ARM_FEATURE_COPROC (FPU_CRYPTO_EXT_ARMV8);
static const arm_feature_set crc_ext_armv8 =
ARM_FEATURE_COPROC (CRC_EXT_ARMV8);
+static const arm_feature_set fpu_neon_ext_v8_1 =
+ ARM_FEATURE_COPROC (FPU_NEON_EXT_ARMV8 | FPU_NEON_EXT_RDMA);
static int mfloat_abi_opt = -1;
/* Record user cpu selection for object attributes. */
{
/* FIXME: 5 = X_PRECISION, should be #define'd where we can use it.
Ditto for 15. */
- if (gen_to_words (words, 5, (long) 15) == 0)
+#define X_PRECISION 5
+#define E_PRECISION 15L
+ if (gen_to_words (words, X_PRECISION, E_PRECISION) == 0)
{
for (i = 0; i < NUM_FLOAT_VALS; i++)
{
return FAIL;
}
+/* Returns TRUE if double precision value V may be cast
+ to single precision without loss of accuracy. */
+
+static bfd_boolean
+is_double_a_single (long int v)
+{
+ int exp = (int) (v >> 52) & 0x7FF;
+ long int mantissa = (v & 0xFFFFFFFFFFFFFl);
+
+ return (exp == 0 || exp == 0x7FF
+ || (exp >= 1023 - 126 && exp <= 1023 + 127))
+ && (mantissa & 0x1FFFFFFFl) == 0;
+}
+
+/* Returns a double precision value casted to single precision
+ (ignoring the least significant bits in exponent and mantissa). */
+
+static int
+double_to_single (long int v)
+{
+ int sign = (int) ((v >> 63) & 1l);
+ int exp = (int) (v >> 52) & 0x7FF;
+ long int mantissa = (v & 0xFFFFFFFFFFFFFl);
+
+ if (exp == 0x7FF)
+ exp = 0xFF;
+ else
+ {
+ exp = exp - 1023 + 127;
+ if (exp >= 0xFF)
+ {
+ /* Infinity. */
+ exp = 0x7F;
+ mantissa = 0;
+ }
+ else if (exp < 0)
+ {
+ /* No denormalized numbers. */
+ exp = 0;
+ mantissa = 0;
+ }
+ }
+ mantissa >>= 29;
+ return (sign << 31) | (exp << 23) | mantissa;
+}
+
enum lit_type
{
CONST_THUMB,
CONST_VEC
};
+static void do_vfp_nsyn_opcode (const char *);
+
/* inst.reloc.exp describes an "=expr" load pseudo-operation.
Determine whether it can be performed with a move instruction; if
it can, convert inst.instruction to that move instruction and
unsigned long tbit;
bfd_boolean thumb_p = (t == CONST_THUMB);
bfd_boolean arm_p = (t == CONST_ARM);
- bfd_boolean vec64_p = (t == CONST_VEC) && !inst.operands[i].issingle;
if (thumb_p)
tbit = (inst.instruction > 0xffff) ? THUMB2_LOAD_BIT : THUMB_LOAD_BIT;
inst.error = _("invalid pseudo operation");
return TRUE;
}
+
if (inst.reloc.exp.X_op != O_constant
&& inst.reloc.exp.X_op != O_symbol
&& inst.reloc.exp.X_op != O_big)
inst.error = _("constant expression expected");
return TRUE;
}
- if ((inst.reloc.exp.X_op == O_constant
- || inst.reloc.exp.X_op == O_big)
- && !inst.operands[i].issingle)
+
+ if (inst.reloc.exp.X_op == O_constant
+ || inst.reloc.exp.X_op == O_big)
{
- if (thumb_p && inst.reloc.exp.X_op == O_constant)
+ offsetT v;
+
+ if (inst.reloc.exp.X_op == O_big)
{
- if (!unified_syntax && (inst.reloc.exp.X_add_number & ~0xFF) == 0)
+ LITTLENUM_TYPE w[X_PRECISION];
+ LITTLENUM_TYPE * l;
+
+ if (inst.reloc.exp.X_add_number == -1)
{
- /* This can be done with a mov(1) instruction. */
- inst.instruction = T_OPCODE_MOV_I8 | (inst.operands[i].reg << 8);
- inst.instruction |= inst.reloc.exp.X_add_number;
- return TRUE;
+ gen_to_words (w, X_PRECISION, E_PRECISION);
+ l = w;
+ /* FIXME: Should we check words w[2..5] ? */
}
+ else
+ l = generic_bignum;
+
+ v = ((l[1] & LITTLENUM_MASK) << LITTLENUM_NUMBER_OF_BITS)
+ | (l[0] & LITTLENUM_MASK);
}
- else if (arm_p && inst.reloc.exp.X_op == O_constant)
+ else
+ v = inst.reloc.exp.X_add_number;
+
+ if (!inst.operands[i].issingle)
{
- int value = encode_arm_immediate (inst.reloc.exp.X_add_number);
- if (value != FAIL)
+ if (thumb_p)
{
- /* This can be done with a mov instruction. */
- inst.instruction &= LITERAL_MASK;
- inst.instruction |= INST_IMMEDIATE | (OPCODE_MOV << DATA_OP_SHIFT);
- inst.instruction |= value & 0xfff;
- return TRUE;
+ if ((v & ~0xFF) == 0)
+ {
+ /* This can be done with a mov(1) instruction. */
+ inst.instruction = T_OPCODE_MOV_I8 | (inst.operands[i].reg << 8);
+ inst.instruction |= v;
+ return TRUE;
+ }
+
+ if (ARM_CPU_HAS_FEATURE (cpu_variant, arm_arch_t2)
+ && ARM_CPU_HAS_FEATURE (cpu_variant, arm_ext_v6t2))
+ {
+ /* Check if on thumb2 it can be done with a mov.w or mvn.w instruction. */
+ unsigned int newimm;
+ bfd_boolean isNegated;
+
+ newimm = encode_thumb32_immediate (v);
+ if (newimm != (unsigned int) FAIL)
+ isNegated = FALSE;
+ else
+ {
+ newimm = encode_thumb32_immediate (~ v);
+ if (newimm != (unsigned int) FAIL)
+ isNegated = TRUE;
+ }
+
+ if (newimm != (unsigned int) FAIL)
+ {
+ inst.instruction = 0xf04f0000 | (inst.operands[i].reg << 8);
+ inst.instruction |= (isNegated?0x200000:0);
+ inst.instruction |= (newimm & 0x800) << 15;
+ inst.instruction |= (newimm & 0x700) << 4;
+ inst.instruction |= (newimm & 0x0ff);
+ return TRUE;
+ }
+ else if ((v & ~0xFFFF) == 0 || (v & ~0xFFFF0000) == 0)
+ {
+ /* The number may be loaded with a movw/movt instruction. */
+ int imm;
+
+ if ((inst.reloc.exp.X_add_number & ~0xFFFF) == 0)
+ {
+ inst.instruction= 0xf2400000;
+ imm = v;
+ }
+ else
+ {
+ inst.instruction = 0xf2c00000;
+ imm = v >> 16;
+ }
+
+ inst.instruction |= (inst.operands[i].reg << 8);
+ inst.instruction |= (imm & 0xf000) << 4;
+ inst.instruction |= (imm & 0x0800) << 15;
+ inst.instruction |= (imm & 0x0700) << 4;
+ inst.instruction |= (imm & 0x00ff);
+ return TRUE;
+ }
+ }
}
+ else if (arm_p)
+ {
+ int value = encode_arm_immediate (v);
- value = encode_arm_immediate (~inst.reloc.exp.X_add_number);
- if (value != FAIL)
+ if (value != FAIL)
+ {
+ /* This can be done with a mov instruction. */
+ inst.instruction &= LITERAL_MASK;
+ inst.instruction |= INST_IMMEDIATE | (OPCODE_MOV << DATA_OP_SHIFT);
+ inst.instruction |= value & 0xfff;
+ return TRUE;
+ }
+
+ value = encode_arm_immediate (~ v);
+ if (value != FAIL)
+ {
+ /* This can be done with a mvn instruction. */
+ inst.instruction &= LITERAL_MASK;
+ inst.instruction |= INST_IMMEDIATE | (OPCODE_MVN << DATA_OP_SHIFT);
+ inst.instruction |= value & 0xfff;
+ return TRUE;
+ }
+ }
+ else if (t == CONST_VEC)
{
- /* This can be done with a mvn instruction. */
- inst.instruction &= LITERAL_MASK;
- inst.instruction |= INST_IMMEDIATE | (OPCODE_MVN << DATA_OP_SHIFT);
- inst.instruction |= value & 0xfff;
- return TRUE;
+ int op = 0;
+ unsigned immbits = 0;
+ unsigned immlo = inst.operands[1].imm;
+ unsigned immhi = inst.operands[1].regisimm
+ ? inst.operands[1].reg
+ : inst.reloc.exp.X_unsigned
+ ? 0
+ : ((bfd_int64_t)((int) immlo)) >> 32;
+ int cmode = neon_cmode_for_move_imm (immlo, immhi, FALSE, &immbits,
+ &op, 64, NT_invtype);
+
+ if (cmode == FAIL)
+ {
+ neon_invert_size (&immlo, &immhi, 64);
+ op = !op;
+ cmode = neon_cmode_for_move_imm (immlo, immhi, FALSE, &immbits,
+ &op, 64, NT_invtype);
+ }
+
+ if (cmode != FAIL)
+ {
+ inst.instruction = (inst.instruction & VLDR_VMOV_SAME)
+ | (1 << 23)
+ | (cmode << 8)
+ | (op << 5)
+ | (1 << 4);
+
+ /* Fill other bits in vmov encoding for both thumb and arm. */
+ if (thumb_mode)
+ inst.instruction |= (0x7 << 29) | (0xF << 24);
+ else
+ inst.instruction |= (0xF << 28) | (0x1 << 25);
+ neon_write_immbits (immbits);
+ return TRUE;
+ }
}
}
- else if (vec64_p)
- {
- int op = 0;
- unsigned immbits = 0;
- unsigned immlo = inst.operands[1].imm;
- unsigned immhi = inst.operands[1].regisimm
- ? inst.operands[1].reg
- : inst.reloc.exp.X_unsigned
- ? 0
- : ((bfd_int64_t)((int) immlo)) >> 32;
- int cmode = neon_cmode_for_move_imm (immlo, immhi, FALSE, &immbits,
- &op, 64, NT_invtype);
- if (cmode == FAIL)
+ if (t == CONST_VEC)
+ {
+ /* Check if vldr Rx, =constant could be optimized to vmov Rx, #constant. */
+ if (inst.operands[i].issingle
+ && is_quarter_float (inst.operands[1].imm)
+ && ARM_CPU_HAS_FEATURE (cpu_variant, fpu_vfp_ext_v3xd))
{
- neon_invert_size (&immlo, &immhi, 64);
- op = !op;
- cmode = neon_cmode_for_move_imm (immlo, immhi, FALSE, &immbits,
- &op, 64, NT_invtype);
+ inst.operands[1].imm =
+ neon_qfloat_bits (v);
+ do_vfp_nsyn_opcode ("fconsts");
+ return TRUE;
}
- if (cmode != FAIL)
+ else if (!inst.operands[1].issingle
+ && ARM_CPU_HAS_FEATURE (cpu_variant, fpu_vfp_ext_v3))
{
- inst.instruction = (inst.instruction & VLDR_VMOV_SAME)
- | (1 << 23)
- | (cmode << 8)
- | (op << 5)
- | (1 << 4);
- /* Fill other bits in vmov encoding for both thumb and arm. */
- if (thumb_mode)
- inst.instruction |= (0x7 << 29) | (0xF << 24);
- else
- inst.instruction |= (0xF << 28) | (0x1 << 25);
- neon_write_immbits (immbits);
- return TRUE;
+ if (is_double_a_single (v)
+ && is_quarter_float (double_to_single (v)))
+ {
+ inst.operands[1].imm =
+ neon_qfloat_bits (double_to_single (v));
+ do_vfp_nsyn_opcode ("fconstd");
+ return TRUE;
+ }
}
}
}
}
}
-static void do_vfp_nsyn_opcode (const char *);
-
static int
do_vfp_nsyn_mrs (void)
{
X(vqdmull, 0x0800d00, N_INV, 0x0800b40), \
X(vqdmulh, 0x0000b00, N_INV, 0x0800c40), \
X(vqrdmulh, 0x1000b00, N_INV, 0x0800d40), \
+ X(vqrdmlah, 0x3000b10, N_INV, 0x0800e40), \
+ X(vqrdmlsh, 0x3000c10, N_INV, 0x0800f40), \
X(vshl, 0x0000400, N_INV, 0x0800510), \
X(vqshl, 0x0000410, N_INV, 0x0800710), \
X(vand, 0x0000110, N_INV, 0x0800030), \
NUF(vrecpsq, 0000f10, 3, (RNQ, oRNQ, RNQ), neon_step),
NUF(vrsqrts, 0200f10, 3, (RNDQ, oRNDQ, RNDQ), neon_step),
NUF(vrsqrtsq, 0200f10, 3, (RNQ, oRNQ, RNQ), neon_step),
+ /* ARM v8.1 extension. */
+ nUF(vqrdmlah, _vqrdmlah, 3, (RNDQ, oRNDQ, RNDQ_RNSC), neon_qdmulh),
+ nUF(vqrdmlahq, _vqrdmlah, 3, (RNQ, oRNQ, RNDQ_RNSC), neon_qdmulh),
+ nUF(vqrdmlsh, _vqrdmlsh, 3, (RNDQ, oRNDQ, RNDQ_RNSC), neon_qdmulh),
+ nUF(vqrdmlshq, _vqrdmlsh, 3, (RNQ, oRNQ, RNDQ_RNSC), neon_qdmulh),
/* Two address, int/float. Types S8 S16 S32 F32. */
NUF(vabsq, 1b10300, 2, (RNQ, RNQ), neon_abs_neg),
void
arm_init_frag (fragS * fragP, int max_chars)
{
+ int frag_thumb_mode;
+
/* If the current ARM vs THUMB mode has not already
been recorded into this frag then do so now. */
if ((fragP->tc_frag_data.thumb_mode & MODE_RECORDED) == 0)
- fragP->tc_frag_data.thumb_mode = thumb_mode | MODE_RECORDED;
+ fragP->tc_frag_data.thumb_mode = thumb_mode | MODE_RECORDED;
+
+ frag_thumb_mode = fragP->tc_frag_data.thumb_mode ^ MODE_RECORDED;
/* Record a mapping symbol for alignment frags. We will delete this
later if the alignment ends up empty. */
mapping_state_2 (MAP_DATA, max_chars);
break;
case rs_align_code:
- mapping_state_2 (thumb_mode ? MAP_THUMB : MAP_ARM, max_chars);
+ mapping_state_2 (frag_thumb_mode ? MAP_THUMB : MAP_ARM, max_chars);
break;
default:
break;
ARM_ARCH_OPT ("armv7-m", ARM_ARCH_V7M, FPU_ARCH_VFP),
ARM_ARCH_OPT ("armv7e-m", ARM_ARCH_V7EM, FPU_ARCH_VFP),
ARM_ARCH_OPT ("armv8-a", ARM_ARCH_V8A, FPU_ARCH_VFP),
+ ARM_ARCH_OPT ("armv8.1-a", ARM_ARCH_V8_1A, FPU_ARCH_VFP),
ARM_ARCH_OPT ("xscale", ARM_ARCH_XSCALE, FPU_ARCH_VFP),
ARM_ARCH_OPT ("iwmmxt", ARM_ARCH_IWMMXT, FPU_ARCH_VFP),
ARM_ARCH_OPT ("iwmmxt2", ARM_ARCH_IWMMXT2,FPU_ARCH_VFP),
| ARM_EXT_DIV),
ARM_FEATURE_CORE_LOW (ARM_EXT_VIRT),
ARM_FEATURE_CORE_LOW (ARM_EXT_V7A)),
+ ARM_EXT_OPT ("rdma", FPU_ARCH_NEON_VFP_ARMV8,
+ ARM_FEATURE_COPROC (FPU_NEON_ARMV8 | FPU_NEON_EXT_RDMA),
+ ARM_FEATURE_CORE_LOW (ARM_EXT_V8)),
ARM_EXT_OPT ("xscale",ARM_FEATURE_COPROC (ARM_CEXT_XSCALE),
ARM_FEATURE_COPROC (ARM_CEXT_XSCALE), ARM_ANY),
{ NULL, 0, ARM_ARCH_NONE, ARM_ARCH_NONE, ARM_ARCH_NONE }
{"neon-fp-armv8", FPU_ARCH_NEON_VFP_ARMV8},
{"crypto-neon-fp-armv8",
FPU_ARCH_CRYPTO_NEON_VFP_ARMV8},
+ {"neon-fp-armv8.1", FPU_ARCH_NEON_VFP_ARMV8_1},
{NULL, ARM_ARCH_NONE}
};