VM: support importc var, ptr/pointer types, cast int <=> ptr/pointer (#12877)

* VM: allow certain hardcoded special var variables at CT
* VM: allow all importc var, cast[int](ptr)
* fix tests tests/vm/tstaticprintseq.nim, tests/cpp/t8241.nim
* VM: == works for ptr/pointer nodes
* bugfix: ==, cast now also works for pointer, not just ptr
* VM supports cast PtrLikeKinds <=> PtrLikeKinds / int
* improve cname handling
* fixup + bug fix
* VM: support cast from ref to int
* address comment: opcLdGlobalDeref => opcLdGlobalDerefFFI
* defensive check against typ == nil

7 files changed, 199 insertions, 35 deletions

diff --git a/compiler/ast.nim b/compiler/ast.nim
index 0daa1f11f..3b9726a13 100644
--- a/compiler/ast.nim
+++ b/compiler/ast.nim
@@ -469,6 +469,7 @@ type
     nfExplicitCall # x.y() was used instead of x.y
     nfExprCall  # this is an attempt to call a regular expression
     nfIsRef     # this node is a 'ref' node; used for the VM
+    nfIsPtr     # this node is a 'ptr' node; used for the VM
     nfPreventCg # this node should be ignored by the codegen
     nfBlockArg  # this a stmtlist appearing in a call (e.g. a do block)
     nfFromTemplate # a top-level node returned from a template
@@ -856,6 +857,9 @@ type
     loc*: TLoc
     annex*: PLib              # additional fields (seldom used, so we use a
                               # reference to another object to save space)
+    when hasFFI:
+      cname*: string          # resolved C declaration name in importc decl, eg:
+                              # proc fun() {.importc: "$1aux".} => cname = funaux
     constraint*: PNode        # additional constraints like 'lit|result'; also
                               # misused for the codegenDecl pragma in the hope
                               # it won't cause problems
@@ -978,12 +982,13 @@ const
                                     tyTuple, tySequence}
   NilableTypes*: TTypeKinds = {tyPointer, tyCString, tyRef, tyPtr,
     tyProc, tyError}
+  PtrLikeKinds*: TTypeKinds = {tyPointer, tyPtr} # for VM
   ExportableSymKinds* = {skVar, skConst, skProc, skFunc, skMethod, skType,
     skIterator,
     skMacro, skTemplate, skConverter, skEnumField, skLet, skStub, skAlias}
   PersistentNodeFlags*: TNodeFlags = {nfBase2, nfBase8, nfBase16,
                                       nfDotSetter, nfDotField,
-                                      nfIsRef, nfPreventCg, nfLL,
+                                      nfIsRef, nfIsPtr, nfPreventCg, nfLL,
                                       nfFromTemplate, nfDefaultRefsParam,
                                       nfExecuteOnReload}
   namePos* = 0
diff --git a/compiler/cgen.nim b/compiler/cgen.nim
index da041bf14..ee4634aad 100644
--- a/compiler/cgen.nim
+++ b/compiler/cgen.nim
@@ -700,7 +700,7 @@ proc loadDynamicLib(m: BModule, lib: PLib) =
 
 proc mangleDynLibProc(sym: PSym): Rope =
   # we have to build this as a single rope in order not to trip the
-  # optimization in genInfixCall
+  # optimization in genInfixCall, see test tests/cpp/t8241.nim
   if sfCompilerProc in sym.flags:
     # NOTE: sym.loc.r is the external name!
     result = rope(sym.name.s)
diff --git a/compiler/evalffi.nim b/compiler/evalffi.nim
index 543810886..cb383a393 100644
--- a/compiler/evalffi.nim
+++ b/compiler/evalffi.nim
@@ -46,19 +46,12 @@ proc getDll(conf: ConfigRef, cache: var TDllCache; dll: string; info: TLineInfo)
 const
   nkPtrLit = nkIntLit # hopefully we can get rid of this hack soon
 
-var myerrno {.importc: "errno", header: "<errno.h>".}: cint ## error variable
-
 proc importcSymbol*(conf: ConfigRef, sym: PSym): PNode =
-  let name = $sym.loc.r
+  let name = sym.cname # $sym.loc.r would point to internal name
   # the AST does not support untyped pointers directly, so we use an nkIntLit
   # that contains the address instead:
   result = newNodeIT(nkPtrLit, sym.info, sym.typ)
-  case name
-  of "stdin":  result.intVal = cast[ByteAddress](system.stdin)
-  of "stdout": result.intVal = cast[ByteAddress](system.stdout)
-  of "stderr": result.intVal = cast[ByteAddress](system.stderr)
-  of "vmErrnoWrapper": result.intVal = cast[ByteAddress](myerrno)
-  else:
+  when true:
     let lib = sym.annex
     if lib != nil and lib.path.kind notin {nkStrLit..nkTripleStrLit}:
       globalError(conf, sym.info, "dynlib needs to be a string lit")
@@ -74,7 +67,7 @@ proc importcSymbol*(conf: ConfigRef, sym: PSym): PNode =
       let dll = if lib.kind == libHeader: libcDll else: lib.path.strVal
       let dllhandle = getDll(conf, gDllCache, dll, sym.info)
       theAddr = dllhandle.symAddr(name)
-    if theAddr.isNil: globalError(conf, sym.info, "cannot import: " & sym.name.s)
+    if theAddr.isNil: globalError(conf, sym.info, "cannot import: " & name)
     result.intVal = cast[ByteAddress](theAddr)
 
 proc mapType(conf: ConfigRef, t: ast.PType): ptr libffi.TType =
diff --git a/compiler/pragmas.nim b/compiler/pragmas.nim
index 726402bb4..1802493f4 100644
--- a/compiler/pragmas.nim
+++ b/compiler/pragmas.nim
@@ -129,6 +129,8 @@ proc setExternName(c: PContext; s: PSym, extname: string, info: TLineInfo) =
       s.loc.r = rope(extname % s.name.s)
     except ValueError:
       localError(c.config, info, "invalid extern name: '" & extname & "'. (Forgot to escape '$'?)")
+  when hasFFI:
+    s.cname = $s.loc.r
   if c.config.cmd == cmdPretty and '$' notin extname:
     # note that '{.importc.}' is transformed into '{.importc: "$1".}'
     s.loc.flags.incl(lfFullExternalName)
diff --git a/compiler/vm.nim b/compiler/vm.nim
index 4403a385d..7d3d608ff 100644
--- a/compiler/vm.nim
+++ b/compiler/vm.nim
@@ -140,6 +140,40 @@ template decodeBx(k: untyped) {.dirty.} =
 template move(a, b: untyped) {.dirty.} = system.shallowCopy(a, b)
 # XXX fix minor 'shallowCopy' overloading bug in compiler
 
+proc derefPtrToReg(address: BiggestInt, typ: PType, r: var TFullReg, isAssign: bool): bool =
+  # nim bug: `isAssign: static bool` doesn't work, giving odd compiler error
+  template fun(field, T, rkind) =
+    if isAssign:
+      cast[ptr T](address)[] = T(r.field)
+    else:
+      # ensureKind(rkind)
+      if r.kind != rkind:
+        myreset(r)
+        r.kind = rkind
+      let val = cast[ptr T](address)[]
+      when T is SomeInteger:
+        r.field = BiggestInt(val)
+      else:
+        r.field = val
+    return true
+
+  ## see also typeinfo.getBiggestInt
+  case typ.kind
+  of tyInt: fun(intVal, int, rkInt)
+  of tyInt8: fun(intVal, int8, rkInt)
+  of tyInt16: fun(intVal, int16, rkInt)
+  of tyInt32: fun(intVal, int32, rkInt)
+  of tyInt64: fun(intVal, int64, rkInt)
+  of tyUInt: fun(intVal, uint, rkInt)
+  of tyUInt8: fun(intVal, uint8, rkInt)
+  of tyUInt16: fun(intVal, uint16, rkInt)
+  of tyUInt32: fun(intVal, uint32, rkInt)
+  of tyUInt64: fun(intVal, uint64, rkInt) # note: differs from typeinfo.getBiggestInt
+  of tyFloat: fun(floatVal, float, rkFloat)
+  of tyFloat32: fun(floatVal, float32, rkFloat)
+  of tyFloat64: fun(floatVal, float64, rkFloat)
+  else: return false
+
 proc createStrKeepNode(x: var TFullReg; keepNode=true) =
   if x.node.isNil or not keepNode:
     x.node = newNode(nkStrLit)
@@ -242,14 +276,20 @@ proc writeField(n: var PNode, x: TFullReg) =
   of rkNodeAddr: n = x.nodeAddr[]
 
 proc putIntoReg(dest: var TFullReg; n: PNode) =
+  template funInt() =
+    dest.kind = rkInt
+    dest.intVal = n.intVal
   case n.kind
   of nkStrLit..nkTripleStrLit:
     dest.kind = rkNode
     createStr(dest)
     dest.node.strVal = n.strVal
-  of nkCharLit..nkUInt64Lit:
-    dest.kind = rkInt
-    dest.intVal = n.intVal
+  of nkIntLit: # use `nkPtrLit` once this is added
+    if dest.kind == rkNode: dest.node = n
+    elif n.typ != nil and n.typ.kind in PtrLikeKinds:
+      dest = TFullReg(kind: rkNode, node: n)
+    else: funInt()
+  of {nkCharLit..nkUInt64Lit} - {nkIntLit}: funInt()
   of nkFloatLit..nkFloat128Lit:
     dest.kind = rkFloat
     dest.floatVal = n.floatVal
@@ -567,6 +607,33 @@ proc rawExecute(c: PCtx, start: int, tos: PStackFrame): TFullReg =
       let rb = instr.regB
       ensureKind(rkFloat)
       regs[ra].floatVal = cast[float64](int64(regs[rb].intVal))
+
+    of opcCastPtrToInt: # RENAME opcCastPtrOrRefToInt
+      decodeBImm(rkInt)
+      case imm
+      of 1: # PtrLikeKinds
+        case regs[rb].kind
+        of rkNode:
+          regs[ra].intVal = cast[int](regs[rb].node.intVal)
+        of rkNodeAddr:
+          regs[ra].intVal = cast[int](regs[rb].nodeAddr)
+        else:
+          stackTrace(c, tos, pc, "opcCastPtrToInt: got " & $regs[rb].kind)
+      of 2: # tyRef
+        regs[ra].intVal = cast[int](regs[rb].node)
+      else: assert false, $imm
+    of opcCastIntToPtr:
+      let rb = instr.regB
+      let typ = regs[ra].node.typ
+      let node2 = newNodeIT(nkIntLit, c.debug[pc], typ)
+      case regs[rb].kind
+      of rkInt: node2.intVal = regs[rb].intVal
+      of rkNode:
+        if regs[rb].node.typ.kind notin PtrLikeKinds:
+          stackTrace(c, tos, pc, "opcCastIntToPtr: regs[rb].node.typ: " & $regs[rb].node.typ.kind)
+        node2.intVal = regs[rb].node.intVal
+      else: stackTrace(c, tos, pc, "opcCastIntToPtr: regs[rb].kind: " & $regs[rb].kind)
+      regs[ra].node = node2
     of opcAsgnComplex:
       asgnComplex(regs[ra], regs[instr.regB])
     of opcFastAsgnComplex:
@@ -651,7 +718,10 @@ proc rawExecute(c: PCtx, start: int, tos: PStackFrame): TFullReg =
       let src = regs[rb].node
       case src.kind
       of nkEmpty..nkNilLit:
-        stackTrace(c, tos, pc, errNilAccess)
+        # for nkPtrLit, this could be supported in the future, use something like:
+        # derefPtrToReg(src.intVal + offsetof(src.typ, rc), typ_field, regs[ra], isAssign = false)
+        # where we compute the offset in bytes for field rc
+        stackTrace(c, tos, pc, errNilAccess & " " & $("kind", src.kind, "typ", typeToString(src.typ), "rc", rc))
       of nkObjConstr:
         let n = src[rc + 1].skipColon
         regs[ra].node = n
@@ -718,10 +788,23 @@ proc rawExecute(c: PCtx, start: int, tos: PStackFrame): TFullReg =
         if regs[rb].node.kind == nkRefTy:
           regs[ra].node = regs[rb].node[0]
         else:
-          ensureKind(rkNode)
-          regs[ra].node = regs[rb].node
+          let node = regs[rb].node
+          let typ = node.typ
+          # see also `nfIsPtr`
+          if node.kind == nkIntLit:
+            var typ2 = typ
+            doAssert typ != nil
+            if typ.kind == tyPtr:
+              typ2 = typ2[0]
+            if not derefPtrToReg(node.intVal, typ2, regs[ra], isAssign = false):
+              # tyObject not supported in this context
+              stackTrace(c, tos, pc, "opcLdDeref unsupported ptr type: " & $(typeToString(typ), typ.kind))
+          else:
+            ## eg: typ.kind = tyObject
+            ensureKind(rkNode)
+            regs[ra].node = regs[rb].node
       else:
-        stackTrace(c, tos, pc, errNilAccess)
+        stackTrace(c, tos, pc, errNilAccess & " kind: " & $regs[rb].kind)
     of opcWrDeref:
       # a[] = c; b unused
       let ra = instr.regA
@@ -742,8 +825,18 @@ proc rawExecute(c: PCtx, start: int, tos: PStackFrame): TFullReg =
       of rkNode:
         if regs[ra].node.kind == nkNilLit:
           stackTrace(c, tos, pc, errNilAccess)
-        regs[ra].node[] = regs[rc].regToNode[]
-        regs[ra].node.flags.incl nfIsRef
+        let node = regs[ra].node
+        let typ = node.typ
+        if nfIsPtr in node.flags or (typ != nil and typ.kind == tyPtr):
+          assert node.kind == nkIntLit, $(node.kind)
+          var typ2 = typ
+          if typ.kind == tyPtr:
+            typ2 = typ2[0]
+          if not derefPtrToReg(node.intVal, typ2, regs[rc], isAssign = true):
+            stackTrace(c, tos, pc, "opcWrDeref unsupported ptr type: " & $(typeToString(typ), typ.kind))
+        else:
+          regs[ra].node[] = regs[rc].regToNode[]
+          regs[ra].node.flags.incl nfIsRef
       else: stackTrace(c, tos, pc, errNilAccess)
     of opcAddInt:
       decodeBC(rkInt)
@@ -918,22 +1011,40 @@ proc rawExecute(c: PCtx, start: int, tos: PStackFrame): TFullReg =
       decodeBC(rkInt)
       regs[ra].intVal = ord(regs[rb].intVal <% regs[rc].intVal)
     of opcEqRef:
+      var ret = false
       decodeBC(rkInt)
+      template getTyp(n): untyped =
+        n.typ.skipTypes(abstractInst)
+      proc ptrEquality(n1: ptr PNode, n2: PNode): bool =
+        ## true if n2.intVal represents a ptr equal to n1
+        let p1 = cast[int](n1)
+        case n2.kind
+        of nkNilLit: return p1 == 0
+        of nkIntLit: # TODO: nkPtrLit
+          # for example, n1.kind == nkFloatLit (ptr float)
+          # the problem is that n1.typ == nil so we can't compare n1.typ and n2.typ
+          # this is the best we can do (pending making sure we assign a valid n1.typ to nodeAddr's)
+          let t2 = n2.getTyp
+          return t2.kind in PtrLikeKinds and n2.intVal == p1
+        else: return false
+
       if regs[rb].kind == rkNodeAddr:
         if regs[rc].kind == rkNodeAddr:
-          regs[ra].intVal = ord(regs[rb].nodeAddr == regs[rc].nodeAddr)
+          ret = regs[rb].nodeAddr == regs[rc].nodeAddr
         else:
-          assert regs[rc].kind == rkNode
-          # we know these cannot be equal
-          regs[ra].intVal = ord(false)
+          ret = ptrEquality(regs[rb].nodeAddr, regs[rc].node)
       elif regs[rc].kind == rkNodeAddr:
-        assert regs[rb].kind == rkNode
-        # we know these cannot be equal
-        regs[ra].intVal = ord(false)
+        ret = ptrEquality(regs[rc].nodeAddr, regs[rb].node)
       else:
-        regs[ra].intVal = ord((regs[rb].node.kind == nkNilLit and
-                              regs[rc].node.kind == nkNilLit) or
-                              regs[rb].node == regs[rc].node)
+        let nb = regs[rb].node
+        let nc = regs[rc].node
+        if nb.kind != nc.kind: discard
+        elif (nb == nc) or (nb.kind == nkNilLit): ret = true
+        elif nb.kind == nkIntLit and nb.intVal == nc.intVal: # TODO: nkPtrLit
+          let tb = nb.getTyp
+          let tc = nc.getTyp
+          ret = tb.kind in PtrLikeKinds and tc.kind == tb.kind
+      regs[ra].intVal = ord(ret)
     of opcEqNimNode:
       decodeBC(rkInt)
       regs[ra].intVal =
@@ -1338,6 +1449,29 @@ proc rawExecute(c: PCtx, start: int, tos: PStackFrame): TFullReg =
       let rb = instr.regBx - wordExcess - 1
       ensureKind(rkNode)
       regs[ra].node = c.globals[rb]
+    of opcLdGlobalDerefFFI:
+      let rb = instr.regBx - wordExcess - 1
+      let node = c.globals[rb]
+      let typ = node.typ
+      doAssert node.kind == nkIntLit, $(node.kind)
+      if typ.kind == tyPtr:
+        ensureKind(rkNode)
+        # use nkPtrLit once this is added
+        let node2 = newNodeIT(nkIntLit, c.debug[pc], typ)
+        node2.intVal = cast[ptr int](node.intVal)[]
+        node2.flags.incl nfIsPtr
+        regs[ra].node = node2
+      elif not derefPtrToReg(node.intVal, typ, regs[ra], isAssign = false):
+        stackTrace(c, tos, pc, "opcLdDeref unsupported type: " & $(typeToString(typ), typ[0].kind))
+    of opcLdGlobalAddrDerefFFI:
+      let rb = instr.regBx - wordExcess - 1
+      let node = c.globals[rb]
+      let typ = node.typ
+      var node2 = newNodeIT(nkIntLit, node.info, typ)
+      node2.intVal = node.intVal
+      node2.flags.incl nfIsPtr
+      ensureKind(rkNode)
+      regs[ra].node = node2
     of opcLdGlobalAddr:
       let rb = instr.regBx - wordExcess - 1
       ensureKind(rkNodeAddr)
diff --git a/compiler/vmdef.nim b/compiler/vmdef.nim
index a5df87ca8..b46d8a293 100644
--- a/compiler/vmdef.nim
+++ b/compiler/vmdef.nim
@@ -63,6 +63,8 @@ type
     opcCastIntToFloat64,    # int and float must be of the same byte size
     opcCastFloatToInt32,    # int and float must be of the same byte size
     opcCastFloatToInt64,    # int and float must be of the same byte size
+    opcCastPtrToInt,
+    opcCastIntToPtr,
     opcFastAsgnComplex,
     opcNodeToReg,
 
@@ -169,6 +171,8 @@ type
     opcAsgnConst, # dest = copy(constants[Bx])
     opcLdGlobal,  # dest = globals[Bx]
     opcLdGlobalAddr, # dest = addr(globals[Bx])
+    opcLdGlobalDerefFFI, # dest = globals[Bx][]
+    opcLdGlobalAddrDerefFFI, # globals[Bx][] = ...
 
     opcLdImmInt,  # dest = immediate value
     opcNBindSym, opcNDynBindSym,
diff --git a/compiler/vmgen.nim b/compiler/vmgen.nim
index 987fc7864..a34ef5d09 100644
--- a/compiler/vmgen.nim
+++ b/compiler/vmgen.nim
@@ -885,8 +885,21 @@ proc genCastIntFloat(c: PCtx; n: PNode; dest: var TDest) =
       c.gABC(n, opcCastFloatToInt64, dest, tmp)
       # narrowing for 64 bits not needed (no extended sign bits available).
     c.freeTemp(tmp)
+  elif src.kind in PtrLikeKinds + {tyRef} and dst.kind == tyInt:
+    let tmp = c.genx(n[1])
+    if dest < 0: dest = c.getTemp(n[0].typ)
+    var imm: BiggestInt = if src.kind in PtrLikeKinds: 1 else: 2
+    c.gABI(n, opcCastPtrToInt, dest, tmp, imm)
+    c.freeTemp(tmp)
+  elif src.kind in PtrLikeKinds + {tyInt} and dst.kind in PtrLikeKinds:
+    let tmp = c.genx(n[1])
+    if dest < 0: dest = c.getTemp(n[0].typ)
+    c.gABx(n, opcSetType, dest, c.genType(dst))
+    c.gABC(n, opcCastIntToPtr, dest, tmp)
+    c.freeTemp(tmp)
   else:
-    globalError(c.config, n.info, "VM is only allowed to 'cast' between integers and/or floats of same size")
+    # todo: support cast from tyInt to tyRef
+    globalError(c.config, n.info, "VM does not support 'cast' from " & $src.kind & " to " & $dst.kind)
 
 proc genVoidABC(c: PCtx, n: PNode, dest: TDest, opcode: TOpcode) =
   unused(c, n, dest)
@@ -1474,11 +1487,17 @@ proc getOwner(c: PCtx): PSym =
   result = c.prc.sym
   if result.isNil: result = c.module
 
+proc importcCondVar*(s: PSym): bool {.inline.} =
+  # see also importcCond
+  if sfImportc in s.flags:
+    return s.kind in {skVar, skLet, skConst}
+
 proc checkCanEval(c: PCtx; n: PNode) =
   # we need to ensure that we don't evaluate 'x' here:
   # proc foo() = var x ...
   let s = n.sym
   if {sfCompileTime, sfGlobal} <= s.flags: return
+  if s.importcCondVar: return
   if s.kind in {skVar, skTemp, skLet, skParam, skResult} and
       not s.isOwnedBy(c.prc.sym) and s.owner != c.module and c.mode != emRepl:
     # little hack ahead for bug #12612: assume gensym'ed variables
@@ -1618,17 +1637,24 @@ proc genRdVar(c: PCtx; n: PNode; dest: var TDest; flags: TGenFlags) =
   assert card(flags * {gfNodeAddr, gfNode}) < 2
   let s = n.sym
   if s.isGlobal:
-    if sfCompileTime in s.flags or c.mode == emRepl:
+    let isImportcVar = importcCondVar(s)
+    if sfCompileTime in s.flags or c.mode == emRepl or isImportcVar:
       discard
     elif s.position == 0:
       cannotEval(c, n)
     if s.position == 0:
-      if importcCond(s): c.importcSym(n.info, s)
+      if importcCond(s) or isImportcVar: c.importcSym(n.info, s)
       else: genGlobalInit(c, n, s)
     if dest < 0: dest = c.getTemp(n.typ)
     assert s.typ != nil
+
     if gfNodeAddr in flags:
-      c.gABx(n, opcLdGlobalAddr, dest, s.position)
+      if isImportcVar:
+        c.gABx(n, opcLdGlobalAddrDerefFFI, dest, s.position)
+      else:
+        c.gABx(n, opcLdGlobalAddr, dest, s.position)
+    elif isImportcVar:
+      c.gABx(n, opcLdGlobalDerefFFI, dest, s.position)
     elif fitsRegister(s.typ) and gfNode notin flags:
       var cc = c.getTemp(n.typ)
       c.gABx(n, opcLdGlobal, cc, s.position)