tcg: Add generic vector expanders

Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>

13 files changed, 2024 insertions, 18 deletions

diff --git a/Makefile.target b/Makefile.target
index 7f30a1e725..6549481096 100644
--- a/Makefile.target
+++ b/Makefile.target
@@ -93,7 +93,7 @@ all: $(PROGS) stap
 # cpu emulator library
 obj-y += exec.o
 obj-y += accel/
-obj-$(CONFIG_TCG) += tcg/tcg.o tcg/tcg-op.o tcg/tcg-op-vec.o
+obj-$(CONFIG_TCG) += tcg/tcg.o tcg/tcg-op.o tcg/tcg-op-vec.o tcg/tcg-op-gvec.o
 obj-$(CONFIG_TCG) += tcg/tcg-common.o tcg/optimize.o
 obj-$(CONFIG_TCG_INTERPRETER) += tcg/tci.o
 obj-$(CONFIG_TCG_INTERPRETER) += disas/tci.o
diff --git a/accel/tcg/Makefile.objs b/accel/tcg/Makefile.objs
index 228cd84fa4..d381a02f34 100644
--- a/accel/tcg/Makefile.objs
+++ b/accel/tcg/Makefile.objs
@@ -1,6 +1,6 @@
 obj-$(CONFIG_SOFTMMU) += tcg-all.o
 obj-$(CONFIG_SOFTMMU) += cputlb.o
-obj-y += tcg-runtime.o
+obj-y += tcg-runtime.o tcg-runtime-gvec.o
 obj-y += cpu-exec.o cpu-exec-common.o translate-all.o
 obj-y += translator.o
 
diff --git a/accel/tcg/tcg-runtime-gvec.c b/accel/tcg/tcg-runtime-gvec.c
new file mode 100644
index 0000000000..e093922225
--- /dev/null
+++ b/accel/tcg/tcg-runtime-gvec.c
@@ -0,0 +1,325 @@
+/*
+ * Generic vectorized operation runtime
+ *
+ * Copyright (c) 2018 Linaro
+ *
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2 of the License, or (at your option) any later version.
+ *
+ * This library is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with this library; if not, see <http://www.gnu.org/licenses/>.
+ */
+
+#include "qemu/osdep.h"
+#include "qemu/host-utils.h"
+#include "cpu.h"
+#include "exec/helper-proto.h"
+#include "tcg-gvec-desc.h"
+
+
+/* Virtually all hosts support 16-byte vectors.  Those that don't can emulate
+ * them via GCC's generic vector extension.  This turns out to be simpler and
+ * more reliable than getting the compiler to autovectorize.
+ *
+ * In tcg-op-gvec.c, we asserted that both the size and alignment of the data
+ * are multiples of 16.
+ *
+ * When the compiler does not support all of the operations we require, the
+ * loops are written so that we can always fall back on the base types.
+ */
+#ifdef CONFIG_VECTOR16
+typedef uint8_t vec8 __attribute__((vector_size(16)));
+typedef uint16_t vec16 __attribute__((vector_size(16)));
+typedef uint32_t vec32 __attribute__((vector_size(16)));
+typedef uint64_t vec64 __attribute__((vector_size(16)));
+
+typedef int8_t svec8 __attribute__((vector_size(16)));
+typedef int16_t svec16 __attribute__((vector_size(16)));
+typedef int32_t svec32 __attribute__((vector_size(16)));
+typedef int64_t svec64 __attribute__((vector_size(16)));
+
+#define DUP16(X)  { X, X, X, X, X, X, X, X, X, X, X, X, X, X, X, X }
+#define DUP8(X)   { X, X, X, X, X, X, X, X }
+#define DUP4(X)   { X, X, X, X }
+#define DUP2(X)   { X, X }
+#else
+typedef uint8_t vec8;
+typedef uint16_t vec16;
+typedef uint32_t vec32;
+typedef uint64_t vec64;
+
+typedef int8_t svec8;
+typedef int16_t svec16;
+typedef int32_t svec32;
+typedef int64_t svec64;
+
+#define DUP16(X)  X
+#define DUP8(X)   X
+#define DUP4(X)   X
+#define DUP2(X)   X
+#endif /* CONFIG_VECTOR16 */
+
+static inline void clear_high(void *d, intptr_t oprsz, uint32_t desc)
+{
+    intptr_t maxsz = simd_maxsz(desc);
+    intptr_t i;
+
+    if (unlikely(maxsz > oprsz)) {
+        for (i = oprsz; i < maxsz; i += sizeof(uint64_t)) {
+            *(uint64_t *)(d + i) = 0;
+        }
+    }
+}
+
+void HELPER(gvec_add8)(void *d, void *a, void *b, uint32_t desc)
+{
+    intptr_t oprsz = simd_oprsz(desc);
+    intptr_t i;
+
+    for (i = 0; i < oprsz; i += sizeof(vec8)) {
+        *(vec8 *)(d + i) = *(vec8 *)(a + i) + *(vec8 *)(b + i);
+    }
+    clear_high(d, oprsz, desc);
+}
+
+void HELPER(gvec_add16)(void *d, void *a, void *b, uint32_t desc)
+{
+    intptr_t oprsz = simd_oprsz(desc);
+    intptr_t i;
+
+    for (i = 0; i < oprsz; i += sizeof(vec16)) {
+        *(vec16 *)(d + i) = *(vec16 *)(a + i) + *(vec16 *)(b + i);
+    }
+    clear_high(d, oprsz, desc);
+}
+
+void HELPER(gvec_add32)(void *d, void *a, void *b, uint32_t desc)
+{
+    intptr_t oprsz = simd_oprsz(desc);
+    intptr_t i;
+
+    for (i = 0; i < oprsz; i += sizeof(vec32)) {
+        *(vec32 *)(d + i) = *(vec32 *)(a + i) + *(vec32 *)(b + i);
+    }
+    clear_high(d, oprsz, desc);
+}
+
+void HELPER(gvec_add64)(void *d, void *a, void *b, uint32_t desc)
+{
+    intptr_t oprsz = simd_oprsz(desc);
+    intptr_t i;
+
+    for (i = 0; i < oprsz; i += sizeof(vec64)) {
+        *(vec64 *)(d + i) = *(vec64 *)(a + i) + *(vec64 *)(b + i);
+    }
+    clear_high(d, oprsz, desc);
+}
+
+void HELPER(gvec_sub8)(void *d, void *a, void *b, uint32_t desc)
+{
+    intptr_t oprsz = simd_oprsz(desc);
+    intptr_t i;
+
+    for (i = 0; i < oprsz; i += sizeof(vec8)) {
+        *(vec8 *)(d + i) = *(vec8 *)(a + i) - *(vec8 *)(b + i);
+    }
+    clear_high(d, oprsz, desc);
+}
+
+void HELPER(gvec_sub16)(void *d, void *a, void *b, uint32_t desc)
+{
+    intptr_t oprsz = simd_oprsz(desc);
+    intptr_t i;
+
+    for (i = 0; i < oprsz; i += sizeof(vec16)) {
+        *(vec16 *)(d + i) = *(vec16 *)(a + i) - *(vec16 *)(b + i);
+    }
+    clear_high(d, oprsz, desc);
+}
+
+void HELPER(gvec_sub32)(void *d, void *a, void *b, uint32_t desc)
+{
+    intptr_t oprsz = simd_oprsz(desc);
+    intptr_t i;
+
+    for (i = 0; i < oprsz; i += sizeof(vec32)) {
+        *(vec32 *)(d + i) = *(vec32 *)(a + i) - *(vec32 *)(b + i);
+    }
+    clear_high(d, oprsz, desc);
+}
+
+void HELPER(gvec_sub64)(void *d, void *a, void *b, uint32_t desc)
+{
+    intptr_t oprsz = simd_oprsz(desc);
+    intptr_t i;
+
+    for (i = 0; i < oprsz; i += sizeof(vec64)) {
+        *(vec64 *)(d + i) = *(vec64 *)(a + i) - *(vec64 *)(b + i);
+    }
+    clear_high(d, oprsz, desc);
+}
+
+void HELPER(gvec_neg8)(void *d, void *a, uint32_t desc)
+{
+    intptr_t oprsz = simd_oprsz(desc);
+    intptr_t i;
+
+    for (i = 0; i < oprsz; i += sizeof(vec8)) {
+        *(vec8 *)(d + i) = -*(vec8 *)(a + i);
+    }
+    clear_high(d, oprsz, desc);
+}
+
+void HELPER(gvec_neg16)(void *d, void *a, uint32_t desc)
+{
+    intptr_t oprsz = simd_oprsz(desc);
+    intptr_t i;
+
+    for (i = 0; i < oprsz; i += sizeof(vec16)) {
+        *(vec16 *)(d + i) = -*(vec16 *)(a + i);
+    }
+    clear_high(d, oprsz, desc);
+}
+
+void HELPER(gvec_neg32)(void *d, void *a, uint32_t desc)
+{
+    intptr_t oprsz = simd_oprsz(desc);
+    intptr_t i;
+
+    for (i = 0; i < oprsz; i += sizeof(vec32)) {
+        *(vec32 *)(d + i) = -*(vec32 *)(a + i);
+    }
+    clear_high(d, oprsz, desc);
+}
+
+void HELPER(gvec_neg64)(void *d, void *a, uint32_t desc)
+{
+    intptr_t oprsz = simd_oprsz(desc);
+    intptr_t i;
+
+    for (i = 0; i < oprsz; i += sizeof(vec64)) {
+        *(vec64 *)(d + i) = -*(vec64 *)(a + i);
+    }
+    clear_high(d, oprsz, desc);
+}
+
+void HELPER(gvec_mov)(void *d, void *a, uint32_t desc)
+{
+    intptr_t oprsz = simd_oprsz(desc);
+
+    memcpy(d, a, oprsz);
+    clear_high(d, oprsz, desc);
+}
+
+void HELPER(gvec_dup64)(void *d, uint32_t desc, uint64_t c)
+{
+    intptr_t oprsz = simd_oprsz(desc);
+    intptr_t i;
+
+    if (c == 0) {
+        oprsz = 0;
+    } else {
+        for (i = 0; i < oprsz; i += sizeof(uint64_t)) {
+            *(uint64_t *)(d + i) = c;
+        }
+    }
+    clear_high(d, oprsz, desc);
+}
+
+void HELPER(gvec_dup32)(void *d, uint32_t desc, uint32_t c)
+{
+    intptr_t oprsz = simd_oprsz(desc);
+    intptr_t i;
+
+    if (c == 0) {
+        oprsz = 0;
+    } else {
+        for (i = 0; i < oprsz; i += sizeof(uint32_t)) {
+            *(uint32_t *)(d + i) = c;
+        }
+    }
+    clear_high(d, oprsz, desc);
+}
+
+void HELPER(gvec_dup16)(void *d, uint32_t desc, uint32_t c)
+{
+    HELPER(gvec_dup32)(d, desc, 0x00010001 * (c & 0xffff));
+}
+
+void HELPER(gvec_dup8)(void *d, uint32_t desc, uint32_t c)
+{
+    HELPER(gvec_dup32)(d, desc, 0x01010101 * (c & 0xff));
+}
+
+void HELPER(gvec_not)(void *d, void *a, uint32_t desc)
+{
+    intptr_t oprsz = simd_oprsz(desc);
+    intptr_t i;
+
+    for (i = 0; i < oprsz; i += sizeof(vec64)) {
+        *(vec64 *)(d + i) = ~*(vec64 *)(a + i);
+    }
+    clear_high(d, oprsz, desc);
+}
+
+void HELPER(gvec_and)(void *d, void *a, void *b, uint32_t desc)
+{
+    intptr_t oprsz = simd_oprsz(desc);
+    intptr_t i;
+
+    for (i = 0; i < oprsz; i += sizeof(vec64)) {
+        *(vec64 *)(d + i) = *(vec64 *)(a + i) & *(vec64 *)(b + i);
+    }
+    clear_high(d, oprsz, desc);
+}
+
+void HELPER(gvec_or)(void *d, void *a, void *b, uint32_t desc)
+{
+    intptr_t oprsz = simd_oprsz(desc);
+    intptr_t i;
+
+    for (i = 0; i < oprsz; i += sizeof(vec64)) {
+        *(vec64 *)(d + i) = *(vec64 *)(a + i) | *(vec64 *)(b + i);
+    }
+    clear_high(d, oprsz, desc);
+}
+
+void HELPER(gvec_xor)(void *d, void *a, void *b, uint32_t desc)
+{
+    intptr_t oprsz = simd_oprsz(desc);
+    intptr_t i;
+
+    for (i = 0; i < oprsz; i += sizeof(vec64)) {
+        *(vec64 *)(d + i) = *(vec64 *)(a + i) ^ *(vec64 *)(b + i);
+    }
+    clear_high(d, oprsz, desc);
+}
+
+void HELPER(gvec_andc)(void *d, void *a, void *b, uint32_t desc)
+{
+    intptr_t oprsz = simd_oprsz(desc);
+    intptr_t i;
+
+    for (i = 0; i < oprsz; i += sizeof(vec64)) {
+        *(vec64 *)(d + i) = *(vec64 *)(a + i) &~ *(vec64 *)(b + i);
+    }
+    clear_high(d, oprsz, desc);
+}
+
+void HELPER(gvec_orc)(void *d, void *a, void *b, uint32_t desc)
+{
+    intptr_t oprsz = simd_oprsz(desc);
+    intptr_t i;
+
+    for (i = 0; i < oprsz; i += sizeof(vec64)) {
+        *(vec64 *)(d + i) = *(vec64 *)(a + i) |~ *(vec64 *)(b + i);
+    }
+    clear_high(d, oprsz, desc);
+}
diff --git a/accel/tcg/tcg-runtime.h b/accel/tcg/tcg-runtime.h
index 1df17d0ba9..76ee41ce58 100644
--- a/accel/tcg/tcg-runtime.h
+++ b/accel/tcg/tcg-runtime.h
@@ -134,3 +134,32 @@ GEN_ATOMIC_HELPERS(xor_fetch)
 GEN_ATOMIC_HELPERS(xchg)
 
 #undef GEN_ATOMIC_HELPERS
+
+DEF_HELPER_FLAGS_3(gvec_mov, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+
+DEF_HELPER_FLAGS_3(gvec_dup8, TCG_CALL_NO_RWG, void, ptr, i32, i32)
+DEF_HELPER_FLAGS_3(gvec_dup16, TCG_CALL_NO_RWG, void, ptr, i32, i32)
+DEF_HELPER_FLAGS_3(gvec_dup32, TCG_CALL_NO_RWG, void, ptr, i32, i32)
+DEF_HELPER_FLAGS_3(gvec_dup64, TCG_CALL_NO_RWG, void, ptr, i32, i64)
+
+DEF_HELPER_FLAGS_4(gvec_add8, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_add16, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_add32, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_add64, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+
+DEF_HELPER_FLAGS_4(gvec_sub8, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_sub16, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_sub32, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_sub64, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+
+DEF_HELPER_FLAGS_3(gvec_neg8, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+DEF_HELPER_FLAGS_3(gvec_neg16, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+DEF_HELPER_FLAGS_3(gvec_neg32, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+DEF_HELPER_FLAGS_3(gvec_neg64, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+
+DEF_HELPER_FLAGS_3(gvec_not, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_and, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_or, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_xor, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_andc, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_orc, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
diff --git a/configure b/configure
index 831ebf248f..00695c7ea7 100755
--- a/configure
+++ b/configure
@@ -5001,6 +5001,50 @@ if compile_prog "" "" ; then
 fi
 
 ########################################
+# See if 16-byte vector operations are supported.
+# Even without a vector unit the compiler may expand these.
+# There is a bug in old GCC for PPC that crashes here.
+# Unfortunately it's the system compiler for Centos 7.
+
+cat > $TMPC << EOF
+typedef unsigned char U1 __attribute__((vector_size(16)));
+typedef unsigned short U2 __attribute__((vector_size(16)));
+typedef unsigned int U4 __attribute__((vector_size(16)));
+typedef unsigned long long U8 __attribute__((vector_size(16)));
+typedef signed char S1 __attribute__((vector_size(16)));
+typedef signed short S2 __attribute__((vector_size(16)));
+typedef signed int S4 __attribute__((vector_size(16)));
+typedef signed long long S8 __attribute__((vector_size(16)));
+static U1 a1, b1;
+static U2 a2, b2;
+static U4 a4, b4;
+static U8 a8, b8;
+static S1 c1;
+static S2 c2;
+static S4 c4;
+static S8 c8;
+static int i;
+int main(void)
+{
+  a1 += b1; a2 += b2; a4 += b4; a8 += b8;
+  a1 -= b1; a2 -= b2; a4 -= b4; a8 -= b8;
+  a1 *= b1; a2 *= b2; a4 *= b4; a8 *= b8;
+  a1 &= b1; a2 &= b2; a4 &= b4; a8 &= b8;
+  a1 |= b1; a2 |= b2; a4 |= b4; a8 |= b8;
+  a1 ^= b1; a2 ^= b2; a4 ^= b4; a8 ^= b8;
+  a1 <<= i; a2 <<= i; a4 <<= i; a8 <<= i;
+  a1 >>= i; a2 >>= i; a4 >>= i; a8 >>= i;
+  c1 >>= i; c2 >>= i; c4 >>= i; c8 >>= i;
+  return 0;
+}
+EOF
+
+vector16=no
+if compile_prog "" "" ; then
+  vector16=yes
+fi
+
+########################################
 # check if getauxval is available.
 
 getauxval=no
@@ -6329,6 +6373,10 @@ if test "$atomic64" = "yes" ; then
   echo "CONFIG_ATOMIC64=y" >> $config_host_mak
 fi
 
+if test "$vector16" = "yes" ; then
+  echo "CONFIG_VECTOR16=y" >> $config_host_mak
+fi
+
 if test "$getauxval" = "yes" ; then
   echo "CONFIG_GETAUXVAL=y" >> $config_host_mak
 fi
diff --git a/tcg/tcg-gvec-desc.h b/tcg/tcg-gvec-desc.h
new file mode 100644
index 0000000000..3b4c2d9c69
--- /dev/null
+++ b/tcg/tcg-gvec-desc.h
@@ -0,0 +1,49 @@
+/*
+ * Generic vector operation descriptor
+ *
+ * Copyright (c) 2018 Linaro
+ *
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2 of the License, or (at your option) any later version.
+ *
+ * This library is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with this library; if not, see <http://www.gnu.org/licenses/>.
+ */
+
+/* ??? These bit widths are set for ARM SVE, maxing out at 256 byte vectors. */
+#define SIMD_OPRSZ_SHIFT   0
+#define SIMD_OPRSZ_BITS    5
+
+#define SIMD_MAXSZ_SHIFT   (SIMD_OPRSZ_SHIFT + SIMD_OPRSZ_BITS)
+#define SIMD_MAXSZ_BITS    5
+
+#define SIMD_DATA_SHIFT    (SIMD_MAXSZ_SHIFT + SIMD_MAXSZ_BITS)
+#define SIMD_DATA_BITS     (32 - SIMD_DATA_SHIFT)
+
+/* Create a descriptor from components.  */
+uint32_t simd_desc(uint32_t oprsz, uint32_t maxsz, int32_t data);
+
+/* Extract the operation size from a descriptor.  */
+static inline intptr_t simd_oprsz(uint32_t desc)
+{
+    return (extract32(desc, SIMD_OPRSZ_SHIFT, SIMD_OPRSZ_BITS) + 1) * 8;
+}
+
+/* Extract the max vector size from a descriptor.  */
+static inline intptr_t simd_maxsz(uint32_t desc)
+{
+    return (extract32(desc, SIMD_MAXSZ_SHIFT, SIMD_MAXSZ_BITS) + 1) * 8;
+}
+
+/* Extract the operation-specific data from a descriptor.  */
+static inline int32_t simd_data(uint32_t desc)
+{
+    return sextract32(desc, SIMD_DATA_SHIFT, SIMD_DATA_BITS);
+}
diff --git a/tcg/tcg-op-gvec.c b/tcg/tcg-op-gvec.c
new file mode 100644
index 0000000000..f8ae75b331
--- /dev/null
+++ b/tcg/tcg-op-gvec.c
@@ -0,0 +1,1309 @@
+/*
+ * Generic vector operation expansion
+ *
+ * Copyright (c) 2018 Linaro
+ *
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2 of the License, or (at your option) any later version.
+ *
+ * This library is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with this library; if not, see <http://www.gnu.org/licenses/>.
+ */
+
+#include "qemu/osdep.h"
+#include "qemu-common.h"
+#include "tcg.h"
+#include "tcg-op.h"
+#include "tcg-op-gvec.h"
+#include "tcg-gvec-desc.h"
+
+#define MAX_UNROLL  4
+
+/* Verify vector size and alignment rules.  OFS should be the OR of all
+   of the operand offsets so that we can check them all at once.  */
+static void check_size_align(uint32_t oprsz, uint32_t maxsz, uint32_t ofs)
+{
+    uint32_t opr_align = oprsz >= 16 ? 15 : 7;
+    uint32_t max_align = maxsz >= 16 || oprsz >= 16 ? 15 : 7;
+    tcg_debug_assert(oprsz > 0);
+    tcg_debug_assert(oprsz <= maxsz);
+    tcg_debug_assert((oprsz & opr_align) == 0);
+    tcg_debug_assert((maxsz & max_align) == 0);
+    tcg_debug_assert((ofs & max_align) == 0);
+}
+
+/* Verify vector overlap rules for two operands.  */
+static void check_overlap_2(uint32_t d, uint32_t a, uint32_t s)
+{
+    tcg_debug_assert(d == a || d + s <= a || a + s <= d);
+}
+
+/* Verify vector overlap rules for three operands.  */
+static void check_overlap_3(uint32_t d, uint32_t a, uint32_t b, uint32_t s)
+{
+    check_overlap_2(d, a, s);
+    check_overlap_2(d, b, s);
+    check_overlap_2(a, b, s);
+}
+
+/* Verify vector overlap rules for four operands.  */
+static void check_overlap_4(uint32_t d, uint32_t a, uint32_t b,
+                            uint32_t c, uint32_t s)
+{
+    check_overlap_2(d, a, s);
+    check_overlap_2(d, b, s);
+    check_overlap_2(d, c, s);
+    check_overlap_2(a, b, s);
+    check_overlap_2(a, c, s);
+    check_overlap_2(b, c, s);
+}
+
+/* Create a descriptor from components.  */
+uint32_t simd_desc(uint32_t oprsz, uint32_t maxsz, int32_t data)
+{
+    uint32_t desc = 0;
+
+    assert(oprsz % 8 == 0 && oprsz <= (8 << SIMD_OPRSZ_BITS));
+    assert(maxsz % 8 == 0 && maxsz <= (8 << SIMD_MAXSZ_BITS));
+    assert(data == sextract32(data, 0, SIMD_DATA_BITS));
+
+    oprsz = (oprsz / 8) - 1;
+    maxsz = (maxsz / 8) - 1;
+    desc = deposit32(desc, SIMD_OPRSZ_SHIFT, SIMD_OPRSZ_BITS, oprsz);
+    desc = deposit32(desc, SIMD_MAXSZ_SHIFT, SIMD_MAXSZ_BITS, maxsz);
+    desc = deposit32(desc, SIMD_DATA_SHIFT, SIMD_DATA_BITS, data);
+
+    return desc;
+}
+
+/* Generate a call to a gvec-style helper with two vector operands.  */
+void tcg_gen_gvec_2_ool(uint32_t dofs, uint32_t aofs,
+                        uint32_t oprsz, uint32_t maxsz, int32_t data,
+                        gen_helper_gvec_2 *fn)
+{
+    TCGv_ptr a0, a1;
+    TCGv_i32 desc = tcg_const_i32(simd_desc(oprsz, maxsz, data));
+
+    a0 = tcg_temp_new_ptr();
+    a1 = tcg_temp_new_ptr();
+
+    tcg_gen_addi_ptr(a0, cpu_env, dofs);
+    tcg_gen_addi_ptr(a1, cpu_env, aofs);
+
+    fn(a0, a1, desc);
+
+    tcg_temp_free_ptr(a0);
+    tcg_temp_free_ptr(a1);
+    tcg_temp_free_i32(desc);
+}
+
+/* Generate a call to a gvec-style helper with three vector operands.  */
+void tcg_gen_gvec_3_ool(uint32_t dofs, uint32_t aofs, uint32_t bofs,
+                        uint32_t oprsz, uint32_t maxsz, int32_t data,
+                        gen_helper_gvec_3 *fn)
+{
+    TCGv_ptr a0, a1, a2;
+    TCGv_i32 desc = tcg_const_i32(simd_desc(oprsz, maxsz, data));
+
+    a0 = tcg_temp_new_ptr();
+    a1 = tcg_temp_new_ptr();
+    a2 = tcg_temp_new_ptr();
+
+    tcg_gen_addi_ptr(a0, cpu_env, dofs);
+    tcg_gen_addi_ptr(a1, cpu_env, aofs);
+    tcg_gen_addi_ptr(a2, cpu_env, bofs);
+
+    fn(a0, a1, a2, desc);
+
+    tcg_temp_free_ptr(a0);
+    tcg_temp_free_ptr(a1);
+    tcg_temp_free_ptr(a2);
+    tcg_temp_free_i32(desc);
+}
+
+/* Generate a call to a gvec-style helper with four vector operands.  */
+void tcg_gen_gvec_4_ool(uint32_t dofs, uint32_t aofs, uint32_t bofs,
+                        uint32_t cofs, uint32_t oprsz, uint32_t maxsz,
+                        int32_t data, gen_helper_gvec_4 *fn)
+{
+    TCGv_ptr a0, a1, a2, a3;
+    TCGv_i32 desc = tcg_const_i32(simd_desc(oprsz, maxsz, data));
+
+    a0 = tcg_temp_new_ptr();
+    a1 = tcg_temp_new_ptr();
+    a2 = tcg_temp_new_ptr();
+    a3 = tcg_temp_new_ptr();
+
+    tcg_gen_addi_ptr(a0, cpu_env, dofs);
+    tcg_gen_addi_ptr(a1, cpu_env, aofs);
+    tcg_gen_addi_ptr(a2, cpu_env, bofs);
+    tcg_gen_addi_ptr(a3, cpu_env, cofs);
+
+    fn(a0, a1, a2, a3, desc);
+
+    tcg_temp_free_ptr(a0);
+    tcg_temp_free_ptr(a1);
+    tcg_temp_free_ptr(a2);
+    tcg_temp_free_ptr(a3);
+    tcg_temp_free_i32(desc);
+}
+
+/* Generate a call to a gvec-style helper with five vector operands.  */
+void tcg_gen_gvec_5_ool(uint32_t dofs, uint32_t aofs, uint32_t bofs,
+                        uint32_t cofs, uint32_t xofs, uint32_t oprsz,
+                        uint32_t maxsz, int32_t data, gen_helper_gvec_5 *fn)
+{
+    TCGv_ptr a0, a1, a2, a3, a4;
+    TCGv_i32 desc = tcg_const_i32(simd_desc(oprsz, maxsz, data));
+
+    a0 = tcg_temp_new_ptr();
+    a1 = tcg_temp_new_ptr();
+    a2 = tcg_temp_new_ptr();
+    a3 = tcg_temp_new_ptr();
+    a4 = tcg_temp_new_ptr();
+
+    tcg_gen_addi_ptr(a0, cpu_env, dofs);
+    tcg_gen_addi_ptr(a1, cpu_env, aofs);
+    tcg_gen_addi_ptr(a2, cpu_env, bofs);
+    tcg_gen_addi_ptr(a3, cpu_env, cofs);
+    tcg_gen_addi_ptr(a4, cpu_env, xofs);
+
+    fn(a0, a1, a2, a3, a4, desc);
+
+    tcg_temp_free_ptr(a0);
+    tcg_temp_free_ptr(a1);
+    tcg_temp_free_ptr(a2);
+    tcg_temp_free_ptr(a3);
+    tcg_temp_free_ptr(a4);
+    tcg_temp_free_i32(desc);
+}
+
+/* Generate a call to a gvec-style helper with three vector operands
+   and an extra pointer operand.  */
+void tcg_gen_gvec_2_ptr(uint32_t dofs, uint32_t aofs,
+                        TCGv_ptr ptr, uint32_t oprsz, uint32_t maxsz,
+                        int32_t data, gen_helper_gvec_2_ptr *fn)
+{
+    TCGv_ptr a0, a1;
+    TCGv_i32 desc = tcg_const_i32(simd_desc(oprsz, maxsz, data));
+
+    a0 = tcg_temp_new_ptr();
+    a1 = tcg_temp_new_ptr();
+
+    tcg_gen_addi_ptr(a0, cpu_env, dofs);
+    tcg_gen_addi_ptr(a1, cpu_env, aofs);
+
+    fn(a0, a1, ptr, desc);
+
+    tcg_temp_free_ptr(a0);
+    tcg_temp_free_ptr(a1);
+    tcg_temp_free_i32(desc);
+}
+
+/* Generate a call to a gvec-style helper with three vector operands
+   and an extra pointer operand.  */
+void tcg_gen_gvec_3_ptr(uint32_t dofs, uint32_t aofs, uint32_t bofs,
+                        TCGv_ptr ptr, uint32_t oprsz, uint32_t maxsz,
+                        int32_t data, gen_helper_gvec_3_ptr *fn)
+{
+    TCGv_ptr a0, a1, a2;
+    TCGv_i32 desc = tcg_const_i32(simd_desc(oprsz, maxsz, data));
+
+    a0 = tcg_temp_new_ptr();
+    a1 = tcg_temp_new_ptr();
+    a2 = tcg_temp_new_ptr();
+
+    tcg_gen_addi_ptr(a0, cpu_env, dofs);
+    tcg_gen_addi_ptr(a1, cpu_env, aofs);
+    tcg_gen_addi_ptr(a2, cpu_env, bofs);
+
+    fn(a0, a1, a2, ptr, desc);
+
+    tcg_temp_free_ptr(a0);
+    tcg_temp_free_ptr(a1);
+    tcg_temp_free_ptr(a2);
+    tcg_temp_free_i32(desc);
+}
+
+/* Generate a call to a gvec-style helper with four vector operands
+   and an extra pointer operand.  */
+void tcg_gen_gvec_4_ptr(uint32_t dofs, uint32_t aofs, uint32_t bofs,
+                        uint32_t cofs, TCGv_ptr ptr, uint32_t oprsz,
+                        uint32_t maxsz, int32_t data,
+                        gen_helper_gvec_4_ptr *fn)
+{
+    TCGv_ptr a0, a1, a2, a3;
+    TCGv_i32 desc = tcg_const_i32(simd_desc(oprsz, maxsz, data));
+
+    a0 = tcg_temp_new_ptr();
+    a1 = tcg_temp_new_ptr();
+    a2 = tcg_temp_new_ptr();
+    a3 = tcg_temp_new_ptr();
+
+    tcg_gen_addi_ptr(a0, cpu_env, dofs);
+    tcg_gen_addi_ptr(a1, cpu_env, aofs);
+    tcg_gen_addi_ptr(a2, cpu_env, bofs);
+    tcg_gen_addi_ptr(a3, cpu_env, cofs);
+
+    fn(a0, a1, a2, a3, ptr, desc);
+
+    tcg_temp_free_ptr(a0);
+    tcg_temp_free_ptr(a1);
+    tcg_temp_free_ptr(a2);
+    tcg_temp_free_ptr(a3);
+    tcg_temp_free_i32(desc);
+}
+
+/* Return true if we want to implement something of OPRSZ bytes
+   in units of LNSZ.  This limits the expansion of inline code.  */
+static inline bool check_size_impl(uint32_t oprsz, uint32_t lnsz)
+{
+    uint32_t lnct = oprsz / lnsz;
+    return lnct >= 1 && lnct <= MAX_UNROLL;
+}
+
+static void expand_clr(uint32_t dofs, uint32_t maxsz);
+
+/* Duplicate C as per VECE.  */
+uint64_t (dup_const)(unsigned vece, uint64_t c)
+{
+    switch (vece) {
+    case MO_8:
+        return 0x0101010101010101ull * (uint8_t)c;
+    case MO_16:
+        return 0x0001000100010001ull * (uint16_t)c;
+    case MO_32:
+        return 0x0000000100000001ull * (uint32_t)c;
+    case MO_64:
+        return c;
+    default:
+        g_assert_not_reached();
+    }
+}
+
+/* Duplicate IN into OUT as per VECE.  */
+static void gen_dup_i32(unsigned vece, TCGv_i32 out, TCGv_i32 in)
+{
+    switch (vece) {
+    case MO_8:
+        tcg_gen_ext8u_i32(out, in);
+        tcg_gen_muli_i32(out, out, 0x01010101);
+        break;
+    case MO_16:
+        tcg_gen_deposit_i32(out, in, in, 16, 16);
+        break;
+    case MO_32:
+        tcg_gen_mov_i32(out, in);
+        break;
+    default:
+        g_assert_not_reached();
+    }
+}
+
+static void gen_dup_i64(unsigned vece, TCGv_i64 out, TCGv_i64 in)
+{
+    switch (vece) {
+    case MO_8:
+        tcg_gen_ext8u_i64(out, in);
+        tcg_gen_muli_i64(out, out, 0x0101010101010101ull);
+        break;
+    case MO_16:
+        tcg_gen_ext16u_i64(out, in);
+        tcg_gen_muli_i64(out, out, 0x0001000100010001ull);
+        break;
+    case MO_32:
+        tcg_gen_deposit_i64(out, in, in, 32, 32);
+        break;
+    case MO_64:
+        tcg_gen_mov_i64(out, in);
+        break;
+    default:
+        g_assert_not_reached();
+    }
+}
+
+/* Set OPRSZ bytes at DOFS to replications of IN_32, IN_64 or IN_C.
+ * Only one of IN_32 or IN_64 may be set;
+ * IN_C is used if IN_32 and IN_64 are unset.
+ */
+static void do_dup(unsigned vece, uint32_t dofs, uint32_t oprsz,
+                   uint32_t maxsz, TCGv_i32 in_32, TCGv_i64 in_64,
+                   uint64_t in_c)
+{
+    TCGType type;
+    TCGv_i64 t_64;
+    TCGv_i32 t_32, t_desc;
+    TCGv_ptr t_ptr;
+    uint32_t i;
+
+    assert(vece <= (in_32 ? MO_32 : MO_64));
+    assert(in_32 == NULL || in_64 == NULL);
+
+    /* If we're storing 0, expand oprsz to maxsz.  */
+    if (in_32 == NULL && in_64 == NULL) {
+        in_c = dup_const(vece, in_c);
+        if (in_c == 0) {
+            oprsz = maxsz;
+        }
+    }
+
+    type = 0;
+    if (TCG_TARGET_HAS_v256 && check_size_impl(oprsz, 32)) {
+        type = TCG_TYPE_V256;
+    } else if (TCG_TARGET_HAS_v128 && check_size_impl(oprsz, 16)) {
+        type = TCG_TYPE_V128;
+    } else if (TCG_TARGET_HAS_v64 && check_size_impl(oprsz, 8)
+               /* Prefer integer when 64-bit host and no variable dup.  */
+               && !(TCG_TARGET_REG_BITS == 64 && in_32 == NULL
+                    && (in_64 == NULL || vece == MO_64))) {
+        type = TCG_TYPE_V64;
+    }
+
+    /* Implement inline with a vector type, if possible.  */
+    if (type != 0) {
+        TCGv_vec t_vec = tcg_temp_new_vec(type);
+
+        if (in_32) {
+            tcg_gen_dup_i32_vec(vece, t_vec, in_32);
+        } else if (in_64) {
+            tcg_gen_dup_i64_vec(vece, t_vec, in_64);
+        } else {
+            switch (vece) {
+            case MO_8:
+                tcg_gen_dup8i_vec(t_vec, in_c);
+                break;
+            case MO_16:
+                tcg_gen_dup16i_vec(t_vec, in_c);
+                break;
+            case MO_32:
+                tcg_gen_dup32i_vec(t_vec, in_c);
+                break;
+            default:
+                tcg_gen_dup64i_vec(t_vec, in_c);
+                break;
+            }
+        }
+
+        i = 0;
+        if (TCG_TARGET_HAS_v256) {
+            for (; i + 32 <= oprsz; i += 32) {
+                tcg_gen_stl_vec(t_vec, cpu_env, dofs + i, TCG_TYPE_V256);
+            }
+        }
+        if (TCG_TARGET_HAS_v128) {
+            for (; i + 16 <= oprsz; i += 16) {
+                tcg_gen_stl_vec(t_vec, cpu_env, dofs + i, TCG_TYPE_V128);
+            }
+        }
+        if (TCG_TARGET_HAS_v64) {
+            for (; i < oprsz; i += 8) {
+                tcg_gen_stl_vec(t_vec, cpu_env, dofs + i, TCG_TYPE_V64);
+            }
+        }
+        tcg_temp_free_vec(t_vec);
+        goto done;
+    }
+
+    /* Otherwise, inline with an integer type, unless "large".  */
+    if (check_size_impl(oprsz, TCG_TARGET_REG_BITS / 8)) {
+        t_64 = NULL;
+        t_32 = NULL;
+
+        if (in_32) {
+            /* We are given a 32-bit variable input.  For a 64-bit host,
+               use a 64-bit operation unless the 32-bit operation would
+               be simple enough.  */
+            if (TCG_TARGET_REG_BITS == 64
+                && (vece != MO_32 || !check_size_impl(oprsz, 4))) {
+                t_64 = tcg_temp_new_i64();
+                tcg_gen_extu_i32_i64(t_64, in_32);
+                gen_dup_i64(vece, t_64, t_64);
+            } else {
+                t_32 = tcg_temp_new_i32();
+                gen_dup_i32(vece, t_32, in_32);
+            }
+        } else if (in_64) {
+            /* We are given a 64-bit variable input.  */
+            t_64 = tcg_temp_new_i64();
+            gen_dup_i64(vece, t_64, in_64);
+        } else {
+            /* We are given a constant input.  */
+            /* For 64-bit hosts, use 64-bit constants for "simple" constants
+               or when we'd need too many 32-bit stores, or when a 64-bit
+               constant is really required.  */
+            if (vece == MO_64
+                || (TCG_TARGET_REG_BITS == 64
+                    && (in_c == 0 || in_c == -1
+                        || !check_size_impl(oprsz, 4)))) {
+                t_64 = tcg_const_i64(in_c);
+            } else {
+                t_32 = tcg_const_i32(in_c);
+            }
+        }
+
+        /* Implement inline if we picked an implementation size above.  */
+        if (t_32) {
+            for (i = 0; i < oprsz; i += 4) {
+                tcg_gen_st_i32(t_32, cpu_env, dofs + i);
+            }
+            tcg_temp_free_i32(t_32);
+            goto done;
+        }
+        if (t_64) {
+            for (i = 0; i < oprsz; i += 8) {
+                tcg_gen_st_i64(t_64, cpu_env, dofs + i);
+            }
+            tcg_temp_free_i64(t_64);
+            goto done;
+        } 
+    }
+
+    /* Otherwise implement out of line.  */
+    t_ptr = tcg_temp_new_ptr();
+    tcg_gen_addi_ptr(t_ptr, cpu_env, dofs);
+    t_desc = tcg_const_i32(simd_desc(oprsz, maxsz, 0));
+
+    if (vece == MO_64) {
+        if (in_64) {
+            gen_helper_gvec_dup64(t_ptr, t_desc, in_64);
+        } else {
+            t_64 = tcg_const_i64(in_c);
+            gen_helper_gvec_dup64(t_ptr, t_desc, t_64);
+            tcg_temp_free_i64(t_64);
+        }
+    } else {
+        typedef void dup_fn(TCGv_ptr, TCGv_i32, TCGv_i32);
+        static dup_fn * const fns[3] = {
+            gen_helper_gvec_dup8,
+            gen_helper_gvec_dup16,
+            gen_helper_gvec_dup32
+        };
+
+        if (in_32) {
+            fns[vece](t_ptr, t_desc, in_32);
+        } else {
+            t_32 = tcg_temp_new_i32();
+            if (in_64) {
+                tcg_gen_extrl_i64_i32(t_32, in_64);
+            } else if (vece == MO_8) {
+                tcg_gen_movi_i32(t_32, in_c & 0xff);
+            } else if (vece == MO_16) {
+                tcg_gen_movi_i32(t_32, in_c & 0xffff);
+            } else {
+                tcg_gen_movi_i32(t_32, in_c);
+            }
+            fns[vece](t_ptr, t_desc, t_32);
+            tcg_temp_free_i32(t_32);
+        }
+    }
+
+    tcg_temp_free_ptr(t_ptr);
+    tcg_temp_free_i32(t_desc);
+    return;
+
+ done:
+    if (oprsz < maxsz) {
+        expand_clr(dofs + oprsz, maxsz - oprsz);
+    }
+}
+
+/* Likewise, but with zero.  */
+static void expand_clr(uint32_t dofs, uint32_t maxsz)
+{
+    do_dup(MO_8, dofs, maxsz, maxsz, NULL, NULL, 0);
+}
+
+/* Expand OPSZ bytes worth of two-operand operations using i32 elements.  */
+static void expand_2_i32(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
+                         void (*fni)(TCGv_i32, TCGv_i32))
+{
+    TCGv_i32 t0 = tcg_temp_new_i32();
+    uint32_t i;
+
+    for (i = 0; i < oprsz; i += 4) {
+        tcg_gen_ld_i32(t0, cpu_env, aofs + i);
+        fni(t0, t0);
+        tcg_gen_st_i32(t0, cpu_env, dofs + i);
+    }
+    tcg_temp_free_i32(t0);
+}
+
+/* Expand OPSZ bytes worth of three-operand operations using i32 elements.  */
+static void expand_3_i32(uint32_t dofs, uint32_t aofs,
+                         uint32_t bofs, uint32_t oprsz, bool load_dest,
+                         void (*fni)(TCGv_i32, TCGv_i32, TCGv_i32))
+{
+    TCGv_i32 t0 = tcg_temp_new_i32();
+    TCGv_i32 t1 = tcg_temp_new_i32();
+    TCGv_i32 t2 = tcg_temp_new_i32();
+    uint32_t i;
+
+    for (i = 0; i < oprsz; i += 4) {
+        tcg_gen_ld_i32(t0, cpu_env, aofs + i);
+        tcg_gen_ld_i32(t1, cpu_env, bofs + i);
+        if (load_dest) {
+            tcg_gen_ld_i32(t2, cpu_env, dofs + i);
+        }
+        fni(t2, t0, t1);
+        tcg_gen_st_i32(t2, cpu_env, dofs + i);
+    }
+    tcg_temp_free_i32(t2);
+    tcg_temp_free_i32(t1);
+    tcg_temp_free_i32(t0);
+}
+
+/* Expand OPSZ bytes worth of three-operand operations using i32 elements.  */
+static void expand_4_i32(uint32_t dofs, uint32_t aofs, uint32_t bofs,
+                         uint32_t cofs, uint32_t oprsz,
+                         void (*fni)(TCGv_i32, TCGv_i32, TCGv_i32, TCGv_i32))
+{
+    TCGv_i32 t0 = tcg_temp_new_i32();
+    TCGv_i32 t1 = tcg_temp_new_i32();
+    TCGv_i32 t2 = tcg_temp_new_i32();
+    TCGv_i32 t3 = tcg_temp_new_i32();
+    uint32_t i;
+
+    for (i = 0; i < oprsz; i += 4) {
+        tcg_gen_ld_i32(t1, cpu_env, aofs + i);
+        tcg_gen_ld_i32(t2, cpu_env, bofs + i);
+        tcg_gen_ld_i32(t3, cpu_env, cofs + i);
+        fni(t0, t1, t2, t3);
+        tcg_gen_st_i32(t0, cpu_env, dofs + i);
+    }
+    tcg_temp_free_i32(t3);
+    tcg_temp_free_i32(t2);
+    tcg_temp_free_i32(t1);
+    tcg_temp_free_i32(t0);
+}
+
+/* Expand OPSZ bytes worth of two-operand operations using i64 elements.  */
+static void expand_2_i64(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
+                         void (*fni)(TCGv_i64, TCGv_i64))
+{
+    TCGv_i64 t0 = tcg_temp_new_i64();
+    uint32_t i;
+
+    for (i = 0; i < oprsz; i += 8) {
+        tcg_gen_ld_i64(t0, cpu_env, aofs + i);
+        fni(t0, t0);
+        tcg_gen_st_i64(t0, cpu_env, dofs + i);
+    }
+    tcg_temp_free_i64(t0);
+}
+
+/* Expand OPSZ bytes worth of three-operand operations using i64 elements.  */
+static void expand_3_i64(uint32_t dofs, uint32_t aofs,
+                         uint32_t bofs, uint32_t oprsz, bool load_dest,
+                         void (*fni)(TCGv_i64, TCGv_i64, TCGv_i64))
+{
+    TCGv_i64 t0 = tcg_temp_new_i64();
+    TCGv_i64 t1 = tcg_temp_new_i64();
+    TCGv_i64 t2 = tcg_temp_new_i64();
+    uint32_t i;
+
+    for (i = 0; i < oprsz; i += 8) {
+        tcg_gen_ld_i64(t0, cpu_env, aofs + i);
+        tcg_gen_ld_i64(t1, cpu_env, bofs + i);
+        if (load_dest) {
+            tcg_gen_ld_i64(t2, cpu_env, dofs + i);
+        }
+        fni(t2, t0, t1);
+        tcg_gen_st_i64(t2, cpu_env, dofs + i);
+    }
+    tcg_temp_free_i64(t2);
+    tcg_temp_free_i64(t1);
+    tcg_temp_free_i64(t0);
+}
+
+/* Expand OPSZ bytes worth of three-operand operations using i64 elements.  */
+static void expand_4_i64(uint32_t dofs, uint32_t aofs, uint32_t bofs,
+                         uint32_t cofs, uint32_t oprsz,
+                         void (*fni)(TCGv_i64, TCGv_i64, TCGv_i64, TCGv_i64))
+{
+    TCGv_i64 t0 = tcg_temp_new_i64();
+    TCGv_i64 t1 = tcg_temp_new_i64();
+    TCGv_i64 t2 = tcg_temp_new_i64();
+    TCGv_i64 t3 = tcg_temp_new_i64();
+    uint32_t i;
+
+    for (i = 0; i < oprsz; i += 8) {
+        tcg_gen_ld_i64(t1, cpu_env, aofs + i);
+        tcg_gen_ld_i64(t2, cpu_env, bofs + i);
+        tcg_gen_ld_i64(t3, cpu_env, cofs + i);
+        fni(t0, t1, t2, t3);
+        tcg_gen_st_i64(t0, cpu_env, dofs + i);
+    }
+    tcg_temp_free_i64(t3);
+    tcg_temp_free_i64(t2);
+    tcg_temp_free_i64(t1);
+    tcg_temp_free_i64(t0);
+}
+
+/* Expand OPSZ bytes worth of two-operand operations using host vectors.  */
+static void expand_2_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
+                         uint32_t oprsz, uint32_t tysz, TCGType type,
+                         void (*fni)(unsigned, TCGv_vec, TCGv_vec))
+{
+    TCGv_vec t0 = tcg_temp_new_vec(type);
+    uint32_t i;
+
+    for (i = 0; i < oprsz; i += tysz) {
+        tcg_gen_ld_vec(t0, cpu_env, aofs + i);
+        fni(vece, t0, t0);
+        tcg_gen_st_vec(t0, cpu_env, dofs + i);
+    }
+    tcg_temp_free_vec(t0);
+}
+
+/* Expand OPSZ bytes worth of three-operand operations using host vectors.  */
+static void expand_3_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
+                         uint32_t bofs, uint32_t oprsz,
+                         uint32_t tysz, TCGType type, bool load_dest,
+                         void (*fni)(unsigned, TCGv_vec, TCGv_vec, TCGv_vec))
+{
+    TCGv_vec t0 = tcg_temp_new_vec(type);
+    TCGv_vec t1 = tcg_temp_new_vec(type);
+    TCGv_vec t2 = tcg_temp_new_vec(type);
+    uint32_t i;
+
+    for (i = 0; i < oprsz; i += tysz) {
+        tcg_gen_ld_vec(t0, cpu_env, aofs + i);
+        tcg_gen_ld_vec(t1, cpu_env, bofs + i);
+        if (load_dest) {
+            tcg_gen_ld_vec(t2, cpu_env, dofs + i);
+        }
+        fni(vece, t2, t0, t1);
+        tcg_gen_st_vec(t2, cpu_env, dofs + i);
+    }
+    tcg_temp_free_vec(t2);
+    tcg_temp_free_vec(t1);
+    tcg_temp_free_vec(t0);
+}
+
+/* Expand OPSZ bytes worth of four-operand operations using host vectors.  */
+static void expand_4_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
+                         uint32_t bofs, uint32_t cofs, uint32_t oprsz,
+                         uint32_t tysz, TCGType type,
+                         void (*fni)(unsigned, TCGv_vec, TCGv_vec,
+                                     TCGv_vec, TCGv_vec))
+{
+    TCGv_vec t0 = tcg_temp_new_vec(type);
+    TCGv_vec t1 = tcg_temp_new_vec(type);
+    TCGv_vec t2 = tcg_temp_new_vec(type);
+    TCGv_vec t3 = tcg_temp_new_vec(type);
+    uint32_t i;
+
+    for (i = 0; i < oprsz; i += tysz) {
+        tcg_gen_ld_vec(t1, cpu_env, aofs + i);
+        tcg_gen_ld_vec(t2, cpu_env, bofs + i);
+        tcg_gen_ld_vec(t3, cpu_env, cofs + i);
+        fni(vece, t0, t1, t2, t3);
+        tcg_gen_st_vec(t0, cpu_env, dofs + i);
+    }
+    tcg_temp_free_vec(t3);
+    tcg_temp_free_vec(t2);
+    tcg_temp_free_vec(t1);
+    tcg_temp_free_vec(t0);
+}
+
+/* Expand a vector two-operand operation.  */
+void tcg_gen_gvec_2(uint32_t dofs, uint32_t aofs,
+                    uint32_t oprsz, uint32_t maxsz, const GVecGen2 *g)
+{
+    check_size_align(oprsz, maxsz, dofs | aofs);
+    check_overlap_2(dofs, aofs, maxsz);
+
+    /* Recall that ARM SVE allows vector sizes that are not a power of 2.
+       Expand with successively smaller host vector sizes.  The intent is
+       that e.g. oprsz == 80 would be expanded with 2x32 + 1x16.  */
+    /* ??? For maxsz > oprsz, the host may be able to use an opr-sized
+       operation, zeroing the balance of the register.  We can then
+       use a max-sized store to implement the clearing without an extra
+       store operation.  This is true for aarch64 and x86_64 hosts.  */
+
+    if (TCG_TARGET_HAS_v256 && g->fniv && check_size_impl(oprsz, 32)
+        && (!g->opc || tcg_can_emit_vec_op(g->opc, TCG_TYPE_V256, g->vece))) {
+        uint32_t some = QEMU_ALIGN_DOWN(oprsz, 32);
+        expand_2_vec(g->vece, dofs, aofs, some, 32, TCG_TYPE_V256, g->fniv);
+        if (some == oprsz) {
+            goto done;
+        }
+        dofs += some;
+        aofs += some;
+        oprsz -= some;
+        maxsz -= some;
+    }
+
+    if (TCG_TARGET_HAS_v128 && g->fniv && check_size_impl(oprsz, 16)
+        && (!g->opc || tcg_can_emit_vec_op(g->opc, TCG_TYPE_V128, g->vece))) {
+        expand_2_vec(g->vece, dofs, aofs, oprsz, 16, TCG_TYPE_V128, g->fniv);
+    } else if (TCG_TARGET_HAS_v64 && !g->prefer_i64
+               && g->fniv && check_size_impl(oprsz, 8)
+               && (!g->opc
+                   || tcg_can_emit_vec_op(g->opc, TCG_TYPE_V64, g->vece))) {
+        expand_2_vec(g->vece, dofs, aofs, oprsz, 8, TCG_TYPE_V64, g->fniv);
+    } else if (g->fni8 && check_size_impl(oprsz, 8)) {
+        expand_2_i64(dofs, aofs, oprsz, g->fni8);
+    } else if (g->fni4 && check_size_impl(oprsz, 4)) {
+        expand_2_i32(dofs, aofs, oprsz, g->fni4);
+    } else {
+        assert(g->fno != NULL);
+        tcg_gen_gvec_2_ool(dofs, aofs, oprsz, maxsz, g->data, g->fno);
+        return;
+    }
+
+ done:
+    if (oprsz < maxsz) {
+        expand_clr(dofs + oprsz, maxsz - oprsz);
+    }
+}
+
+/* Expand a vector three-operand operation.  */
+void tcg_gen_gvec_3(uint32_t dofs, uint32_t aofs, uint32_t bofs,
+                    uint32_t oprsz, uint32_t maxsz, const GVecGen3 *g)
+{
+    check_size_align(oprsz, maxsz, dofs | aofs | bofs);
+    check_overlap_3(dofs, aofs, bofs, maxsz);
+
+    /* Recall that ARM SVE allows vector sizes that are not a power of 2.
+       Expand with successively smaller host vector sizes.  The intent is
+       that e.g. oprsz == 80 would be expanded with 2x32 + 1x16.  */
+
+    if (TCG_TARGET_HAS_v256 && g->fniv && check_size_impl(oprsz, 32)
+        && (!g->opc || tcg_can_emit_vec_op(g->opc, TCG_TYPE_V256, g->vece))) {
+        uint32_t some = QEMU_ALIGN_DOWN(oprsz, 32);
+        expand_3_vec(g->vece, dofs, aofs, bofs, some, 32, TCG_TYPE_V256,
+                     g->load_dest, g->fniv);
+        if (some == oprsz) {
+            goto done;
+        }
+        dofs += some;
+        aofs += some;
+        bofs += some;
+        oprsz -= some;
+        maxsz -= some;
+    }
+
+    if (TCG_TARGET_HAS_v128 && g->fniv && check_size_impl(oprsz, 16)
+        && (!g->opc || tcg_can_emit_vec_op(g->opc, TCG_TYPE_V128, g->vece))) {
+        expand_3_vec(g->vece, dofs, aofs, bofs, oprsz, 16, TCG_TYPE_V128,
+                     g->load_dest, g->fniv);
+    } else if (TCG_TARGET_HAS_v64 && !g->prefer_i64
+               && g->fniv && check_size_impl(oprsz, 8)
+               && (!g->opc
+                   || tcg_can_emit_vec_op(g->opc, TCG_TYPE_V64, g->vece))) {
+        expand_3_vec(g->vece, dofs, aofs, bofs, oprsz, 8, TCG_TYPE_V64,
+                     g->load_dest, g->fniv);
+    } else if (g->fni8 && check_size_impl(oprsz, 8)) {
+        expand_3_i64(dofs, aofs, bofs, oprsz, g->load_dest, g->fni8);
+    } else if (g->fni4 && check_size_impl(oprsz, 4)) {
+        expand_3_i32(dofs, aofs, bofs, oprsz, g->load_dest, g->fni4);
+    } else {
+        assert(g->fno != NULL);
+        tcg_gen_gvec_3_ool(dofs, aofs, bofs, oprsz, maxsz, g->data, g->fno);
+    }
+
+ done:
+    if (oprsz < maxsz) {
+        expand_clr(dofs + oprsz, maxsz - oprsz);
+    }
+}
+
+/* Expand a vector four-operand operation.  */
+void tcg_gen_gvec_4(uint32_t dofs, uint32_t aofs, uint32_t bofs, uint32_t cofs,
+                    uint32_t oprsz, uint32_t maxsz, const GVecGen4 *g)
+{
+    check_size_align(oprsz, maxsz, dofs | aofs | bofs | cofs);
+    check_overlap_4(dofs, aofs, bofs, cofs, maxsz);
+
+    /* Recall that ARM SVE allows vector sizes that are not a power of 2.
+       Expand with successively smaller host vector sizes.  The intent is
+       that e.g. oprsz == 80 would be expanded with 2x32 + 1x16.  */
+
+    if (TCG_TARGET_HAS_v256 && g->fniv && check_size_impl(oprsz, 32)
+        && (!g->opc || tcg_can_emit_vec_op(g->opc, TCG_TYPE_V256, g->vece))) {
+        uint32_t some = QEMU_ALIGN_DOWN(oprsz, 32);
+        expand_4_vec(g->vece, dofs, aofs, bofs, cofs, some,
+                     32, TCG_TYPE_V256, g->fniv);
+        if (some == oprsz) {
+            goto done;
+        }
+        dofs += some;
+        aofs += some;
+        bofs += some;
+        cofs += some;
+        oprsz -= some;
+        maxsz -= some;
+    }
+
+    if (TCG_TARGET_HAS_v128 && g->fniv && check_size_impl(oprsz, 16)
+        && (!g->opc || tcg_can_emit_vec_op(g->opc, TCG_TYPE_V128, g->vece))) {
+        expand_4_vec(g->vece, dofs, aofs, bofs, cofs, oprsz,
+                     16, TCG_TYPE_V128, g->fniv);
+    } else if (TCG_TARGET_HAS_v64 && !g->prefer_i64
+               && g->fniv && check_size_impl(oprsz, 8)
+                && (!g->opc
+                    || tcg_can_emit_vec_op(g->opc, TCG_TYPE_V64, g->vece))) {
+        expand_4_vec(g->vece, dofs, aofs, bofs, cofs, oprsz,
+                     8, TCG_TYPE_V64, g->fniv);
+    } else if (g->fni8 && check_size_impl(oprsz, 8)) {
+        expand_4_i64(dofs, aofs, bofs, cofs, oprsz, g->fni8);
+    } else if (g->fni4 && check_size_impl(oprsz, 4)) {
+        expand_4_i32(dofs, aofs, bofs, cofs, oprsz, g->fni4);
+    } else {
+        assert(g->fno != NULL);
+        tcg_gen_gvec_4_ool(dofs, aofs, bofs, cofs,
+                           oprsz, maxsz, g->data, g->fno);
+        return;
+    }
+
+ done:
+    if (oprsz < maxsz) {
+        expand_clr(dofs + oprsz, maxsz - oprsz);
+    }
+}
+
+/*
+ * Expand specific vector operations.
+ */
+
+static void vec_mov2(unsigned vece, TCGv_vec a, TCGv_vec b)
+{
+    tcg_gen_mov_vec(a, b);
+}
+
+void tcg_gen_gvec_mov(unsigned vece, uint32_t dofs, uint32_t aofs,
+                      uint32_t oprsz, uint32_t maxsz)
+{
+    static const GVecGen2 g = {
+        .fni8 = tcg_gen_mov_i64,
+        .fniv = vec_mov2,
+        .fno = gen_helper_gvec_mov,
+        .prefer_i64 = TCG_TARGET_REG_BITS == 64,
+    };
+    if (dofs != aofs) {
+        tcg_gen_gvec_2(dofs, aofs, oprsz, maxsz, &g);
+    } else {
+        check_size_align(oprsz, maxsz, dofs);
+        if (oprsz < maxsz) {
+            expand_clr(dofs + oprsz, maxsz - oprsz);
+        }
+    }
+}
+
+void tcg_gen_gvec_dup_i32(unsigned vece, uint32_t dofs, uint32_t oprsz,
+                          uint32_t maxsz, TCGv_i32 in)
+{
+    check_size_align(oprsz, maxsz, dofs);
+    tcg_debug_assert(vece <= MO_32);
+    do_dup(vece, dofs, oprsz, maxsz, in, NULL, 0);
+}
+
+void tcg_gen_gvec_dup_i64(unsigned vece, uint32_t dofs, uint32_t oprsz,
+                          uint32_t maxsz, TCGv_i64 in)
+{
+    check_size_align(oprsz, maxsz, dofs);
+    tcg_debug_assert(vece <= MO_64);
+    do_dup(vece, dofs, oprsz, maxsz, NULL, in, 0);
+}
+
+void tcg_gen_gvec_dup_mem(unsigned vece, uint32_t dofs, uint32_t aofs,
+                          uint32_t oprsz, uint32_t maxsz)
+{
+    if (vece <= MO_32) {
+        TCGv_i32 in = tcg_temp_new_i32();
+        switch (vece) {
+        case MO_8:
+            tcg_gen_ld8u_i32(in, cpu_env, aofs);
+            break;
+        case MO_16:
+            tcg_gen_ld16u_i32(in, cpu_env, aofs);
+            break;
+        case MO_32:
+            tcg_gen_ld_i32(in, cpu_env, aofs);
+            break;
+        }
+        tcg_gen_gvec_dup_i32(vece, dofs, oprsz, maxsz, in);
+        tcg_temp_free_i32(in);
+    } else if (vece == MO_64) {
+        TCGv_i64 in = tcg_temp_new_i64();
+        tcg_gen_ld_i64(in, cpu_env, aofs);
+        tcg_gen_gvec_dup_i64(MO_64, dofs, oprsz, maxsz, in);
+        tcg_temp_free_i64(in);
+    } else {
+        /* 128-bit duplicate.  */
+        /* ??? Dup to 256-bit vector.  */
+        int i;
+
+        tcg_debug_assert(vece == 4);
+        tcg_debug_assert(oprsz >= 16);
+        if (TCG_TARGET_HAS_v128) {
+            TCGv_vec in = tcg_temp_new_vec(TCG_TYPE_V128);
+
+            tcg_gen_ld_vec(in, cpu_env, aofs);
+            for (i = 0; i < oprsz; i += 16) {
+                tcg_gen_st_vec(in, cpu_env, dofs + i);
+            }
+            tcg_temp_free_vec(in);
+        } else {
+            TCGv_i64 in0 = tcg_temp_new_i64();
+            TCGv_i64 in1 = tcg_temp_new_i64();
+
+            tcg_gen_ld_i64(in0, cpu_env, aofs);
+            tcg_gen_ld_i64(in1, cpu_env, aofs + 8);
+            for (i = 0; i < oprsz; i += 16) {
+                tcg_gen_st_i64(in0, cpu_env, dofs + i);
+                tcg_gen_st_i64(in1, cpu_env, dofs + i + 8);
+            }
+            tcg_temp_free_i64(in0);
+            tcg_temp_free_i64(in1);
+        }
+    }
+}
+
+void tcg_gen_gvec_dup64i(uint32_t dofs, uint32_t oprsz,
+                         uint32_t maxsz, uint64_t x)
+{
+    check_size_align(oprsz, maxsz, dofs);
+    do_dup(MO_64, dofs, oprsz, maxsz, NULL, NULL, x);
+}
+
+void tcg_gen_gvec_dup32i(uint32_t dofs, uint32_t oprsz,
+                         uint32_t maxsz, uint32_t x)
+{
+    check_size_align(oprsz, maxsz, dofs);
+    do_dup(MO_32, dofs, oprsz, maxsz, NULL, NULL, x);
+}
+
+void tcg_gen_gvec_dup16i(uint32_t dofs, uint32_t oprsz,
+                         uint32_t maxsz, uint16_t x)
+{
+    check_size_align(oprsz, maxsz, dofs);
+    do_dup(MO_16, dofs, oprsz, maxsz, NULL, NULL, x);
+}
+
+void tcg_gen_gvec_dup8i(uint32_t dofs, uint32_t oprsz,
+                         uint32_t maxsz, uint8_t x)
+{
+    check_size_align(oprsz, maxsz, dofs);
+    do_dup(MO_8, dofs, oprsz, maxsz, NULL, NULL, x);
+}
+
+void tcg_gen_gvec_not(unsigned vece, uint32_t dofs, uint32_t aofs,
+                      uint32_t oprsz, uint32_t maxsz)
+{
+    static const GVecGen2 g = {
+        .fni8 = tcg_gen_not_i64,
+        .fniv = tcg_gen_not_vec,
+        .fno = gen_helper_gvec_not,
+        .prefer_i64 = TCG_TARGET_REG_BITS == 64,
+    };
+    tcg_gen_gvec_2(dofs, aofs, oprsz, maxsz, &g);
+}
+
+/* Perform a vector addition using normal addition and a mask.  The mask
+   should be the sign bit of each lane.  This 6-operation form is more
+   efficient than separate additions when there are 4 or more lanes in
+   the 64-bit operation.  */
+static void gen_addv_mask(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b, TCGv_i64 m)
+{
+    TCGv_i64 t1 = tcg_temp_new_i64();
+    TCGv_i64 t2 = tcg_temp_new_i64();
+    TCGv_i64 t3 = tcg_temp_new_i64();
+
+    tcg_gen_andc_i64(t1, a, m);
+    tcg_gen_andc_i64(t2, b, m);
+    tcg_gen_xor_i64(t3, a, b);
+    tcg_gen_add_i64(d, t1, t2);
+    tcg_gen_and_i64(t3, t3, m);
+    tcg_gen_xor_i64(d, d, t3);
+
+    tcg_temp_free_i64(t1);
+    tcg_temp_free_i64(t2);
+    tcg_temp_free_i64(t3);
+}
+
+void tcg_gen_vec_add8_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
+{
+    TCGv_i64 m = tcg_const_i64(dup_const(MO_8, 0x80));
+    gen_addv_mask(d, a, b, m);
+    tcg_temp_free_i64(m);
+}
+
+void tcg_gen_vec_add16_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
+{
+    TCGv_i64 m = tcg_const_i64(dup_const(MO_16, 0x8000));
+    gen_addv_mask(d, a, b, m);
+    tcg_temp_free_i64(m);
+}
+
+void tcg_gen_vec_add32_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
+{
+    TCGv_i64 t1 = tcg_temp_new_i64();
+    TCGv_i64 t2 = tcg_temp_new_i64();
+
+    tcg_gen_andi_i64(t1, a, ~0xffffffffull);
+    tcg_gen_add_i64(t2, a, b);
+    tcg_gen_add_i64(t1, t1, b);
+    tcg_gen_deposit_i64(d, t1, t2, 0, 32);
+
+    tcg_temp_free_i64(t1);
+    tcg_temp_free_i64(t2);
+}
+
+void tcg_gen_gvec_add(unsigned vece, uint32_t dofs, uint32_t aofs,
+                      uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
+{
+    static const GVecGen3 g[4] = {
+        { .fni8 = tcg_gen_vec_add8_i64,
+          .fniv = tcg_gen_add_vec,
+          .fno = gen_helper_gvec_add8,
+          .opc = INDEX_op_add_vec,
+          .vece = MO_8 },
+        { .fni8 = tcg_gen_vec_add16_i64,
+          .fniv = tcg_gen_add_vec,
+          .fno = gen_helper_gvec_add16,
+          .opc = INDEX_op_add_vec,
+          .vece = MO_16 },
+        { .fni4 = tcg_gen_add_i32,
+          .fniv = tcg_gen_add_vec,
+          .fno = gen_helper_gvec_add32,
+          .opc = INDEX_op_add_vec,
+          .vece = MO_32 },
+        { .fni8 = tcg_gen_add_i64,
+          .fniv = tcg_gen_add_vec,
+          .fno = gen_helper_gvec_add64,
+          .opc = INDEX_op_add_vec,
+          .prefer_i64 = TCG_TARGET_REG_BITS == 64,
+          .vece = MO_64 },
+    };
+
+    tcg_debug_assert(vece <= MO_64);
+    tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
+}
+
+/* Perform a vector subtraction using normal subtraction and a mask.
+   Compare gen_addv_mask above.  */
+static void gen_subv_mask(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b, TCGv_i64 m)
+{
+    TCGv_i64 t1 = tcg_temp_new_i64();
+    TCGv_i64 t2 = tcg_temp_new_i64();
+    TCGv_i64 t3 = tcg_temp_new_i64();
+
+    tcg_gen_or_i64(t1, a, m);
+    tcg_gen_andc_i64(t2, b, m);
+    tcg_gen_eqv_i64(t3, a, b);
+    tcg_gen_sub_i64(d, t1, t2);
+    tcg_gen_and_i64(t3, t3, m);
+    tcg_gen_xor_i64(d, d, t3);
+
+    tcg_temp_free_i64(t1);
+    tcg_temp_free_i64(t2);
+    tcg_temp_free_i64(t3);
+}
+
+void tcg_gen_vec_sub8_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
+{
+    TCGv_i64 m = tcg_const_i64(dup_const(MO_8, 0x80));
+    gen_subv_mask(d, a, b, m);
+    tcg_temp_free_i64(m);
+}
+
+void tcg_gen_vec_sub16_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
+{
+    TCGv_i64 m = tcg_const_i64(dup_const(MO_16, 0x8000));
+    gen_subv_mask(d, a, b, m);
+    tcg_temp_free_i64(m);
+}
+
+void tcg_gen_vec_sub32_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
+{
+    TCGv_i64 t1 = tcg_temp_new_i64();
+    TCGv_i64 t2 = tcg_temp_new_i64();
+
+    tcg_gen_andi_i64(t1, b, ~0xffffffffull);
+    tcg_gen_sub_i64(t2, a, b);
+    tcg_gen_sub_i64(t1, a, t1);
+    tcg_gen_deposit_i64(d, t1, t2, 0, 32);
+
+    tcg_temp_free_i64(t1);
+    tcg_temp_free_i64(t2);
+}
+
+void tcg_gen_gvec_sub(unsigned vece, uint32_t dofs, uint32_t aofs,
+                      uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
+{
+    static const GVecGen3 g[4] = {
+        { .fni8 = tcg_gen_vec_sub8_i64,
+          .fniv = tcg_gen_sub_vec,
+          .fno = gen_helper_gvec_sub8,
+          .opc = INDEX_op_sub_vec,
+          .vece = MO_8 },
+        { .fni8 = tcg_gen_vec_sub16_i64,
+          .fniv = tcg_gen_sub_vec,
+          .fno = gen_helper_gvec_sub16,
+          .opc = INDEX_op_sub_vec,
+          .vece = MO_16 },
+        { .fni4 = tcg_gen_sub_i32,
+          .fniv = tcg_gen_sub_vec,
+          .fno = gen_helper_gvec_sub32,
+          .opc = INDEX_op_sub_vec,
+          .vece = MO_32 },
+        { .fni8 = tcg_gen_sub_i64,
+          .fniv = tcg_gen_sub_vec,
+          .fno = gen_helper_gvec_sub64,
+          .opc = INDEX_op_sub_vec,
+          .prefer_i64 = TCG_TARGET_REG_BITS == 64,
+          .vece = MO_64 },
+    };
+
+    tcg_debug_assert(vece <= MO_64);
+    tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
+}
+
+/* Perform a vector negation using normal negation and a mask.
+   Compare gen_subv_mask above.  */
+static void gen_negv_mask(TCGv_i64 d, TCGv_i64 b, TCGv_i64 m)
+{
+    TCGv_i64 t2 = tcg_temp_new_i64();
+    TCGv_i64 t3 = tcg_temp_new_i64();
+
+    tcg_gen_andc_i64(t3, m, b);
+    tcg_gen_andc_i64(t2, b, m);
+    tcg_gen_sub_i64(d, m, t2);
+    tcg_gen_xor_i64(d, d, t3);
+
+    tcg_temp_free_i64(t2);
+    tcg_temp_free_i64(t3);
+}
+
+void tcg_gen_vec_neg8_i64(TCGv_i64 d, TCGv_i64 b)
+{
+    TCGv_i64 m = tcg_const_i64(dup_const(MO_8, 0x80));
+    gen_negv_mask(d, b, m);
+    tcg_temp_free_i64(m);
+}
+
+void tcg_gen_vec_neg16_i64(TCGv_i64 d, TCGv_i64 b)
+{
+    TCGv_i64 m = tcg_const_i64(dup_const(MO_16, 0x8000));
+    gen_negv_mask(d, b, m);
+    tcg_temp_free_i64(m);
+}
+
+void tcg_gen_vec_neg32_i64(TCGv_i64 d, TCGv_i64 b)
+{
+    TCGv_i64 t1 = tcg_temp_new_i64();
+    TCGv_i64 t2 = tcg_temp_new_i64();
+
+    tcg_gen_andi_i64(t1, b, ~0xffffffffull);
+    tcg_gen_neg_i64(t2, b);
+    tcg_gen_neg_i64(t1, t1);
+    tcg_gen_deposit_i64(d, t1, t2, 0, 32);
+
+    tcg_temp_free_i64(t1);
+    tcg_temp_free_i64(t2);
+}
+
+void tcg_gen_gvec_neg(unsigned vece, uint32_t dofs, uint32_t aofs,
+                      uint32_t oprsz, uint32_t maxsz)
+{
+    static const GVecGen2 g[4] = {
+        { .fni8 = tcg_gen_vec_neg8_i64,
+          .fniv = tcg_gen_neg_vec,
+          .fno = gen_helper_gvec_neg8,
+          .opc = INDEX_op_neg_vec,
+          .vece = MO_8 },
+        { .fni8 = tcg_gen_vec_neg16_i64,
+          .fniv = tcg_gen_neg_vec,
+          .fno = gen_helper_gvec_neg16,
+          .opc = INDEX_op_neg_vec,
+          .vece = MO_16 },
+        { .fni4 = tcg_gen_neg_i32,
+          .fniv = tcg_gen_neg_vec,
+          .fno = gen_helper_gvec_neg32,
+          .opc = INDEX_op_neg_vec,
+          .vece = MO_32 },
+        { .fni8 = tcg_gen_neg_i64,
+          .fniv = tcg_gen_neg_vec,
+          .fno = gen_helper_gvec_neg64,
+          .opc = INDEX_op_neg_vec,
+          .prefer_i64 = TCG_TARGET_REG_BITS == 64,
+          .vece = MO_64 },
+    };
+
+    tcg_debug_assert(vece <= MO_64);
+    tcg_gen_gvec_2(dofs, aofs, oprsz, maxsz, &g[vece]);
+}
+
+void tcg_gen_gvec_and(unsigned vece, uint32_t dofs, uint32_t aofs,
+                      uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
+{
+    static const GVecGen3 g = {
+        .fni8 = tcg_gen_and_i64,
+        .fniv = tcg_gen_and_vec,
+        .fno = gen_helper_gvec_and,
+        .opc = INDEX_op_and_vec,
+        .prefer_i64 = TCG_TARGET_REG_BITS == 64,
+    };
+    tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
+}
+
+void tcg_gen_gvec_or(unsigned vece, uint32_t dofs, uint32_t aofs,
+                     uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
+{
+    static const GVecGen3 g = {
+        .fni8 = tcg_gen_or_i64,
+        .fniv = tcg_gen_or_vec,
+        .fno = gen_helper_gvec_or,
+        .opc = INDEX_op_or_vec,
+        .prefer_i64 = TCG_TARGET_REG_BITS == 64,
+    };
+    tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
+}
+
+void tcg_gen_gvec_xor(unsigned vece, uint32_t dofs, uint32_t aofs,
+                      uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
+{
+    static const GVecGen3 g = {
+        .fni8 = tcg_gen_xor_i64,
+        .fniv = tcg_gen_xor_vec,
+        .fno = gen_helper_gvec_xor,
+        .opc = INDEX_op_xor_vec,
+        .prefer_i64 = TCG_TARGET_REG_BITS == 64,
+    };
+    tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
+}
+
+void tcg_gen_gvec_andc(unsigned vece, uint32_t dofs, uint32_t aofs,
+                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
+{
+    static const GVecGen3 g = {
+        .fni8 = tcg_gen_andc_i64,
+        .fniv = tcg_gen_andc_vec,
+        .fno = gen_helper_gvec_andc,
+        .opc = INDEX_op_andc_vec,
+        .prefer_i64 = TCG_TARGET_REG_BITS == 64,
+    };
+    tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
+}
+
+void tcg_gen_gvec_orc(unsigned vece, uint32_t dofs, uint32_t aofs,
+                      uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
+{
+    static const GVecGen3 g = {
+        .fni8 = tcg_gen_orc_i64,
+        .fniv = tcg_gen_orc_vec,
+        .fno = gen_helper_gvec_orc,
+        .opc = INDEX_op_orc_vec,
+        .prefer_i64 = TCG_TARGET_REG_BITS == 64,
+    };
+    tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
+}
diff --git a/tcg/tcg-op-gvec.h b/tcg/tcg-op-gvec.h
new file mode 100644
index 0000000000..5a7d640a9d
--- /dev/null
+++ b/tcg/tcg-op-gvec.h
@@ -0,0 +1,198 @@
+/*
+ * Generic vector operation expansion
+ *
+ * Copyright (c) 2018 Linaro
+ *
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2 of the License, or (at your option) any later version.
+ *
+ * This library is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with this library; if not, see <http://www.gnu.org/licenses/>.
+ */
+
+/*
+ * "Generic" vectors.  All operands are given as offsets from ENV,
+ * and therefore cannot also be allocated via tcg_global_mem_new_*.
+ * OPRSZ is the byte size of the vector upon which the operation is performed.
+ * MAXSZ is the byte size of the full vector; bytes beyond OPSZ are cleared.
+ *
+ * All sizes must be 8 or any multiple of 16.
+ * When OPRSZ is 8, the alignment may be 8, otherwise must be 16.
+ * Operands may completely, but not partially, overlap.
+ */
+
+/* Expand a call to a gvec-style helper, with pointers to two vector
+   operands, and a descriptor (see tcg-gvec-desc.h).  */
+typedef void gen_helper_gvec_2(TCGv_ptr, TCGv_ptr, TCGv_i32);
+void tcg_gen_gvec_2_ool(uint32_t dofs, uint32_t aofs,
+                        uint32_t oprsz, uint32_t maxsz, int32_t data,
+                        gen_helper_gvec_2 *fn);
+
+/* Similarly, passing an extra pointer (e.g. env or float_status).  */
+typedef void gen_helper_gvec_2_ptr(TCGv_ptr, TCGv_ptr, TCGv_ptr, TCGv_i32);
+void tcg_gen_gvec_2_ptr(uint32_t dofs, uint32_t aofs,
+                        TCGv_ptr ptr, uint32_t oprsz, uint32_t maxsz,
+                        int32_t data, gen_helper_gvec_2_ptr *fn);
+
+/* Similarly, with three vector operands.  */
+typedef void gen_helper_gvec_3(TCGv_ptr, TCGv_ptr, TCGv_ptr, TCGv_i32);
+void tcg_gen_gvec_3_ool(uint32_t dofs, uint32_t aofs, uint32_t bofs,
+                        uint32_t oprsz, uint32_t maxsz, int32_t data,
+                        gen_helper_gvec_3 *fn);
+
+/* Similarly, with four vector operands.  */
+typedef void gen_helper_gvec_4(TCGv_ptr, TCGv_ptr, TCGv_ptr,
+                               TCGv_ptr, TCGv_i32);
+void tcg_gen_gvec_4_ool(uint32_t dofs, uint32_t aofs, uint32_t bofs,
+                        uint32_t cofs, uint32_t oprsz, uint32_t maxsz,
+                        int32_t data, gen_helper_gvec_4 *fn);
+
+/* Similarly, with five vector operands.  */
+typedef void gen_helper_gvec_5(TCGv_ptr, TCGv_ptr, TCGv_ptr, TCGv_ptr,
+                               TCGv_ptr, TCGv_i32);
+void tcg_gen_gvec_5_ool(uint32_t dofs, uint32_t aofs, uint32_t bofs,
+                        uint32_t cofs, uint32_t xofs, uint32_t oprsz,
+                        uint32_t maxsz, int32_t data, gen_helper_gvec_5 *fn);
+
+typedef void gen_helper_gvec_3_ptr(TCGv_ptr, TCGv_ptr, TCGv_ptr,
+                                   TCGv_ptr, TCGv_i32);
+void tcg_gen_gvec_3_ptr(uint32_t dofs, uint32_t aofs, uint32_t bofs,
+                        TCGv_ptr ptr, uint32_t oprsz, uint32_t maxsz,
+                        int32_t data, gen_helper_gvec_3_ptr *fn);
+
+typedef void gen_helper_gvec_4_ptr(TCGv_ptr, TCGv_ptr, TCGv_ptr,
+                                   TCGv_ptr, TCGv_ptr, TCGv_i32);
+void tcg_gen_gvec_4_ptr(uint32_t dofs, uint32_t aofs, uint32_t bofs,
+                        uint32_t cofs, TCGv_ptr ptr, uint32_t oprsz,
+                        uint32_t maxsz, int32_t data,
+                        gen_helper_gvec_4_ptr *fn);
+
+/* Expand a gvec operation.  Either inline or out-of-line depending on
+   the actual vector size and the operations supported by the host.  */
+typedef struct {
+    /* Expand inline as a 64-bit or 32-bit integer.
+       Only one of these will be non-NULL.  */
+    void (*fni8)(TCGv_i64, TCGv_i64);
+    void (*fni4)(TCGv_i32, TCGv_i32);
+    /* Expand inline with a host vector type.  */
+    void (*fniv)(unsigned, TCGv_vec, TCGv_vec);
+    /* Expand out-of-line helper w/descriptor.  */
+    gen_helper_gvec_2 *fno;
+    /* The opcode, if any, to which this corresponds.  */
+    TCGOpcode opc;
+    /* The data argument to the out-of-line helper.  */
+    int32_t data;
+    /* The vector element size, if applicable.  */
+    uint8_t vece;
+    /* Prefer i64 to v64.  */
+    bool prefer_i64;
+} GVecGen2;
+
+typedef struct {
+    /* Expand inline as a 64-bit or 32-bit integer.
+       Only one of these will be non-NULL.  */
+    void (*fni8)(TCGv_i64, TCGv_i64, TCGv_i64);
+    void (*fni4)(TCGv_i32, TCGv_i32, TCGv_i32);
+    /* Expand inline with a host vector type.  */
+    void (*fniv)(unsigned, TCGv_vec, TCGv_vec, TCGv_vec);
+    /* Expand out-of-line helper w/descriptor.  */
+    gen_helper_gvec_3 *fno;
+    /* The opcode, if any, to which this corresponds.  */
+    TCGOpcode opc;
+    /* The data argument to the out-of-line helper.  */
+    int32_t data;
+    /* The vector element size, if applicable.  */
+    uint8_t vece;
+    /* Prefer i64 to v64.  */
+    bool prefer_i64;
+    /* Load dest as a 3rd source operand.  */
+    bool load_dest;
+} GVecGen3;
+
+typedef struct {
+    /* Expand inline as a 64-bit or 32-bit integer.
+       Only one of these will be non-NULL.  */
+    void (*fni8)(TCGv_i64, TCGv_i64, TCGv_i64, TCGv_i64);
+    void (*fni4)(TCGv_i32, TCGv_i32, TCGv_i32, TCGv_i32);
+    /* Expand inline with a host vector type.  */
+    void (*fniv)(unsigned, TCGv_vec, TCGv_vec, TCGv_vec, TCGv_vec);
+    /* Expand out-of-line helper w/descriptor.  */
+    gen_helper_gvec_4 *fno;
+    /* The opcode, if any, to which this corresponds.  */
+    TCGOpcode opc;
+    /* The data argument to the out-of-line helper.  */
+    int32_t data;
+    /* The vector element size, if applicable.  */
+    uint8_t vece;
+    /* Prefer i64 to v64.  */
+    bool prefer_i64;
+} GVecGen4;
+
+void tcg_gen_gvec_2(uint32_t dofs, uint32_t aofs,
+                    uint32_t oprsz, uint32_t maxsz, const GVecGen2 *);
+void tcg_gen_gvec_3(uint32_t dofs, uint32_t aofs, uint32_t bofs,
+                    uint32_t oprsz, uint32_t maxsz, const GVecGen3 *);
+void tcg_gen_gvec_4(uint32_t dofs, uint32_t aofs, uint32_t bofs, uint32_t cofs,
+                    uint32_t oprsz, uint32_t maxsz, const GVecGen4 *);
+
+/* Expand a specific vector operation.  */
+
+void tcg_gen_gvec_mov(unsigned vece, uint32_t dofs, uint32_t aofs,
+                      uint32_t oprsz, uint32_t maxsz);
+void tcg_gen_gvec_not(unsigned vece, uint32_t dofs, uint32_t aofs,
+                      uint32_t oprsz, uint32_t maxsz);
+void tcg_gen_gvec_neg(unsigned vece, uint32_t dofs, uint32_t aofs,
+                      uint32_t oprsz, uint32_t maxsz);
+
+void tcg_gen_gvec_add(unsigned vece, uint32_t dofs, uint32_t aofs,
+                      uint32_t bofs, uint32_t oprsz, uint32_t maxsz);
+void tcg_gen_gvec_sub(unsigned vece, uint32_t dofs, uint32_t aofs,
+                      uint32_t bofs, uint32_t oprsz, uint32_t maxsz);
+
+void tcg_gen_gvec_and(unsigned vece, uint32_t dofs, uint32_t aofs,
+                      uint32_t bofs, uint32_t oprsz, uint32_t maxsz);
+void tcg_gen_gvec_or(unsigned vece, uint32_t dofs, uint32_t aofs,
+                     uint32_t bofs, uint32_t oprsz, uint32_t maxsz);
+void tcg_gen_gvec_xor(unsigned vece, uint32_t dofs, uint32_t aofs,
+                      uint32_t bofs, uint32_t oprsz, uint32_t maxsz);
+void tcg_gen_gvec_andc(unsigned vece, uint32_t dofs, uint32_t aofs,
+                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz);
+void tcg_gen_gvec_orc(unsigned vece, uint32_t dofs, uint32_t aofs,
+                      uint32_t bofs, uint32_t oprsz, uint32_t maxsz);
+
+void tcg_gen_gvec_dup_mem(unsigned vece, uint32_t dofs, uint32_t aofs,
+                          uint32_t s, uint32_t m);
+void tcg_gen_gvec_dup_i32(unsigned vece, uint32_t dofs, uint32_t s,
+                          uint32_t m, TCGv_i32);
+void tcg_gen_gvec_dup_i64(unsigned vece, uint32_t dofs, uint32_t s,
+                          uint32_t m, TCGv_i64);
+
+void tcg_gen_gvec_dup8i(uint32_t dofs, uint32_t s, uint32_t m, uint8_t x);
+void tcg_gen_gvec_dup16i(uint32_t dofs, uint32_t s, uint32_t m, uint16_t x);
+void tcg_gen_gvec_dup32i(uint32_t dofs, uint32_t s, uint32_t m, uint32_t x);
+void tcg_gen_gvec_dup64i(uint32_t dofs, uint32_t s, uint32_t m, uint64_t x);
+
+/*
+ * 64-bit vector operations.  Use these when the register has been allocated
+ * with tcg_global_mem_new_i64, and so we cannot also address it via pointer.
+ * OPRSZ = MAXSZ = 8.
+ */
+
+void tcg_gen_vec_neg8_i64(TCGv_i64 d, TCGv_i64 a);
+void tcg_gen_vec_neg16_i64(TCGv_i64 d, TCGv_i64 a);
+void tcg_gen_vec_neg32_i64(TCGv_i64 d, TCGv_i64 a);
+
+void tcg_gen_vec_add8_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b);
+void tcg_gen_vec_add16_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b);
+void tcg_gen_vec_add32_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b);
+
+void tcg_gen_vec_sub8_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b);
+void tcg_gen_vec_sub16_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b);
+void tcg_gen_vec_sub32_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b);
diff --git a/tcg/tcg-op-vec.c b/tcg/tcg-op-vec.c
index 9e4678878b..ac5b69ccf6 100644
--- a/tcg/tcg-op-vec.c
+++ b/tcg/tcg-op-vec.c
@@ -73,7 +73,8 @@ static void vec_gen_op2(TCGOpcode opc, unsigned vece, TCGv_vec r, TCGv_vec a)
     TCGTemp *at = tcgv_vec_temp(a);
     TCGType type = rt->base_type;
 
-    tcg_debug_assert(at->base_type == type);
+    /* Must enough inputs for the output.  */
+    tcg_debug_assert(at->base_type >= type);
     vec_gen_2(opc, type, vece, temp_arg(rt), temp_arg(at));
 }
 
@@ -85,8 +86,9 @@ static void vec_gen_op3(TCGOpcode opc, unsigned vece,
     TCGTemp *bt = tcgv_vec_temp(b);
     TCGType type = rt->base_type;
 
-    tcg_debug_assert(at->base_type == type);
-    tcg_debug_assert(bt->base_type == type);
+    /* Must enough inputs for the output.  */
+    tcg_debug_assert(at->base_type >= type);
+    tcg_debug_assert(bt->base_type >= type);
     vec_gen_3(opc, type, vece, temp_arg(rt), temp_arg(at), temp_arg(bt));
 }
 
@@ -99,7 +101,7 @@ void tcg_gen_mov_vec(TCGv_vec r, TCGv_vec a)
 
 #define MO_REG  (TCG_TARGET_REG_BITS == 64 ? MO_64 : MO_32)
 
-static void tcg_gen_dupi_vec(TCGv_vec r, unsigned vece, TCGArg a)
+static void do_dupi_vec(TCGv_vec r, unsigned vece, TCGArg a)
 {
     TCGTemp *rt = tcgv_vec_temp(r);
     vec_gen_2(INDEX_op_dupi_vec, rt->base_type, vece, temp_arg(rt), a);
@@ -108,14 +110,14 @@ static void tcg_gen_dupi_vec(TCGv_vec r, unsigned vece, TCGArg a)
 TCGv_vec tcg_const_zeros_vec(TCGType type)
 {
     TCGv_vec ret = tcg_temp_new_vec(type);
-    tcg_gen_dupi_vec(ret, MO_REG, 0);
+    do_dupi_vec(ret, MO_REG, 0);
     return ret;
 }
 
 TCGv_vec tcg_const_ones_vec(TCGType type)
 {
     TCGv_vec ret = tcg_temp_new_vec(type);
-    tcg_gen_dupi_vec(ret, MO_REG, -1);
+    do_dupi_vec(ret, MO_REG, -1);
     return ret;
 }
 
@@ -134,9 +136,9 @@ TCGv_vec tcg_const_ones_vec_matching(TCGv_vec m)
 void tcg_gen_dup64i_vec(TCGv_vec r, uint64_t a)
 {
     if (TCG_TARGET_REG_BITS == 32 && a == deposit64(a, 32, 32, a)) {
-        tcg_gen_dupi_vec(r, MO_32, a);
+        do_dupi_vec(r, MO_32, a);
     } else if (TCG_TARGET_REG_BITS == 64 || a == (uint64_t)(int32_t)a) {
-        tcg_gen_dupi_vec(r, MO_64, a);
+        do_dupi_vec(r, MO_64, a);
     } else {
         TCGv_i64 c = tcg_const_i64(a);
         tcg_gen_dup_i64_vec(MO_64, r, c);
@@ -146,17 +148,22 @@ void tcg_gen_dup64i_vec(TCGv_vec r, uint64_t a)
 
 void tcg_gen_dup32i_vec(TCGv_vec r, uint32_t a)
 {
-    tcg_gen_dupi_vec(r, MO_REG, ((TCGArg)-1 / 0xffffffffu) * a);
+    do_dupi_vec(r, MO_REG, dup_const(MO_32, a));
 }
 
 void tcg_gen_dup16i_vec(TCGv_vec r, uint32_t a)
 {
-    tcg_gen_dupi_vec(r, MO_REG, ((TCGArg)-1 / 0xffff) * (a & 0xffff));
+    do_dupi_vec(r, MO_REG, dup_const(MO_16, a));
 }
 
 void tcg_gen_dup8i_vec(TCGv_vec r, uint32_t a)
 {
-    tcg_gen_dupi_vec(r, MO_REG, ((TCGArg)-1 / 0xff) * (a & 0xff));
+    do_dupi_vec(r, MO_REG, dup_const(MO_8, a));
+}
+
+void tcg_gen_dupi_vec(unsigned vece, TCGv_vec r, uint64_t a)
+{
+    do_dupi_vec(r, MO_REG, dup_const(vece, a));
 }
 
 void tcg_gen_dup_i64_vec(unsigned vece, TCGv_vec r, TCGv_i64 a)
@@ -167,14 +174,14 @@ void tcg_gen_dup_i64_vec(unsigned vece, TCGv_vec r, TCGv_i64 a)
 
     if (TCG_TARGET_REG_BITS == 64) {
         TCGArg ai = tcgv_i64_arg(a);
-        vec_gen_2(INDEX_op_dup_vec, type, MO_64, ri, ai);
+        vec_gen_2(INDEX_op_dup_vec, type, vece, ri, ai);
     } else if (vece == MO_64) {
         TCGArg al = tcgv_i32_arg(TCGV_LOW(a));
         TCGArg ah = tcgv_i32_arg(TCGV_HIGH(a));
         vec_gen_3(INDEX_op_dup2_vec, type, MO_64, ri, al, ah);
     } else {
         TCGArg ai = tcgv_i32_arg(TCGV_LOW(a));
-        vec_gen_2(INDEX_op_dup_vec, type, MO_64, ri, ai);
+        vec_gen_2(INDEX_op_dup_vec, type, vece, ri, ai);
     }
 }
 
diff --git a/tcg/tcg-op.h b/tcg/tcg-op.h
index a684ab5890..f8ba63340e 100644
--- a/tcg/tcg-op.h
+++ b/tcg/tcg-op.h
@@ -914,6 +914,7 @@ void tcg_gen_dup8i_vec(TCGv_vec, uint32_t);
 void tcg_gen_dup16i_vec(TCGv_vec, uint32_t);
 void tcg_gen_dup32i_vec(TCGv_vec, uint32_t);
 void tcg_gen_dup64i_vec(TCGv_vec, uint64_t);
+void tcg_gen_dupi_vec(unsigned vece, TCGv_vec, uint64_t);
 void tcg_gen_add_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b);
 void tcg_gen_sub_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b);
 void tcg_gen_and_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b);
diff --git a/tcg/tcg-opc.h b/tcg/tcg-opc.h
index b851ad4bca..801b0b1e16 100644
--- a/tcg/tcg-opc.h
+++ b/tcg/tcg-opc.h
@@ -228,6 +228,12 @@ DEF(andc_vec, 1, 2, 0, IMPLVEC | IMPL(TCG_TARGET_HAS_andc_vec))
 DEF(orc_vec, 1, 2, 0, IMPLVEC | IMPL(TCG_TARGET_HAS_orc_vec))
 DEF(not_vec, 1, 1, 0, IMPLVEC | IMPL(TCG_TARGET_HAS_not_vec))
 
+DEF(last_generic, 0, 0, 0, TCG_OPF_NOT_PRESENT)
+
+#if TCG_TARGET_MAYBE_vec
+#include "tcg-target.opc.h"
+#endif
+
 #undef TLADDR_ARGS
 #undef DATA64_ARGS
 #undef IMPL
diff --git a/tcg/tcg.c b/tcg/tcg.c
index 42f0acdf8e..0862cff58a 100644
--- a/tcg/tcg.c
+++ b/tcg/tcg.c
@@ -1403,10 +1403,10 @@ bool tcg_op_supported(TCGOpcode op)
     case INDEX_op_orc_vec:
         return have_vec && TCG_TARGET_HAS_orc_vec;
 
-    case NB_OPS:
-        break;
+    default:
+        tcg_debug_assert(op > INDEX_op_last_generic && op < NB_OPS);
+        return true;
     }
-    g_assert_not_reached();
 }
 
 /* Note: we convert the 64 bit args to 32 bit and do some alignment
@@ -3733,3 +3733,10 @@ void tcg_register_jit(void *buf, size_t buf_size)
 {
 }
 #endif /* ELF_HOST_MACHINE */
+
+#if !TCG_TARGET_MAYBE_vec
+void tcg_expand_vec_op(TCGOpcode o, TCGType t, unsigned e, TCGArg a0, ...)
+{
+    g_assert_not_reached();
+}
+#endif
diff --git a/tcg/tcg.h b/tcg/tcg.h
index dce483b0ee..ec8f1bc72e 100644
--- a/tcg/tcg.h
+++ b/tcg/tcg.h
@@ -1207,6 +1207,33 @@ uintptr_t tcg_qemu_tb_exec(CPUArchState *env, uint8_t *tb_ptr);
 
 void tcg_register_jit(void *buf, size_t buf_size);
 
+#if TCG_TARGET_MAYBE_vec
+/* Return zero if the tuple (opc, type, vece) is unsupportable;
+   return > 0 if it is directly supportable;
+   return < 0 if we must call tcg_expand_vec_op.  */
+int tcg_can_emit_vec_op(TCGOpcode, TCGType, unsigned);
+#else
+static inline int tcg_can_emit_vec_op(TCGOpcode o, TCGType t, unsigned ve)
+{
+    return 0;
+}
+#endif
+
+/* Expand the tuple (opc, type, vece) on the given arguments.  */
+void tcg_expand_vec_op(TCGOpcode, TCGType, unsigned, TCGArg, ...);
+
+/* Replicate a constant C accoring to the log2 of the element size.  */
+uint64_t dup_const(unsigned vece, uint64_t c);
+
+#define dup_const(VECE, C)                                         \
+    (__builtin_constant_p(VECE)                                    \
+     ? (  (VECE) == MO_8  ? 0x0101010101010101ull * (uint8_t)(C)   \
+        : (VECE) == MO_16 ? 0x0001000100010001ull * (uint16_t)(C)  \
+        : (VECE) == MO_32 ? 0x0000000100000001ull * (uint32_t)(C)  \
+        : dup_const(VECE, C))                                      \
+     : dup_const(VECE, C))
+
+
 /*
  * Memory helpers that will be used by TCG generated code.
  */