define vec64_uq 1
#define vec64_sb vec64_ub
#define vec64_sh vec64_uh
#define vec64_sw vec64_uw
#define vec64_sq vec64_uq
typedef union {
uint8_t ub[8];
uint16_t uh[4];
uint32_t uw[2];
uint8_t ub[vec64_ub];
uint16_t uh[vec64_uh];
uint32_t uw[vec64_uw];
uint64_t uq;
int8_t sb[8];
int16_t sh[4];
int32_t sw[2];
int8_t sb[vec64_ub];
int16_t sh[vec64_uh];
int32_t sw[vec64_uw];
int64_t sq;
} vec64;
@ -94,6 +104,21 @@ uint64_t HELPER(packed_shuffle_i64)(uint64_t src1, uint64_t src2, uint64_t src3)
return ret.uq;
}
#define GEN_HELPER_PACKED(name, type, code) \
uint64_t HELPER(glue(name, type))(uint64_t src1, uint64_t src2) \
{ \
size_t i = 0; \
vec64 s1 = { .uq = src1 }, s2 = { .uq = src2 }, dst; \
for (; i < glue(vec64_, type); i++) { \
code \
} \
return dst.uq; \
}
#define GEN_HELPER_PACKED_MINMAX(name, type, op) \
GEN_HELPER_PACKED(name, type, { \
dst.type[i] = op(s1.type[i], s2.type[i]); \
})
// FIXME: not tested
uint64_t HELPER(pcmpeqb)(uint64_t src1, uint64_t src2)
{
@ -114,65 +139,19 @@ uint64_t HELPER(pcmpeqb)(uint64_t src1, uint64_t src2)
return ret.uq;
}
uint64_t HELPER(pminub)(uint64_t src1, uint64_t src2)
{
unsigned int i;
vec64 s1, s2, ret;
GEN_HELPER_PACKED_MINMAX(pmin, ub, MIN)
GEN_HELPER_PACKED_MINMAX(pmin, sb, MIN)
GEN_HELPER_PACKED_MINMAX(pmin, uh, MIN)
GEN_HELPER_PACKED_MINMAX(pmin, sh, MIN)
GEN_HELPER_PACKED_MINMAX(pmin, uw, MIN)
GEN_HELPER_PACKED_MINMAX(pmin, sw, MIN)
s1.uq = src1;
s2.uq = src2;
for (i = 0; i < 8; i++) {
ret.ub[i] = MIN(s1.ub[i], s2.ub[i]);
}
return ret.uq;
}
uint64_t HELPER(pminsh)(uint64_t src1, uint64_t src2)
{
unsigned int i;
vec64 s1, s2, ret;
s1.uq = src1;
s2.uq = src2;
for (i = 0; i < 4; i++) {
ret.sh[i] = MIN(s1.sh[i], s2.sh[i]);
}
return ret.uq;
}
uint64_t HELPER(pmaxub)(uint64_t src1, uint64_t src2)
{
unsigned int i;
vec64 s1, s2, ret;
s1.uq = src1;
s2.uq = src2;
for (i = 0; i < 8; i++) {
ret.ub[i] = MAX(s1.ub[i], s2.ub[i]);
}
return ret.uq;
}
uint64_t HELPER(pmaxsh)(uint64_t src1, uint64_t src2)
{
unsigned int i;
vec64 s1, s2, ret;
s1.uq = src1;
s2.uq = src2;
for (i = 0; i < 4; i++) {
ret.sh[i] = MAX(s1.sh[i], s2.sh[i]);
}
return ret.uq;
}
GEN_HELPER_PACKED_MINMAX(pmax, ub, MAX)
GEN_HELPER_PACKED_MINMAX(pmax, sb, MAX)
GEN_HELPER_PACKED_MINMAX(pmax, uh, MAX)
GEN_HELPER_PACKED_MINMAX(pmax, sh, MAX)
GEN_HELPER_PACKED_MINMAX(pmax, uw, MAX)
GEN_HELPER_PACKED_MINMAX(pmax, sw, MAX)
uint64_t HELPER(pmovmskb)(uint64_t src1, uint64_t src2)
{

22
target/e2k/translate/alc.c
View File

@ -2933,10 +2933,18 @@ static void gen_op(DisasContext *ctx, Instr *instr)
case OP_MOVFI: gen_movfi(instr); break;
case OP_MOVIF: gen_movif(instr); break;
case OP_MOVTD: gen_movtd(instr); break;
case OP_PMAXSH: gen_alopf1_ddd(instr, gen_helper_pmaxsh); break;
case OP_PMAXUB: gen_alopf1_ddd(instr, gen_helper_pmaxub); break;
case OP_PMINSH: gen_alopf1_ddd(instr, gen_helper_pminsh); break;
case OP_PMINUB: gen_alopf1_ddd(instr, gen_helper_pminub); break;
case OP_PMINSB: gen_alopf1_ddd(instr, gen_helper_pminsb); break;
case OP_PMINUH: gen_alopf1_ddd(instr, gen_helper_pminuh); break;
case OP_PMINSH: gen_alopf1_ddd(instr, gen_helper_pminsh); break;
case OP_PMINUW: gen_alopf1_ddd(instr, gen_helper_pminuw); break;
case OP_PMINSW: gen_alopf1_ddd(instr, gen_helper_pminsw); break;
case OP_PMAXUB: gen_alopf1_ddd(instr, gen_helper_pmaxub); break;
case OP_PMAXSB: gen_alopf1_ddd(instr, gen_helper_pmaxsb); break;
case OP_PMAXUH: gen_alopf1_ddd(instr, gen_helper_pmaxuh); break;
case OP_PMAXSH: gen_alopf1_ddd(instr, gen_helper_pmaxsh); break;
case OP_PMAXUW: gen_alopf1_ddd(instr, gen_helper_pmaxuw); break;
case OP_PMAXSW: gen_alopf1_ddd(instr, gen_helper_pmaxsw); break;
case OP_GETTAGS: gen_gettag_i32(instr); break;
case OP_GETTAGD: gen_gettag_i64(instr); break;
case OP_PUTTAGS: gen_puttag_i32(instr); break;
@ -3289,10 +3297,6 @@ static void gen_op(DisasContext *ctx, Instr *instr)
case OP_PMULUBHH:
case OP_FSTOIFS:
case OP_FDTOIFD:
case OP_PMINUW:
case OP_PMINSW:
case OP_PMAXUW:
case OP_PMAXSW:
case OP_MPSADBH:
case OP_PACKUSWH:
case OP_PCMPEQD:
@ -3300,10 +3304,6 @@ static void gen_op(DisasContext *ctx, Instr *instr)
case OP_PFHADDS:
case OP_PFHSUBS:
case OP_PFADDSUBS:
case OP_PMINSB:
case OP_PMINUH:
case OP_PMAXSB:
case OP_PMAXUH:
case OP_PFSTOIFS:
case OP_PFDTOIFD:
case OP_PHADDH: