VM: refactoring [backport] (#25280)

Note to @narimiran backport because IC requires it.

compiler/vm.nim

@@ -480,9 +480,9 @@ proc opConv(c: PCtx; dest: var TFullReg, src: TFullReg, desttyp, srctyp: PType):
     else:
       asgnComplex(dest, src)
 
-proc compile(c: PCtx, s: PSym): int =
+proc compile(c: PCtx, s: PSym): VmProcInfo =
   result = vmgen.genProc(c, s)
-  when debugEchoCode: c.echoCode result
+  when debugEchoCode: c.echoCode result.pc
   #c.echoCode
 
 template handleJmpBack() {.dirty.} =
@@ -1435,13 +1435,13 @@ proc rawExecute(c: PCtx, start: int, tos: PStackFrame): TFullReg =
         else:
           globalError(c.config, c.debug[pc], "VM not built with FFI support")
       elif prc.kind != skTemplate:
-        let newPc = compile(c, prc)
+        let procInfo = compile(c, prc)
         # tricky: a recursion is also a jump back, so we use the same
         # logic as for loops:
-        if newPc < pc: handleJmpBack()
+        if procInfo.pc < pc: handleJmpBack()
         #echo "new pc ", newPc, " calling: ", prc.name.s
         var newFrame = PStackFrame(prc: prc, comesFrom: pc, next: tos)
-        newSeq(newFrame.slots, prc.offset+ord(isClosure))
+        newSeq(newFrame.slots, procInfo.usedRegisters+ord(isClosure))
         # setup slot for proc result:
         let ret {.cursor.} = prc.typ.returnType
         # hot spot ahead!
@@ -1467,7 +1467,7 @@ proc rawExecute(c: PCtx, start: int, tos: PStackFrame): TFullReg =
         tos = newFrame
         updateRegsAlias
         # -1 for the following 'inc pc'
-        pc = newPc-1
+        pc = procInfo.pc-1
       else:
         # for 'getAst' support we need to support template expansion here:
         let genSymOwner = if tos.next != nil and tos.next.prc != nil:
@@ -2351,8 +2351,7 @@ proc execProc*(c: PCtx; sym: PSym; args: openArray[PNode]): PNode =
       let start = genProc(c, sym)
 
       var tos = PStackFrame(prc: sym, comesFrom: 0, next: nil)
-      let maxSlots = sym.offset
-      newSeq(tos.slots, maxSlots)
+      newSeq(tos.slots, start.usedRegisters)
 
       # setup parameters:
       if not isEmptyType(sym.typ.returnType) or sym.kind == skMacro:
@@ -2361,7 +2360,7 @@ proc execProc*(c: PCtx; sym: PSym; args: openArray[PNode]): PNode =
       for i in 0..<sym.typ.paramsLen:
         putIntoReg(tos.slots[i+1], args[i])
 
-      result = rawExecute(c, start, tos).regToNode
+      result = rawExecute(c, start.pc, tos).regToNode
   else:
     result = nil
     localError(c.config, sym.info,
@@ -2550,8 +2549,7 @@ proc evalMacroCall*(module: PSym; idgen: IdGenerator; g: ModuleGraph; templInstC
     return errorNode(idgen, module, n)
 
   var tos = PStackFrame(prc: sym, comesFrom: 0, next: nil)
-  let maxSlots = sym.offset
-  newSeq(tos.slots, maxSlots)
+  newSeq(tos.slots, start.usedRegisters)
   # setup arguments:
   var L = n.safeLen
   if L == 0: L = 1
@@ -2578,7 +2576,7 @@ proc evalMacroCall*(module: PSym; idgen: IdGenerator; g: ModuleGraph; templInstC
                  " generic parameter(s)")
   # temporary storage:
   #for i in L..<maxSlots: tos.slots[i] = newNode(nkEmpty)
-  result = rawExecute(c, start, tos).regToNode
+  result = rawExecute(c, start.pc, tos).regToNode
   if result.info.line < 0: result.info = n.info
   if cyclicTree(result): globalError(c.config, n.info, "macro produced a cyclic tree")
   dec(g.config.evalMacroCounter)

compiler/vmdef.nim

@@ -243,6 +243,11 @@ type
   VmCallback* = proc (args: VmArgs) {.closure.}
 
   PCtx* = ref TCtx
+
+  VmProcInfo* = object
+    pc*: int32
+    usedRegisters*: int32
+
   TCtx* = object of TPassContext # code gen context
     code*: seq[TInstr]
     debug*: seq[TLineInfo]  # line info for every instruction; kept separate
@@ -271,7 +276,7 @@ type
     profiler*: Profiler
     templInstCounter*: ref int # gives every template instantiation a unique ID, needed here for getAst
     vmstateDiff*: seq[(PSym, PNode)] # we remember the "diff" to global state here (feature for IC)
-    procToCodePos*: Table[int, int]
+    procToCodePos*: Table[int, VmProcInfo]
     cannotEval*: bool
     locals*: IntSet
 
@@ -293,6 +298,9 @@ type
 
   PEvalContext* = PCtx
 
+const
+  NoVmProcInfo* = VmProcInfo(pc: 0'i32, usedRegisters: -1'i32)
+
 proc newCtx*(module: PSym; cache: IdentCache; g: ModuleGraph; idgen: IdGenerator): PCtx =
   PCtx(code: @[], debug: @[],
     globals: newNode(nkStmtListExpr), constants: newNode(nkStmtList), types: @[],

compiler/vmgen.nim

@@ -2128,7 +2128,7 @@ proc genTupleConstr(c: PCtx, n: PNode, dest: var TDest) =
         c.preventFalseAlias(it, opcWrObj, dest, i.TRegister, tmp)
         c.freeTemp(tmp)
 
-proc genProc*(c: PCtx; s: PSym): int
+proc genProc*(c: PCtx; s: PSym): VmProcInfo
 
 proc toKey(s: PSym): string =
   result = ""
@@ -2412,27 +2412,19 @@ proc optimizeJumps(c: PCtx; start: int) =
         c.finalJumpTarget(i, d - i)
     else: discard
 
-proc genProc(c: PCtx; s: PSym): int =
-  let
-    pos = c.procToCodePos.getOrDefault(s.id)
-    wasNotGenProcBefore = pos == 0
-    noRegistersAllocated = s.offset == -1
-  if wasNotGenProcBefore or noRegistersAllocated:
-    # xxx: the noRegisterAllocated check is required in order to avoid issues
-    #      where nimsuggest can crash due as a macro with pos will be loaded
-    #      but it doesn't have offsets for register allocations see:
-    #      https://github.com/nim-lang/Nim/issues/18385
-    #      Improvements and further use of IC should remove the need for this.
+proc genProc(c: PCtx; s: PSym): VmProcInfo =
+  result = c.procToCodePos.getOrDefault(s.id, NoVmProcInfo)
+  if result.usedRegisters < 0:
     #if s.name.s == "outterMacro" or s.name.s == "innerProc":
     #  echo "GENERATING CODE FOR ", s.name.s
     let last = c.code.len-1
-    var eofInstr: TInstr = default(TInstr)
+    var eofInstr = default(TInstr)
     if last >= 0 and c.code[last].opcode == opcEof:
       eofInstr = c.code[last]
       c.code.setLen(last)
       c.debug.setLen(last)
     #c.removeLastEof
-    result = c.code.len+1 # skip the jump instruction
+    result.pc = (c.code.len+1).int32 # skip the jump instruction
     c.procToCodePos[s.id] = result
     # thanks to the jmp we can add top level statements easily and also nest
     # procs easily:
@@ -2457,12 +2449,12 @@ proc genProc(c: PCtx; s: PSym): int =
     c.gABC(body, opcRet)
     c.patch(procStart)
     c.gABC(body, opcEof, eofInstr.regA)
-    c.optimizeJumps(result)
-    s.offset = c.prc.regInfo.len.int32
+    c.optimizeJumps(result.pc)
+    result.usedRegisters = c.prc.regInfo.len.int32
+    c.procToCodePos[s.id] = result
     #if s.name.s == "main" or s.name.s == "[]":
     #  echo renderTree(body)
     #  c.echoCode(result)
     c.prc = oldPrc
   else:
-    c.prc.regInfo.setLen s.offset
-    result = pos
+    c.prc.regInfo.setLen result.usedRegisters