fix post increment of floats (dlang#21348)

Author: WalterBright

compiler/src/dmd/backend/arm/cod2.d

@@ -1957,6 +1957,7 @@ void floatPost(ref CGstate cg, ref CodeBuilder cdb,elem* e,ref regm_t pretregs)
     elem* e2 = e.E2;
     const tym_t ty1 = tybasic(e.E1.Ety);
     const sz = _tysize[ty1];
+    assert(sz <= 8);            // TODO AArch64 16 byte floats
     const ftype = INSTR.szToFtype(sz);
 
     regm_t retregs;
@@ -1988,6 +1989,7 @@ void floatPost(ref CGstate cg, ref CodeBuilder cdb,elem* e,ref regm_t pretregs)
             retregs = INSTR.FLOATREGS;
         reg = allocreg(cdb,retregs,ty1);
         loadFromEA(cs,reg,sz == 8 ? 8 : 4,sz);
+        cdb.gen(&cs);
     }
 
     if (regvar && pretregs == mPSW)
@@ -2000,16 +2002,17 @@ void floatPost(ref CGstate cg, ref CodeBuilder cdb,elem* e,ref regm_t pretregs)
         regm_t vretregs = INSTR.FLOATREGS & ~mask(cs.reg);
         reg_t vreg = allocreg(cdb,vretregs,ty1);
         double value = sz == 8 ? e2.Vdouble : e2.Vfloat;
+        uint opx = e.Eoper == OPpostinc ? 0 : 1;
         loadFloatRegConst(cdb,vreg,value,sz);                   // FMOV vreg,value
 
         switch (e.Eoper)
         {
             case OPpostinc:
-                cdb.gen1(INSTR.fadd_float(ftype,reg,vreg,reg)); // FADD Rd,Rn,Rm
+                cdb.gen1(INSTR.fadd_float(ftype,vreg,reg,reg)); // FADD Rd,Rn,Rm
                 break;
 
             case OPpostdec:
-                cdb.gen1(INSTR.fsub_float(ftype,reg,vreg,reg)); // FSUB Rd,Rn,Rm
+                cdb.gen1(INSTR.fsub_float(ftype,vreg,reg,reg)); // FSUB Rd,Rn,Rm
                 break;
 
             default:
@@ -2036,11 +2039,11 @@ void floatPost(ref CGstate cg, ref CodeBuilder cdb,elem* e,ref regm_t pretregs)
     switch (e.Eoper)
     {
         case OPpostinc:
-            cdb.gen1(INSTR.fadd_float(ftype,reg,vreg,reg)); // FADD Rd,Rn,Rm
+            cdb.gen1(INSTR.fadd_float(ftype,vreg,reg,reg)); // FADD Rd,Rn,Rm
             break;
 
         case OPpostdec:
-            cdb.gen1(INSTR.fsub_float(ftype,reg,vreg,reg)); // FSUB Rd,Rn,Rm
+            cdb.gen1(INSTR.fsub_float(ftype,vreg,reg,reg)); // FSUB Rd,Rn,Rm
             break;
 
         default:

compiler/src/dmd/backend/arm/cod4.d

@@ -567,19 +567,19 @@ void floatOpAss(ref CodeBuilder cdb,elem* e,ref regm_t pretregs)
             // FADD/FSUB (extended register)
             // http://www.scs.stanford.edu/~zyedidia/arm64/encodingindex.html#floatdp2
             case OPaddass:
-                cdb.gen1(INSTR.fadd_float(ftype,Rm,Rn,Rd));     // FADD Rd,Rn,Rm
+                cdb.gen1(INSTR.fadd_float(ftype,Rn,Rm,Rd));     // FADD Rd,Rn,Rm
                 break;
 
             case OPminass:
-                cdb.gen1(INSTR.fsub_float(ftype,Rm,Rn,Rd));     // FSUB Rd,Rn,Rm
+                cdb.gen1(INSTR.fsub_float(ftype,Rn,Rm,Rd));     // FSUB Rd,Rn,Rm
                 break;
 
             case OPmulass:
-                cdb.gen1(INSTR.fmul_float(ftype,Rm,Rn,Rd));     // FMUL Rd,Rn,Rm
+                cdb.gen1(INSTR.fmul_float(ftype,Rn,Rm,Rd));     // FMUL Rd,Rn,Rm
                 break;
 
             case OPdivass:
-                cdb.gen1(INSTR.fdiv_float(ftype,Rm,Rn,Rd));     // FDIV Rd,Rn,Rm
+                cdb.gen1(INSTR.fdiv_float(ftype,Rn,Rm,Rd));     // FDIV Rd,Rn,Rm
                 break;
 
             default:
@@ -953,7 +953,7 @@ void cdcmp(ref CGstate cg, ref CodeBuilder cdb,elem* e,ref regm_t pretregs)
         else
         {
             uint ftype = INSTR.szToFtype(sz);
-            cdb.gen1(INSTR.fcmpe_float(ftype,Vm,Vn));    // FCMPE Vn,Vm
+            cdb.gen1(INSTR.fcmpe_float(ftype,Vn,Vm));    // FCMPE Vn,Vm
         }
         goto L3;
     }

compiler/src/dmd/backend/elem.d

@@ -1488,7 +1488,7 @@ elem* el_convert(ref GlobalOptimizer go, elem* e)
              */
             if (tyreal(e.Ety) &&       // don't bother with imaginary or complex
                 e.E2.Eoper == OPconst && el_toldoubled(e.E2) == 2.0L &&
-                !go.AArch64)           // doesn't do the *2 optimization
+                !go.AArch64)           // TODO AArch64 do the *2 optimization
             {
                 e.E1 = el_convert(go, e.E1);
                 /* Don't call el_convert(e.E2), we want it to stay as a constant