load 128 bit floating point constants from global (#21279)

compiler/src/dmd/backend/backconfig.d

@@ -313,6 +313,7 @@ static if (0)
     configv.vasm = vasm;
     configv.verbose = verbose;
 
+    go.AArch64 = arm;
     if (optimize)
         go_flag(go, cast(char*)"-o".ptr);
 

compiler/src/dmd/backend/elem.d

@@ -1144,7 +1144,7 @@ int el_countCommas(const(elem)* e)
  * Convert floating point constant to a read-only symbol.
  * Needed iff floating point code can't load immediate constants.
  */
-@trusted
+private @trusted
 elem* el_convfloat(ref GlobalOptimizer go, elem* e)
 {
     //printf("el_convfloat()\n"); elem_print(e);
@@ -1232,12 +1232,75 @@ elem* el_convfloat(ref GlobalOptimizer go, elem* e)
     return e;
 }
 
+/************************************
+ * Convert AArch64 128 bit floating point constant to a read-only symbol.
+ * Params:
+ *      go = optimizer state
+ *      e  = floating point constant
+ * Returns:
+ *      read-only constant variable
+ */
+private @trusted
+elem* el_convreal(ref GlobalOptimizer go, elem* e)
+{
+    //printf("el_convfloat()\n"); elem_print(e);
+    ubyte[32] buffer = void;
+
+    tym_t ty = e.Ety;
+    int sz = tysize(ty);
+    assert(sz <= buffer.length);
+    void* p;
+    switch (tybasic(ty))
+    {
+        case TYldouble:
+        case TYildouble:
+            /* The size, alignment, and padding of long doubles may be different
+             * from host to target
+             */
+            p = buffer.ptr;
+            // TODO AArch64 these are supposed to be 128 bit floats, not 80 bit
+            memset(buffer.ptr, 0, sz);                      // ensure padding is 0
+            memcpy(buffer.ptr, &e.Vldouble, 10);
+            break;
+
+        case TYcldouble:
+            p = buffer.ptr;
+            memset(buffer.ptr, 0, sz);
+            memcpy(buffer.ptr, &e.Vcldouble.re, 10);
+            memcpy(buffer.ptr + tysize(TYldouble), &e.Vcldouble.im, 10);
+            break;
+
+        default:
+            return e;   // not necessary
+    }
+
+    static if (0)
+    {
+        printf("%gL+%gLi\n", cast(double)e.Vcldouble.re, cast(double)e.Vcldouble.im);
+        printf("el_convfloat() %g %g sz=%d\n", e.Vcdouble.re, e.Vcdouble.im, sz);
+        printf("el_convfloat(): sz = %d\n", sz);
+        ushort* p = cast(ushort*)&e.Vcldouble;
+        for (int i = 0; i < sz/2; i++) printf("%04x ", p[i]);
+        printf("\n");
+    }
+
+    Symbol* s  = out_readonly_sym(ty, p, sz);
+    el_free(e);
+
+    elem* ep = el_ptr(s);
+    elem* ec = el_una(OPind, ty, ep);
+
+    go.changes++;
+    //printf("s: %s %d:x%x\n", s.Sident, s.Sseg, s.Soffset);
+    return ec;
+}
+
 /************************************
  * Convert vector constant to a read-only symbol.
  * Needed iff vector code can't load immediate constants.
  */
 
-@trusted
+private @trusted
 elem* el_convxmm(ref GlobalOptimizer go, elem* e)
 {
     // Do not convert if the constants can be loaded with the special XMM instructions
@@ -1272,7 +1335,7 @@ elem* el_convxmm(ref GlobalOptimizer go, elem* e)
  * stored in the static data segment.
  */
 
-@trusted
+private @trusted
 elem* el_convstring(elem* e)
 {
     //printf("el_convstring()\n");
@@ -1386,6 +1449,7 @@ void shrinkLongDoubleConstantIfPossible(elem* e)
 /*************************
  * Run through a tree converting it to CODGEN.
  */
+public
 @trusted
 elem* el_convert(ref GlobalOptimizer go, elem* e)
 {
@@ -1400,8 +1464,13 @@ elem* el_convert(ref GlobalOptimizer go, elem* e)
         case OPconst:
             if (tyvector(e.Ety))
                 e = el_convxmm(go, e);
-            else if (tyfloating(e.Ety) && config.inline8087)
-                e = el_convfloat(go, e);
+            else if (tyfloating(e.Ety))
+            {
+                if (config.inline8087)
+                    e = el_convfloat(go, e);
+                else if (go.AArch64)
+                    e = el_convreal(go, e);
+            }
             break;
 
         case OPstring:

compiler/src/dmd/backend/goh.d

@@ -74,6 +74,7 @@ struct DefNode
  */
 struct GlobalOptimizer
 {
+    bool AArch64;       // AArch64 is the target
     mftype mfoptim;
     uint changes;       // # of optimizations performed