All of lore.kernel.org
 help / color / mirror / Atom feed
From: Richard Henderson <richard.henderson@linaro.org>
To: qemu-devel@nongnu.org
Cc: Peter Maydell <peter.maydell@linaro.org>, qemu-arm@nongnu.org
Subject: [PATCH v7 05/92] target/arm: Split out saturating/rounding shifts from neon
Date: Mon, 24 May 2021 18:02:31 -0700	[thread overview]
Message-ID: <20210525010358.152808-6-richard.henderson@linaro.org> (raw)
In-Reply-To: <20210525010358.152808-1-richard.henderson@linaro.org>

Split these operations out into a header that can be shared
between neon and sve.  The "sat" pointer acts both as a boolean
for control of saturating behavior and controls the difference
in behavior between neon and sve -- QC bit or no QC bit.

Widen the shift operand in the new helpers, as the SVE2 insns treat
the whole input element as significant.  For the neon uses, truncate
the shift to int8_t while passing the parameter.

Implement right-shift rounding as

    tmp = src >> (shift - 1);
    dst = (tmp >> 1) + (tmp & 1);

This is the same number of instructions as the current

    tmp = 1 << (shift - 1);
    dst = (src + tmp) >> shift;

without any possibility of intermediate overflow.

Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
v2: Widen the shift operand (laurent desnouges)
v7: Add null checks to suqrshl (pm215).
---
 target/arm/vec_internal.h | 138 +++++++++++
 target/arm/neon_helper.c  | 507 +++++++-------------------------------
 2 files changed, 221 insertions(+), 424 deletions(-)

diff --git a/target/arm/vec_internal.h b/target/arm/vec_internal.h
index e3eb3e7a6b..5b78e79329 100644
--- a/target/arm/vec_internal.h
+++ b/target/arm/vec_internal.h
@@ -30,4 +30,142 @@ static inline void clear_tail(void *vd, uintptr_t opr_sz, uintptr_t max_sz)
     }
 }
 
+static inline int32_t do_sqrshl_bhs(int32_t src, int32_t shift, int bits,
+                                    bool round, uint32_t *sat)
+{
+    if (shift <= -bits) {
+        /* Rounding the sign bit always produces 0. */
+        if (round) {
+            return 0;
+        }
+        return src >> 31;
+    } else if (shift < 0) {
+        if (round) {
+            src >>= -shift - 1;
+            return (src >> 1) + (src & 1);
+        }
+        return src >> -shift;
+    } else if (shift < bits) {
+        int32_t val = src << shift;
+        if (bits == 32) {
+            if (!sat || val >> shift == src) {
+                return val;
+            }
+        } else {
+            int32_t extval = sextract32(val, 0, bits);
+            if (!sat || val == extval) {
+                return extval;
+            }
+        }
+    } else if (!sat || src == 0) {
+        return 0;
+    }
+
+    *sat = 1;
+    return (1u << (bits - 1)) - (src >= 0);
+}
+
+static inline uint32_t do_uqrshl_bhs(uint32_t src, int32_t shift, int bits,
+                                     bool round, uint32_t *sat)
+{
+    if (shift <= -(bits + round)) {
+        return 0;
+    } else if (shift < 0) {
+        if (round) {
+            src >>= -shift - 1;
+            return (src >> 1) + (src & 1);
+        }
+        return src >> -shift;
+    } else if (shift < bits) {
+        uint32_t val = src << shift;
+        if (bits == 32) {
+            if (!sat || val >> shift == src) {
+                return val;
+            }
+        } else {
+            uint32_t extval = extract32(val, 0, bits);
+            if (!sat || val == extval) {
+                return extval;
+            }
+        }
+    } else if (!sat || src == 0) {
+        return 0;
+    }
+
+    *sat = 1;
+    return MAKE_64BIT_MASK(0, bits);
+}
+
+static inline int32_t do_suqrshl_bhs(int32_t src, int32_t shift, int bits,
+                                     bool round, uint32_t *sat)
+{
+    if (sat && src < 0) {
+        *sat = 1;
+        return 0;
+    }
+    return do_uqrshl_bhs(src, shift, bits, round, sat);
+}
+
+static inline int64_t do_sqrshl_d(int64_t src, int64_t shift,
+                                  bool round, uint32_t *sat)
+{
+    if (shift <= -64) {
+        /* Rounding the sign bit always produces 0. */
+        if (round) {
+            return 0;
+        }
+        return src >> 63;
+    } else if (shift < 0) {
+        if (round) {
+            src >>= -shift - 1;
+            return (src >> 1) + (src & 1);
+        }
+        return src >> -shift;
+    } else if (shift < 64) {
+        int64_t val = src << shift;
+        if (!sat || val >> shift == src) {
+            return val;
+        }
+    } else if (!sat || src == 0) {
+        return 0;
+    }
+
+    *sat = 1;
+    return src < 0 ? INT64_MIN : INT64_MAX;
+}
+
+static inline uint64_t do_uqrshl_d(uint64_t src, int64_t shift,
+                                   bool round, uint32_t *sat)
+{
+    if (shift <= -(64 + round)) {
+        return 0;
+    } else if (shift < 0) {
+        if (round) {
+            src >>= -shift - 1;
+            return (src >> 1) + (src & 1);
+        }
+        return src >> -shift;
+    } else if (shift < 64) {
+        uint64_t val = src << shift;
+        if (!sat || val >> shift == src) {
+            return val;
+        }
+    } else if (!sat || src == 0) {
+        return 0;
+    }
+
+    *sat = 1;
+    return UINT64_MAX;
+}
+
+static inline int64_t do_suqrshl_d(int64_t src, int64_t shift,
+                                   bool round, uint32_t *sat)
+{
+    if (sat && src < 0) {
+        *sat = 1;
+        return 0;
+    }
+    return do_uqrshl_d(src, shift, round, sat);
+}
+
 #endif /* TARGET_ARM_VEC_INTERNALS_H */
diff --git a/target/arm/neon_helper.c b/target/arm/neon_helper.c
index b637265691..338b9189d5 100644
--- a/target/arm/neon_helper.c
+++ b/target/arm/neon_helper.c
@@ -11,6 +11,7 @@
 #include "cpu.h"
 #include "exec/helper-proto.h"
 #include "fpu/softfloat.h"
+#include "vec_internal.h"
 
 #define SIGNBIT (uint32_t)0x80000000
 #define SIGNBIT64 ((uint64_t)1 << 63)
@@ -576,496 +577,154 @@ NEON_POP(pmax_s16, neon_s16, 2)
 NEON_POP(pmax_u16, neon_u16, 2)
 #undef NEON_FN
 
-#define NEON_FN(dest, src1, src2) do { \
-    int8_t tmp; \
-    tmp = (int8_t)src2; \
-    if (tmp >= (ssize_t)sizeof(src1) * 8 || \
-        tmp <= -(ssize_t)sizeof(src1) * 8) { \
-        dest = 0; \
-    } else if (tmp < 0) { \
-        dest = src1 >> -tmp; \
-    } else { \
-        dest = src1 << tmp; \
-    }} while (0)
+#define NEON_FN(dest, src1, src2) \
+    (dest = do_uqrshl_bhs(src1, (int8_t)src2, 16, false, NULL))
 NEON_VOP(shl_u16, neon_u16, 2)
 #undef NEON_FN
 
-#define NEON_FN(dest, src1, src2) do { \
-    int8_t tmp; \
-    tmp = (int8_t)src2; \
-    if (tmp >= (ssize_t)sizeof(src1) * 8) { \
-        dest = 0; \
-    } else if (tmp <= -(ssize_t)sizeof(src1) * 8) { \
-        dest = src1 >> (sizeof(src1) * 8 - 1); \
-    } else if (tmp < 0) { \
-        dest = src1 >> -tmp; \
-    } else { \
-        dest = src1 << tmp; \
-    }} while (0)
+#define NEON_FN(dest, src1, src2) \
+    (dest = do_sqrshl_bhs(src1, (int8_t)src2, 16, false, NULL))
 NEON_VOP(shl_s16, neon_s16, 2)
 #undef NEON_FN
 
-#define NEON_FN(dest, src1, src2) do { \
-    int8_t tmp; \
-    tmp = (int8_t)src2; \
-    if ((tmp >= (ssize_t)sizeof(src1) * 8) \
-        || (tmp <= -(ssize_t)sizeof(src1) * 8)) { \
-        dest = 0; \
-    } else if (tmp < 0) { \
-        dest = (src1 + (1 << (-1 - tmp))) >> -tmp; \
-    } else { \
-        dest = src1 << tmp; \
-    }} while (0)
+#define NEON_FN(dest, src1, src2) \
+    (dest = do_sqrshl_bhs(src1, (int8_t)src2, 8, true, NULL))
 NEON_VOP(rshl_s8, neon_s8, 4)
+#undef NEON_FN
+
+#define NEON_FN(dest, src1, src2) \
+    (dest = do_sqrshl_bhs(src1, (int8_t)src2, 16, true, NULL))
 NEON_VOP(rshl_s16, neon_s16, 2)
 #undef NEON_FN
 
-/* The addition of the rounding constant may overflow, so we use an
- * intermediate 64 bit accumulator.  */
-uint32_t HELPER(neon_rshl_s32)(uint32_t valop, uint32_t shiftop)
+uint32_t HELPER(neon_rshl_s32)(uint32_t val, uint32_t shift)
 {
-    int32_t dest;
-    int32_t val = (int32_t)valop;
-    int8_t shift = (int8_t)shiftop;
-    if ((shift >= 32) || (shift <= -32)) {
-        dest = 0;
-    } else if (shift < 0) {
-        int64_t big_dest = ((int64_t)val + (1 << (-1 - shift)));
-        dest = big_dest >> -shift;
-    } else {
-        dest = val << shift;
-    }
-    return dest;
+    return do_sqrshl_bhs(val, (int8_t)shift, 32, true, NULL);
 }
 
-/* Handling addition overflow with 64 bit input values is more
- * tricky than with 32 bit values.  */
-uint64_t HELPER(neon_rshl_s64)(uint64_t valop, uint64_t shiftop)
+uint64_t HELPER(neon_rshl_s64)(uint64_t val, uint64_t shift)
 {
-    int8_t shift = (int8_t)shiftop;
-    int64_t val = valop;
-    if ((shift >= 64) || (shift <= -64)) {
-        val = 0;
-    } else if (shift < 0) {
-        val >>= (-shift - 1);
-        if (val == INT64_MAX) {
-            /* In this case, it means that the rounding constant is 1,
-             * and the addition would overflow. Return the actual
-             * result directly.  */
-            val = 0x4000000000000000LL;
-        } else {
-            val++;
-            val >>= 1;
-        }
-    } else {
-        val <<= shift;
-    }
-    return val;
+    return do_sqrshl_d(val, (int8_t)shift, true, NULL);
 }
 
-#define NEON_FN(dest, src1, src2) do { \
-    int8_t tmp; \
-    tmp = (int8_t)src2; \
-    if (tmp >= (ssize_t)sizeof(src1) * 8 || \
-        tmp < -(ssize_t)sizeof(src1) * 8) { \
-        dest = 0; \
-    } else if (tmp == -(ssize_t)sizeof(src1) * 8) { \
-        dest = src1 >> (-tmp - 1); \
-    } else if (tmp < 0) { \
-        dest = (src1 + (1 << (-1 - tmp))) >> -tmp; \
-    } else { \
-        dest = src1 << tmp; \
-    }} while (0)
+#define NEON_FN(dest, src1, src2) \
+    (dest = do_uqrshl_bhs(src1, (int8_t)src2, 8, true, NULL))
 NEON_VOP(rshl_u8, neon_u8, 4)
+#undef NEON_FN
+
+#define NEON_FN(dest, src1, src2) \
+    (dest = do_uqrshl_bhs(src1, (int8_t)src2, 16, true, NULL))
 NEON_VOP(rshl_u16, neon_u16, 2)
 #undef NEON_FN
 
-/* The addition of the rounding constant may overflow, so we use an
- * intermediate 64 bit accumulator.  */
-uint32_t HELPER(neon_rshl_u32)(uint32_t val, uint32_t shiftop)
+uint32_t HELPER(neon_rshl_u32)(uint32_t val, uint32_t shift)
 {
-    uint32_t dest;
-    int8_t shift = (int8_t)shiftop;
-    if (shift >= 32 || shift < -32) {
-        dest = 0;
-    } else if (shift == -32) {
-        dest = val >> 31;
-    } else if (shift < 0) {
-        uint64_t big_dest = ((uint64_t)val + (1 << (-1 - shift)));
-        dest = big_dest >> -shift;
-    } else {
-        dest = val << shift;
-    }
-    return dest;
+    return do_uqrshl_bhs(val, (int8_t)shift, 32, true, NULL);
 }
 
-/* Handling addition overflow with 64 bit input values is more
- * tricky than with 32 bit values.  */
-uint64_t HELPER(neon_rshl_u64)(uint64_t val, uint64_t shiftop)
+uint64_t HELPER(neon_rshl_u64)(uint64_t val, uint64_t shift)
 {
-    int8_t shift = (uint8_t)shiftop;
-    if (shift >= 64 || shift < -64) {
-        val = 0;
-    } else if (shift == -64) {
-        /* Rounding a 1-bit result just preserves that bit.  */
-        val >>= 63;
-    } else if (shift < 0) {
-        val >>= (-shift - 1);
-        if (val == UINT64_MAX) {
-            /* In this case, it means that the rounding constant is 1,
-             * and the addition would overflow. Return the actual
-             * result directly.  */
-            val = 0x8000000000000000ULL;
-        } else {
-            val++;
-            val >>= 1;
-        }
-    } else {
-        val <<= shift;
-    }
-    return val;
+    return do_uqrshl_d(val, (int8_t)shift, true, NULL);
 }
 
-#define NEON_FN(dest, src1, src2) do { \
-    int8_t tmp; \
-    tmp = (int8_t)src2; \
-    if (tmp >= (ssize_t)sizeof(src1) * 8) { \
-        if (src1) { \
-            SET_QC(); \
-            dest = ~0; \
-        } else { \
-            dest = 0; \
-        } \
-    } else if (tmp <= -(ssize_t)sizeof(src1) * 8) { \
-        dest = 0; \
-    } else if (tmp < 0) { \
-        dest = src1 >> -tmp; \
-    } else { \
-        dest = src1 << tmp; \
-        if ((dest >> tmp) != src1) { \
-            SET_QC(); \
-            dest = ~0; \
-        } \
-    }} while (0)
+#define NEON_FN(dest, src1, src2) \
+    (dest = do_uqrshl_bhs(src1, (int8_t)src2, 8, false, env->vfp.qc))
 NEON_VOP_ENV(qshl_u8, neon_u8, 4)
+#undef NEON_FN
+
+#define NEON_FN(dest, src1, src2) \
+    (dest = do_uqrshl_bhs(src1, (int8_t)src2, 16, false, env->vfp.qc))
 NEON_VOP_ENV(qshl_u16, neon_u16, 2)
-NEON_VOP_ENV(qshl_u32, neon_u32, 1)
 #undef NEON_FN
 
-uint64_t HELPER(neon_qshl_u64)(CPUARMState *env, uint64_t val, uint64_t shiftop)
+uint32_t HELPER(neon_qshl_u32)(CPUARMState *env, uint32_t val, uint32_t shift)
 {
-    int8_t shift = (int8_t)shiftop;
-    if (shift >= 64) {
-        if (val) {
-            val = ~(uint64_t)0;
-            SET_QC();
-        }
-    } else if (shift <= -64) {
-        val = 0;
-    } else if (shift < 0) {
-        val >>= -shift;
-    } else {
-        uint64_t tmp = val;
-        val <<= shift;
-        if ((val >> shift) != tmp) {
-            SET_QC();
-            val = ~(uint64_t)0;
-        }
-    }
-    return val;
+    return do_uqrshl_bhs(val, (int8_t)shift, 32, false, env->vfp.qc);
 }
 
-#define NEON_FN(dest, src1, src2) do { \
-    int8_t tmp; \
-    tmp = (int8_t)src2; \
-    if (tmp >= (ssize_t)sizeof(src1) * 8) { \
-        if (src1) { \
-            SET_QC(); \
-            dest = (uint32_t)(1 << (sizeof(src1) * 8 - 1)); \
-            if (src1 > 0) { \
-                dest--; \
-            } \
-        } else { \
-            dest = src1; \
-        } \
-    } else if (tmp <= -(ssize_t)sizeof(src1) * 8) { \
-        dest = src1 >> 31; \
-    } else if (tmp < 0) { \
-        dest = src1 >> -tmp; \
-    } else { \
-        dest = src1 << tmp; \
-        if ((dest >> tmp) != src1) { \
-            SET_QC(); \
-            dest = (uint32_t)(1 << (sizeof(src1) * 8 - 1)); \
-            if (src1 > 0) { \
-                dest--; \
-            } \
-        } \
-    }} while (0)
+uint64_t HELPER(neon_qshl_u64)(CPUARMState *env, uint64_t val, uint64_t shift)
+{
+    return do_uqrshl_d(val, (int8_t)shift, false, env->vfp.qc);
+}
+
+#define NEON_FN(dest, src1, src2) \
+    (dest = do_sqrshl_bhs(src1, (int8_t)src2, 8, false, env->vfp.qc))
 NEON_VOP_ENV(qshl_s8, neon_s8, 4)
+#undef NEON_FN
+
+#define NEON_FN(dest, src1, src2) \
+    (dest = do_sqrshl_bhs(src1, (int8_t)src2, 16, false, env->vfp.qc))
 NEON_VOP_ENV(qshl_s16, neon_s16, 2)
-NEON_VOP_ENV(qshl_s32, neon_s32, 1)
 #undef NEON_FN
 
-uint64_t HELPER(neon_qshl_s64)(CPUARMState *env, uint64_t valop, uint64_t shiftop)
+uint32_t HELPER(neon_qshl_s32)(CPUARMState *env, uint32_t val, uint32_t shift)
 {
-    int8_t shift = (uint8_t)shiftop;
-    int64_t val = valop;
-    if (shift >= 64) {
-        if (val) {
-            SET_QC();
-            val = (val >> 63) ^ ~SIGNBIT64;
-        }
-    } else if (shift <= -64) {
-        val >>= 63;
-    } else if (shift < 0) {
-        val >>= -shift;
-    } else {
-        int64_t tmp = val;
-        val <<= shift;
-        if ((val >> shift) != tmp) {
-            SET_QC();
-            val = (tmp >> 63) ^ ~SIGNBIT64;
-        }
-    }
-    return val;
+    return do_sqrshl_bhs(val, (int8_t)shift, 32, false, env->vfp.qc);
 }
 
-#define NEON_FN(dest, src1, src2) do { \
-    if (src1 & (1 << (sizeof(src1) * 8 - 1))) { \
-        SET_QC(); \
-        dest = 0; \
-    } else { \
-        int8_t tmp; \
-        tmp = (int8_t)src2; \
-        if (tmp >= (ssize_t)sizeof(src1) * 8) { \
-            if (src1) { \
-                SET_QC(); \
-                dest = ~0; \
-            } else { \
-                dest = 0; \
-            } \
-        } else if (tmp <= -(ssize_t)sizeof(src1) * 8) { \
-            dest = 0; \
-        } else if (tmp < 0) { \
-            dest = src1 >> -tmp; \
-        } else { \
-            dest = src1 << tmp; \
-            if ((dest >> tmp) != src1) { \
-                SET_QC(); \
-                dest = ~0; \
-            } \
-        } \
-    }} while (0)
-NEON_VOP_ENV(qshlu_s8, neon_u8, 4)
-NEON_VOP_ENV(qshlu_s16, neon_u16, 2)
+uint64_t HELPER(neon_qshl_s64)(CPUARMState *env, uint64_t val, uint64_t shift)
+{
+    return do_sqrshl_d(val, (int8_t)shift, false, env->vfp.qc);
+}
+
+#define NEON_FN(dest, src1, src2) \
+    (dest = do_suqrshl_bhs(src1, (int8_t)src2, 8, false, env->vfp.qc))
+NEON_VOP_ENV(qshlu_s8, neon_s8, 4)
 #undef NEON_FN
 
-uint32_t HELPER(neon_qshlu_s32)(CPUARMState *env, uint32_t valop, uint32_t shiftop)
+#define NEON_FN(dest, src1, src2) \
+    (dest = do_suqrshl_bhs(src1, (int8_t)src2, 16, false, env->vfp.qc))
+NEON_VOP_ENV(qshlu_s16, neon_s16, 2)
+#undef NEON_FN
+
+uint32_t HELPER(neon_qshlu_s32)(CPUARMState *env, uint32_t val, uint32_t shift)
 {
-    if ((int32_t)valop < 0) {
-        SET_QC();
-        return 0;
-    }
-    return helper_neon_qshl_u32(env, valop, shiftop);
+    return do_suqrshl_bhs(val, (int8_t)shift, 32, false, env->vfp.qc);
 }
 
-uint64_t HELPER(neon_qshlu_s64)(CPUARMState *env, uint64_t valop, uint64_t shiftop)
+uint64_t HELPER(neon_qshlu_s64)(CPUARMState *env, uint64_t val, uint64_t shift)
 {
-    if ((int64_t)valop < 0) {
-        SET_QC();
-        return 0;
-    }
-    return helper_neon_qshl_u64(env, valop, shiftop);
+    return do_suqrshl_d(val, (int8_t)shift, false, env->vfp.qc);
 }
 
-#define NEON_FN(dest, src1, src2) do { \
-    int8_t tmp; \
-    tmp = (int8_t)src2; \
-    if (tmp >= (ssize_t)sizeof(src1) * 8) { \
-        if (src1) { \
-            SET_QC(); \
-            dest = ~0; \
-        } else { \
-            dest = 0; \
-        } \
-    } else if (tmp < -(ssize_t)sizeof(src1) * 8) { \
-        dest = 0; \
-    } else if (tmp == -(ssize_t)sizeof(src1) * 8) { \
-        dest = src1 >> (sizeof(src1) * 8 - 1); \
-    } else if (tmp < 0) { \
-        dest = (src1 + (1 << (-1 - tmp))) >> -tmp; \
-    } else { \
-        dest = src1 << tmp; \
-        if ((dest >> tmp) != src1) { \
-            SET_QC(); \
-            dest = ~0; \
-        } \
-    }} while (0)
+#define NEON_FN(dest, src1, src2) \
+    (dest = do_uqrshl_bhs(src1, (int8_t)src2, 8, true, env->vfp.qc))
 NEON_VOP_ENV(qrshl_u8, neon_u8, 4)
+#undef NEON_FN
+
+#define NEON_FN(dest, src1, src2) \
+    (dest = do_uqrshl_bhs(src1, (int8_t)src2, 16, true, env->vfp.qc))
 NEON_VOP_ENV(qrshl_u16, neon_u16, 2)
 #undef NEON_FN
 
-/* The addition of the rounding constant may overflow, so we use an
- * intermediate 64 bit accumulator.  */
-uint32_t HELPER(neon_qrshl_u32)(CPUARMState *env, uint32_t val, uint32_t shiftop)
+uint32_t HELPER(neon_qrshl_u32)(CPUARMState *env, uint32_t val, uint32_t shift)
 {
-    uint32_t dest;
-    int8_t shift = (int8_t)shiftop;
-    if (shift >= 32) {
-        if (val) {
-            SET_QC();
-            dest = ~0;
-        } else {
-            dest = 0;
-        }
-    } else if (shift < -32) {
-        dest = 0;
-    } else if (shift == -32) {
-        dest = val >> 31;
-    } else if (shift < 0) {
-        uint64_t big_dest = ((uint64_t)val + (1 << (-1 - shift)));
-        dest = big_dest >> -shift;
-    } else {
-        dest = val << shift;
-        if ((dest >> shift) != val) {
-            SET_QC();
-            dest = ~0;
-        }
-    }
-    return dest;
+    return do_uqrshl_bhs(val, (int8_t)shift, 32, true, env->vfp.qc);
 }
 
-/* Handling addition overflow with 64 bit input values is more
- * tricky than with 32 bit values.  */
-uint64_t HELPER(neon_qrshl_u64)(CPUARMState *env, uint64_t val, uint64_t shiftop)
+uint64_t HELPER(neon_qrshl_u64)(CPUARMState *env, uint64_t val, uint64_t shift)
 {
-    int8_t shift = (int8_t)shiftop;
-    if (shift >= 64) {
-        if (val) {
-            SET_QC();
-            val = ~0;
-        }
-    } else if (shift < -64) {
-        val = 0;
-    } else if (shift == -64) {
-        val >>= 63;
-    } else if (shift < 0) {
-        val >>= (-shift - 1);
-        if (val == UINT64_MAX) {
-            /* In this case, it means that the rounding constant is 1,
-             * and the addition would overflow. Return the actual
-             * result directly.  */
-            val = 0x8000000000000000ULL;
-        } else {
-            val++;
-            val >>= 1;
-        }
-    } else { \
-        uint64_t tmp = val;
-        val <<= shift;
-        if ((val >> shift) != tmp) {
-            SET_QC();
-            val = ~0;
-        }
-    }
-    return val;
+    return do_uqrshl_d(val, (int8_t)shift, true, env->vfp.qc);
 }
 
-#define NEON_FN(dest, src1, src2) do { \
-    int8_t tmp; \
-    tmp = (int8_t)src2; \
-    if (tmp >= (ssize_t)sizeof(src1) * 8) { \
-        if (src1) { \
-            SET_QC(); \
-            dest = (typeof(dest))(1 << (sizeof(src1) * 8 - 1)); \
-            if (src1 > 0) { \
-                dest--; \
-            } \
-        } else { \
-            dest = 0; \
-        } \
-    } else if (tmp <= -(ssize_t)sizeof(src1) * 8) { \
-        dest = 0; \
-    } else if (tmp < 0) { \
-        dest = (src1 + (1 << (-1 - tmp))) >> -tmp; \
-    } else { \
-        dest = src1 << tmp; \
-        if ((dest >> tmp) != src1) { \
-            SET_QC(); \
-            dest = (uint32_t)(1 << (sizeof(src1) * 8 - 1)); \
-            if (src1 > 0) { \
-                dest--; \
-            } \
-        } \
-    }} while (0)
+#define NEON_FN(dest, src1, src2) \
+    (dest = do_sqrshl_bhs(src1, (int8_t)src2, 8, true, env->vfp.qc))
 NEON_VOP_ENV(qrshl_s8, neon_s8, 4)
+#undef NEON_FN
+
+#define NEON_FN(dest, src1, src2) \
+    (dest = do_sqrshl_bhs(src1, (int8_t)src2, 16, true, env->vfp.qc))
 NEON_VOP_ENV(qrshl_s16, neon_s16, 2)
 #undef NEON_FN
 
-/* The addition of the rounding constant may overflow, so we use an
- * intermediate 64 bit accumulator.  */
-uint32_t HELPER(neon_qrshl_s32)(CPUARMState *env, uint32_t valop, uint32_t shiftop)
+uint32_t HELPER(neon_qrshl_s32)(CPUARMState *env, uint32_t val, uint32_t shift)
 {
-    int32_t dest;
-    int32_t val = (int32_t)valop;
-    int8_t shift = (int8_t)shiftop;
-    if (shift >= 32) {
-        if (val) {
-            SET_QC();
-            dest = (val >> 31) ^ ~SIGNBIT;
-        } else {
-            dest = 0;
-        }
-    } else if (shift <= -32) {
-        dest = 0;
-    } else if (shift < 0) {
-        int64_t big_dest = ((int64_t)val + (1 << (-1 - shift)));
-        dest = big_dest >> -shift;
-    } else {
-        dest = val << shift;
-        if ((dest >> shift) != val) {
-            SET_QC();
-            dest = (val >> 31) ^ ~SIGNBIT;
-        }
-    }
-    return dest;
+    return do_sqrshl_bhs(val, (int8_t)shift, 32, true, env->vfp.qc);
 }
 
-/* Handling addition overflow with 64 bit input values is more
- * tricky than with 32 bit values.  */
-uint64_t HELPER(neon_qrshl_s64)(CPUARMState *env, uint64_t valop, uint64_t shiftop)
+uint64_t HELPER(neon_qrshl_s64)(CPUARMState *env, uint64_t val, uint64_t shift)
 {
-    int8_t shift = (uint8_t)shiftop;
-    int64_t val = valop;
-
-    if (shift >= 64) {
-        if (val) {
-            SET_QC();
-            val = (val >> 63) ^ ~SIGNBIT64;
-        }
-    } else if (shift <= -64) {
-        val = 0;
-    } else if (shift < 0) {
-        val >>= (-shift - 1);
-        if (val == INT64_MAX) {
-            /* In this case, it means that the rounding constant is 1,
-             * and the addition would overflow. Return the actual
-             * result directly.  */
-            val = 0x4000000000000000ULL;
-        } else {
-            val++;
-            val >>= 1;
-        }
-    } else {
-        int64_t tmp = val;
-        val <<= shift;
-        if ((val >> shift) != tmp) {
-            SET_QC();
-            val = (tmp >> 63) ^ ~SIGNBIT64;
-        }
-    }
-    return val;
+    return do_sqrshl_d(val, (int8_t)shift, true, env->vfp.qc);
 }
 
 uint32_t HELPER(neon_add_u8)(uint32_t a, uint32_t b)
-- 
2.25.1



  parent reply	other threads:[~2021-05-25  1:09 UTC|newest]

Thread overview: 95+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2021-05-25  1:02 [PATCH v7 00/92] target/arm: Implement SVE2 Richard Henderson
2021-05-25  1:02 ` [PATCH v7 01/92] target/arm: Add ID_AA64ZFR0 fields and isar_feature_aa64_sve2 Richard Henderson
2021-05-25  1:02 ` [PATCH v7 02/92] target/arm: Implement SVE2 Integer Multiply - Unpredicated Richard Henderson
2021-05-25  1:02 ` [PATCH v7 03/92] target/arm: Implement SVE2 integer pairwise add and accumulate long Richard Henderson
2021-05-25  1:02 ` [PATCH v7 04/92] target/arm: Implement SVE2 integer unary operations (predicated) Richard Henderson
2021-05-25  1:02 ` Richard Henderson [this message]
2021-05-25  1:02 ` [PATCH v7 06/92] target/arm: Implement SVE2 saturating/rounding bitwise shift left (predicated) Richard Henderson
2021-05-25  1:02 ` [PATCH v7 07/92] target/arm: Implement SVE2 integer halving add/subtract (predicated) Richard Henderson
2021-05-25  1:02 ` [PATCH v7 08/92] target/arm: Implement SVE2 integer pairwise arithmetic Richard Henderson
2021-05-25  1:02 ` [PATCH v7 09/92] target/arm: Implement SVE2 saturating add/subtract (predicated) Richard Henderson
2021-05-25  1:02 ` [PATCH v7 10/92] target/arm: Implement SVE2 integer add/subtract long Richard Henderson
2021-05-25  1:02 ` [PATCH v7 11/92] target/arm: Implement SVE2 integer add/subtract interleaved long Richard Henderson
2021-05-25  1:02 ` [PATCH v7 12/92] target/arm: Implement SVE2 integer add/subtract wide Richard Henderson
2021-05-25  1:02 ` [PATCH v7 13/92] target/arm: Implement SVE2 integer multiply long Richard Henderson
2021-05-25  1:02 ` [PATCH v7 14/92] target/arm: Implement SVE2 PMULLB, PMULLT Richard Henderson
2021-05-25  1:02 ` [PATCH v7 15/92] target/arm: Implement SVE2 bitwise shift left long Richard Henderson
2021-05-25  1:02 ` [PATCH v7 16/92] target/arm: Implement SVE2 bitwise exclusive-or interleaved Richard Henderson
2021-05-25  1:02 ` [PATCH v7 17/92] target/arm: Implement SVE2 bitwise permute Richard Henderson
2021-05-25  1:02 ` [PATCH v7 18/92] target/arm: Implement SVE2 complex integer add Richard Henderson
2021-05-25  1:02 ` [PATCH v7 19/92] target/arm: Implement SVE2 integer absolute difference and accumulate long Richard Henderson
2021-05-25  1:02 ` [PATCH v7 20/92] target/arm: Implement SVE2 integer add/subtract long with carry Richard Henderson
2021-05-25  1:02 ` [PATCH v7 21/92] target/arm: Implement SVE2 bitwise shift right and accumulate Richard Henderson
2021-05-25  1:02 ` [PATCH v7 22/92] target/arm: Implement SVE2 bitwise shift and insert Richard Henderson
2021-05-25  1:02 ` [PATCH v7 23/92] target/arm: Implement SVE2 integer absolute difference and accumulate Richard Henderson
2021-05-25  1:02 ` [PATCH v7 24/92] target/arm: Implement SVE2 saturating extract narrow Richard Henderson
2021-05-25  1:02 ` [PATCH v7 25/92] target/arm: Implement SVE2 floating-point pairwise Richard Henderson
2021-05-25  1:02 ` [PATCH v7 26/92] target/arm: Implement SVE2 SHRN, RSHRN Richard Henderson
2021-05-25  1:02 ` [PATCH v7 27/92] target/arm: Implement SVE2 SQSHRUN, SQRSHRUN Richard Henderson
2021-05-25  1:02 ` [PATCH v7 28/92] target/arm: Implement SVE2 UQSHRN, UQRSHRN Richard Henderson
2021-05-25  1:02 ` [PATCH v7 29/92] target/arm: Implement SVE2 SQSHRN, SQRSHRN Richard Henderson
2021-05-25  1:02 ` [PATCH v7 30/92] target/arm: Implement SVE2 WHILEGT, WHILEGE, WHILEHI, WHILEHS Richard Henderson
2021-05-25  1:02 ` [PATCH v7 31/92] target/arm: Implement SVE2 WHILERW, WHILEWR Richard Henderson
2021-05-25  1:02 ` [PATCH v7 32/92] target/arm: Implement SVE2 bitwise ternary operations Richard Henderson
2021-05-25  1:02 ` [PATCH v7 33/92] target/arm: Implement SVE2 MATCH, NMATCH Richard Henderson
2021-05-25  1:03 ` [PATCH v7 34/92] target/arm: Implement SVE2 saturating multiply-add long Richard Henderson
2021-05-25  1:03 ` [PATCH v7 35/92] target/arm: Implement SVE2 saturating multiply-add high Richard Henderson
2021-05-25  1:03 ` [PATCH v7 36/92] target/arm: Implement SVE2 integer multiply-add long Richard Henderson
2021-05-25  1:03 ` [PATCH v7 37/92] target/arm: Implement SVE2 complex integer multiply-add Richard Henderson
2021-05-25  1:03 ` [PATCH v7 38/92] target/arm: Implement SVE2 ADDHNB, ADDHNT Richard Henderson
2021-05-25  1:03 ` [PATCH v7 39/92] target/arm: Implement SVE2 RADDHNB, RADDHNT Richard Henderson
2021-05-25  1:03 ` [PATCH v7 40/92] target/arm: Implement SVE2 SUBHNB, SUBHNT Richard Henderson
2021-05-25  1:03 ` [PATCH v7 41/92] target/arm: Implement SVE2 RSUBHNB, RSUBHNT Richard Henderson
2021-05-25  1:03 ` [PATCH v7 42/92] target/arm: Implement SVE2 HISTCNT, HISTSEG Richard Henderson
2021-05-25  1:03 ` [PATCH v7 43/92] target/arm: Implement SVE2 XAR Richard Henderson
2021-05-25  1:03 ` [PATCH v7 44/92] target/arm: Implement SVE2 scatter store insns Richard Henderson
2021-05-25  1:03 ` [PATCH v7 45/92] target/arm: Implement SVE2 gather load insns Richard Henderson
2021-05-25  1:03 ` [PATCH v7 46/92] target/arm: Implement SVE2 FMMLA Richard Henderson
2021-05-25  1:03 ` [PATCH v7 47/92] target/arm: Implement SVE2 SPLICE, EXT Richard Henderson
2021-05-25  1:03 ` [PATCH v7 48/92] target/arm: Use correct output type for gvec_sdot_*_b Richard Henderson
2021-05-25  1:03 ` [PATCH v7 49/92] target/arm: Pass separate addend to {U, S}DOT helpers Richard Henderson
2021-05-25  1:03 ` [PATCH v7 50/92] target/arm: Pass separate addend to FCMLA helpers Richard Henderson
2021-05-25  1:03 ` [PATCH v7 51/92] target/arm: Split out formats for 2 vectors + 1 index Richard Henderson
2021-05-25  1:03 ` [PATCH v7 52/92] target/arm: Split out formats for 3 " Richard Henderson
2021-05-25  1:03 ` [PATCH v7 53/92] target/arm: Implement SVE2 integer multiply (indexed) Richard Henderson
2021-05-25  1:03 ` [PATCH v7 54/92] target/arm: Implement SVE2 integer multiply-add (indexed) Richard Henderson
2021-05-25  1:03 ` [PATCH v7 55/92] target/arm: Implement SVE2 saturating multiply-add high (indexed) Richard Henderson
2021-05-25  1:03 ` [PATCH v7 56/92] target/arm: Implement SVE2 saturating multiply-add (indexed) Richard Henderson
2021-05-25  1:03 ` [PATCH v7 57/92] target/arm: Implement SVE2 saturating multiply (indexed) Richard Henderson
2021-05-25  1:03 ` [PATCH v7 58/92] target/arm: Implement SVE2 signed saturating doubling multiply high Richard Henderson
2021-05-25  1:03 ` [PATCH v7 59/92] target/arm: Implement SVE2 saturating multiply high (indexed) Richard Henderson
2021-05-25  1:03 ` [PATCH v7 60/92] target/arm: Implement SVE2 multiply-add long (indexed) Richard Henderson
2021-05-25  1:03 ` [PATCH v7 61/92] target/arm: Implement SVE2 integer multiply " Richard Henderson
2021-05-25  1:03 ` [PATCH v7 62/92] target/arm: Implement SVE2 complex integer multiply-add (indexed) Richard Henderson
2021-05-25  1:03 ` [PATCH v7 63/92] target/arm: Implement SVE2 complex integer dot product Richard Henderson
2021-05-25  1:03 ` [PATCH v7 64/92] target/arm: Macroize helper_gvec_{s,u}dot_{b,h} Richard Henderson
2021-05-25  1:03 ` [PATCH v7 65/92] target/arm: Macroize helper_gvec_{s,u}dot_idx_{b,h} Richard Henderson
2021-05-25  1:03 ` [PATCH v7 66/92] target/arm: Implement SVE mixed sign dot product (indexed) Richard Henderson
2021-05-25  1:03 ` [PATCH v7 67/92] target/arm: Implement SVE mixed sign dot product Richard Henderson
2021-05-25  1:03 ` [PATCH v7 68/92] target/arm: Implement SVE2 crypto unary operations Richard Henderson
2021-05-25  1:03 ` [PATCH v7 69/92] target/arm: Implement SVE2 crypto destructive binary operations Richard Henderson
2021-05-25  1:03 ` [PATCH v7 70/92] target/arm: Implement SVE2 crypto constructive " Richard Henderson
2021-05-25  1:03 ` [PATCH v7 71/92] target/arm: Implement SVE2 TBL, TBX Richard Henderson
2021-05-25  1:03 ` [PATCH v7 72/92] target/arm: Implement SVE2 FCVTNT Richard Henderson
2021-05-25  1:03 ` [PATCH v7 73/92] target/arm: Implement SVE2 FCVTLT Richard Henderson
2021-05-25  1:03 ` [PATCH v7 74/92] target/arm: Implement SVE2 FCVTXNT, FCVTX Richard Henderson
2021-05-25  1:03 ` [PATCH v7 75/92] target/arm: Implement SVE2 FLOGB Richard Henderson
2021-05-25  1:03 ` [PATCH v7 76/92] target/arm: Share table of sve load functions Richard Henderson
2021-05-25  1:03 ` [PATCH v7 77/92] target/arm: Tidy do_ldrq Richard Henderson
2021-05-25  1:03 ` [PATCH v7 78/92] target/arm: Implement SVE2 LD1RO Richard Henderson
2021-05-25  1:03 ` [PATCH v7 79/92] target/arm: Implement 128-bit ZIP, UZP, TRN Richard Henderson
2021-05-25  1:03 ` [PATCH v7 80/92] target/arm: Implement SVE2 bitwise shift immediate Richard Henderson
2021-05-25  1:03 ` [PATCH v7 81/92] target/arm: Move endian adjustment macros to vec_internal.h Richard Henderson
2021-05-25  1:03 ` [PATCH v7 82/92] target/arm: Implement SVE2 fp multiply-add long Richard Henderson
2021-05-25  1:03 ` [PATCH v7 83/92] target/arm: Implement aarch64 SUDOT, USDOT Richard Henderson
2021-05-25  1:03 ` [PATCH v7 84/92] target/arm: Split out do_neon_ddda_fpst Richard Henderson
2021-05-25  1:03 ` [PATCH v7 85/92] target/arm: Remove unused fpst from VDOT_scalar Richard Henderson
2021-05-25  1:03 ` [PATCH v7 86/92] target/arm: Fix decode for VDOT (indexed) Richard Henderson
2021-05-25  1:03 ` [PATCH v7 87/92] target/arm: Split out do_neon_ddda Richard Henderson
2021-05-25  1:03 ` [PATCH v7 88/92] target/arm: Split decode of VSDOT and VUDOT Richard Henderson
2021-05-25  1:03 ` [PATCH v7 89/92] target/arm: Implement aarch32 VSUDOT, VUSDOT Richard Henderson
2021-05-25  1:03 ` [PATCH v7 90/92] target/arm: Implement integer matrix multiply accumulate Richard Henderson
2021-05-25  1:03 ` [PATCH v7 91/92] linux-user/aarch64: Enable hwcap bits for sve2 and related extensions Richard Henderson
2021-05-25  1:03 ` [PATCH v7 92/92] target/arm: Enable SVE2 " Richard Henderson
2021-05-25  2:37 ` [PATCH v7 00/92] target/arm: Implement SVE2 no-reply
2021-05-25 12:33 ` Peter Maydell

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20210525010358.152808-6-richard.henderson@linaro.org \
    --to=richard.henderson@linaro.org \
    --cc=peter.maydell@linaro.org \
    --cc=qemu-arm@nongnu.org \
    --cc=qemu-devel@nongnu.org \
    --subject='Re: [PATCH v7 05/92] target/arm: Split out saturating/rounding shifts from neon' \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link

This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.