add floating point fadd code gen

dlang · Feb 5, 2025 · 9fdcf2a · 9fdcf2a
1 parent 0118e7c
commit 9fdcf2a
Show file tree

Hide file tree

Showing 6 changed files with 82 additions and 18 deletions.
diff --git a/compiler/src/dmd/backend/arm/cod1.d b/compiler/src/dmd/backend/arm/cod1.d
@@ -2274,7 +2274,7 @@ static if (1)
     int sz = _tysize[tym];
     cs.Iflags = 0;
     flags = outretregs & mPSW;             /* save original                */
-    forregs = outretregs & cgstate.allregs;     // XMMREGS ?
+    forregs = outretregs & (cgstate.allregs | INSTR.FLOATREGS);     // XMMREGS ?
     //if (outretregs & mSTACK)
         //forregs |= DOUBLEREGS;
     if (e.Eoper == OPconst)

diff --git a/compiler/src/dmd/backend/arm/cod2.d b/compiler/src/dmd/backend/arm/cod2.d
@@ -59,6 +59,7 @@ import dmd.backend.divcoeff : choose_multiplier, udiv_coefficients;
 void cdorth(ref CGstate cg, ref CodeBuilder cdb,elem* e,ref regm_t pretregs)
 {
     //printf("cdorth(e = %p, pretregs = %s)\n",e,regm_str(pretregs));
+    //elem_print(e);
 
     elem* e1 = e.E1;
     elem* e2 = e.E2;
@@ -82,7 +83,15 @@ void cdorth(ref CGstate cg, ref CodeBuilder cdb,elem* e,ref regm_t pretregs)
     reg_t Rn = findreg(retregs1);
 
     regm_t retregs2 = posregs & ~retregs1;
+//printf("retregs1: %s retregs2: %s\n", regm_str(retregs1), regm_str(retregs2));
+static if (0)
+{
     scodelem(cg, cdb, e2, retregs2, retregs1, false);
+}
+else
+{
+    retregs2 = mask(33);
+}
     reg_t Rm = findreg(retregs2);
 
     regm_t retregs = pretregs & posregs;
@@ -94,21 +103,18 @@ void cdorth(ref CGstate cg, ref CodeBuilder cdb,elem* e,ref regm_t pretregs)
 
     if (tyfloating(ty1))
     {
+        uint ftype = sz == 2 ? 3 :
+                     sz == 4 ? 0 : 1;
         switch (e.Eoper)
         {
             // FADD/FSUB (extended register)
             // http://www.scs.stanford.edu/~zyedidia/arm64/encodingindex.html#addsub_ext
             case OPadd:
+                cdb.gen1(INSTR.fadd_float(ftype,Rm,Rn,Rd));     // FADD Rd,Rn,Rm
+                break;
+
             case OPmin:
-                uint sf = sz == 8;
-                uint op = e.Eoper == OPadd ? 0 : 1;
-                uint S = PSW != 0;
-                uint opt = 0;
-                uint option = tyToExtend(ty);
-                uint imm3 = 0;
-                cdb.gen1(INSTR.addsub_ext(sf, op, S, opt, Rm, option, imm3, Rn, Rd));
-                PSW = 0;
-                pretregs &= ~mPSW;
+                cdb.gen1(INSTR.fsub_float(ftype,Rm,Rn,Rd));     // FSUB Rd,Rn,Rm
                 break;
 
             default:

diff --git a/compiler/src/dmd/backend/arm/instr.d b/compiler/src/dmd/backend/arm/instr.d
@@ -737,8 +737,56 @@ struct INSTR
     /* Floating-point compare
      * Floating-point immediate
      * Floating-point condistional compare
-     * Floating-point data-processing (2 source)
-     * Floating-point conditional select
+     */
+
+    /* Floating-point data-processing (2 source) https://www.scs.stanford.edu/~zyedidia/arm64/encodingindex.html#floatdp2
+     */
+    static uint floatdp2(uint M, uint S, uint ftype, reg_t Vm, uint opcode, reg_t Vn, reg_t Vd)
+    {
+        assert(Vm >= 32 && Vn >= 32 && Vd >= 32);
+        reg_t Rm = Vm & 31;
+        reg_t Rn = Vn & 31;
+        reg_t Rd = Vd & 31;
+        return (M << 31) | (S << 29) | (0x1E << 24) | (ftype << 22) | (1 << 21) | (Rm << 16) | (opcode << 12) | (2 << 10) | (Rn << 5) | Rd;
+    }
+
+    /* FMUL (scalar) https://www.scs.stanford.edu/~zyedidia/arm64/fmul_float.html
+     */
+    static uint fmul_float(uint ftype, reg_t Vm, reg_t Vn, reg_t Vd) { return floatdp2(0,0,ftype,Vm,0,Vn,Vd); }
+
+    /* FDIV (scalar) https://www.scs.stanford.edu/~zyedidia/arm64/fdiv_float.html
+     */
+    static uint fdiv_float(uint ftype, reg_t Vm, reg_t Vn, reg_t Vd) { return floatdp2(0,0,ftype,Vm,1,Vn,Vd); }
+
+    /* FADD (scalar) https://www.scs.stanford.edu/~zyedidia/arm64/fadd_float.html
+     */
+    static uint fadd_float(uint ftype, reg_t Vm, reg_t Vn, reg_t Vd) { return floatdp2(0,0,ftype,Vm,2,Vn,Vd); }
+
+    /* FSUB (scalar) https://www.scs.stanford.edu/~zyedidia/arm64/fsub_float.html
+     */
+    static uint fsub_float(uint ftype, reg_t Vm, reg_t Vn, reg_t Vd) { return floatdp2(0,0,ftype,Vm,3,Vn,Vd); }
+
+    /* FMAX (scalar) https://www.scs.stanford.edu/~zyedidia/arm64/fmax_float.html
+     */
+    static uint fmax_float(uint ftype, reg_t Vm, reg_t Vn, reg_t Vd) { return floatdp2(0,0,ftype,Vm,4,Vn,Vd); }
+
+    /* FMIN (scalar) https://www.scs.stanford.edu/~zyedidia/arm64/fmin_float.html
+     */
+    static uint fmin_float(uint ftype, reg_t Vm, reg_t Vn, reg_t Vd) { return floatdp2(0,0,ftype,Vm,5,Vn,Vd); }
+
+    /* FMAXNM (scalar) https://www.scs.stanford.edu/~zyedidia/arm64/fmaxnm_float.html
+     */
+    static uint fmaxnm_float(uint ftype, reg_t Vm, reg_t Vn, reg_t Vd) { return floatdp2(0,0,ftype,Vm,6,Vn,Vd); }
+
+    /* FMINNM (scalar) https://www.scs.stanford.edu/~zyedidia/arm64/fminnm_float.html
+     */
+    static uint fminnm_float(uint ftype, reg_t Vm, reg_t Vn, reg_t Vd) { return floatdp2(0,0,ftype,Vm,7,Vn,Vd); }
+
+    /* FNMUL (scalar) https://www.scs.stanford.edu/~zyedidia/arm64/fnmul_float.html
+     */
+    static uint fnmul_float(uint ftype, reg_t Vm, reg_t Vn, reg_t Vd) { return floatdp2(0,0,ftype,Vm,8,Vn,Vd); }
+
+    /* Floating-point conditional select
      * Floating-point data-processing (3 source)
      */
 

diff --git a/compiler/src/dmd/backend/debugprint.d b/compiler/src/dmd/backend/debugprint.d
@@ -153,7 +153,10 @@ const(char)* tym_str(tym_t ty)
     const tyb = tybasic(ty);
     if (tyb >= TYMAX)
     {
-        printf("TY %x\n",cast(int)ty);
+        if (tyb == TYMAX)
+            printf("TY TYMAX\n");
+        else
+            printf("TY %x\n",cast(int)ty);
         assert(0);
     }
     strcat(p, "TY");

diff --git a/compiler/src/dmd/backend/x86/cgcod.d b/compiler/src/dmd/backend/x86/cgcod.d
@@ -40,6 +40,7 @@ import dmd.backend.ty;
 import dmd.backend.type;
 
 import dmd.backend.arm.disasmarm;
+import dmd.backend.arm.instr;
 
 import dmd.backend.x86.code_x86;
 import dmd.backend.x86.disasm86;
@@ -1620,7 +1621,7 @@ static if (0)
 }
         tym = tybasic(tym);
         uint size = _tysize[tym];
-        outretregs &= mES | cgstate.allregs | XMMREGS;
+        outretregs &= mES | cgstate.allregs | XMMREGS | INSTR.FLOATREGS;
         regm_t retregs = outretregs;
         regm_t[] lastRetregs = cgstate.lastRetregs[];
 
@@ -1630,7 +1631,7 @@ static if (0)
         if ((retregs & cgstate.regcon.mvar) == retregs) // if exactly in reg vars
         {
             reg_t outreg;
-            if (size <= REGSIZE || (retregs & XMMREGS))
+            if (size <= REGSIZE || (retregs & XMMREGS) || (retregs & INSTR.FLOATREGS))
             {
                 outreg = findreg(retregs);
                 assert(retregs == mask(outreg)); /* no more bits are set */
@@ -3002,7 +3003,7 @@ const(char)* regm_str(regm_t rm)
                 {
                     char[4] buf = void;
                     char c = j < 32 ? 'r' : 'f';
-                    sprintf(buf.ptr, "c%u", c, j);
+                    sprintf(buf.ptr, "%c%u", c, j);
                     strcat(p, buf.ptr);
                 }
             }

diff --git a/compiler/src/dmd/backend/x86/cod3.d b/compiler/src/dmd/backend/x86/cod3.d
@@ -1142,7 +1142,8 @@ static if (NTEXCEPTIONS)
         case BC.retexp:
             reg_t reg1, reg2;
             retregs = allocretregs(cgstate, e.Ety, e.ET, funcsym_p.ty(), reg1, reg2);
-            //printf("allocretregs returns %s\n", regm_str(mask(reg1) | mask(reg2)));
+	    //printf("reg1: %d, reg2: %d\n", reg1, reg2);
+            //printf("allocretregs returns %llx %s\n", retregs, regm_str(retregs));
 
             reg_t lreg = NOREG;
             reg_t mreg = NOREG;
@@ -1377,7 +1378,7 @@ static if (NTEXCEPTIONS)
 @trusted
 regm_t allocretregs(ref CGstate cgstate, const tym_t ty, type* t, const tym_t tyf, out reg_t reg1, out reg_t reg2)
 {
-    //printf("allocretregs() ty: %s\n", tym_str(ty));
+    printf("allocretregs() ty: %s\n", tym_str(ty));
     reg1 = reg2 = NOREG;
     auto AArch64 = cgstate.AArch64;
 
@@ -1504,6 +1505,7 @@ regm_t allocretregs(ref CGstate cgstate, const tym_t ty, type* t, const tym_t ty
     {
         if (tym == TYMAX)
             return NOREG;
+printf("tym: %s\n", tym_str(tym));
         switch (tysize(tym))
         {
         case 1:
@@ -1525,6 +1527,8 @@ regm_t allocretregs(ref CGstate cgstate, const tym_t ty, type* t, const tym_t ty
                 assert(tyfb == TYjfunc && I32);
                 return ST01;
             }
+            else if (AArch64 && tyfloating(tym))
+                return rralloc.fpt();
             else if (tysimd(tym))
             {
                 return rralloc.xmm();
@@ -1559,6 +1563,8 @@ regm_t allocretregs(ref CGstate cgstate, const tym_t ty, type* t, const tym_t ty
     reg1 = allocreg(ty1);
     reg2 = allocreg(ty2);
 
+    //printf("reg1: %d reg2: %d NOREG: %d\n", reg1, reg2, NOREG);
+    //printf("reg1: %llx reg2: %llx ~NOREG: %llx\n", mask(reg1), mask(reg2), ~mask(NOREG));
     return (mask(reg1) | mask(reg2)) & ~mask(NOREG);
 }