macros for turning off bounds checks. fixes #1392

`@inbounds expr` allows bounds checks to be omitted for code syntactically
inside the argument expression.

adds @inbounds to matmul, comprehensions, and vectorized binary operators.

Author: JeffBezanson

base/array.jl

@@ -934,7 +934,7 @@ for f in (:+, :-, :div, :mod, :&, :|, :$)
         function ($f){S,T}(A::StridedArray{S}, B::StridedArray{T})
             F = Array(promote_type(S,T), promote_shape(size(A),size(B)))
             for i=1:length(A)
-                F[i] = ($f)(A[i], B[i])
+                @inbounds F[i] = ($f)(A[i], B[i])
             end
             return F
         end
@@ -945,14 +945,14 @@ for f in (:+, :-, :.*, :./, :div, :mod, :&, :|, :$)
         function ($f){T}(A::Number, B::StridedArray{T})
             F = similar(B, promote_array_type(typeof(A),T))
             for i=1:length(B)
-                F[i] = ($f)(A, B[i])
+                @inbounds F[i] = ($f)(A, B[i])
             end
             return F
         end
         function ($f){T}(A::StridedArray{T}, B::Number)
             F = similar(A, promote_array_type(typeof(B),T))
             for i=1:length(A)
-                F[i] = ($f)(A[i], B)
+                @inbounds F[i] = ($f)(A[i], B)
             end
             return F
         end
@@ -961,7 +961,7 @@ for f in (:+, :-, :.*, :./, :div, :mod, :&, :|, :$)
             F = Array(promote_type(S,T), promote_shape(size(A),size(B)))
             i = 1
             for b in B
-                F[i] = ($f)(A[i], b)
+                @inbounds F[i] = ($f)(A[i], b)
                 i += 1
             end
             return F
@@ -970,7 +970,7 @@ for f in (:+, :-, :.*, :./, :div, :mod, :&, :|, :$)
             F = Array(promote_type(S,T), promote_shape(size(A),size(B)))
             i = 1
             for a in A
-                F[i] = ($f)(a, B[i])
+                @inbounds F[i] = ($f)(a, B[i])
                 i += 1
             end
             return F
@@ -1000,23 +1000,23 @@ function complex{S<:Real,T<:Real}(A::Array{S}, B::Array{T})
     if size(A) != size(B); error("argument dimensions must match"); end
     F = similar(A, typeof(complex(zero(S),zero(T))))
     for i=1:length(A)
-        F[i] = complex(A[i], B[i])
+        @inbounds F[i] = complex(A[i], B[i])
     end
     return F
 end
 
 function complex{T<:Real}(A::Real, B::Array{T})
     F = similar(B, typeof(complex(A,zero(T))))
     for i=1:length(B)
-        F[i] = complex(A, B[i])
+        @inbounds F[i] = complex(A, B[i])
     end
     return F
 end
 
 function complex{T<:Real}(A::Array{T}, B::Real)
     F = similar(A, typeof(complex(zero(T),B)))
     for i=1:length(A)
-        F[i] = complex(A[i], B)
+        @inbounds F[i] = complex(A[i], B)
     end
     return F
 end

base/base.jl

@@ -127,6 +127,18 @@ function precompile(f, args::Tuple)
     end
 end
 
+macro boundscheck(yesno,blk)
+    quote
+        $(Expr(:boundscheck,yesno))
+        $(esc(blk))
+        $(Expr(:boundscheck,0))
+    end
+end
+
+macro inbounds(blk)
+    :(@boundscheck false $(esc(blk)))
+end
+
 # NOTE: Base shares Array with Core so we can add definitions to it
 
 Array{T,N}(::Type{T}, d::NTuple{N,Int}) =

base/exports.jl

@@ -1215,4 +1215,6 @@ export
     @show,
     @printf,
     @sprintf,
-    @deprecate
+    @deprecate,
+    @boundscheck,
+    @inbounds

base/linalg/matmul.jl

@@ -407,6 +407,7 @@ function generic_matmatmul{T,S,R}(C::StridedVecOrMat{R}, tA, tB, A::StridedVecOr
     if mA == nA == nB == 2; return matmul2x2(C, tA, tB, A, B); end
     if mA == nA == nB == 3; return matmul3x3(C, tA, tB, A, B); end
 
+    @inbounds begin
     if isbits(R)
         tile_size = int(ifloor(sqrt(tilebufsize/sizeof(R))))
         sz = (tile_size, tile_size)
@@ -559,6 +560,7 @@ function generic_matmatmul{T,S,R}(C::StridedVecOrMat{R}, tA, tB, A::StridedVecOr
             end
         end
     end
+    end
     return C
 end
 

src/alloc.c

@@ -84,7 +84,7 @@ jl_sym_t *anonymous_sym;  jl_sym_t *underscore_sym;
 jl_sym_t *abstracttype_sym; jl_sym_t *bitstype_sym;
 jl_sym_t *compositetype_sym; jl_sym_t *type_goto_sym;
 jl_sym_t *global_sym; jl_sym_t *tuple_sym;
-jl_sym_t *dot_sym;
+jl_sym_t *dot_sym;    jl_sym_t *boundscheck_sym;
 
 typedef struct {
     int64_t a;

src/cgutils.cpp

@@ -371,8 +371,10 @@ static Value *emit_bounds_check(Value *i, Value *len, jl_codectx_t *ctx)
 {
     Value *im1 = builder.CreateSub(i, ConstantInt::get(T_size, 1));
 #if CHECK_BOUNDS==1
-    Value *ok = builder.CreateICmpULT(im1, len);
-    raise_exception_unless(ok, jlboundserr_var, ctx);
+    if (ctx->boundsCheck.empty() || ctx->boundsCheck.back()==true) {
+        Value *ok = builder.CreateICmpULT(im1, len);
+        raise_exception_unless(ok, jlboundserr_var, ctx);
+    }
 #endif
     return im1;
 }
@@ -681,9 +683,13 @@ static Value *emit_array_nd_index(Value *a, jl_value_t *ex, size_t nd, jl_value_
 {
     Value *i = ConstantInt::get(T_size, 0);
     Value *stride = ConstantInt::get(T_size, 1);
+    bool bc = ctx->boundsCheck.empty() || ctx->boundsCheck.back()==true;
 #if CHECK_BOUNDS==1
-    BasicBlock *failBB = BasicBlock::Create(getGlobalContext(), "oob");
-    BasicBlock *endBB = BasicBlock::Create(getGlobalContext(), "idxend");
+    BasicBlock *failBB=NULL, *endBB=NULL;
+    if (bc) {
+        failBB = BasicBlock::Create(getGlobalContext(), "oob");
+        endBB = BasicBlock::Create(getGlobalContext(), "idxend");
+    }
 #endif
     for(size_t k=0; k < nidxs; k++) {
         Value *ii = emit_unbox(T_size, T_psize, emit_unboxed(args[k], ctx));
@@ -693,28 +699,32 @@ static Value *emit_array_nd_index(Value *a, jl_value_t *ex, size_t nd, jl_value_
             Value *d =
                 k >= nd ? ConstantInt::get(T_size, 1) : emit_arraysize(a, ex, k+1, ctx);
 #if CHECK_BOUNDS==1
-            BasicBlock *okBB = BasicBlock::Create(getGlobalContext(), "ib");
-            // if !(i < d) goto error
-            builder.CreateCondBr(builder.CreateICmpULT(ii, d), okBB, failBB);
-            ctx->f->getBasicBlockList().push_back(okBB);
-            builder.SetInsertPoint(okBB);
+            if (bc) {
+                BasicBlock *okBB = BasicBlock::Create(getGlobalContext(), "ib");
+                // if !(i < d) goto error
+                builder.CreateCondBr(builder.CreateICmpULT(ii, d), okBB, failBB);
+                ctx->f->getBasicBlockList().push_back(okBB);
+                builder.SetInsertPoint(okBB);
+            }
 #endif
             stride = builder.CreateMul(stride, d);
         }
     }
 #if CHECK_BOUNDS==1
-    Value *alen = emit_arraylen(a, ex, ctx);
-    // if !(i < alen) goto error
-    builder.CreateCondBr(builder.CreateICmpULT(i, alen), endBB, failBB);
-
-    ctx->f->getBasicBlockList().push_back(failBB);
-    builder.SetInsertPoint(failBB);
-    builder.CreateCall2(jlthrow_line_func, builder.CreateLoad(jlboundserr_var),
-                        ConstantInt::get(T_int32, ctx->lineno));
-    builder.CreateUnreachable();
-
-    ctx->f->getBasicBlockList().push_back(endBB);
-    builder.SetInsertPoint(endBB);
+    if (bc) {
+        Value *alen = emit_arraylen(a, ex, ctx);
+        // if !(i < alen) goto error
+        builder.CreateCondBr(builder.CreateICmpULT(i, alen), endBB, failBB);
+
+        ctx->f->getBasicBlockList().push_back(failBB);
+        builder.SetInsertPoint(failBB);
+        builder.CreateCall2(jlthrow_line_func, builder.CreateLoad(jlboundserr_var),
+                            ConstantInt::get(T_int32, ctx->lineno));
+        builder.CreateUnreachable();
+
+        ctx->f->getBasicBlockList().push_back(endBB);
+        builder.SetInsertPoint(endBB);
+    }
 #endif
 
     return i;

src/codegen.cpp

@@ -402,6 +402,7 @@ typedef struct {
     bool vaStack;      // varargs stack-allocated
     int nReqArgs;
     int lineno;
+    std::vector<bool> boundsCheck;
 } jl_codectx_t;
 
 static Value *emit_expr(jl_value_t *expr, jl_codectx_t *ctx, bool boxed=true,
@@ -2057,6 +2058,23 @@ static Value *emit_expr(jl_value_t *expr, jl_codectx_t *ctx, bool isboxed,
 #endif
         builder.SetInsertPoint(tryblk);
     }
+    else if (head == boundscheck_sym) {
+        if (jl_array_len(ex->args) > 0) {
+            jl_value_t *arg = args[0];
+            if (arg == jl_true) {
+                ctx->boundsCheck.push_back(true);
+            }
+            else if (arg == jl_false) {
+                ctx->boundsCheck.push_back(false);
+            }
+            else {
+                if (!ctx->boundsCheck.empty())
+                    ctx->boundsCheck.pop_back();
+            }
+        }
+        if (valuepos)
+            return literal_pointer_val((jl_value_t*)jl_nothing);
+    }
     else {
         if (!strcmp(head->name, "$"))
             jl_error("syntax: prefix $ in non-quoted expression");
@@ -2211,6 +2229,7 @@ static Function *emit_function(jl_lambda_info_t *lam, bool cstyle)
     ctx.funcName = lam->name->name;
     ctx.vaName = NULL;
     ctx.vaStack = false;
+    ctx.boundsCheck.push_back(true);
 
     // step 2. process var-info lists to see what vars are captured, need boxing
     jl_array_t *largs = jl_lam_args(ast);

src/interpreter.c

@@ -349,6 +349,9 @@ static jl_value_t *eval(jl_value_t *e, jl_value_t **locals, size_t nl)
             jl_errorf("syntax: %s", jl_string_data(args[0]));
         jl_throw(args[0]);
     }
+    else if (ex->head == boundscheck_sym) {
+        return (jl_value_t*)jl_nothing;
+    }
     jl_errorf("unsupported or misplaced expression %s", ex->head->name);
     return (jl_value_t*)jl_nothing;
 }

src/jltypes.c

@@ -2791,4 +2791,5 @@ void jl_init_types(void)
     tuple_sym = jl_symbol("tuple");
     kw_sym = jl_symbol("kw");
     dot_sym = jl_symbol(".");
+    boundscheck_sym = jl_symbol("boundscheck");
 }

src/julia-syntax.scm

@@ -1565,7 +1565,9 @@
 	(if (null? ranges)
 	    `(block (= ,oneresult ,expr)
 		    (type_goto ,initlabl)
+		    (boundscheck false)
 		    (call (top setindex!) ,result ,oneresult ,ri)
+		    (boundscheck 0)
 		    (= ,ri (call (top +) ,ri 1)))
 	    `(for ,(car ranges)
 		  ,(construct-loops (cdr ranges)))))
@@ -1593,9 +1595,10 @@
     (typed_comprehension atype expr . ranges)
     (if (any (lambda (x) (eq? x ':)) ranges)
 	(lower-nd-comprehension atype expr ranges)
-    (let ( (result (gensy))
-	   (ri (gensy))
-	   (rs (map (lambda (x) (gensy)) ranges)) )
+    (let ((result    (gensy))
+	  (oneresult (gensy))
+	  (ri (gensy))
+	  (rs (map (lambda (x) (gensy)) ranges)) )
 
       ;; compute the dimensions of the result
       (define (compute-dims ranges)
@@ -1605,7 +1608,10 @@
       ;; construct loops to cycle over all dimensions of an n-d comprehension
       (define (construct-loops ranges rs)
 	(if (null? ranges)
-	    `(block (call (top setindex!) ,result ,expr ,ri)
+	    `(block (= ,oneresult ,expr)
+		    (boundscheck false)
+                    (call (top setindex!) ,result ,oneresult ,ri)
+		    (boundscheck 0)
 		    (= ,ri (call (top +) ,ri 1)))
 	    `(for (= ,(cadr (car ranges)) ,(car rs))
 		  ,(construct-loops (cdr ranges) (cdr rs)))))