diff mbox

[for-2.0,2/2] target-arm: A64: Add saturating accumulate ops (USQADD/SUQADD)

Message ID 1395170635-21281-3-git-send-email-peter.maydell@linaro.org
State Superseded
Headers show

Commit Message

Peter Maydell March 18, 2014, 7:23 p.m. UTC
From: Alex Bennée <alex.bennee@linaro.org>

Add the saturating accumulate operations USQADD and SUQADD
to the A64 instruction set. This completes coverage of A64 Neon.
These operations (which are unsigned + signed -> signed and
signed + unsigned -> unsigned) don't exist in the A32/T32
instruction set, so require a complete new set of helper functions.

Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
---
 target-arm/helper.h        |  20 ++++--
 target-arm/neon_helper.c   | 165 +++++++++++++++++++++++++++++++++++++++++++++
 target-arm/translate-a64.c | 109 ++++++++++++++++++++++++++++--
 3 files changed, 284 insertions(+), 10 deletions(-)

Comments

Richard Henderson March 18, 2014, 9:10 p.m. UTC | #1
On 03/18/2014 12:23 PM, Peter Maydell wrote:
> +#define USATACC(bits, shift) \
> +    do { \
> +        va = (int##bits##_t)((a >> shift) & ((1 << bits) - 1));         \
> +        vb = (uint##bits##_t)((b >> shift) & ((1 << bits) - 1));        \

The masking seems redundant with the cast.
Perhaps better to use extract32/sextract32?

> +        r |= (uint32_t) (vr & ((1 << bits) - 1)) << shift;              \

Perhaps better with deposit32?  That implies an unnecessary mask of R, but
perhaps the compiler can clean that up...


Otherwise I don't actually see anything wrong,

Reviewed-by: Richard Henderson <rth@twiddle.net>


r~
Peter Maydell March 18, 2014, 10:42 p.m. UTC | #2
On 18 March 2014 21:10, Richard Henderson <rth@twiddle.net> wrote:
> On 03/18/2014 12:23 PM, Peter Maydell wrote:
>> +#define USATACC(bits, shift) \
>> +    do { \
>> +        va = (int##bits##_t)((a >> shift) & ((1 << bits) - 1));         \
>> +        vb = (uint##bits##_t)((b >> shift) & ((1 << bits) - 1));        \
>
> The masking seems redundant with the cast.
> Perhaps better to use extract32/sextract32?
>
>> +        r |= (uint32_t) (vr & ((1 << bits) - 1)) << shift;              \
>
> Perhaps better with deposit32?  That implies an unnecessary mask of R, but
> perhaps the compiler can clean that up...

Yeah, it looks much nicer with extract32/deposit32, so I'll just
switch it (and SSATACC) to use those.

thanks
-- PMM
diff mbox

Patch

diff --git a/target-arm/helper.h b/target-arm/helper.h
index b006fd5..366c1b3 100644
--- a/target-arm/helper.h
+++ b/target-arm/helper.h
@@ -186,12 +186,20 @@  DEF_HELPER_FLAGS_2(rints, TCG_CALL_NO_RWG, f32, f32, ptr)
 DEF_HELPER_FLAGS_2(rintd, TCG_CALL_NO_RWG, f64, f64, ptr)
 
 /* neon_helper.c */
-DEF_HELPER_3(neon_qadd_u8, i32, env, i32, i32)
-DEF_HELPER_3(neon_qadd_s8, i32, env, i32, i32)
-DEF_HELPER_3(neon_qadd_u16, i32, env, i32, i32)
-DEF_HELPER_3(neon_qadd_s16, i32, env, i32, i32)
-DEF_HELPER_3(neon_qadd_u32, i32, env, i32, i32)
-DEF_HELPER_3(neon_qadd_s32, i32, env, i32, i32)
+DEF_HELPER_FLAGS_3(neon_qadd_u8, TCG_CALL_NO_RWG, i32, env, i32, i32)
+DEF_HELPER_FLAGS_3(neon_qadd_s8, TCG_CALL_NO_RWG, i32, env, i32, i32)
+DEF_HELPER_FLAGS_3(neon_qadd_u16, TCG_CALL_NO_RWG, i32, env, i32, i32)
+DEF_HELPER_FLAGS_3(neon_qadd_s16, TCG_CALL_NO_RWG, i32, env, i32, i32)
+DEF_HELPER_FLAGS_3(neon_qadd_u32, TCG_CALL_NO_RWG, i32, env, i32, i32)
+DEF_HELPER_FLAGS_3(neon_qadd_s32, TCG_CALL_NO_RWG, i32, env, i32, i32)
+DEF_HELPER_FLAGS_3(neon_uqadd_s8, TCG_CALL_NO_RWG, i32, env, i32, i32)
+DEF_HELPER_FLAGS_3(neon_uqadd_s16, TCG_CALL_NO_RWG, i32, env, i32, i32)
+DEF_HELPER_FLAGS_3(neon_uqadd_s32, TCG_CALL_NO_RWG, i32, env, i32, i32)
+DEF_HELPER_FLAGS_3(neon_uqadd_s64, TCG_CALL_NO_RWG, i64, env, i64, i64)
+DEF_HELPER_FLAGS_3(neon_sqadd_u8, TCG_CALL_NO_RWG, i32, env, i32, i32)
+DEF_HELPER_FLAGS_3(neon_sqadd_u16, TCG_CALL_NO_RWG, i32, env, i32, i32)
+DEF_HELPER_FLAGS_3(neon_sqadd_u32, TCG_CALL_NO_RWG, i32, env, i32, i32)
+DEF_HELPER_FLAGS_3(neon_sqadd_u64, TCG_CALL_NO_RWG, i64, env, i64, i64)
 DEF_HELPER_3(neon_qsub_u8, i32, env, i32, i32)
 DEF_HELPER_3(neon_qsub_s8, i32, env, i32, i32)
 DEF_HELPER_3(neon_qsub_u16, i32, env, i32, i32)
diff --git a/target-arm/neon_helper.c b/target-arm/neon_helper.c
index e23f224..3c65b8e 100644
--- a/target-arm/neon_helper.c
+++ b/target-arm/neon_helper.c
@@ -236,6 +236,171 @@  uint64_t HELPER(neon_qadd_s64)(CPUARMState *env, uint64_t src1, uint64_t src2)
     return res;
 }
 
+/* Unsigned saturating accumulate of signed value
+ *
+ * Op1/Rn is treated as signed
+ * Op2/Rd is treated as unsigned
+ *
+ * Explicit casting is used to ensure the correct sign extension of
+ * inputs. The result is treated as a unsigned value and saturated as such.
+ *
+ * We use a macro for the 8/16 bit cases which expects signed integers of va,
+ * vb, and vr for interim calculation and an unsigned 32 bit result value r.
+ */
+
+#define USATACC(bits, shift) \
+    do { \
+        va = (int##bits##_t)((a >> shift) & ((1 << bits) - 1));         \
+        vb = (uint##bits##_t)((b >> shift) & ((1 << bits) - 1));        \
+        vr = va + vb;                                                   \
+        if (vr > UINT##bits##_MAX) {                                    \
+            SET_QC();                                                   \
+            vr = UINT##bits##_MAX;                                      \
+        } else if (vr < 0) {                                            \
+            SET_QC();                                                   \
+            vr = 0;                                                     \
+        }                                                               \
+        r |= (uint32_t) vr << shift;                                    \
+   } while (0)
+
+uint32_t HELPER(neon_uqadd_s8)(CPUARMState *env, uint32_t a, uint32_t b)
+{
+    int16_t va, vb, vr;
+    uint32_t r = 0;
+
+    USATACC(8, 0);
+    USATACC(8, 8);
+    USATACC(8, 16);
+    USATACC(8, 24);
+    return r;
+}
+
+uint32_t HELPER(neon_uqadd_s16)(CPUARMState *env, uint32_t a, uint32_t b)
+{
+    int32_t va, vb, vr;
+    uint64_t r = 0;
+
+    USATACC(16, 0);
+    USATACC(16, 16);
+    return r;
+}
+
+#undef USATACC
+
+uint32_t HELPER(neon_uqadd_s32)(CPUARMState *env, uint32_t a, uint32_t b)
+{
+    int64_t va = (int32_t)a;
+    int64_t vb = (uint32_t)b;
+    int64_t vr = va + vb;
+    if (vr > UINT32_MAX) {
+        SET_QC();
+        vr = UINT32_MAX;
+    } else if (vr < 0) {
+        SET_QC();
+        vr = 0;
+    }
+    return vr;
+}
+
+uint64_t HELPER(neon_uqadd_s64)(CPUARMState *env, uint64_t a, uint64_t b)
+{
+    uint64_t res;
+    res = a + b;
+    /* We only need to look at the pattern of SIGN bits to detect
+     * +ve/-ve saturation
+     */
+    if (~a & b & ~res & SIGNBIT64) {
+        SET_QC();
+        res = UINT64_MAX;
+    } else if (a & ~b & res & SIGNBIT64) {
+        SET_QC();
+        res = 0;
+    }
+    return res;
+}
+
+/* Signed saturating accumulate of unsigned value
+ *
+ * Op1/Rn is treated as unsigned
+ * Op2/Rd is treated as signed
+ *
+ * The result is treated as a signed value and saturated as such
+ *
+ * We use a macro for the 8/16 bit cases which expects signed integers of va,
+ * vb, and vr for interim calculation and an unsigned 32 bit result value r.
+ */
+
+#define SSATACC(bits, shift) \
+    do { \
+        va = (uint##bits##_t)((a >> shift) & ((1 << bits) - 1));        \
+        vb = (int##bits##_t)((b >> shift) & ((1 << bits) - 1));         \
+        vr = va + vb;                                                   \
+        if (vr > INT##bits##_MAX) {                                     \
+            SET_QC();                                                   \
+            vr = INT##bits##_MAX;                                       \
+        } else if (vr < INT##bits##_MIN) {                              \
+            SET_QC();                                                   \
+            vr = INT##bits##_MIN;                                       \
+        }                                                               \
+        r |= (uint32_t) (vr & ((1 << bits) - 1)) << shift;              \
+    } while (0)
+
+uint32_t HELPER(neon_sqadd_u8)(CPUARMState *env, uint32_t a, uint32_t b)
+{
+    int16_t va, vb, vr;
+    uint32_t r = 0;
+
+    SSATACC(8, 0);
+    SSATACC(8, 8);
+    SSATACC(8, 16);
+    SSATACC(8, 24);
+    return r;
+}
+
+uint32_t HELPER(neon_sqadd_u16)(CPUARMState *env, uint32_t a, uint32_t b)
+{
+    int32_t va, vb, vr;
+    uint32_t r = 0;
+
+    SSATACC(16, 0);
+    SSATACC(16, 16);
+
+    return r;
+}
+
+#undef SSATACC
+
+uint32_t HELPER(neon_sqadd_u32)(CPUARMState *env, uint32_t a, uint32_t b)
+{
+    int64_t res;
+    int64_t op1 = (uint32_t)a;
+    int64_t op2 = (int32_t)b;
+    res = op1 + op2;
+    if (res > INT32_MAX) {
+        SET_QC();
+        res = INT32_MAX;
+    } else if (res < INT32_MIN) {
+        SET_QC();
+        res = INT32_MIN;
+    }
+    return res;
+}
+
+uint64_t HELPER(neon_sqadd_u64)(CPUARMState *env, uint64_t a, uint64_t b)
+{
+    uint64_t res;
+    res = a + b;
+    /* We only need to look at the pattern of SIGN bits to detect an overflow */
+    if (((a & res)
+         | (~b & res)
+         | (a & ~b)) & SIGNBIT64) {
+        SET_QC();
+        res = INT64_MAX;
+    }
+    return res;
+}
+
+
 #define NEON_USAT(dest, src1, src2, type) do { \
     uint32_t tmp = (uint32_t)src1 - (uint32_t)src2; \
     if (tmp != (type)tmp) { \
diff --git a/target-arm/translate-a64.c b/target-arm/translate-a64.c
index 18659d7..9f06450 100644
--- a/target-arm/translate-a64.c
+++ b/target-arm/translate-a64.c
@@ -7321,6 +7321,101 @@  static void handle_2misc_narrow(DisasContext *s, bool scalar,
     }
 }
 
+/* Remaining saturating accumulating ops */
+static void handle_2misc_satacc(DisasContext *s, bool is_scalar, bool is_u,
+                                bool is_q, int size, int rn, int rd)
+{
+    bool is_double = (size == 3);
+
+    if (is_double) {
+        TCGv_i64 tcg_rn = tcg_temp_new_i64();
+        TCGv_i64 tcg_rd = tcg_temp_new_i64();
+        int pass;
+
+        for (pass = 0; pass < (is_scalar ? 1 : 2); pass++) {
+            read_vec_element(s, tcg_rn, rn, pass, MO_64);
+            read_vec_element(s, tcg_rd, rd, pass, MO_64);
+
+            if (is_u) { /* USQADD */
+                gen_helper_neon_uqadd_s64(tcg_rd, cpu_env, tcg_rn, tcg_rd);
+            } else { /* SUQADD */
+                gen_helper_neon_sqadd_u64(tcg_rd, cpu_env, tcg_rn, tcg_rd);
+            }
+            write_vec_element(s, tcg_rd, rd, pass, MO_64);
+        }
+        if (is_scalar) {
+            clear_vec_high(s, rd);
+        }
+
+        tcg_temp_free_i64(tcg_rd);
+        tcg_temp_free_i64(tcg_rn);
+    } else {
+        TCGv_i32 tcg_rn = tcg_temp_new_i32();
+        TCGv_i32 tcg_rd = tcg_temp_new_i32();
+        int pass, maxpasses;
+
+        if (is_scalar) {
+            maxpasses = 1;
+        } else {
+            maxpasses = is_q ? 4 : 2;
+        }
+
+        for (pass = 0; pass < maxpasses; pass++) {
+            if (is_scalar) {
+                read_vec_element_i32(s, tcg_rn, rn, pass, size);
+                read_vec_element_i32(s, tcg_rd, rd, pass, size);
+            } else {
+                read_vec_element_i32(s, tcg_rn, rn, pass, MO_32);
+                read_vec_element_i32(s, tcg_rd, rd, pass, MO_32);
+            }
+
+            if (is_u) { /* USQADD */
+                switch (size) {
+                case 0:
+                    gen_helper_neon_uqadd_s8(tcg_rd, cpu_env, tcg_rn, tcg_rd);
+                    break;
+                case 1:
+                    gen_helper_neon_uqadd_s16(tcg_rd, cpu_env, tcg_rn, tcg_rd);
+                    break;
+                case 2:
+                    gen_helper_neon_uqadd_s32(tcg_rd, cpu_env, tcg_rn, tcg_rd);
+                    break;
+                default:
+                    g_assert_not_reached();
+                }
+            } else { /* SUQADD */
+                switch (size) {
+                case 0:
+                    gen_helper_neon_sqadd_u8(tcg_rd, cpu_env, tcg_rn, tcg_rd);
+                    break;
+                case 1:
+                    gen_helper_neon_sqadd_u16(tcg_rd, cpu_env, tcg_rn, tcg_rd);
+                    break;
+                case 2:
+                    gen_helper_neon_sqadd_u32(tcg_rd, cpu_env, tcg_rn, tcg_rd);
+                    break;
+                default:
+                    g_assert_not_reached();
+                }
+            }
+
+            if (is_scalar) {
+                TCGv_i64 tcg_zero = tcg_const_i64(0);
+                write_vec_element(s, tcg_zero, rd, 0, MO_64);
+                tcg_temp_free_i64(tcg_zero);
+            }
+            write_vec_element_i32(s, tcg_rd, rd, pass, MO_32);
+        }
+
+        if (!is_q) {
+            clear_vec_high(s, rd);
+        }
+
+        tcg_temp_free_i32(tcg_rd);
+        tcg_temp_free_i32(tcg_rn);
+    }
+}
+
 /* C3.6.12 AdvSIMD scalar two reg misc
  *  31 30  29 28       24 23  22 21       17 16    12 11 10 9    5 4    0
  * +-----+---+-----------+------+-----------+--------+-----+------+------+
@@ -7340,6 +7435,9 @@  static void disas_simd_scalar_two_reg_misc(DisasContext *s, uint32_t insn)
     TCGv_ptr tcg_fpstatus;
 
     switch (opcode) {
+    case 0x3: /* USQADD / SUQADD*/
+        handle_2misc_satacc(s, true, u, false, size, rn, rd);
+        return;
     case 0x7: /* SQABS / SQNEG */
         break;
     case 0xa: /* CMLT */
@@ -7427,10 +7525,7 @@  static void disas_simd_scalar_two_reg_misc(DisasContext *s, uint32_t insn)
         }
         break;
     default:
-        /* Other categories of encoding in this class:
-         *  + SUQADD/USQADD/SQABS/SQNEG : size 8, 16, 32 or 64
-         */
-        unsupported_encoding(s, insn);
+        unallocated_encoding(s);
         return;
     }
 
@@ -9194,6 +9289,12 @@  static void disas_simd_two_reg_misc(DisasContext *s, uint32_t insn)
         }
         break;
     case 0x3: /* SUQADD, USQADD */
+        if (size == 3 && !is_q) {
+            unallocated_encoding(s);
+            return;
+        }
+        handle_2misc_satacc(s, false, u, is_q, size, rn, rd);
+        return;
     case 0x7: /* SQABS, SQNEG */
         if (size == 3 && !is_q) {
             unallocated_encoding(s);