[PULL 06/23] target/arm: Convert Neon 3-reg-diff saturating doubling multiplies

[Peter Maydell]

Convert the Neon 3-reg-diff insns VQDMULL, VQDMLAL and VQDMLSL:
these are all saturating doubling long multiplies with a possible
accumulate step.

These are the last insns in the group which use the pass-over-each
elements loop, so we can delete that code.

Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
---
 target/arm/neon-dp.decode       |  6 +++
 target/arm/translate-neon.inc.c | 82 +++++++++++++++++++++++++++++++++
 target/arm/translate.c          | 59 ++----------------------
 3 files changed, 92 insertions(+), 55 deletions(-)

diff --git a/target/arm/neon-dp.decode b/target/arm/neon-dp.decode
index 1da492df146..65ea30d3edf 100644
--- a/target/arm/neon-dp.decode
+++ b/target/arm/neon-dp.decode
@@ -454,10 +454,16 @@ Vimm_1r          1111 001 . 1 . 000 ... .... cmode:4 0 . 
op:1 1 .... @1reg_imm
     VMLAL_S_3d   1111 001 0 1 . .. .... .... 1000 . 0 . 0 .... @3diff
     VMLAL_U_3d   1111 001 1 1 . .. .... .... 1000 . 0 . 0 .... @3diff
 
+    VQDMLAL_3d   1111 001 0 1 . .. .... .... 1001 . 0 . 0 .... @3diff
+
     VMLSL_S_3d   1111 001 0 1 . .. .... .... 1010 . 0 . 0 .... @3diff
     VMLSL_U_3d   1111 001 1 1 . .. .... .... 1010 . 0 . 0 .... @3diff
 
+    VQDMLSL_3d   1111 001 0 1 . .. .... .... 1011 . 0 . 0 .... @3diff
+
     VMULL_S_3d   1111 001 0 1 . .. .... .... 1100 . 0 . 0 .... @3diff
     VMULL_U_3d   1111 001 1 1 . .. .... .... 1100 . 0 . 0 .... @3diff
+
+    VQDMULL_3d   1111 001 0 1 . .. .... .... 1101 . 0 . 0 .... @3diff
   ]
 }
diff --git a/target/arm/translate-neon.inc.c b/target/arm/translate-neon.inc.c
index c435f0685d4..083e3af8c56 100644
--- a/target/arm/translate-neon.inc.c
+++ b/target/arm/translate-neon.inc.c
@@ -2222,3 +2222,85 @@ DO_VMLAL(VMLAL_S,mull_s,add)
 DO_VMLAL(VMLAL_U,mull_u,add)
 DO_VMLAL(VMLSL_S,mull_s,sub)
 DO_VMLAL(VMLSL_U,mull_u,sub)
+
+static void gen_VQDMULL_16(TCGv_i64 rd, TCGv_i32 rn, TCGv_i32 rm)
+{
+    gen_helper_neon_mull_s16(rd, rn, rm);
+    gen_helper_neon_addl_saturate_s32(rd, cpu_env, rd, rd);
+}
+
+static void gen_VQDMULL_32(TCGv_i64 rd, TCGv_i32 rn, TCGv_i32 rm)
+{
+    gen_mull_s32(rd, rn, rm);
+    gen_helper_neon_addl_saturate_s64(rd, cpu_env, rd, rd);
+}
+
+static bool trans_VQDMULL_3d(DisasContext *s, arg_3diff *a)
+{
+    static NeonGenTwoOpWidenFn * const opfn[] = {
+        NULL,
+        gen_VQDMULL_16,
+        gen_VQDMULL_32,
+        NULL,
+    };
+
+    return do_long_3d(s, a, opfn[a->size], NULL);
+}
+
+static void gen_VQDMLAL_acc_16(TCGv_i64 rd, TCGv_i64 rn, TCGv_i64 rm)
+{
+    gen_helper_neon_addl_saturate_s32(rd, cpu_env, rn, rm);
+}
+
+static void gen_VQDMLAL_acc_32(TCGv_i64 rd, TCGv_i64 rn, TCGv_i64 rm)
+{
+    gen_helper_neon_addl_saturate_s64(rd, cpu_env, rn, rm);
+}
+
+static bool trans_VQDMLAL_3d(DisasContext *s, arg_3diff *a)
+{
+    static NeonGenTwoOpWidenFn * const opfn[] = {
+        NULL,
+        gen_VQDMULL_16,
+        gen_VQDMULL_32,
+        NULL,
+    };
+    static NeonGenTwo64OpFn * const accfn[] = {
+        NULL,
+        gen_VQDMLAL_acc_16,
+        gen_VQDMLAL_acc_32,
+        NULL,
+    };
+
+    return do_long_3d(s, a, opfn[a->size], accfn[a->size]);
+}
+
+static void gen_VQDMLSL_acc_16(TCGv_i64 rd, TCGv_i64 rn, TCGv_i64 rm)
+{
+    gen_helper_neon_negl_u32(rm, rm);
+    gen_helper_neon_addl_saturate_s32(rd, cpu_env, rn, rm);
+}
+
+static void gen_VQDMLSL_acc_32(TCGv_i64 rd, TCGv_i64 rn, TCGv_i64 rm)
+{
+    tcg_gen_neg_i64(rm, rm);
+    gen_helper_neon_addl_saturate_s64(rd, cpu_env, rn, rm);
+}
+
+static bool trans_VQDMLSL_3d(DisasContext *s, arg_3diff *a)
+{
+    static NeonGenTwoOpWidenFn * const opfn[] = {
+        NULL,
+        gen_VQDMULL_16,
+        gen_VQDMULL_32,
+        NULL,
+    };
+    static NeonGenTwo64OpFn * const accfn[] = {
+        NULL,
+        gen_VQDMLSL_acc_16,
+        gen_VQDMLSL_acc_32,
+        NULL,
+    };
+
+    return do_long_3d(s, a, opfn[a->size], accfn[a->size]);
+}
diff --git a/target/arm/translate.c b/target/arm/translate.c
index a2c47d19f21..88e91845c02 100644
--- a/target/arm/translate.c
+++ b/target/arm/translate.c
@@ -5247,11 +5247,11 @@ static int disas_neon_data_insn(DisasContext *s, 
uint32_t insn)
                     {0, 0, 0, 7}, /* VSUBHN: handled by decodetree */
                     {0, 0, 0, 7}, /* VABDL */
                     {0, 0, 0, 7}, /* VMLAL */
-                    {0, 0, 0, 9}, /* VQDMLAL */
+                    {0, 0, 0, 7}, /* VQDMLAL */
                     {0, 0, 0, 7}, /* VMLSL */
-                    {0, 0, 0, 9}, /* VQDMLSL */
+                    {0, 0, 0, 7}, /* VQDMLSL */
                     {0, 0, 0, 7}, /* Integer VMULL */
-                    {0, 0, 0, 9}, /* VQDMULL */
+                    {0, 0, 0, 7}, /* VQDMULL */
                     {0, 0, 0, 0xa}, /* Polynomial VMULL */
                     {0, 0, 0, 7}, /* Reserved: always UNDEF */
                 };
@@ -5282,58 +5282,7 @@ static int disas_neon_data_insn(DisasContext *s, 
uint32_t insn)
                     }
                     return 0;
                 }
-
-                /* Avoid overlapping operands.  Wide source operands are
-                   always aligned so will never overlap with wide
-                   destinations in problematic ways.  */
-                if (rd == rm) {
-                    tmp = neon_load_reg(rm, 1);
-                    neon_store_scratch(2, tmp);
-                } else if (rd == rn) {
-                    tmp = neon_load_reg(rn, 1);
-                    neon_store_scratch(2, tmp);
-                }
-                tmp3 = NULL;
-                for (pass = 0; pass < 2; pass++) {
-                    if (pass == 1 && rd == rn) {
-                        tmp = neon_load_scratch(2);
-                    } else {
-                        tmp = neon_load_reg(rn, pass);
-                    }
-                    if (pass == 1 && rd == rm) {
-                        tmp2 = neon_load_scratch(2);
-                    } else {
-                        tmp2 = neon_load_reg(rm, pass);
-                    }
-                    switch (op) {
-                    case 9: case 11: case 13:
-                        /* VQDMLAL, VQDMLSL, VQDMULL */
-                        gen_neon_mull(cpu_V0, tmp, tmp2, size, u);
-                        break;
-                    default: /* 15 is RESERVED: caught earlier  */
-                        abort();
-                    }
-                    if (op == 13) {
-                        /* VQDMULL */
-                        gen_neon_addl_saturate(cpu_V0, cpu_V0, size);
-                        neon_store_reg64(cpu_V0, rd + pass);
-                    } else {
-                        /* Accumulate.  */
-                        neon_load_reg64(cpu_V1, rd + pass);
-                        switch (op) {
-                        case 9: case 11: /* VQDMLAL, VQDMLSL */
-                            gen_neon_addl_saturate(cpu_V0, cpu_V0, size);
-                            if (op == 11) {
-                                gen_neon_negl(cpu_V0, size);
-                            }
-                            gen_neon_addl_saturate(cpu_V0, cpu_V1, size);
-                            break;
-                        default:
-                            abort();
-                        }
-                        neon_store_reg64(cpu_V0, rd + pass);
-                    }
-                }
+                abort(); /* all others handled by decodetree */
             } else {
                 /* Two registers and a scalar. NB that for ops of this form
                  * the ARM ARM labels bit 24 as Q, but it is in our variable
-- 
2.20.1