[PATCH v5 15/81] target/arm: Implement SVE2 bitwise shift left long

From: Richard Henderson <richard.henderson@linaro.org>
To: qemu-devel@nongnu.org
Cc: qemu-arm@nongnu.org
Subject: [PATCH v5 15/81] target/arm: Implement SVE2 bitwise shift left long
Date: Fri, 16 Apr 2021 14:01:34 -0700	[thread overview]
Message-ID: <20210416210240.1591291-16-richard.henderson@linaro.org> (raw)
In-Reply-To: <20210416210240.1591291-1-richard.henderson@linaro.org>

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
 target/arm/helper-sve.h    |   8 ++
 target/arm/sve.decode      |   8 ++
 target/arm/sve_helper.c    |  26 ++++++
 target/arm/translate-sve.c | 159 +++++++++++++++++++++++++++++++++++++
 4 files changed, 201 insertions(+)

diff --git a/target/arm/helper-sve.h b/target/arm/helper-sve.h
index bf3e533eb4..740939e7a8 100644
--- a/target/arm/helper-sve.h
+++ b/target/arm/helper-sve.h
@@ -2364,3 +2364,11 @@ DEF_HELPER_FLAGS_4(sve2_umull_zzz_d, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
 
 DEF_HELPER_FLAGS_4(sve2_pmull_h, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
 DEF_HELPER_FLAGS_4(sve2_pmull_d, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+
+DEF_HELPER_FLAGS_3(sve2_sshll_h, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+DEF_HELPER_FLAGS_3(sve2_sshll_s, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+DEF_HELPER_FLAGS_3(sve2_sshll_d, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+
+DEF_HELPER_FLAGS_3(sve2_ushll_h, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+DEF_HELPER_FLAGS_3(sve2_ushll_s, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+DEF_HELPER_FLAGS_3(sve2_ushll_d, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
diff --git a/target/arm/sve.decode b/target/arm/sve.decode
index 016c15ebb6..a3191eba7b 100644
--- a/target/arm/sve.decode
+++ b/target/arm/sve.decode
@@ -1207,3 +1207,11 @@ SMULLB_zzz      01000101 .. 0 ..... 011 100 ..... .....  @rd_rn_rm
 SMULLT_zzz      01000101 .. 0 ..... 011 101 ..... .....  @rd_rn_rm
 UMULLB_zzz      01000101 .. 0 ..... 011 110 ..... .....  @rd_rn_rm
 UMULLT_zzz      01000101 .. 0 ..... 011 111 ..... .....  @rd_rn_rm
+
+## SVE2 bitwise shift left long
+
+# Note bit23 == 0 is handled by esz > 0 in do_sve2_shll_tb.
+SSHLLB          01000101 .. 0 ..... 1010 00 ..... .....  @rd_rn_tszimm_shl
+SSHLLT          01000101 .. 0 ..... 1010 01 ..... .....  @rd_rn_tszimm_shl
+USHLLB          01000101 .. 0 ..... 1010 10 ..... .....  @rd_rn_tszimm_shl
+USHLLT          01000101 .. 0 ..... 1010 11 ..... .....  @rd_rn_tszimm_shl
diff --git a/target/arm/sve_helper.c b/target/arm/sve_helper.c
index f30f3722af..73aa670a77 100644
--- a/target/arm/sve_helper.c
+++ b/target/arm/sve_helper.c
@@ -625,6 +625,8 @@ DO_ZPZZ(sve2_sqrshl_zpzz_h, int16_t, H1_2, do_sqrshl_h)
 DO_ZPZZ(sve2_sqrshl_zpzz_s, int32_t, H1_4, do_sqrshl_s)
 DO_ZPZZ_D(sve2_sqrshl_zpzz_d, int64_t, do_sqrshl_d)
 
+#undef do_sqrshl_d
+
 #define do_uqrshl_b(n, m) \
    ({ uint32_t discard; do_uqrshl_bhs(n, (int8_t)m, 8, true, &discard); })
 #define do_uqrshl_h(n, m) \
@@ -639,6 +641,8 @@ DO_ZPZZ(sve2_uqrshl_zpzz_h, uint16_t, H1_2, do_uqrshl_h)
 DO_ZPZZ(sve2_uqrshl_zpzz_s, uint32_t, H1_4, do_uqrshl_s)
 DO_ZPZZ_D(sve2_uqrshl_zpzz_d, uint64_t, do_uqrshl_d)
 
+#undef do_uqrshl_d
+
 #define DO_HADD_BHS(n, m)  (((int64_t)n + m) >> 1)
 #define DO_HADD_D(n, m)    ((n >> 1) + (m >> 1) + (n & m & 1))
 
@@ -1217,6 +1221,28 @@ DO_ZZZ_WTB(sve2_usubw_d, uint64_t, uint32_t,     , H1_4, DO_SUB)
 
 #undef DO_ZZZ_WTB
 
+#define DO_ZZI_SHLL(NAME, TYPEW, TYPEN, HW, HN) \
+void HELPER(NAME)(void *vd, void *vn, uint32_t desc)           \
+{                                                              \
+    intptr_t i, opr_sz = simd_oprsz(desc);                     \
+    intptr_t sel = (simd_data(desc) & 1) * sizeof(TYPEN);      \
+    int shift = simd_data(desc) >> 1;                          \
+    for (i = 0; i < opr_sz; i += sizeof(TYPEW)) {              \
+        TYPEW nn = *(TYPEN *)(vn + HN(i + sel));               \
+        *(TYPEW *)(vd + HW(i)) = nn << shift;                  \
+    }                                                          \
+}
+
+DO_ZZI_SHLL(sve2_sshll_h, int16_t, int8_t, H1_2, H1)
+DO_ZZI_SHLL(sve2_sshll_s, int32_t, int16_t, H1_4, H1_2)
+DO_ZZI_SHLL(sve2_sshll_d, int64_t, int32_t,     , H1_4)
+
+DO_ZZI_SHLL(sve2_ushll_h, uint16_t, uint8_t, H1_2, H1)
+DO_ZZI_SHLL(sve2_ushll_s, uint32_t, uint16_t, H1_4, H1_2)
+DO_ZZI_SHLL(sve2_ushll_d, uint64_t, uint32_t,     , H1_4)
+
+#undef DO_ZZI_SHLL
+
 /* Two-operand reduction expander, controlled by a predicate.
  * The difference between TYPERED and TYPERET has to do with
  * sign-extension.  E.g. for SMAX, TYPERED must be signed,
diff --git a/target/arm/translate-sve.c b/target/arm/translate-sve.c
index fbdccc1c68..afd208212b 100644
--- a/target/arm/translate-sve.c
+++ b/target/arm/translate-sve.c
@@ -6071,3 +6071,162 @@ DO_SVE2_ZZZ_WTB(UADDWB, uaddw, false)
 DO_SVE2_ZZZ_WTB(UADDWT, uaddw, true)
 DO_SVE2_ZZZ_WTB(USUBWB, usubw, false)
 DO_SVE2_ZZZ_WTB(USUBWT, usubw, true)
+
+static void gen_sshll_vec(unsigned vece, TCGv_vec d, TCGv_vec n, int64_t imm)
+{
+    int top = imm & 1;
+    int shl = imm >> 1;
+    int halfbits = 4 << vece;
+
+    if (top) {
+        if (shl == halfbits) {
+            TCGv_vec t = tcg_temp_new_vec_matching(d);
+            tcg_gen_dupi_vec(vece, t, MAKE_64BIT_MASK(halfbits, halfbits));
+            tcg_gen_and_vec(vece, d, n, t);
+            tcg_temp_free_vec(t);
+        } else {
+            tcg_gen_sari_vec(vece, d, n, halfbits);
+            tcg_gen_shli_vec(vece, d, d, shl);
+        }
+    } else {
+        tcg_gen_shli_vec(vece, d, n, halfbits);
+        tcg_gen_sari_vec(vece, d, d, halfbits - shl);
+    }
+}
+
+static void gen_ushll_i64(unsigned vece, TCGv_i64 d, TCGv_i64 n, int imm)
+{
+    int halfbits = 4 << vece;
+    int top = imm & 1;
+    int shl = (imm >> 1);
+    int shift;
+    uint64_t mask;
+
+    mask = MAKE_64BIT_MASK(0, halfbits);
+    mask <<= shl;
+    mask = dup_const(vece, mask);
+
+    shift = shl - top * halfbits;
+    if (shift < 0) {
+        tcg_gen_shri_i64(d, n, -shift);
+    } else {
+        tcg_gen_shri_i64(d, n, shift);
+    }
+    tcg_gen_andi_i64(d, d, mask);
+}
+
+static void gen_ushll16_i64(TCGv_i64 d, TCGv_i64 n, int64_t imm)
+{
+    gen_ushll_i64(MO_16, d, n, imm);
+}
+
+static void gen_ushll32_i64(TCGv_i64 d, TCGv_i64 n, int64_t imm)
+{
+    gen_ushll_i64(MO_32, d, n, imm);
+}
+
+static void gen_ushll64_i64(TCGv_i64 d, TCGv_i64 n, int64_t imm)
+{
+    gen_ushll_i64(MO_64, d, n, imm);
+}
+
+static void gen_ushll_vec(unsigned vece, TCGv_vec d, TCGv_vec n, int64_t imm)
+{
+    int halfbits = 4 << vece;
+    int top = imm & 1;
+    int shl = imm >> 1;
+
+    if (top) {
+        if (shl == halfbits) {
+            TCGv_vec t = tcg_temp_new_vec_matching(d);
+            tcg_gen_dupi_vec(vece, t, MAKE_64BIT_MASK(halfbits, halfbits));
+            tcg_gen_and_vec(vece, d, n, t);
+            tcg_temp_free_vec(t);
+        } else {
+            tcg_gen_shri_vec(vece, d, n, halfbits);
+            tcg_gen_shli_vec(vece, d, d, shl);
+        }
+    } else {
+        if (shl == 0) {
+            TCGv_vec t = tcg_temp_new_vec_matching(d);
+            tcg_gen_dupi_vec(vece, t, MAKE_64BIT_MASK(0, halfbits));
+            tcg_gen_and_vec(vece, d, n, t);
+            tcg_temp_free_vec(t);
+        } else {
+            tcg_gen_shli_vec(vece, d, n, halfbits);
+            tcg_gen_shri_vec(vece, d, d, halfbits - shl);
+        }
+    }
+}
+
+static bool do_sve2_shll_tb(DisasContext *s, arg_rri_esz *a,
+                            bool sel, bool uns)
+{
+    static const TCGOpcode sshll_list[] = {
+        INDEX_op_shli_vec, INDEX_op_sari_vec, 0
+    };
+    static const TCGOpcode ushll_list[] = {
+        INDEX_op_shli_vec, INDEX_op_shri_vec, 0
+    };
+    static const GVecGen2i ops[2][3] = {
+        { { .fniv = gen_sshll_vec,
+            .opt_opc = sshll_list,
+            .fno = gen_helper_sve2_sshll_h,
+            .vece = MO_16 },
+          { .fniv = gen_sshll_vec,
+            .opt_opc = sshll_list,
+            .fno = gen_helper_sve2_sshll_s,
+            .vece = MO_32 },
+          { .fniv = gen_sshll_vec,
+            .opt_opc = sshll_list,
+            .fno = gen_helper_sve2_sshll_d,
+            .vece = MO_64 } },
+        { { .fni8 = gen_ushll16_i64,
+            .fniv = gen_ushll_vec,
+            .opt_opc = ushll_list,
+            .fno = gen_helper_sve2_ushll_h,
+            .vece = MO_16 },
+          { .fni8 = gen_ushll32_i64,
+            .fniv = gen_ushll_vec,
+            .opt_opc = ushll_list,
+            .fno = gen_helper_sve2_ushll_s,
+            .vece = MO_32 },
+          { .fni8 = gen_ushll64_i64,
+            .fniv = gen_ushll_vec,
+            .opt_opc = ushll_list,
+            .fno = gen_helper_sve2_ushll_d,
+            .vece = MO_64 } },
+    };
+
+    if (a->esz < 0 || a->esz > 2 || !dc_isar_feature(aa64_sve2, s)) {
+        return false;
+    }
+    if (sve_access_check(s)) {
+        unsigned vsz = vec_full_reg_size(s);
+        tcg_gen_gvec_2i(vec_full_reg_offset(s, a->rd),
+                        vec_full_reg_offset(s, a->rn),
+                        vsz, vsz, (a->imm << 1) | sel,
+                        &ops[uns][a->esz]);
+    }
+    return true;
+}
+
+static bool trans_SSHLLB(DisasContext *s, arg_rri_esz *a)
+{
+    return do_sve2_shll_tb(s, a, false, false);
+}
+
+static bool trans_SSHLLT(DisasContext *s, arg_rri_esz *a)
+{
+    return do_sve2_shll_tb(s, a, true, false);
+}
+
+static bool trans_USHLLB(DisasContext *s, arg_rri_esz *a)
+{
+    return do_sve2_shll_tb(s, a, false, true);
+}
+
+static bool trans_USHLLT(DisasContext *s, arg_rri_esz *a)
+{
+    return do_sve2_shll_tb(s, a, true, true);
+}
-- 
2.25.1


