]> Git Repo - qemu.git/blobdiff - tcg/tcg-op.c
tcg/aarch64: Implement field extraction opcodes
[qemu.git] / tcg / tcg-op.c
index bb2bfeef3c339b065465004f1df7a354cf906487..1927e531234b1a106f784266cea2d4bba01b7cab 100644 (file)
@@ -533,10 +533,11 @@ void tcg_gen_deposit_i32(TCGv_i32 ret, TCGv_i32 arg1, TCGv_i32 arg2,
     TCGv_i32 t1;
 
     tcg_debug_assert(ofs < 32);
+    tcg_debug_assert(len > 0);
     tcg_debug_assert(len <= 32);
     tcg_debug_assert(ofs + len <= 32);
 
-    if (ofs == 0 && len == 32) {
+    if (len == 32) {
         tcg_gen_mov_i32(ret, arg2);
         return;
     }
@@ -560,6 +561,189 @@ void tcg_gen_deposit_i32(TCGv_i32 ret, TCGv_i32 arg1, TCGv_i32 arg2,
     tcg_temp_free_i32(t1);
 }
 
+void tcg_gen_deposit_z_i32(TCGv_i32 ret, TCGv_i32 arg,
+                           unsigned int ofs, unsigned int len)
+{
+    tcg_debug_assert(ofs < 32);
+    tcg_debug_assert(len > 0);
+    tcg_debug_assert(len <= 32);
+    tcg_debug_assert(ofs + len <= 32);
+
+    if (ofs + len == 32) {
+        tcg_gen_shli_i32(ret, arg, ofs);
+    } else if (ofs == 0) {
+        tcg_gen_andi_i32(ret, arg, (1u << len) - 1);
+    } else if (TCG_TARGET_HAS_deposit_i32
+               && TCG_TARGET_deposit_i32_valid(ofs, len)) {
+        TCGv_i32 zero = tcg_const_i32(0);
+        tcg_gen_op5ii_i32(INDEX_op_deposit_i32, ret, zero, arg, ofs, len);
+        tcg_temp_free_i32(zero);
+    } else {
+        /* To help two-operand hosts we prefer to zero-extend first,
+           which allows ARG to stay live.  */
+        switch (len) {
+        case 16:
+            if (TCG_TARGET_HAS_ext16u_i32) {
+                tcg_gen_ext16u_i32(ret, arg);
+                tcg_gen_shli_i32(ret, ret, ofs);
+                return;
+            }
+            break;
+        case 8:
+            if (TCG_TARGET_HAS_ext8u_i32) {
+                tcg_gen_ext8u_i32(ret, arg);
+                tcg_gen_shli_i32(ret, ret, ofs);
+                return;
+            }
+            break;
+        }
+        /* Otherwise prefer zero-extension over AND for code size.  */
+        switch (ofs + len) {
+        case 16:
+            if (TCG_TARGET_HAS_ext16u_i32) {
+                tcg_gen_shli_i32(ret, arg, ofs);
+                tcg_gen_ext16u_i32(ret, ret);
+                return;
+            }
+            break;
+        case 8:
+            if (TCG_TARGET_HAS_ext8u_i32) {
+                tcg_gen_shli_i32(ret, arg, ofs);
+                tcg_gen_ext8u_i32(ret, ret);
+                return;
+            }
+            break;
+        }
+        tcg_gen_andi_i32(ret, arg, (1u << len) - 1);
+        tcg_gen_shli_i32(ret, ret, ofs);
+    }
+}
+
+void tcg_gen_extract_i32(TCGv_i32 ret, TCGv_i32 arg,
+                         unsigned int ofs, unsigned int len)
+{
+    tcg_debug_assert(ofs < 32);
+    tcg_debug_assert(len > 0);
+    tcg_debug_assert(len <= 32);
+    tcg_debug_assert(ofs + len <= 32);
+
+    /* Canonicalize certain special cases, even if extract is supported.  */
+    if (ofs + len == 32) {
+        tcg_gen_shri_i32(ret, arg, 32 - len);
+        return;
+    }
+    if (ofs == 0) {
+        tcg_gen_andi_i32(ret, arg, (1u << len) - 1);
+        return;
+    }
+
+    if (TCG_TARGET_HAS_extract_i32
+        && TCG_TARGET_extract_i32_valid(ofs, len)) {
+        tcg_gen_op4ii_i32(INDEX_op_extract_i32, ret, arg, ofs, len);
+        return;
+    }
+
+    /* Assume that zero-extension, if available, is cheaper than a shift.  */
+    switch (ofs + len) {
+    case 16:
+        if (TCG_TARGET_HAS_ext16u_i32) {
+            tcg_gen_ext16u_i32(ret, arg);
+            tcg_gen_shri_i32(ret, ret, ofs);
+            return;
+        }
+        break;
+    case 8:
+        if (TCG_TARGET_HAS_ext8u_i32) {
+            tcg_gen_ext8u_i32(ret, arg);
+            tcg_gen_shri_i32(ret, ret, ofs);
+            return;
+        }
+        break;
+    }
+
+    /* ??? Ideally we'd know what values are available for immediate AND.
+       Assume that 8 bits are available, plus the special case of 16,
+       so that we get ext8u, ext16u.  */
+    switch (len) {
+    case 1 ... 8: case 16:
+        tcg_gen_shri_i32(ret, arg, ofs);
+        tcg_gen_andi_i32(ret, ret, (1u << len) - 1);
+        break;
+    default:
+        tcg_gen_shli_i32(ret, arg, 32 - len - ofs);
+        tcg_gen_shri_i32(ret, ret, 32 - len);
+        break;
+    }
+}
+
+void tcg_gen_sextract_i32(TCGv_i32 ret, TCGv_i32 arg,
+                          unsigned int ofs, unsigned int len)
+{
+    tcg_debug_assert(ofs < 32);
+    tcg_debug_assert(len > 0);
+    tcg_debug_assert(len <= 32);
+    tcg_debug_assert(ofs + len <= 32);
+
+    /* Canonicalize certain special cases, even if extract is supported.  */
+    if (ofs + len == 32) {
+        tcg_gen_sari_i32(ret, arg, 32 - len);
+        return;
+    }
+    if (ofs == 0) {
+        switch (len) {
+        case 16:
+            tcg_gen_ext16s_i32(ret, arg);
+            return;
+        case 8:
+            tcg_gen_ext8s_i32(ret, arg);
+            return;
+        }
+    }
+
+    if (TCG_TARGET_HAS_sextract_i32
+        && TCG_TARGET_extract_i32_valid(ofs, len)) {
+        tcg_gen_op4ii_i32(INDEX_op_sextract_i32, ret, arg, ofs, len);
+        return;
+    }
+
+    /* Assume that sign-extension, if available, is cheaper than a shift.  */
+    switch (ofs + len) {
+    case 16:
+        if (TCG_TARGET_HAS_ext16s_i32) {
+            tcg_gen_ext16s_i32(ret, arg);
+            tcg_gen_sari_i32(ret, ret, ofs);
+            return;
+        }
+        break;
+    case 8:
+        if (TCG_TARGET_HAS_ext8s_i32) {
+            tcg_gen_ext8s_i32(ret, arg);
+            tcg_gen_sari_i32(ret, ret, ofs);
+            return;
+        }
+        break;
+    }
+    switch (len) {
+    case 16:
+        if (TCG_TARGET_HAS_ext16s_i32) {
+            tcg_gen_shri_i32(ret, arg, ofs);
+            tcg_gen_ext16s_i32(ret, ret);
+            return;
+        }
+        break;
+    case 8:
+        if (TCG_TARGET_HAS_ext8s_i32) {
+            tcg_gen_shri_i32(ret, arg, ofs);
+            tcg_gen_ext8s_i32(ret, ret);
+            return;
+        }
+        break;
+    }
+
+    tcg_gen_shli_i32(ret, arg, 32 - len - ofs);
+    tcg_gen_sari_i32(ret, ret, 32 - len);
+}
+
 void tcg_gen_movcond_i32(TCGCond cond, TCGv_i32 ret, TCGv_i32 c1,
                          TCGv_i32 c2, TCGv_i32 v1, TCGv_i32 v2)
 {
@@ -678,6 +862,33 @@ void tcg_gen_muls2_i32(TCGv_i32 rl, TCGv_i32 rh, TCGv_i32 arg1, TCGv_i32 arg2)
     }
 }
 
+void tcg_gen_mulsu2_i32(TCGv_i32 rl, TCGv_i32 rh, TCGv_i32 arg1, TCGv_i32 arg2)
+{
+    if (TCG_TARGET_REG_BITS == 32) {
+        TCGv_i32 t0 = tcg_temp_new_i32();
+        TCGv_i32 t1 = tcg_temp_new_i32();
+        TCGv_i32 t2 = tcg_temp_new_i32();
+        tcg_gen_mulu2_i32(t0, t1, arg1, arg2);
+        /* Adjust for negative input for the signed arg1.  */
+        tcg_gen_sari_i32(t2, arg1, 31);
+        tcg_gen_and_i32(t2, t2, arg2);
+        tcg_gen_sub_i32(rh, t1, t2);
+        tcg_gen_mov_i32(rl, t0);
+        tcg_temp_free_i32(t0);
+        tcg_temp_free_i32(t1);
+        tcg_temp_free_i32(t2);
+    } else {
+        TCGv_i64 t0 = tcg_temp_new_i64();
+        TCGv_i64 t1 = tcg_temp_new_i64();
+        tcg_gen_ext_i32_i64(t0, arg1);
+        tcg_gen_extu_i32_i64(t1, arg2);
+        tcg_gen_mul_i64(t0, t0, t1);
+        tcg_gen_extr_i64_i32(rl, rh, t0);
+        tcg_temp_free_i64(t0);
+        tcg_temp_free_i64(t1);
+    }
+}
+
 void tcg_gen_ext8s_i32(TCGv_i32 ret, TCGv_i32 arg)
 {
     if (TCG_TARGET_HAS_ext8s_i32) {
@@ -790,7 +1001,7 @@ void tcg_gen_ld8u_i64(TCGv_i64 ret, TCGv_ptr arg2, tcg_target_long offset)
 void tcg_gen_ld8s_i64(TCGv_i64 ret, TCGv_ptr arg2, tcg_target_long offset)
 {
     tcg_gen_ld8s_i32(TCGV_LOW(ret), arg2, offset);
-    tcg_gen_sari_i32(TCGV_HIGH(ret), TCGV_HIGH(ret), 31);
+    tcg_gen_sari_i32(TCGV_HIGH(ret), TCGV_LOW(ret), 31);
 }
 
 void tcg_gen_ld16u_i64(TCGv_i64 ret, TCGv_ptr arg2, tcg_target_long offset)
@@ -1566,10 +1777,11 @@ void tcg_gen_deposit_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2,
     TCGv_i64 t1;
 
     tcg_debug_assert(ofs < 64);
+    tcg_debug_assert(len > 0);
     tcg_debug_assert(len <= 64);
     tcg_debug_assert(ofs + len <= 64);
 
-    if (ofs == 0 && len == 64) {
+    if (len == 64) {
         tcg_gen_mov_i64(ret, arg2);
         return;
     }
@@ -1608,6 +1820,289 @@ void tcg_gen_deposit_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2,
     tcg_temp_free_i64(t1);
 }
 
+void tcg_gen_deposit_z_i64(TCGv_i64 ret, TCGv_i64 arg,
+                           unsigned int ofs, unsigned int len)
+{
+    tcg_debug_assert(ofs < 64);
+    tcg_debug_assert(len > 0);
+    tcg_debug_assert(len <= 64);
+    tcg_debug_assert(ofs + len <= 64);
+
+    if (ofs + len == 64) {
+        tcg_gen_shli_i64(ret, arg, ofs);
+    } else if (ofs == 0) {
+        tcg_gen_andi_i64(ret, arg, (1ull << len) - 1);
+    } else if (TCG_TARGET_HAS_deposit_i64
+               && TCG_TARGET_deposit_i64_valid(ofs, len)) {
+        TCGv_i64 zero = tcg_const_i64(0);
+        tcg_gen_op5ii_i64(INDEX_op_deposit_i64, ret, zero, arg, ofs, len);
+        tcg_temp_free_i64(zero);
+    } else {
+        if (TCG_TARGET_REG_BITS == 32) {
+            if (ofs >= 32) {
+                tcg_gen_deposit_z_i32(TCGV_HIGH(ret), TCGV_LOW(arg),
+                                      ofs - 32, len);
+                tcg_gen_movi_i32(TCGV_LOW(ret), 0);
+                return;
+            }
+            if (ofs + len <= 32) {
+                tcg_gen_deposit_z_i32(TCGV_LOW(ret), TCGV_LOW(arg), ofs, len);
+                tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
+                return;
+            }
+        }
+        /* To help two-operand hosts we prefer to zero-extend first,
+           which allows ARG to stay live.  */
+        switch (len) {
+        case 32:
+            if (TCG_TARGET_HAS_ext32u_i64) {
+                tcg_gen_ext32u_i64(ret, arg);
+                tcg_gen_shli_i64(ret, ret, ofs);
+                return;
+            }
+            break;
+        case 16:
+            if (TCG_TARGET_HAS_ext16u_i64) {
+                tcg_gen_ext16u_i64(ret, arg);
+                tcg_gen_shli_i64(ret, ret, ofs);
+                return;
+            }
+            break;
+        case 8:
+            if (TCG_TARGET_HAS_ext8u_i64) {
+                tcg_gen_ext8u_i64(ret, arg);
+                tcg_gen_shli_i64(ret, ret, ofs);
+                return;
+            }
+            break;
+        }
+        /* Otherwise prefer zero-extension over AND for code size.  */
+        switch (ofs + len) {
+        case 32:
+            if (TCG_TARGET_HAS_ext32u_i64) {
+                tcg_gen_shli_i64(ret, arg, ofs);
+                tcg_gen_ext32u_i64(ret, ret);
+                return;
+            }
+            break;
+        case 16:
+            if (TCG_TARGET_HAS_ext16u_i64) {
+                tcg_gen_shli_i64(ret, arg, ofs);
+                tcg_gen_ext16u_i64(ret, ret);
+                return;
+            }
+            break;
+        case 8:
+            if (TCG_TARGET_HAS_ext8u_i64) {
+                tcg_gen_shli_i64(ret, arg, ofs);
+                tcg_gen_ext8u_i64(ret, ret);
+                return;
+            }
+            break;
+        }
+        tcg_gen_andi_i64(ret, arg, (1ull << len) - 1);
+        tcg_gen_shli_i64(ret, ret, ofs);
+    }
+}
+
+void tcg_gen_extract_i64(TCGv_i64 ret, TCGv_i64 arg,
+                         unsigned int ofs, unsigned int len)
+{
+    tcg_debug_assert(ofs < 64);
+    tcg_debug_assert(len > 0);
+    tcg_debug_assert(len <= 64);
+    tcg_debug_assert(ofs + len <= 64);
+
+    /* Canonicalize certain special cases, even if extract is supported.  */
+    if (ofs + len == 64) {
+        tcg_gen_shri_i64(ret, arg, 64 - len);
+        return;
+    }
+    if (ofs == 0) {
+        tcg_gen_andi_i64(ret, arg, (1ull << len) - 1);
+        return;
+    }
+
+    if (TCG_TARGET_REG_BITS == 32) {
+        /* Look for a 32-bit extract within one of the two words.  */
+        if (ofs >= 32) {
+            tcg_gen_extract_i32(TCGV_LOW(ret), TCGV_HIGH(arg), ofs - 32, len);
+            tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
+            return;
+        }
+        if (ofs + len <= 32) {
+            tcg_gen_extract_i32(TCGV_LOW(ret), TCGV_LOW(arg), ofs, len);
+            tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
+            return;
+        }
+        /* The field is split across two words.  One double-word
+           shift is better than two double-word shifts.  */
+        goto do_shift_and;
+    }
+
+    if (TCG_TARGET_HAS_extract_i64
+        && TCG_TARGET_extract_i64_valid(ofs, len)) {
+        tcg_gen_op4ii_i64(INDEX_op_extract_i64, ret, arg, ofs, len);
+        return;
+    }
+
+    /* Assume that zero-extension, if available, is cheaper than a shift.  */
+    switch (ofs + len) {
+    case 32:
+        if (TCG_TARGET_HAS_ext32u_i64) {
+            tcg_gen_ext32u_i64(ret, arg);
+            tcg_gen_shri_i64(ret, ret, ofs);
+            return;
+        }
+        break;
+    case 16:
+        if (TCG_TARGET_HAS_ext16u_i64) {
+            tcg_gen_ext16u_i64(ret, arg);
+            tcg_gen_shri_i64(ret, ret, ofs);
+            return;
+        }
+        break;
+    case 8:
+        if (TCG_TARGET_HAS_ext8u_i64) {
+            tcg_gen_ext8u_i64(ret, arg);
+            tcg_gen_shri_i64(ret, ret, ofs);
+            return;
+        }
+        break;
+    }
+
+    /* ??? Ideally we'd know what values are available for immediate AND.
+       Assume that 8 bits are available, plus the special cases of 16 and 32,
+       so that we get ext8u, ext16u, and ext32u.  */
+    switch (len) {
+    case 1 ... 8: case 16: case 32:
+    do_shift_and:
+        tcg_gen_shri_i64(ret, arg, ofs);
+        tcg_gen_andi_i64(ret, ret, (1ull << len) - 1);
+        break;
+    default:
+        tcg_gen_shli_i64(ret, arg, 64 - len - ofs);
+        tcg_gen_shri_i64(ret, ret, 64 - len);
+        break;
+    }
+}
+
+void tcg_gen_sextract_i64(TCGv_i64 ret, TCGv_i64 arg,
+                          unsigned int ofs, unsigned int len)
+{
+    tcg_debug_assert(ofs < 64);
+    tcg_debug_assert(len > 0);
+    tcg_debug_assert(len <= 64);
+    tcg_debug_assert(ofs + len <= 64);
+
+    /* Canonicalize certain special cases, even if sextract is supported.  */
+    if (ofs + len == 64) {
+        tcg_gen_sari_i64(ret, arg, 64 - len);
+        return;
+    }
+    if (ofs == 0) {
+        switch (len) {
+        case 32:
+            tcg_gen_ext32s_i64(ret, arg);
+            return;
+        case 16:
+            tcg_gen_ext16s_i64(ret, arg);
+            return;
+        case 8:
+            tcg_gen_ext8s_i64(ret, arg);
+            return;
+        }
+    }
+
+    if (TCG_TARGET_REG_BITS == 32) {
+        /* Look for a 32-bit extract within one of the two words.  */
+        if (ofs >= 32) {
+            tcg_gen_sextract_i32(TCGV_LOW(ret), TCGV_HIGH(arg), ofs - 32, len);
+        } else if (ofs + len <= 32) {
+            tcg_gen_sextract_i32(TCGV_LOW(ret), TCGV_LOW(arg), ofs, len);
+        } else if (ofs == 0) {
+            tcg_gen_mov_i32(TCGV_LOW(ret), TCGV_LOW(arg));
+            tcg_gen_sextract_i32(TCGV_HIGH(ret), TCGV_HIGH(arg), 0, len - 32);
+            return;
+        } else if (len > 32) {
+            TCGv_i32 t = tcg_temp_new_i32();
+            /* Extract the bits for the high word normally.  */
+            tcg_gen_sextract_i32(t, TCGV_HIGH(arg), ofs + 32, len - 32);
+            /* Shift the field down for the low part.  */
+            tcg_gen_shri_i64(ret, arg, ofs);
+            /* Overwrite the shift into the high part.  */
+            tcg_gen_mov_i32(TCGV_HIGH(ret), t);
+            tcg_temp_free_i32(t);
+            return;
+        } else {
+            /* Shift the field down for the low part, such that the
+               field sits at the MSB.  */
+            tcg_gen_shri_i64(ret, arg, ofs + len - 32);
+            /* Shift the field down from the MSB, sign extending.  */
+            tcg_gen_sari_i32(TCGV_LOW(ret), TCGV_LOW(ret), 32 - len);
+        }
+        /* Sign-extend the field from 32 bits.  */
+        tcg_gen_sari_i32(TCGV_HIGH(ret), TCGV_LOW(ret), 31);
+        return;
+    }
+
+    if (TCG_TARGET_HAS_sextract_i64
+        && TCG_TARGET_extract_i64_valid(ofs, len)) {
+        tcg_gen_op4ii_i64(INDEX_op_sextract_i64, ret, arg, ofs, len);
+        return;
+    }
+
+    /* Assume that sign-extension, if available, is cheaper than a shift.  */
+    switch (ofs + len) {
+    case 32:
+        if (TCG_TARGET_HAS_ext32s_i64) {
+            tcg_gen_ext32s_i64(ret, arg);
+            tcg_gen_sari_i64(ret, ret, ofs);
+            return;
+        }
+        break;
+    case 16:
+        if (TCG_TARGET_HAS_ext16s_i64) {
+            tcg_gen_ext16s_i64(ret, arg);
+            tcg_gen_sari_i64(ret, ret, ofs);
+            return;
+        }
+        break;
+    case 8:
+        if (TCG_TARGET_HAS_ext8s_i64) {
+            tcg_gen_ext8s_i64(ret, arg);
+            tcg_gen_sari_i64(ret, ret, ofs);
+            return;
+        }
+        break;
+    }
+    switch (len) {
+    case 32:
+        if (TCG_TARGET_HAS_ext32s_i64) {
+            tcg_gen_shri_i64(ret, arg, ofs);
+            tcg_gen_ext32s_i64(ret, ret);
+            return;
+        }
+        break;
+    case 16:
+        if (TCG_TARGET_HAS_ext16s_i64) {
+            tcg_gen_shri_i64(ret, arg, ofs);
+            tcg_gen_ext16s_i64(ret, ret);
+            return;
+        }
+        break;
+    case 8:
+        if (TCG_TARGET_HAS_ext8s_i64) {
+            tcg_gen_shri_i64(ret, arg, ofs);
+            tcg_gen_ext8s_i64(ret, ret);
+            return;
+        }
+        break;
+    }
+    tcg_gen_shli_i64(ret, arg, 64 - len - ofs);
+    tcg_gen_sari_i64(ret, ret, 64 - len);
+}
+
 void tcg_gen_movcond_i64(TCGCond cond, TCGv_i64 ret, TCGv_i64 c1,
                          TCGv_i64 c2, TCGv_i64 v1, TCGv_i64 v2)
 {
@@ -1748,6 +2243,22 @@ void tcg_gen_muls2_i64(TCGv_i64 rl, TCGv_i64 rh, TCGv_i64 arg1, TCGv_i64 arg2)
     }
 }
 
+void tcg_gen_mulsu2_i64(TCGv_i64 rl, TCGv_i64 rh, TCGv_i64 arg1, TCGv_i64 arg2)
+{
+    TCGv_i64 t0 = tcg_temp_new_i64();
+    TCGv_i64 t1 = tcg_temp_new_i64();
+    TCGv_i64 t2 = tcg_temp_new_i64();
+    tcg_gen_mulu2_i64(t0, t1, arg1, arg2);
+    /* Adjust for negative input for the signed arg1.  */
+    tcg_gen_sari_i64(t2, arg1, 63);
+    tcg_gen_and_i64(t2, t2, arg2);
+    tcg_gen_sub_i64(rh, t1, t2);
+    tcg_gen_mov_i64(rl, t0);
+    tcg_temp_free_i64(t0);
+    tcg_temp_free_i64(t1);
+    tcg_temp_free_i64(t2);
+}
+
 /* Size changing operations.  */
 
 void tcg_gen_extrl_i64_i32(TCGv_i32 ret, TCGv_i64 arg)
This page took 0.041611 seconds and 4 git commands to generate.