target/arm: Implement SVE2 saturating multiply-add high

SVE2 has two additional sizes of the operation and unlike NEON,
there is no saturation flag.  Create new entry points for SVE2
that do not set QC.

Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20210525010358.152808-36-richard.henderson@linaro.org
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>

1 files changed, 155 insertions, 6 deletions

diff --git a/target/arm/vec_helper.c b/target/arm/vec_helper.c
index b0ce597060..c56337e724 100644
--- a/target/arm/vec_helper.c
+++ b/target/arm/vec_helper.c
@@ -22,6 +22,7 @@
 #include "exec/helper-proto.h"
 #include "tcg/tcg-gvec-desc.h"
 #include "fpu/softfloat.h"
+#include "qemu/int128.h"
 #include "vec_internal.h"
 
 /* Note that vector data is stored in host-endian 64-bit chunks,
@@ -36,19 +37,59 @@
 #define H4(x)  (x)
 #endif
 
-/* Signed saturating rounding doubling multiply-accumulate high half, 16-bit */
-static int16_t do_sqrdmlah_h(int16_t src1, int16_t src2, int16_t src3,
-                             bool neg, bool round, uint32_t *sat)
+/* Signed saturating rounding doubling multiply-accumulate high half, 8-bit */
+static int8_t do_sqrdmlah_b(int8_t src1, int8_t src2, int8_t src3,
+                            bool neg, bool round)
 {
     /*
      * Simplify:
-     * = ((a3 << 16) + ((e1 * e2) << 1) + (1 << 15)) >> 16
-     * = ((a3 << 15) + (e1 * e2) + (1 << 14)) >> 15
+     * = ((a3 << 8) + ((e1 * e2) << 1) + (round << 7)) >> 8
+     * = ((a3 << 7) + (e1 * e2) + (round << 6)) >> 7
      */
     int32_t ret = (int32_t)src1 * src2;
     if (neg) {
         ret = -ret;
     }
+    ret += ((int32_t)src3 << 7) + (round << 6);
+    ret >>= 7;
+
+    if (ret != (int8_t)ret) {
+        ret = (ret < 0 ? INT8_MIN : INT8_MAX);
+    }
+    return ret;
+}
+
+void HELPER(sve2_sqrdmlah_b)(void *vd, void *vn, void *vm,
+                             void *va, uint32_t desc)
+{
+    intptr_t i, opr_sz = simd_oprsz(desc);
+    int8_t *d = vd, *n = vn, *m = vm, *a = va;
+
+    for (i = 0; i < opr_sz; ++i) {
+        d[i] = do_sqrdmlah_b(n[i], m[i], a[i], false, true);
+    }
+}
+
+void HELPER(sve2_sqrdmlsh_b)(void *vd, void *vn, void *vm,
+                             void *va, uint32_t desc)
+{
+    intptr_t i, opr_sz = simd_oprsz(desc);
+    int8_t *d = vd, *n = vn, *m = vm, *a = va;
+
+    for (i = 0; i < opr_sz; ++i) {
+        d[i] = do_sqrdmlah_b(n[i], m[i], a[i], true, true);
+    }
+}
+
+/* Signed saturating rounding doubling multiply-accumulate high half, 16-bit */
+static int16_t do_sqrdmlah_h(int16_t src1, int16_t src2, int16_t src3,
+                             bool neg, bool round, uint32_t *sat)
+{
+    /* Simplify similarly to do_sqrdmlah_b above.  */
+    int32_t ret = (int32_t)src1 * src2;
+    if (neg) {
+        ret = -ret;
+    }
     ret += ((int32_t)src3 << 15) + (round << 14);
     ret >>= 15;
 
@@ -133,11 +174,35 @@ void HELPER(neon_sqrdmulh_h)(void *vd, void *vn, void *vm,
     clear_tail(d, opr_sz, simd_maxsz(desc));
 }
 
+void HELPER(sve2_sqrdmlah_h)(void *vd, void *vn, void *vm,
+                             void *va, uint32_t desc)
+{
+    intptr_t i, opr_sz = simd_oprsz(desc);
+    int16_t *d = vd, *n = vn, *m = vm, *a = va;
+    uint32_t discard;
+
+    for (i = 0; i < opr_sz / 2; ++i) {
+        d[i] = do_sqrdmlah_h(n[i], m[i], a[i], false, true, &discard);
+    }
+}
+
+void HELPER(sve2_sqrdmlsh_h)(void *vd, void *vn, void *vm,
+                             void *va, uint32_t desc)
+{
+    intptr_t i, opr_sz = simd_oprsz(desc);
+    int16_t *d = vd, *n = vn, *m = vm, *a = va;
+    uint32_t discard;
+
+    for (i = 0; i < opr_sz / 2; ++i) {
+        d[i] = do_sqrdmlah_h(n[i], m[i], a[i], true, true, &discard);
+    }
+}
+
 /* Signed saturating rounding doubling multiply-accumulate high half, 32-bit */
 static int32_t do_sqrdmlah_s(int32_t src1, int32_t src2, int32_t src3,
                              bool neg, bool round, uint32_t *sat)
 {
-    /* Simplify similarly to int_qrdmlah_s16 above.  */
+    /* Simplify similarly to do_sqrdmlah_b above.  */
     int64_t ret = (int64_t)src1 * src2;
     if (neg) {
         ret = -ret;
@@ -220,6 +285,90 @@ void HELPER(neon_sqrdmulh_s)(void *vd, void *vn, void *vm,
     clear_tail(d, opr_sz, simd_maxsz(desc));
 }
 
+void HELPER(sve2_sqrdmlah_s)(void *vd, void *vn, void *vm,
+                             void *va, uint32_t desc)
+{
+    intptr_t i, opr_sz = simd_oprsz(desc);
+    int32_t *d = vd, *n = vn, *m = vm, *a = va;
+    uint32_t discard;
+
+    for (i = 0; i < opr_sz / 4; ++i) {
+        d[i] = do_sqrdmlah_s(n[i], m[i], a[i], false, true, &discard);
+    }
+}
+
+void HELPER(sve2_sqrdmlsh_s)(void *vd, void *vn, void *vm,
+                             void *va, uint32_t desc)
+{
+    intptr_t i, opr_sz = simd_oprsz(desc);
+    int32_t *d = vd, *n = vn, *m = vm, *a = va;
+    uint32_t discard;
+
+    for (i = 0; i < opr_sz / 4; ++i) {
+        d[i] = do_sqrdmlah_s(n[i], m[i], a[i], true, true, &discard);
+    }
+}
+
+/* Signed saturating rounding doubling multiply-accumulate high half, 64-bit */
+static int64_t do_sat128_d(Int128 r)
+{
+    int64_t ls = int128_getlo(r);
+    int64_t hs = int128_gethi(r);
+
+    if (unlikely(hs != (ls >> 63))) {
+        return hs < 0 ? INT64_MIN : INT64_MAX;
+    }
+    return ls;
+}
+
+static int64_t do_sqrdmlah_d(int64_t n, int64_t m, int64_t a,
+                             bool neg, bool round)
+{
+    uint64_t l, h;
+    Int128 r, t;
+
+    /* As in do_sqrdmlah_b, but with 128-bit arithmetic. */
+    muls64(&l, &h, m, n);
+    r = int128_make128(l, h);
+    if (neg) {
+        r = int128_neg(r);
+    }
+    if (a) {
+        t = int128_exts64(a);
+        t = int128_lshift(t, 63);
+        r = int128_add(r, t);
+    }
+    if (round) {
+        t = int128_exts64(1ll << 62);
+        r = int128_add(r, t);
+    }
+    r = int128_rshift(r, 63);
+
+    return do_sat128_d(r);
+}
+
+void HELPER(sve2_sqrdmlah_d)(void *vd, void *vn, void *vm,
+                             void *va, uint32_t desc)
+{
+    intptr_t i, opr_sz = simd_oprsz(desc);
+    int64_t *d = vd, *n = vn, *m = vm, *a = va;
+
+    for (i = 0; i < opr_sz / 8; ++i) {
+        d[i] = do_sqrdmlah_d(n[i], m[i], a[i], false, true);
+    }
+}
+
+void HELPER(sve2_sqrdmlsh_d)(void *vd, void *vn, void *vm,
+                             void *va, uint32_t desc)
+{
+    intptr_t i, opr_sz = simd_oprsz(desc);
+    int64_t *d = vd, *n = vn, *m = vm, *a = va;
+
+    for (i = 0; i < opr_sz / 8; ++i) {
+        d[i] = do_sqrdmlah_d(n[i], m[i], a[i], true, true);
+    }
+}
+
 /* Integer 8 and 16-bit dot-product.
  *
  * Note that for the loops herein, host endianness does not matter