[mlir][arith] Fix overflow bug in arith::CeilDivSIOp::fold

The folder for arith::CeilDivSIOp should only be applied when it can be
guaranteed that no overflow would happen. The current implementation
works fine when both dividends are positive and the only arithmetic
operation is the division itself.

However, in cases where at least one of the dividends is negative, the
division is split into multiple operations, e.g.: `- ( -a / b)`. That's
additional 2 operations on top of the actual division that can overflow
- the folder should check all 3 ops for overflow. The current logic
doesn't do that - it effectively only the last operation (i.e. the
division). It breaks when using e.g. MININT values (e.g. -128 for
8-bit integers) - negating such values overflows.

This PR makes sure that no folding happens if any of the intermediate
arithmetic operations overflows.

Author: banach-space

mlir/lib/Dialect/Arith/IR/ArithOps.cpp

@@ -701,22 +701,34 @@ OpFoldResult arith::CeilDivSIOp::fold(FoldAdaptor adaptor) {
           // Both positive, return ceil(a, b).
           return signedCeilNonnegInputs(a, b, overflowOrDiv0);
         }
+
+        bool overflowNegA = false;
+        bool overflowNegB = false;
+        bool overflowNegDiv = false;
+        bool overflowDiv = false;
         if (!aGtZero && !bGtZero) {
           // Both negative, return ceil(-a, -b).
-          APInt posA = zero.ssub_ov(a, overflowOrDiv0);
-          APInt posB = zero.ssub_ov(b, overflowOrDiv0);
-          return signedCeilNonnegInputs(posA, posB, overflowOrDiv0);
+          APInt posA = zero.ssub_ov(a, overflowNegA);
+          APInt posB = zero.ssub_ov(b, overflowNegB);
+          APInt res = signedCeilNonnegInputs(posA, posB, overflowDiv);
+          overflowOrDiv0 = (overflowNegA || overflowNegB || overflowDiv);
+          return res;
         }
         if (!aGtZero && bGtZero) {
           // A is negative, b is positive, return - ( -a / b).
-          APInt posA = zero.ssub_ov(a, overflowOrDiv0);
-          APInt div = posA.sdiv_ov(b, overflowOrDiv0);
-          return zero.ssub_ov(div, overflowOrDiv0);
+          APInt posA = zero.ssub_ov(a, overflowNegA);
+          APInt div = posA.sdiv_ov(b, overflowDiv);
+          APInt res = zero.ssub_ov(div, overflowNegDiv);
+          overflowOrDiv0 = (overflowNegA || overflowDiv || overflowNegDiv);
+          return res;
         }
         // A is positive, b is negative, return - (a / -b).
-        APInt posB = zero.ssub_ov(b, overflowOrDiv0);
-        APInt div = a.sdiv_ov(posB, overflowOrDiv0);
-        return zero.ssub_ov(div, overflowOrDiv0);
+        APInt posB = zero.ssub_ov(b, overflowNegB);
+        APInt div = a.sdiv_ov(posB, overflowDiv);
+        APInt res = zero.ssub_ov(div, overflowNegDiv);
+
+        overflowOrDiv0 = (overflowNegB || overflowDiv || overflowNegDiv);
+        return res;
       });
 
   return overflowOrDiv0 ? Attribute() : result;

mlir/test/Transforms/constant-fold.mlir

@@ -478,6 +478,32 @@ func.func @simple_arith.ceildivsi() -> (i32, i32, i32, i32, i32) {
 
 // -----
 
+// CHECK-LABEL: func @simple_arith.ceildivsi_overflow
+func.func @simple_arith.ceildivsi_overflow() -> (i8, i16, i32) {
+  // The negative values below are MININTs for the corresponding bit-width. The
+  // folder will try to negate them (so that the division operartes on two
+  // positive numbers), but that would cause overflow (negating MININT
+  // overflows). Hence folding should not happen and the original ceildivsi is
+  // preserved.
+
+  // CHECK-COUNT-3:  arith.ceildivsi
+  %0 = arith.constant 7 : i8
+  %min_int_i8 = arith.constant -128 : i8
+  %2 = arith.ceildivsi %min_int_i8, %0 : i8
+
+  %3 = arith.constant 7 : i16
+  %min_int_i16 = arith.constant -32768 : i16
+  %5 = arith.ceildivsi %min_int_i16, %3 : i16
+
+  %6 = arith.constant 7 : i32
+  %min_int_i32 = arith.constant -2147483648 : i32
+  %8 = arith.ceildivsi %min_int_i32, %6 : i32
+
+  return %2, %5, %8 : i8, i16, i32
+}
+
+// -----
+
 // CHECK-LABEL: func @simple_arith.ceildivui
 func.func @simple_arith.ceildivui() -> (i32, i32, i32, i32, i32) {
   // CHECK-DAG: [[C0:%.+]] = arith.constant 0