tcg/ppc: Implement cmpsel_vec

Do not allow cmpsel_vec to be expanded early, so that we can
make the correct decision wrt the sense of the comparison.

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
This commit is contained in:
Richard Henderson 2024-09-10 04:43:54 +00:00
parent 782cffa4ce
commit d0dabf9ec5
3 changed files with 54 additions and 9 deletions

View File

@ -33,6 +33,7 @@ C_O1_I2(r, r, rU)
C_O1_I2(r, r, rZW)
C_O1_I2(v, v, v)
C_O1_I3(v, v, v, v)
C_O1_I4(v, v, v, v, v)
C_O1_I4(r, r, rC, rZ, rZ)
C_O1_I4(r, r, r, ri, ri)
C_O2_I1(r, r, r)

View File

@ -3573,6 +3573,7 @@ int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
case INDEX_op_rotli_vec:
return vece <= MO_32 || have_isa_2_07 ? -1 : 0;
case INDEX_op_cmp_vec:
case INDEX_op_cmpsel_vec:
return vece <= MO_32 || have_isa_2_07 ? 1 : 0;
case INDEX_op_neg_vec:
return vece >= MO_32 && have_isa_3_00;
@ -3719,6 +3720,33 @@ static void tcg_out_not_vec(TCGContext *s, TCGReg a0, TCGReg a1)
tcg_out32(s, VNOR | VRT(a0) | VRA(a1) | VRB(a1));
}
static void tcg_out_or_vec(TCGContext *s, TCGReg a0, TCGReg a1, TCGReg a2)
{
tcg_out32(s, VOR | VRT(a0) | VRA(a1) | VRB(a2));
}
static void tcg_out_and_vec(TCGContext *s, TCGReg a0, TCGReg a1, TCGReg a2)
{
tcg_out32(s, VAND | VRT(a0) | VRA(a1) | VRB(a2));
}
static void tcg_out_andc_vec(TCGContext *s, TCGReg a0, TCGReg a1, TCGReg a2)
{
tcg_out32(s, VANDC | VRT(a0) | VRA(a1) | VRB(a2));
}
static void tcg_out_bitsel_vec(TCGContext *s, TCGReg d,
TCGReg c, TCGReg t, TCGReg f)
{
if (TCG_TARGET_HAS_bitsel_vec) {
tcg_out32(s, XXSEL | VRT(d) | VRC(c) | VRB(t) | VRA(f));
} else {
tcg_out_and_vec(s, TCG_VEC_TMP2, t, c);
tcg_out_andc_vec(s, d, f, c);
tcg_out_or_vec(s, d, d, TCG_VEC_TMP2);
}
}
static bool tcg_out_cmp_vec_noinv(TCGContext *s, unsigned vece, TCGReg a0,
TCGReg a1, TCGReg a2, TCGCond cond)
{
@ -3798,6 +3826,18 @@ static void tcg_out_cmp_vec(TCGContext *s, unsigned vece, TCGReg a0,
}
}
static void tcg_out_cmpsel_vec(TCGContext *s, unsigned vece, TCGReg a0,
TCGReg c1, TCGReg c2, TCGReg v3, TCGReg v4,
TCGCond cond)
{
if (tcg_out_cmp_vec_noinv(s, vece, TCG_VEC_TMP1, c1, c2, cond)) {
TCGReg swap = v3;
v3 = v4;
v4 = swap;
}
tcg_out_bitsel_vec(s, a0, TCG_VEC_TMP1, v3, v4);
}
static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
unsigned vecl, unsigned vece,
const TCGArg args[TCG_MAX_OP_ARGS],
@ -3889,17 +3929,17 @@ static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
insn = sarv_op[vece];
break;
case INDEX_op_and_vec:
insn = VAND;
break;
tcg_out_and_vec(s, a0, a1, a2);
return;
case INDEX_op_or_vec:
insn = VOR;
break;
tcg_out_or_vec(s, a0, a1, a2);
return;
case INDEX_op_xor_vec:
insn = VXOR;
break;
case INDEX_op_andc_vec:
insn = VANDC;
break;
tcg_out_andc_vec(s, a0, a1, a2);
return;
case INDEX_op_not_vec:
tcg_out_not_vec(s, a0, a1);
return;
@ -3919,9 +3959,11 @@ static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
case INDEX_op_cmp_vec:
tcg_out_cmp_vec(s, vece, a0, a1, a2, args[3]);
return;
case INDEX_op_cmpsel_vec:
tcg_out_cmpsel_vec(s, vece, a0, a1, a2, args[3], args[4], args[5]);
return;
case INDEX_op_bitsel_vec:
tcg_out32(s, XXSEL | VRT(a0) | VRC(a1) | VRB(a2) | VRA(args[3]));
tcg_out_bitsel_vec(s, a0, a1, a2, args[3]);
return;
case INDEX_op_dup2_vec:
@ -4287,6 +4329,8 @@ static TCGConstraintSetIndex tcg_target_op_def(TCGOpcode op)
case INDEX_op_bitsel_vec:
case INDEX_op_ppc_msum_vec:
return C_O1_I3(v, v, v, v);
case INDEX_op_cmpsel_vec:
return C_O1_I4(v, v, v, v, v);
default:
g_assert_not_reached();

View File

@ -172,7 +172,7 @@ typedef enum {
#define TCG_TARGET_HAS_sat_vec 1
#define TCG_TARGET_HAS_minmax_vec 1
#define TCG_TARGET_HAS_bitsel_vec have_vsx
#define TCG_TARGET_HAS_cmpsel_vec 0
#define TCG_TARGET_HAS_cmpsel_vec 1
#define TCG_TARGET_HAS_tst_vec 0
#define TCG_TARGET_DEFAULT_MO (0)