target/arm: Implement SVE2 HISTCNT, HISTSEG

Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Stephen Long <steplong@quicinc.com>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20210525010358.152808-43-richard.henderson@linaro.org
Message-Id: <20200416173109.8856-1-steplong@quicinc.com>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>

1 files changed, 131 insertions, 0 deletions

diff --git a/target/arm/sve_helper.c b/target/arm/sve_helper.c
index 891f6ff..662ed80 100644
--- a/target/arm/sve_helper.c
+++ b/target/arm/sve_helper.c
@@ -7071,3 +7071,134 @@ DO_PPZZ_MATCH(sve2_nmatch_ppzz_b, MO_8, true)
 DO_PPZZ_MATCH(sve2_nmatch_ppzz_h, MO_16, true)
 
 #undef DO_PPZZ_MATCH
+
+void HELPER(sve2_histcnt_s)(void *vd, void *vn, void *vm, void *vg,
+                            uint32_t desc)
+{
+    ARMVectorReg scratch;
+    intptr_t i, j;
+    intptr_t opr_sz = simd_oprsz(desc);
+    uint32_t *d = vd, *n = vn, *m = vm;
+    uint8_t *pg = vg;
+
+    if (d == n) {
+        n = memcpy(&scratch, n, opr_sz);
+        if (d == m) {
+            m = n;
+        }
+    } else if (d == m) {
+        m = memcpy(&scratch, m, opr_sz);
+    }
+
+    for (i = 0; i < opr_sz; i += 4) {
+        uint64_t count = 0;
+        uint8_t pred;
+
+        pred = pg[H1(i >> 3)] >> (i & 7);
+        if (pred & 1) {
+            uint32_t nn = n[H4(i >> 2)];
+
+            for (j = 0; j <= i; j += 4) {
+                pred = pg[H1(j >> 3)] >> (j & 7);
+                if ((pred & 1) && nn == m[H4(j >> 2)]) {
+                    ++count;
+                }
+            }
+        }
+        d[H4(i >> 2)] = count;
+    }
+}
+
+void HELPER(sve2_histcnt_d)(void *vd, void *vn, void *vm, void *vg,
+                            uint32_t desc)
+{
+    ARMVectorReg scratch;
+    intptr_t i, j;
+    intptr_t opr_sz = simd_oprsz(desc);
+    uint64_t *d = vd, *n = vn, *m = vm;
+    uint8_t *pg = vg;
+
+    if (d == n) {
+        n = memcpy(&scratch, n, opr_sz);
+        if (d == m) {
+            m = n;
+        }
+    } else if (d == m) {
+        m = memcpy(&scratch, m, opr_sz);
+    }
+
+    for (i = 0; i < opr_sz / 8; ++i) {
+        uint64_t count = 0;
+        if (pg[H1(i)] & 1) {
+            uint64_t nn = n[i];
+            for (j = 0; j <= i; ++j) {
+                if ((pg[H1(j)] & 1) && nn == m[j]) {
+                    ++count;
+                }
+            }
+        }
+        d[i] = count;
+    }
+}
+
+/*
+ * Returns the number of bytes in m0 and m1 that match n.
+ * Unlike do_match2 we don't just need true/false, we need an exact count.
+ * This requires two extra logical operations.
+ */
+static inline uint64_t do_histseg_cnt(uint8_t n, uint64_t m0, uint64_t m1)
+{
+    const uint64_t mask = dup_const(MO_8, 0x7f);
+    uint64_t cmp0, cmp1;
+
+    cmp1 = dup_const(MO_8, n);
+    cmp0 = cmp1 ^ m0;
+    cmp1 = cmp1 ^ m1;
+
+    /*
+     * 1: clear msb of each byte to avoid carry to next byte (& mask)
+     * 2: carry in to msb if byte != 0 (+ mask)
+     * 3: set msb if cmp has msb set (| cmp)
+     * 4: set ~msb to ignore them (| mask)
+     * We now have 0xff for byte != 0 or 0x7f for byte == 0.
+     * 5: invert, resulting in 0x80 if and only if byte == 0.
+     */
+    cmp0 = ~(((cmp0 & mask) + mask) | cmp0 | mask);
+    cmp1 = ~(((cmp1 & mask) + mask) | cmp1 | mask);
+
+    /*
+     * Combine the two compares in a way that the bits do
+     * not overlap, and so preserves the count of set bits.
+     * If the host has an efficient instruction for ctpop,
+     * then ctpop(x) + ctpop(y) has the same number of
+     * operations as ctpop(x | (y >> 1)).  If the host does
+     * not have an efficient ctpop, then we only want to
+     * use it once.
+     */
+    return ctpop64(cmp0 | (cmp1 >> 1));
+}
+
+void HELPER(sve2_histseg)(void *vd, void *vn, void *vm, uint32_t desc)
+{
+    intptr_t i, j;
+    intptr_t opr_sz = simd_oprsz(desc);
+
+    for (i = 0; i < opr_sz; i += 16) {
+        uint64_t n0 = *(uint64_t *)(vn + i);
+        uint64_t m0 = *(uint64_t *)(vm + i);
+        uint64_t n1 = *(uint64_t *)(vn + i + 8);
+        uint64_t m1 = *(uint64_t *)(vm + i + 8);
+        uint64_t out0 = 0;
+        uint64_t out1 = 0;
+
+        for (j = 0; j < 64; j += 8) {
+            uint64_t cnt0 = do_histseg_cnt(n0 >> j, m0, m1);
+            uint64_t cnt1 = do_histseg_cnt(n1 >> j, m0, m1);
+            out0 |= cnt0 << j;
+            out1 |= cnt1 << j;
+        }
+
+        *(uint64_t *)(vd + i) = out0;
+        *(uint64_t *)(vd + i + 8) = out1;
+    }
+}