arm64: implement FSQRT 2d_2d, 4s_4s, 2s_2s
AFAICS this completes the AArch64 SIMD implementation, except for the
crypto instructions.
This changes the type of Iop_Sqrt64x2 and Iop_Sqrt32x4 so as to take a
rounding mode argument. This will (temporarily, of course) break all
of the other targets that implement vector fsqrt.
git-svn-id: svn://svn.valgrind.org/vex/trunk@3120 8f6e269a-dfd6-0310-a8e1-e2731360e62c
diff --git a/priv/guest_arm64_toIR.c b/priv/guest_arm64_toIR.c
index a5b7d21..14c27f1 100644
--- a/priv/guest_arm64_toIR.c
+++ b/priv/guest_arm64_toIR.c
@@ -12437,6 +12437,21 @@
return True;
}
+ if (bitU == 1 && size >= X10 && opcode == BITS5(1,1,1,1,1)) {
+ /* -------- 1,1x,11111: FSQRT 2d_2d, 4s_4s, 2s_2s -------- */
+ Bool isD = (size & 1) == 1;
+ IROp op = isD ? Iop_Sqrt64Fx2 : Iop_Sqrt32Fx4;
+ if (bitQ == 0 && isD) return False; // implied 1d case
+ IRTemp resV = newTempV128();
+ assign(resV, binop(op, mkexpr(mk_get_IR_rounding_mode()),
+ getQReg128(nn)));
+ putQReg128(dd, math_MAYBE_ZERO_HI64(bitQ, resV));
+ const HChar* arr = bitQ == 0 ? "2s" : (size == X11 ? "2d" : "4s");
+ DIP("%s %s.%s, %s.%s\n", "fsqrt",
+ nameQReg128(dd), arr, nameQReg128(nn), arr);
+ return True;
+ }
+
return False;
# undef INSN
}
diff --git a/priv/host_arm64_defs.c b/priv/host_arm64_defs.c
index fc0984a..b886e81 100644
--- a/priv/host_arm64_defs.c
+++ b/priv/host_arm64_defs.c
@@ -749,6 +749,8 @@
case ARM64vecu_FRECPE32x4: *nm = "frecpe"; *ar = "4s"; return;
case ARM64vecu_FRSQRTE64x2: *nm = "frsqrte"; *ar = "2d"; return;
case ARM64vecu_FRSQRTE32x4: *nm = "frsqrte"; *ar = "4s"; return;
+ case ARM64vecu_FSQRT64x2: *nm = "fsqrt"; *ar = "2d"; return;
+ case ARM64vecu_FSQRT32x4: *nm = "fsqrt"; *ar = "4s"; return;
default: vpanic("showARM64VecUnaryOp");
}
}
@@ -4869,6 +4871,9 @@
011 01110 11 1 00001 110110 n d FRECPE Vd.2d, Vn.2d
011 01110 10 1 00001 110110 n d FRECPE Vd.4s, Vn.4s
+
+ 011 01110 11 1 00001 111110 n d FSQRT Vd.2d, Vn.2d
+ 011 01110 10 1 00001 111110 n d FSQRT Vd.4s, Vn.4s
*/
UInt vD = qregEnc(i->ARM64in.VUnaryV.dst);
UInt vN = qregEnc(i->ARM64in.VUnaryV.arg);
@@ -4960,6 +4965,12 @@
case ARM64vecu_FRSQRTE32x4:
*p++ = X_3_8_5_6_5_5(X011, X01110101, X00001, X110110, vN, vD);
break;
+ case ARM64vecu_FSQRT64x2:
+ *p++ = X_3_8_5_6_5_5(X011, X01110111, X00001, X111110, vN, vD);
+ break;
+ case ARM64vecu_FSQRT32x4:
+ *p++ = X_3_8_5_6_5_5(X011, X01110101, X00001, X111110, vN, vD);
+ break;
default:
goto bad;
}
diff --git a/priv/host_arm64_defs.h b/priv/host_arm64_defs.h
index fce9f83..039fce1 100644
--- a/priv/host_arm64_defs.h
+++ b/priv/host_arm64_defs.h
@@ -414,6 +414,7 @@
ARM64vecu_URSQRTE32x4,
ARM64vecu_FRECPE64x2, ARM64vecu_FRECPE32x4,
ARM64vecu_FRSQRTE64x2, ARM64vecu_FRSQRTE32x4,
+ ARM64vecu_FSQRT64x2, ARM64vecu_FSQRT32x4,
ARM64vecu_INVALID
}
ARM64VecUnaryOp;
diff --git a/priv/host_arm64_isel.c b/priv/host_arm64_isel.c
index a6e507d..b6e42f1 100644
--- a/priv/host_arm64_isel.c
+++ b/priv/host_arm64_isel.c
@@ -2363,6 +2363,17 @@
if (e->tag == Iex_Binop) {
switch (e->Iex.Binop.op) {
+ case Iop_Sqrt32Fx4:
+ case Iop_Sqrt64Fx2: {
+ HReg arg = iselV128Expr(env, e->Iex.Binop.arg2);
+ HReg res = newVRegV(env);
+ set_FPCR_rounding_mode(env, e->Iex.Binop.arg1);
+ ARM64VecUnaryOp op
+ = e->Iex.Binop.op == Iop_Sqrt32Fx4
+ ? ARM64vecu_FSQRT32x4 : ARM64vecu_FSQRT64x2;
+ addInstr(env, ARM64Instr_VUnaryV(op, res, arg));
+ return res;
+ }
case Iop_64HLtoV128: {
HReg res = newVRegV(env);
HReg argL = iselIntExpr_R(env, e->Iex.Binop.arg1);
diff --git a/priv/ir_defs.c b/priv/ir_defs.c
index 0b70f39..596379e 100644
--- a/priv/ir_defs.c
+++ b/priv/ir_defs.c
@@ -2859,6 +2859,10 @@
case Iop_RSqrtEst32Ux4:
UNARY(Ity_V128, Ity_V128);
+ case Iop_Sqrt64Fx2:
+ case Iop_Sqrt32Fx4:
+ BINARY(ity_RMode,Ity_V128, Ity_V128);
+
case Iop_64HLtoV128:
BINARY(Ity_I64,Ity_I64, Ity_V128);
@@ -3019,8 +3023,8 @@
case Iop_RecipEst64Fx2: case Iop_RSqrtEst64Fx2:
case Iop_RecipEst32Ux4:
case Iop_RSqrtEst32F0x4:
- case Iop_Sqrt32Fx4: case Iop_Sqrt32F0x4:
- case Iop_Sqrt64Fx2: case Iop_Sqrt64F0x2:
+ case Iop_Sqrt32F0x4:
+ case Iop_Sqrt64F0x2:
case Iop_CmpNEZ8x16: case Iop_CmpNEZ16x8:
case Iop_CmpNEZ32x4: case Iop_CmpNEZ64x2:
case Iop_Cnt8x16:
diff --git a/pub/libvex_ir.h b/pub/libvex_ir.h
index 7178c7e..145caa4 100644
--- a/pub/libvex_ir.h
+++ b/pub/libvex_ir.h
@@ -1294,9 +1294,11 @@
/* unary */
Iop_Abs32Fx4,
- Iop_Sqrt32Fx4,
Iop_Neg32Fx4,
+ /* binary :: IRRoundingMode(I32) x V128 -> V128 */
+ Iop_Sqrt32Fx4,
+
/* Vector Reciprocal Estimate finds an approximate reciprocal of each
element in the operand vector, and places the results in the
destination vector. */
@@ -1359,9 +1361,11 @@
/* unary */
Iop_Abs64Fx2,
- Iop_Sqrt64Fx2,
Iop_Neg64Fx2,
+ /* binary :: IRRoundingMode(I32) x V128 -> V128 */
+ Iop_Sqrt64Fx2,
+
/* see 32Fx4 variants for description */
Iop_RecipEst64Fx2, // unary
Iop_RecipStep64Fx2, // binary