[PATCH v2 09/10] softfloat: Use aarch64 assembly for {add, sub}{192, 256}

[Richard Henderson <richard.henderson@linaro.org>]

The compiler cannot chain more than two additions together.
Use inline assembly for 3 or 4 additions.

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
 include/fpu/softfloat-macros.h | 14 ++++++++++++++
 fpu/softfloat.c                | 27 +++++++++++++++++++++++++++
 2 files changed, 41 insertions(+)

diff --git a/include/fpu/softfloat-macros.h b/include/fpu/softfloat-macros.h
index 99fa124e56..969a486fd2 100644
--- a/include/fpu/softfloat-macros.h
+++ b/include/fpu/softfloat-macros.h
@@ -442,6 +442,13 @@ static inline void
         "adc %3, %0"
         : "=&r"(*z0Ptr), "=&r"(*z1Ptr), "=&r"(*z2Ptr)
         : "rm"(b0), "rm"(b1), "rm"(b2), "0"(a0), "1"(a1), "2"(a2));
+#elif defined(__aarch64__)
+    asm("adds %2, %x5, %x8\n\t"
+        "adcs %1, %x4, %x7\n\t"
+        "adc  %0, %x3, %x6"
+        : "=&r"(*z0Ptr), "=&r"(*z1Ptr), "=&r"(*z2Ptr)
+        : "rZ"(a0), "rZ"(a1), "rZ"(a2), "rZ"(b0), "rZ"(b1), "rZ"(b2)
+        : "cc");
 #else
     uint64_t z0, z1, z2;
     int8_t carry0, carry1;
@@ -507,6 +514,13 @@ static inline void
         "sbb %3, %0"
         : "=&r"(*z0Ptr), "=&r"(*z1Ptr), "=&r"(*z2Ptr)
         : "rm"(b0), "rm"(b1), "rm"(b2), "0"(a0), "1"(a1), "2"(a2));
+#elif defined(__aarch64__)
+    asm("subs %2, %x5, %x8\n\t"
+        "sbcs %1, %x4, %x7\n\t"
+        "sbc  %0, %x3, %x6"
+        : "=&r"(*z0Ptr), "=&r"(*z1Ptr), "=&r"(*z2Ptr)
+        : "rZ"(a0), "rZ"(a1), "rZ"(a2), "rZ"(b0), "rZ"(b1), "rZ"(b2)
+        : "cc");
 #else
     uint64_t z0, z1, z2;
     int8_t borrow0, borrow1;
diff --git a/fpu/softfloat.c b/fpu/softfloat.c
index fdf5bde69e..07dc17caad 100644
--- a/fpu/softfloat.c
+++ b/fpu/softfloat.c
@@ -7358,6 +7358,18 @@ static void sub256(UInt256 *r, UInt256 *a, UInt256 *b)
         : "=&r"(r->w[0]), "=&r"(r->w[1]), "=&r"(r->w[2]), "=&r"(r->w[3])
         : "rme"(b->w[0]), "rme"(b->w[1]), "rme"(b->w[2]), "rme"(b->w[3]),
             "0"(a->w[0]),   "1"(a->w[1]),   "2"(a->w[2]),   "3"(a->w[3]));
+#elif defined(__aarch64__)
+    asm("subs %[r3], %x[a3], %x[b3]\n\t"
+        "sbcs %[r2], %x[a2], %x[b2]\n\t"
+        "sbcs %[r1], %x[a1], %x[b1]\n\t"
+        "sbc  %[r0], %x[a0], %x[b0]"
+        : [r0] "=&r"(r->w[0]), [r1] "=&r"(r->w[1]),
+          [r2] "=&r"(r->w[2]), [r3] "=&r"(r->w[3])
+        : [a0] "rZ"(a->w[0]), [a1] "rZ"(a->w[1]),
+          [a2] "rZ"(a->w[2]), [a3] "rZ"(a->w[3]),
+          [b0] "rZ"(b->w[0]), [b1] "rZ"(b->w[1]),
+          [b2] "rZ"(b->w[2]), [b3] "rZ"(b->w[3])
+        : "cc");
 #else
     bool borrow = false;
 
@@ -7388,6 +7400,13 @@ static void neg256(UInt256 *a)
         : "=&r"(a->w[0]), "=&r"(a->w[1]), "=&r"(a->w[2]), "+rm"(a->w[3])
         : "rme"(a->w[0]), "rme"(a->w[1]), "rme"(a->w[2]),
           "0"(0), "1"(0), "2"(0));
+#elif defined(__aarch64__)
+    asm("negs %3, %3\n\t"
+        "ngcs %2, %2\n\t"
+        "ngcs %1, %1\n\t"
+        "ngc  %0, %0"
+        : "+r"(a->w[0]), "+r"(a->w[1]), "+r"(a->w[2]), "+r"(a->w[3])
+        : : "cc");
 #else
     /*
      * Recall that -X - 1 = ~X, and that since this is negation,
@@ -7430,6 +7449,14 @@ static void add256(UInt256 *a, UInt256 *b)
         "adc %4, %0"
         :  "+r"(a->w[0]),  "+r"(a->w[1]),  "+r"(a->w[2]),  "+r"(a->w[3])
         : "rme"(b->w[0]), "rme"(b->w[1]), "rme"(b->w[2]), "rme"(b->w[3]));
+#elif defined(__aarch64__)
+    asm("adds %3, %3, %x7\n\t"
+        "adcs %2, %2, %x6\n\t"
+        "adcs %1, %1, %x5\n\t"
+        "adc  %0, %0, %x4"
+        : "+r"(a->w[0]), "+r"(a->w[1]), "+r"(a->w[2]), "+r"(a->w[3])
+        : "rZ"(b->w[0]), "rZ"(b->w[1]), "rZ"(b->w[2]), "rZ"(b->w[3])
+        : "cc");
 #else
     bool carry = false;
 
-- 
2.25.1