Fix NULLIF()'s handling of read-write expanded objects.
authorTom Lane <[email protected]>
Mon, 25 Nov 2024 23:08:58 +0000 (18:08 -0500)
committerTom Lane <[email protected]>
Mon, 25 Nov 2024 23:09:10 +0000 (18:09 -0500)
If passed a read-write expanded object pointer, the EEOP_NULLIF
code would hand that same pointer to the equality function
and then (unless equality was reported) also return the same
pointer as its value.  This is no good, because a function that
receives a read-write expanded object pointer is fully entitled
to scribble on or even delete the object, thus corrupting the
NULLIF output.  (This problem is likely unobservable with the
equality functions provided in core Postgres, but it's easy to
demonstrate with one coded in plpgsql.)

To fix, make sure the pointer passed to the equality function
is read-only.  We can still return the original read-write
pointer as the NULLIF result, allowing optimization of later
operations.

Per bug #18722 from Alexander Lakhin.  This has been wrong
since we invented expanded objects, so back- to all
supported branches.

Discussion: https://postgr.es/m/18722-fd9e645448cc78b4@postgresql.org

src/backend/executor/execExpr.c
src/backend/executor/execExprInterp.c
src/backend/jit/llvm/llvmjit_expr.c
src/include/executor/execExpr.h
src/test/regress/expected/case.out
src/test/regress/sql/case.sql

index ab89ad9afbf85132861e850fc2ae1be9c003c337..1e3b93a69d87102accca4bcbf03fd775f20dc42b 100644 (file)
@@ -1188,6 +1188,14 @@ ExecInitExprRec(Expr *node, ExprState *state,
                             op->args, op->opfuncid, op->inputcollid,
                             state);
 
+               /*
+                * If first argument is of varlena type, we'll need to ensure
+                * that the value passed to the comparison function is a
+                * read-only pointer.
+                */
+               scratch.d.func.make_ro =
+                   (get_typlen(exprType((Node *) linitial(op->args))) == -1);
+
                /*
                 * Change opcode of call instruction to EEOP_NULLIF.
                 *
index aa68c115ba9a6478f0260ba8c8d6283aa15f1145..a55ad7725662911fe33ecce6ddad3ff6357a951d 100644 (file)
@@ -1295,12 +1295,24 @@ ExecInterpExpr(ExprState *state, ExprContext *econtext, bool *isnull)
             * The arguments are already evaluated into fcinfo->args.
             */
            FunctionCallInfo fcinfo = op->d.func.fcinfo_data;
+           Datum       save_arg0 = fcinfo->args[0].value;
 
            /* if either argument is NULL they can't be equal */
            if (!fcinfo->args[0].isnull && !fcinfo->args[1].isnull)
            {
                Datum       result;
 
+               /*
+                * If first argument is of varlena type, it might be an
+                * expanded datum.  We need to ensure that the value passed to
+                * the comparison function is a read-only pointer.  However,
+                * if we end by returning the first argument, that will be the
+                * original read-write pointer if it was read-write.
+                */
+               if (op->d.func.make_ro)
+                   fcinfo->args[0].value =
+                       MakeExpandedObjectReadOnlyInternal(save_arg0);
+
                fcinfo->isnull = false;
                result = op->d.func.fn_addr(fcinfo);
 
@@ -1315,7 +1327,7 @@ ExecInterpExpr(ExprState *state, ExprContext *econtext, bool *isnull)
            }
 
            /* Arguments aren't equal, so return the first one */
-           *op->resvalue = fcinfo->args[0].value;
+           *op->resvalue = save_arg0;
            *op->resnull = fcinfo->args[0].isnull;
 
            EEO_NEXT();
index 306aea82d3b7b3942174196dc427221101c0efa9..b9fdd444a252284f49bf6a1bdab862d630876939 100644 (file)
@@ -1562,6 +1562,9 @@ llvm_compile_expr(ExprState *state)
 
                    v_fcinfo = l_ptr_const(fcinfo, l_ptr(StructFunctionCallInfoData));
 
+                   /* save original arg[0] */
+                   v_arg0 = l_funcvalue(b, v_fcinfo, 0);
+
                    /* if either argument is NULL they can't be equal */
                    v_argnull0 = l_funcnull(b, v_fcinfo, 0);
                    v_argnull1 = l_funcnull(b, v_fcinfo, 1);
@@ -1578,7 +1581,6 @@ llvm_compile_expr(ExprState *state)
 
                    /* one (or both) of the arguments are null, return arg[0] */
                    LLVMPositionBuilderAtEnd(b, b_hasnull);
-                   v_arg0 = l_funcvalue(b, v_fcinfo, 0);
                    LLVMBuildStore(b, v_argnull0, v_resnullp);
                    LLVMBuildStore(b, v_arg0, v_resvaluep);
                    LLVMBuildBr(b, opblocks[opno + 1]);
@@ -1586,12 +1588,35 @@ llvm_compile_expr(ExprState *state)
                    /* build block to invoke function and check result */
                    LLVMPositionBuilderAtEnd(b, b_nonull);
 
+                   /*
+                    * If first argument is of varlena type, it might be an
+                    * expanded datum.  We need to ensure that the value
+                    * passed to the comparison function is a read-only
+                    * pointer.  However, if we end by returning the first
+                    * argument, that will be the original read-write pointer
+                    * if it was read-write.
+                    */
+                   if (op->d.func.make_ro)
+                   {
+                       LLVMValueRef v_params[1];
+                       LLVMValueRef v_arg0_ro;
+
+                       v_params[0] = v_arg0;
+                       v_arg0_ro =
+                           l_call(b,
+                                  llvm_pg_var_func_type("MakeExpandedObjectReadOnlyInternal"),
+                                  llvm_pg_func(mod, "MakeExpandedObjectReadOnlyInternal"),
+                                  v_params, lengthof(v_params), "");
+                       LLVMBuildStore(b, v_arg0_ro,
+                                      l_funcvaluep(b, v_fcinfo, 0));
+                   }
+
                    v_retval = BuildV1Call(context, b, mod, fcinfo, &v_fcinfo_isnull);
 
                    /*
-                    * If result not null, and arguments are equal return null
-                    * (same result as if there'd been NULLs, hence reuse
-                    * b_hasnull).
+                    * If result not null and arguments are equal return null,
+                    * else return arg[0] (same result as if there'd been
+                    * NULLs, hence reuse b_hasnull).
                     */
                    v_argsequal = LLVMBuildAnd(b,
                                               LLVMBuildICmp(b, LLVMIntEQ,
index d983a9a7fed8f87ac7b924fc4f942c910bc98612..ef1fa37716de560e0032f99303e3112abc44ea97 100644 (file)
@@ -356,6 +356,7 @@ typedef struct ExprEvalStep
            /* faster to access without additional indirection: */
            PGFunction  fn_addr;    /* actual call address */
            int         nargs;  /* number of arguments */
+           bool        make_ro;    /* make arg0 R/O (used only for NULLIF) */
        }           func;
 
        /* for EEOP_BOOL_*_STEP */
index f5136c17abbf0eaed00ab76ece3d9ee933dd9190..efee7fc43173b4c52f6d65fb0b14936191edfcd6 100644 (file)
@@ -397,6 +397,14 @@ SELECT CASE make_ad(1,2)
  right
 (1 row)
 
+-- While we're here, also test handling of a NULLIF arg that is a read/write
+-- object (bug #18722)
+SELECT NULLIF(make_ad(1,2), array[2,3]::arrdomain);
+ nullif 
+--------
+ {1,2}
+(1 row)
+
 ROLLBACK;
 -- Test interaction of CASE with ArrayCoerceExpr (bug #15471)
 BEGIN;
index 83fe43be6b84f1b96db4285e2e9eeca037ff261f..388d4c6f52835ebee50168b0abd4b18cabfcfde9 100644 (file)
@@ -242,6 +242,11 @@ SELECT CASE make_ad(1,2)
   WHEN array[1,2]::arrdomain THEN 'right'
   END;
 
+-- While we're here, also test handling of a NULLIF arg that is a read/write
+-- object (bug #18722)
+
+SELECT NULLIF(make_ad(1,2), array[2,3]::arrdomain);
+
 ROLLBACK;
 
 -- Test interaction of CASE with ArrayCoerceExpr (bug #15471)