[Qemu-devel] [PATCH for-4.1 v2 07/13] tcg: Add INDEX_op_dup_mem_vec

[Richard Henderson]

Allow the backend to expand dup from memory directly, instead of
forcing the value into a temp first.  This is especially important
if integer/vector register moves do not exist.

Signed-off-by: Richard Henderson <address@hidden>
---
 tcg/aarch64/tcg-target.h |  1 +
 tcg/i386/tcg-target.h    |  1 +
 tcg/ppc/tcg-target.h     |  1 +
 tcg/tcg-op.h             |  1 +
 tcg/tcg-opc.h            |  1 +
 tcg/tcg.h                |  1 +
 tcg/tcg-op-gvec.c        | 88 +++++++++++++++++++++-------------------
 tcg/tcg-op-vec.c         | 11 +++++
 tcg/tcg.c                |  2 +
 9 files changed, 66 insertions(+), 41 deletions(-)

diff --git a/tcg/aarch64/tcg-target.h b/tcg/aarch64/tcg-target.h
index 2d93cf404e..8ce99fc9c8 100644
--- a/tcg/aarch64/tcg-target.h
+++ b/tcg/aarch64/tcg-target.h
@@ -137,6 +137,7 @@ typedef enum {
 #define TCG_TARGET_HAS_mul_vec          1
 #define TCG_TARGET_HAS_sat_vec          1
 #define TCG_TARGET_HAS_minmax_vec       1
+#define TCG_TARGET_HAS_dupm_vec         0
 
 #define TCG_TARGET_DEFAULT_MO (0)
 #define TCG_TARGET_HAS_MEMORY_BSWAP     1
diff --git a/tcg/i386/tcg-target.h b/tcg/i386/tcg-target.h
index 7995fe3eab..8e8d59f4f4 100644
--- a/tcg/i386/tcg-target.h
+++ b/tcg/i386/tcg-target.h
@@ -187,6 +187,7 @@ extern bool have_avx2;
 #define TCG_TARGET_HAS_mul_vec          1
 #define TCG_TARGET_HAS_sat_vec          1
 #define TCG_TARGET_HAS_minmax_vec       1
+#define TCG_TARGET_HAS_dupm_vec         0
 
 #define TCG_TARGET_deposit_i32_valid(ofs, len) \
     (((ofs) == 0 && (len) == 8) || ((ofs) == 8 && (len) == 8) || \
diff --git a/tcg/ppc/tcg-target.h b/tcg/ppc/tcg-target.h
index 683eb807ae..5143ee853a 100644
--- a/tcg/ppc/tcg-target.h
+++ b/tcg/ppc/tcg-target.h
@@ -152,6 +152,7 @@ extern bool have_isa_3_00;
 #define TCG_TARGET_HAS_mul_vec          1
 #define TCG_TARGET_HAS_sat_vec          1
 #define TCG_TARGET_HAS_minmax_vec       1
+#define TCG_TARGET_HAS_dupm_vec         0
 
 void flush_icache_range(uintptr_t start, uintptr_t stop);
 void tb_target_set_jmp_target(uintptr_t, uintptr_t, uintptr_t);
diff --git a/tcg/tcg-op.h b/tcg/tcg-op.h
index d3e51b15af..64cd3f58ef 100644
--- a/tcg/tcg-op.h
+++ b/tcg/tcg-op.h
@@ -950,6 +950,7 @@ void tcg_gen_atomic_umax_fetch_i64(TCGv_i64, TCGv, 
TCGv_i64, TCGArg, TCGMemOp);
 void tcg_gen_mov_vec(TCGv_vec, TCGv_vec);
 void tcg_gen_dup_i32_vec(unsigned vece, TCGv_vec, TCGv_i32);
 void tcg_gen_dup_i64_vec(unsigned vece, TCGv_vec, TCGv_i64);
+void tcg_gen_dup_mem_vec(unsigned vece, TCGv_vec, TCGv_ptr, tcg_target_long);
 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);
diff --git a/tcg/tcg-opc.h b/tcg/tcg-opc.h
index 4e0238ad1a..b8ad147377 100644
--- a/tcg/tcg-opc.h
+++ b/tcg/tcg-opc.h
@@ -211,6 +211,7 @@ DEF(qemu_st_i64, 0, TLADDR_ARGS + DATA64_ARGS, 1,
 
 DEF(mov_vec, 1, 1, 0, TCG_OPF_VECTOR | TCG_OPF_NOT_PRESENT)
 DEF(dupi_vec, 1, 0, 1, TCG_OPF_VECTOR | TCG_OPF_NOT_PRESENT)
+DEF(dupm_vec, 1, 1, 1, TCG_OPF_VECTOR | IMPL(TCG_TARGET_HAS_dupm_vec))
 
 DEF(dup_vec, 1, 1, 0, IMPLVEC)
 DEF(dup2_vec, 1, 2, 0, IMPLVEC | IMPL(TCG_TARGET_REG_BITS == 32))
diff --git a/tcg/tcg.h b/tcg/tcg.h
index 32b7cf3489..f7c12de75a 100644
--- a/tcg/tcg.h
+++ b/tcg/tcg.h
@@ -185,6 +185,7 @@ typedef uint64_t TCGRegSet;
 #define TCG_TARGET_HAS_mul_vec          0
 #define TCG_TARGET_HAS_sat_vec          0
 #define TCG_TARGET_HAS_minmax_vec       0
+#define TCG_TARGET_HAS_dupm_vec         0
 #else
 #define TCG_TARGET_MAYBE_vec            1
 #endif
diff --git a/tcg/tcg-op-gvec.c b/tcg/tcg-op-gvec.c
index 0996ef0812..59ab516bf0 100644
--- a/tcg/tcg-op-gvec.c
+++ b/tcg/tcg-op-gvec.c
@@ -390,6 +390,40 @@ static TCGType choose_vector_type(TCGOpcode op, unsigned 
vece, uint32_t size,
     return 0;
 }
 
+static void do_dup_store(TCGType type, uint32_t dofs, uint32_t oprsz,
+                         uint32_t maxsz, TCGv_vec t_vec)
+{
+    uint32_t i = 0;
+
+    switch (type) {
+    case TCG_TYPE_V256:
+        /* Recall that ARM SVE allows vector sizes that are not a
+         * power of 2, but always a multiple of 16.  The intent is
+         * that e.g. size == 80 would be expanded with 2x32 + 1x16.
+         */
+        for (; i + 32 <= oprsz; i += 32) {
+            tcg_gen_stl_vec(t_vec, cpu_env, dofs + i, TCG_TYPE_V256);
+        }
+        /* fallthru */
+    case TCG_TYPE_V128:
+        for (; i + 16 <= oprsz; i += 16) {
+            tcg_gen_stl_vec(t_vec, cpu_env, dofs + i, TCG_TYPE_V128);
+        }
+        break;
+    case TCG_TYPE_V64:
+        for (; i < oprsz; i += 8) {
+            tcg_gen_stl_vec(t_vec, cpu_env, dofs + i, TCG_TYPE_V64);
+        }
+        break;
+    default:
+        g_assert_not_reached();
+    }
+
+    if (oprsz < maxsz) {
+        expand_clr(dofs + oprsz, maxsz - oprsz);
+    }
+}
+
 /* 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.
@@ -429,49 +463,11 @@ static void do_dup(unsigned vece, uint32_t dofs, uint32_t 
oprsz,
         } 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;
-            }
+            tcg_gen_dupi_vec(vece, t_vec, in_c);
         }
-
-        i = 0;
-        switch (type) {
-        case TCG_TYPE_V256:
-            /* Recall that ARM SVE allows vector sizes that are not a
-             * power of 2, but always a multiple of 16.  The intent is
-             * that e.g. size == 80 would be expanded with 2x32 + 1x16.
-             */
-            for (; i + 32 <= oprsz; i += 32) {
-                tcg_gen_stl_vec(t_vec, cpu_env, dofs + i, TCG_TYPE_V256);
-            }
-            /* fallthru */
-        case TCG_TYPE_V128:
-            for (; i + 16 <= oprsz; i += 16) {
-                tcg_gen_stl_vec(t_vec, cpu_env, dofs + i, TCG_TYPE_V128);
-            }
-            break;
-        case TCG_TYPE_V64:
-            for (; i < oprsz; i += 8) {
-                tcg_gen_stl_vec(t_vec, cpu_env, dofs + i, TCG_TYPE_V64);
-            }
-            break;
-        default:
-            g_assert_not_reached();
-        }
-
+        do_dup_store(type, dofs, oprsz, maxsz, t_vec);
         tcg_temp_free_vec(t_vec);
-        goto done;
+        return;
     }
 
     /* Otherwise, inline with an integer type, unless "large".  */
@@ -1287,6 +1283,16 @@ void tcg_gen_gvec_dup_i64(unsigned vece, uint32_t dofs, 
uint32_t oprsz,
 void tcg_gen_gvec_dup_mem(unsigned vece, uint32_t dofs, uint32_t aofs,
                           uint32_t oprsz, uint32_t maxsz)
 {
+    if (TCG_TARGET_HAS_dupm_vec) {
+        TCGType type = choose_vector_type(INDEX_op_dupm_vec, vece, oprsz, 0);
+        if (type != 0) {
+            TCGv_vec t_vec = tcg_temp_new_vec(type);
+            tcg_gen_dup_mem_vec(vece, t_vec, cpu_env, aofs);
+            do_dup_store(type, dofs, oprsz, maxsz, t_vec);
+            tcg_temp_free_vec(t_vec);
+            return;
+        }
+    }
     if (vece <= MO_32) {
         TCGv_i32 in = tcg_temp_new_i32();
         switch (vece) {
diff --git a/tcg/tcg-op-vec.c b/tcg/tcg-op-vec.c
index cfb18682b1..ce7987b858 100644
--- a/tcg/tcg-op-vec.c
+++ b/tcg/tcg-op-vec.c
@@ -194,6 +194,17 @@ void tcg_gen_dup_i32_vec(unsigned vece, TCGv_vec r, 
TCGv_i32 a)
     vec_gen_2(INDEX_op_dup_vec, type, vece, ri, ai);
 }
 
+void tcg_gen_dup_mem_vec(unsigned vece, TCGv_vec r, TCGv_ptr b,
+                         tcg_target_long ofs)
+{
+    TCGArg ri = tcgv_vec_arg(r);
+    TCGArg bi = tcgv_ptr_arg(b);
+    TCGTemp *rt = arg_temp(ri);
+    TCGType type = rt->base_type;
+
+    vec_gen_3(INDEX_op_dupm_vec, type, vece, ri, bi, ofs);
+}
+
 static void vec_gen_ldst(TCGOpcode opc, TCGv_vec r, TCGv_ptr b, TCGArg o)
 {
     TCGArg ri = tcgv_vec_arg(r);
diff --git a/tcg/tcg.c b/tcg/tcg.c
index b5389ea767..e0d771d610 100644
--- a/tcg/tcg.c
+++ b/tcg/tcg.c
@@ -1623,6 +1623,8 @@ bool tcg_op_supported(TCGOpcode op)
     case INDEX_op_smax_vec:
     case INDEX_op_umax_vec:
         return have_vec && TCG_TARGET_HAS_minmax_vec;
+    case INDEX_op_dupm_vec:
+        return have_vec && TCG_TARGET_HAS_dupm_vec;
 
     default:
         tcg_debug_assert(op > INDEX_op_last_generic && op < NB_OPS);
-- 
2.17.2