catchable defects (#13626)

* allow defects to be caught even for --exceptions:goto (WIP)
* implemented the new --panics:on|off switch; refs https://github.com/nim-lang/RFCs/issues/180
* new implementation for integer overflow checking
* produce a warning if a user-defined exception type inherits from Exception directly
* applied Timothee's suggestions; improved the documentation and replace the term 'checked runtime check' by 'panic'
* fixes #13627
* don't inherit from Exception directly

lib/nimbase.h

@@ -545,4 +545,26 @@ typedef int Nim_and_C_compiler_disagree_on_target_architecture[sizeof(NI) == siz
 #define NIM_CHECK_SIZE(typ, sz) \
   _Static_assert(sizeof(typ) == sz, "Nim & C disagree on type size")
 
+/* these exist to make the codegen logic simpler */
+#define nimModInt(a, b, res) (((*res) = (a) % (b)), 0)
+#define nimModInt64(a, b, res) (((*res) = (a) % (b)), 0)
+
+/* these exist because we cannot have .compilerProcs that are importc'ed
+   by a different name */
+
+#define nimAddInt64(a, b, res) __builtin_saddll_overflow(a, b, (long long int*)res)
+#define nimSubInt64(a, b, res) __builtin_ssubll_overflow(a, b, (long long int*)res)
+#define nimMulInt64(a, b, res) __builtin_smulll_overflow(a, b, (long long int*)res)
+
+#if NIM_INTBITS == 32
+  #define nimAddInt(a, b, res) __builtin_sadd_overflow(a, b, res)
+  #define nimSubInt(a, b, res) __builtin_ssub_overflow(a, b, res)
+  #define nimMulInt(a, b, res) __builtin_smul_overflow(a, b, res)
+#else
+  /* map it to the 'long long' variant */
+  #define nimAddInt(a, b, res) __builtin_saddll_overflow(a, b, (long long int*)res)
+  #define nimSubInt(a, b, res) __builtin_ssubll_overflow(a, b, (long long int*)res)
+  #define nimMulInt(a, b, res) __builtin_smulll_overflow(a, b, (long long int*)res)
+#endif
+
 #endif /* NIMBASE_H */

lib/system.nim

@@ -2221,7 +2221,10 @@ when notJSnotNims:
 
   # we cannot compile this with stack tracing on
   # as it would recurse endlessly!
-  include "system/arithm"
+  when defined(nimNewIntegerOps):
+    include "system/integerops"
+  else:
+    include "system/arithm"
   {.pop.}
 
 

lib/system/arithm.nim

@@ -409,13 +409,13 @@ when not declared(mulInt):
 # We avoid setting the FPU control word here for compatibility with libraries
 # written in other languages.
 
-proc raiseFloatInvalidOp {.noinline.} =
+proc raiseFloatInvalidOp {.compilerproc, noinline.} =
   sysFatal(FloatInvalidOpError, "FPU operation caused a NaN result")
 
 proc nanCheck(x: float64) {.compilerproc, inline.} =
   if x != x: raiseFloatInvalidOp()
 
-proc raiseFloatOverflow(x: float64) {.noinline.} =
+proc raiseFloatOverflow(x: float64) {.compilerproc, noinline.} =
   if x > 0.0:
     sysFatal(FloatOverflowError, "FPU operation caused an overflow")
   else:

lib/system/chcks.nim

@@ -28,6 +28,19 @@ proc raiseIndexError() {.compilerproc, noinline.} =
 proc raiseFieldError(f: string) {.compilerproc, noinline.} =
   sysFatal(FieldError, f)
 
+proc raiseRangeErrorI(i, a, b: BiggestInt) {.compilerproc, noinline.} =
+  sysFatal(RangeError, "value out of range: " & $i & " notin " & $a & " .. " & $b)
+
+proc raiseRangeErrorF(i, a, b: float) {.compilerproc, noinline.} =
+  sysFatal(RangeError, "value out of range: " & $i & " notin " & $a & " .. " & $b)
+
+proc raiseRangeErrorU(i, a, b: uint64) {.compilerproc, noinline.} =
+  # todo: better error reporting
+  sysFatal(RangeError, "value out of range")
+
+proc raiseObjectConversionError() {.compilerproc, noinline.} =
+  sysFatal(ObjectConversionError, "invalid object conversion")
+
 proc chckIndx(i, a, b: int): int =
   if i >= a and i <= b:
     return i
@@ -116,6 +129,5 @@ when not defined(nimV2):
     return true
 
 when defined(nimV2):
-  proc nimFieldDiscriminantCheckV2(oldDiscVal, newDiscVal: uint8) {.compilerproc.} =
-    if oldDiscVal != newDiscVal:
-      sysFatal(FieldError, "assignment to discriminant changes object branch")
+  proc raiseObjectCaseTransition() {.compilerproc.} =
+    sysFatal(FieldError, "assignment to discriminant changes object branch")

lib/system/fatal.nim

@@ -24,7 +24,7 @@ when hostOS == "standalone":
     rawoutput(message)
     panic(arg)
 
-elif (defined(nimQuirky) or gotoBasedExceptions) and not defined(nimscript):
+elif (defined(nimQuirky) or defined(nimPanics)) and not defined(nimscript):
   import ansi_c
 
   proc name(t: typedesc): string {.magic: "TypeTrait".}
@@ -46,20 +46,9 @@ elif (defined(nimQuirky) or gotoBasedExceptions) and not defined(nimscript):
 
 else:
   proc sysFatal(exceptn: typedesc, message: string) {.inline, noreturn.} =
-    when declared(owned):
-      var e: owned(ref exceptn)
-    else:
-      var e: ref exceptn
-    new(e)
-    e.msg = message
-    raise e
+    raise (ref exceptn)(msg: message)
 
   proc sysFatal(exceptn: typedesc, message, arg: string) {.inline, noreturn.} =
-    when declared(owned):
-      var e: owned(ref exceptn)
-    else:
-      var e: ref exceptn
-    new(e)
-    e.msg = message & arg
-    raise e
+    raise (ref exceptn)(msg: message & arg)
+
 {.pop.}

lib/system/gc.nim

@@ -441,13 +441,15 @@ proc newObj(typ: PNimType, size: int): pointer {.compilerRtl.} =
   zeroMem(result, size)
   when defined(memProfiler): nimProfile(size)
 
+{.push overflowChecks: on.}
 proc newSeq(typ: PNimType, len: int): pointer {.compilerRtl.} =
   # `newObj` already uses locks, so no need for them here.
-  let size = addInt(mulInt(len, typ.base.size), GenericSeqSize)
+  let size = len * typ.base.size + GenericSeqSize
   result = newObj(typ, size)
   cast[PGenericSeq](result).len = len
   cast[PGenericSeq](result).reserved = len
   when defined(memProfiler): nimProfile(size)
+{.pop.}
 
 proc newObjRC1(typ: PNimType, size: int): pointer {.compilerRtl.} =
   # generates a new object and sets its reference counter to 1
@@ -476,12 +478,14 @@ proc newObjRC1(typ: PNimType, size: int): pointer {.compilerRtl.} =
   sysAssert(allocInv(gch.region), "newObjRC1 end")
   when defined(memProfiler): nimProfile(size)
 
+{.push overflowChecks: on.}
 proc newSeqRC1(typ: PNimType, len: int): pointer {.compilerRtl.} =
-  let size = addInt(mulInt(len, typ.base.size), GenericSeqSize)
+  let size = len * typ.base.size + GenericSeqSize
   result = newObjRC1(typ, size)
   cast[PGenericSeq](result).len = len
   cast[PGenericSeq](result).reserved = len
   when defined(memProfiler): nimProfile(size)
+{.pop.}
 
 proc growObj(old: pointer, newsize: int, gch: var GcHeap): pointer =
   collectCT(gch)

lib/system/gc_common.nim

@@ -79,7 +79,7 @@ template decTypeSize(cell, t) =
     if t.kind in {tyString, tySequence}:
       let cap = cast[PGenericSeq](cellToUsr(cell)).space
       let size = if t.kind == tyString: cap+1+GenericSeqSize
-                 else: addInt(mulInt(cap, t.base.size), GenericSeqSize)
+                 else: cap * t.base.size + GenericSeqSize
       atomicDec t.sizes, size+sizeof(Cell)
     else:
       atomicDec t.sizes, t.base.size+sizeof(Cell)

lib/system/gc_ms.nim

@@ -305,20 +305,22 @@ proc newObjRC1(typ: PNimType, size: int): pointer {.compilerRtl.} =
   when defined(memProfiler): nimProfile(size)
 
 when not defined(nimSeqsV2):
+  {.push overflowChecks: on.}
   proc newSeq(typ: PNimType, len: int): pointer {.compilerRtl.} =
     # `newObj` already uses locks, so no need for them here.
-    let size = addInt(mulInt(len, typ.base.size), GenericSeqSize)
+    let size = len * typ.base.size + GenericSeqSize
     result = newObj(typ, size)
     cast[PGenericSeq](result).len = len
     cast[PGenericSeq](result).reserved = len
     when defined(memProfiler): nimProfile(size)
 
   proc newSeqRC1(typ: PNimType, len: int): pointer {.compilerRtl.} =
-    let size = addInt(mulInt(len, typ.base.size), GenericSeqSize)
+    let size = len * typ.base.size + GenericSeqSize
     result = newObj(typ, size)
     cast[PGenericSeq](result).len = len
     cast[PGenericSeq](result).reserved = len
     when defined(memProfiler): nimProfile(size)
+  {.pop.}
 
   proc growObj(old: pointer, newsize: int, gch: var GcHeap): pointer =
     collectCT(gch, newsize + sizeof(Cell))

lib/system/gc_regions.nim

@@ -334,20 +334,21 @@ proc newObjNoInit(typ: PNimType, size: int): pointer {.compilerRtl.} =
   result = rawNewObj(tlRegion, typ, size)
   when defined(memProfiler): nimProfile(size)
 
+{.push overflowChecks: on.}
 proc newSeq(typ: PNimType, len: int): pointer {.compilerRtl.} =
-  let size = roundup(addInt(mulInt(len, typ.base.size), GenericSeqSize),
-                     MemAlign)
+  let size = roundup(len * typ.base.size + GenericSeqSize, MemAlign)
   result = rawNewSeq(tlRegion, typ, size)
   zeroMem(result, size)
   cast[PGenericSeq](result).len = len
   cast[PGenericSeq](result).reserved = len
 
 proc newStr(typ: PNimType, len: int; init: bool): pointer {.compilerRtl.} =
-  let size = roundup(addInt(len, GenericSeqSize), MemAlign)
+  let size = roundup(len + GenericSeqSize, MemAlign)
   result = rawNewSeq(tlRegion, typ, size)
   if init: zeroMem(result, size)
   cast[PGenericSeq](result).len = 0
   cast[PGenericSeq](result).reserved = len
+{.pop.}
 
 proc newObjRC1(typ: PNimType, size: int): pointer {.compilerRtl.} =
   result = rawNewObj(tlRegion, typ, size)

lib/system/integerops.nim

@@ -0,0 +1,132 @@
+#
+#
+#            Nim's Runtime Library
+#        (c) Copyright 2020 Andreas Rumpf
+#
+#    See the file "copying.txt", included in this
+#    distribution, for details about the copyright.
+#
+
+# Integer arithmetic with overflow checking. Uses
+# intrinsics or inline assembler.
+
+proc raiseOverflow {.compilerproc, noinline.} =
+  # a single proc to reduce code size to a minimum
+  sysFatal(OverflowError, "over- or underflow")
+
+proc raiseDivByZero {.compilerproc, noinline.} =
+  sysFatal(DivByZeroError, "division by zero")
+
+{.pragma: nimbaseH, importc, nodecl, noSideEffect, compilerproc.}
+
+when defined(gcc) or defined(clang):
+  # take the #define from nimbase.h
+
+  proc nimAddInt(a, b: int, res: ptr int): bool {.nimbaseH.}
+  proc nimSubInt(a, b: int, res: ptr int): bool {.nimbaseH.}
+  proc nimMulInt(a, b: int, res: ptr int): bool {.nimbaseH.}
+
+  proc nimAddInt64(a, b: int64; res: ptr int64): bool {.nimbaseH.}
+  proc nimSubInt64(a, b: int64; res: ptr int64): bool {.nimbaseH.}
+  proc nimMulInt64(a, b: int64; res: ptr int64): bool {.nimbaseH.}
+
+# unary minus and 'abs' not required here anymore and are directly handled
+# in the code generator.
+# 'nimModInt' does exist in nimbase.h without check as we moved the
+# check for 0 to the codgen.
+proc nimModInt(a, b: int; res: ptr int): bool {.nimbaseH.}
+
+proc nimModInt64(a, b: int64; res: ptr int64): bool {.nimbaseH.}
+
+# Platform independent versions.
+
+template addImplFallback(name, T, U) {.dirty.} =
+  when not declared(name):
+    proc name(a, b: T; res: ptr T): bool {.compilerproc, inline.} =
+      let r = cast[T](cast[U](a) + cast[U](b))
+      if (r xor a) >= T(0) or (r xor b) >= T(0):
+        res[] = r
+      else:
+        result = true
+
+addImplFallback(nimAddInt, int, uint)
+addImplFallback(nimAddInt64, int64, uint64)
+
+template subImplFallback(name, T, U) {.dirty.} =
+  when not declared(name):
+    proc name(a, b: T; res: ptr T): bool {.compilerproc, inline.} =
+      let r = cast[T](cast[U](a) - cast[U](b))
+      if (r xor a) >= 0 or (r xor not b) >= 0:
+        res[] = r
+      else:
+        result = true
+
+subImplFallback(nimSubInt, int, uint)
+subImplFallback(nimSubInt64, int64, uint64)
+
+template mulImplFallback(name, T, U, conv) {.dirty.} =
+  #
+  # This code has been inspired by Python's source code.
+  # The native int product x*y is either exactly right or *way* off, being
+  # just the last n bits of the true product, where n is the number of bits
+  # in an int (the delivered product is the true product plus i*2**n for
+  # some integer i).
+  #
+  # The native float64 product x*y is subject to three
+  # rounding errors: on a sizeof(int)==8 box, each cast to double can lose
+  # info, and even on a sizeof(int)==4 box, the multiplication can lose info.
+  # But, unlike the native int product, it's not in *range* trouble:  even
+  # if sizeof(int)==32 (256-bit ints), the product easily fits in the
+  # dynamic range of a float64. So the leading 50 (or so) bits of the float64
+  # product are correct.
+  #
+  # We check these two ways against each other, and declare victory if
+  # they're approximately the same. Else, because the native int product is
+  # the only one that can lose catastrophic amounts of information, it's the
+  # native int product that must have overflowed.
+  #
+  when not declared(name):
+    proc name(a, b: T; res: ptr T): bool {.compilerproc, inline.} =
+      let r = cast[T](cast[U](a) * cast[U](b))
+      let floatProd = conv(a) * conv(b)
+      let resAsFloat = conv(r)
+      # Fast path for normal case: small multiplicands, and no info
+      # is lost in either method.
+      if resAsFloat == floatProd:
+        res[] = r
+      else:
+        # Somebody somewhere lost info. Close enough, or way off? Note
+        # that a != 0 and b != 0 (else resAsFloat == floatProd == 0).
+        # The difference either is or isn't significant compared to the
+        # true value (of which floatProd is a good approximation).
+
+        # abs(diff)/abs(prod) <= 1/32 iff
+        #   32 * abs(diff) <= abs(prod) -- 5 good bits is "close enough"
+        if 32.0 * abs(resAsFloat - floatProd) <= abs(floatProd):
+          res[] = r
+        else:
+          result = true
+
+mulImplFallback(nimMulInt, int, uint, toFloat)
+mulImplFallback(nimMulInt64, int64, uint64, toBiggestFloat)
+
+
+template divImplFallback(name, T) {.dirty.} =
+  proc name(a, b: T; res: ptr T): bool {.compilerproc, inline.} =
+    # we moved the b == 0 case out into the codegen.
+    if a == low(T) and b == T(-1):
+      result = true
+    else:
+      res[] = a div b
+
+divImplFallback(nimDivInt, int)
+divImplFallback(nimDivInt64, int64)
+
+proc raiseFloatInvalidOp {.compilerproc, noinline.} =
+  sysFatal(FloatInvalidOpError, "FPU operation caused a NaN result")
+
+proc raiseFloatOverflow(x: float64) {.compilerproc, noinline.} =
+  if x > 0.0:
+    sysFatal(FloatOverflowError, "FPU operation caused an overflow")
+  else:
+    sysFatal(FloatUnderflowError, "FPU operations caused an underflow")

lib/system/mm/boehm.nim

@@ -101,10 +101,12 @@ proc newObj(typ: PNimType, size: int): pointer {.compilerproc.} =
   else: result = alloc(size)
   if typ.finalizer != nil:
     boehmRegisterFinalizer(result, boehmgc_finalizer, typ.finalizer, nil, nil)
+{.push overflowChecks: on.}
 proc newSeq(typ: PNimType, len: int): pointer {.compilerproc.} =
-  result = newObj(typ, addInt(mulInt(len, typ.base.size), GenericSeqSize))
+  result = newObj(typ, len * typ.base.size + GenericSeqSize)
   cast[PGenericSeq](result).len = len
   cast[PGenericSeq](result).reserved = len
+{.pop.}
 
 proc growObj(old: pointer, newsize: int): pointer =
   result = realloc(old, newsize)

lib/system/mm/none.nim

@@ -19,10 +19,12 @@ proc newObj(typ: PNimType, size: int): pointer {.compilerproc.} =
 proc newObjNoInit(typ: PNimType, size: int): pointer =
   result = alloc(size)
 
+{.push overflowChecks: on.}
 proc newSeq(typ: PNimType, len: int): pointer {.compilerproc.} =
-  result = newObj(typ, addInt(mulInt(len, typ.base.size), GenericSeqSize))
+  result = newObj(typ, len * typ.base.size + GenericSeqSize)
   cast[PGenericSeq](result).len = len
   cast[PGenericSeq](result).reserved = len
+{.pop.}
 
 proc growObj(old: pointer, newsize: int): pointer =
   result = realloc(old, newsize)

lib/system/mmdisp.nim

@@ -102,18 +102,20 @@ else:
     include "system/gc"
 
 when not declared(nimNewSeqOfCap) and not defined(nimSeqsV2):
+  {.push overflowChecks: on.}
   proc nimNewSeqOfCap(typ: PNimType, cap: int): pointer {.compilerproc.} =
     when defined(gcRegions):
-      let s = mulInt(cap, typ.base.size)  # newStr already adds GenericSeqSize
+      let s = cap * typ.base.size  # newStr already adds GenericSeqSize
       result = newStr(typ, s, ntfNoRefs notin typ.base.flags)
     else:
-      let s = addInt(mulInt(cap, typ.base.size), GenericSeqSize)
+      let s = cap * typ.base.size + GenericSeqSize
       when declared(newObjNoInit):
         result = if ntfNoRefs in typ.base.flags: newObjNoInit(typ, s) else: newObj(typ, s)
       else:
         result = newObj(typ, s)
       cast[PGenericSeq](result).len = 0
       cast[PGenericSeq](result).reserved = cap
+  {.pop.}
 
 {.pop.}