On Mon, May 13, 2013 at 1:40 PM, Marc Glisse <marc.gli...@inria.fr> wrote: > On Mon, 13 May 2013, Richard Biener wrote: > >> On Sat, May 11, 2013 at 11:38 AM, Marc Glisse <marc.gli...@inria.fr> >> wrote: >>> >>> Second try. >>> >>> I removed the fold_single_bit_test thing (I thought I'd handle it, so I >>> started by the easy part, and never did the rest). >>> >>> Adapting invert_truthvalue_loc for vectors, I thought: calling >>> fold_truth_not_expr and build1 if it fails is just the same as >>> fold_build1. >>> Except that it wasn't: fold_unary_loc fold_convert to boolean before >>> calling >>> fold_truth_not_expr and then back to the required type. And instead of >>> simply changing the type of an EQ_EXPR, fold_convert introduces a >>> NOP_EXPR >>> (one that STRIP_NOPS doesn't remove), which hides the comparison from >>> many >>> other parts of the front-end (affects warnings) and folding. I hesitated >>> between removing this cast and enhancing fold_convert, and chose the one >>> that removes code. As a side benefit, I got an XPASS :-) >>> >>> >>> Passes bootstrap+testsuite on x86_64-linux-gnu. >>> >>> 2013-05-11 Marc Glisse <marc.gli...@inria.fr> >>> >>> >>> gcc/ >>> * fold-const.c (fold_negate_expr): Handle vectors. >>> (fold_truth_not_expr): Make it static. >>> (invert_truthvalue_loc): Handle vectors. Do not call >>> fold_truth_not_expr directly. >>> (fold_unary_loc) <BIT_NOT_EXPR>: Handle vector comparisons. >>> <TRUTH_NOT_EXPR>: Do not cast to boolean. >>> (fold_comparison): Handle vector constants. >>> (fold_ternary_loc) <VEC_COND_EXPR>: Adapt more COND_EXPR >>> optimizations. >>> * tree.h (fold_truth_not_expr): Remove declaration. >>> >>> >>> gcc/testsuite/ >>> * g++.dg/ext/vector22.C: New testcase. >>> * gcc.dg/binop-xor3.c: Remove xfail. >>> >>> -- >>> Marc Glisse >>> Index: gcc/fold-const.c >>> =================================================================== >>> --- gcc/fold-const.c (revision 198796) >>> +++ gcc/fold-const.c (working copy) >>> @@ -519,21 +519,21 @@ fold_negate_expr (location_t loc, tree t >>> { >>> tree type = TREE_TYPE (t); >>> tree tem; >>> >>> switch (TREE_CODE (t)) >>> { >>> /* Convert - (~A) to A + 1. */ >>> case BIT_NOT_EXPR: >>> if (INTEGRAL_TYPE_P (type)) >>> return fold_build2_loc (loc, PLUS_EXPR, type, TREE_OPERAND (t, >>> 0), >>> - build_int_cst (type, 1)); >>> + build_one_cst (type)); >>> break; >>> >>> case INTEGER_CST: >>> tem = fold_negate_const (t, type); >>> if (TREE_OVERFLOW (tem) == TREE_OVERFLOW (t) >>> || !TYPE_OVERFLOW_TRAPS (type)) >>> return tem; >>> break; >>> >>> case REAL_CST: >>> @@ -3078,21 +3078,21 @@ omit_two_operands_loc (location_t loc, t >>> } >>> >>> >>> /* Return a simplified tree node for the truth-negation of ARG. This >>> never alters ARG itself. We assume that ARG is an operation that >>> returns a truth value (0 or 1). >>> >>> FIXME: one would think we would fold the result, but it causes >>> problems with the dominator optimizer. */ >>> >>> -tree >>> +static tree >>> fold_truth_not_expr (location_t loc, tree arg) >>> { >>> tree type = TREE_TYPE (arg); >>> enum tree_code code = TREE_CODE (arg); >>> location_t loc1, loc2; >>> >>> /* If this is a comparison, we can simply invert it, except for >>> floating-point non-equality comparisons, in which case we just >>> enclose a TRUTH_NOT_EXPR around what we have. */ >>> >>> @@ -3215,38 +3215,33 @@ fold_truth_not_expr (location_t loc, tre >>> return build1_loc (loc, CLEANUP_POINT_EXPR, type, >>> invert_truthvalue_loc (loc1, TREE_OPERAND (arg, >>> 0))); >>> >>> default: >>> return NULL_TREE; >>> } >>> } >>> >>> /* Return a simplified tree node for the truth-negation of ARG. This >>> never alters ARG itself. We assume that ARG is an operation that >>> - returns a truth value (0 or 1). >>> - >>> - FIXME: one would think we would fold the result, but it causes >>> - problems with the dominator optimizer. */ >>> + returns a truth value (0 or 1 for scalars, 0 or -1 for vectors). */ >>> >>> tree >>> invert_truthvalue_loc (location_t loc, tree arg) >>> { >>> - tree tem; >>> - >>> if (TREE_CODE (arg) == ERROR_MARK) >>> return arg; >>> >>> - tem = fold_truth_not_expr (loc, arg); >>> - if (!tem) >>> - tem = build1_loc (loc, TRUTH_NOT_EXPR, TREE_TYPE (arg), arg); >>> - >>> - return tem; >>> + tree type = TREE_TYPE (arg); >>> + return fold_build1_loc (loc, VECTOR_TYPE_P (type) >>> + ? BIT_NOT_EXPR >>> + : TRUTH_NOT_EXPR, >>> + type, arg); >>> } >>> >>> /* Given a bit-wise operation CODE applied to ARG0 and ARG1, see if both >>> operands are another bit-wise operation with a common input. If so, >>> distribute the bit operations to save an operation and possibly two >>> if >>> constants are involved. For example, convert >>> (A | B) & (A | C) into A | (B & C) >>> Further simplification will occur if B and C are constants. >>> >>> If this optimization cannot be done, 0 will be returned. */ >>> @@ -8274,28 +8269,34 @@ fold_unary_loc (location_t loc, enum tre >>> { >>> elem = VECTOR_CST_ELT (arg0, i); >>> elem = fold_unary_loc (loc, BIT_NOT_EXPR, TREE_TYPE (type), >>> elem); >>> if (elem == NULL_TREE) >>> break; >>> elements[i] = elem; >>> } >>> if (i == count) >>> return build_vector (type, elements); >>> } >>> + else if (COMPARISON_CLASS_P (arg0) && VECTOR_INTEGER_TYPE_P >>> (type)) >>> + { >>> + tree op_type = TREE_TYPE (TREE_OPERAND (arg0, 0)); >>> + enum tree_code subcode = invert_tree_comparison (TREE_CODE >>> (arg0), >>> + HONOR_NANS (TYPE_MODE (op_type))); >>> + if (subcode != ERROR_MARK) >>> + return build2_loc (loc, subcode, type, TREE_OPERAND (arg0, >>> 0), >>> + TREE_OPERAND (arg0, 1)); >>> + } >>> + >> >> >> I wonder why you restrict this to VECTOR_INTEGER_TYPE_P - for >> TYPE_PRECISION == 1 type this should work, too. > > > If TYPE_PRECISION == 1, wouldn't it be better to turn BIT_NOT_EXPR into > TRUTH_NOT_EXPR? Then it will be handled by fold_truth_not_expr.
Hmm, not sure - on GIMPLE we are no longer having the TRUTH_* tree codes, so we don't want to fold BIT_* to TRUTH_*. >> Also there should >> never be a comparison resulting in a non-integer vector type, no? > > > Yes, I was going to write VECTOR_TYPE_P, and adding integer seemed more > explicit, but I can go back. Works for me. >>> return NULL_TREE; >>> >>> case TRUTH_NOT_EXPR: >>> - /* The argument to invert_truthvalue must have Boolean type. */ >>> - if (TREE_CODE (TREE_TYPE (arg0)) != BOOLEAN_TYPE) >>> - arg0 = fold_convert_loc (loc, boolean_type_node, arg0); >>> - >>> /* Note that the operand of this must be an int >>> and its values must be 0 or 1. >>> ("true" is a fixed value perhaps depending on the language, >>> but we don't handle values other than 1 correctly yet.) */ >>> tem = fold_truth_not_expr (loc, arg0); >>> if (!tem) >>> return NULL_TREE; >>> return fold_convert_loc (loc, type, tem); >>> >>> case REALPART_EXPR: >>> @@ -9579,21 +9580,21 @@ fold_comparison (location_t loc, enum tr >>> { >>> tree cmp_type = TREE_TYPE (TREE_OPERAND (arg0, 0)); >>> return fold_build2_loc (loc, code, type, >>> fold_convert_loc (loc, cmp_type, >>> TREE_OPERAND (arg1, 0)), >>> TREE_OPERAND (arg0, 0)); >>> } >>> >>> /* Fold ~X op C as X op' ~C, where op' is the swapped comparison. */ >>> if (TREE_CODE (arg0) == BIT_NOT_EXPR >>> - && TREE_CODE (arg1) == INTEGER_CST) >>> + && (TREE_CODE (arg1) == INTEGER_CST || TREE_CODE (arg1) == >>> VECTOR_CST)) >>> { >>> tree cmp_type = TREE_TYPE (TREE_OPERAND (arg0, 0)); >>> return fold_build2_loc (loc, swap_tree_comparison (code), type, >>> TREE_OPERAND (arg0, 0), >>> fold_build1_loc (loc, BIT_NOT_EXPR, cmp_type, >>> fold_convert_loc (loc, cmp_type, >>> arg1))); >>> } >>> >>> return NULL_TREE; >>> } >>> @@ -14030,61 +14031,67 @@ fold_ternary_loc (location_t loc, enum t >>> return tem; >>> } >>> >>> if (COMPARISON_CLASS_P (arg0) >>> && operand_equal_for_comparison_p (TREE_OPERAND (arg0, 0), >>> op2, >>> TREE_OPERAND (arg0, 1)) >>> && !HONOR_SIGNED_ZEROS (TYPE_MODE (TREE_TYPE (op2)))) >>> { >>> location_t loc0 = expr_location_or (arg0, loc); >>> - tem = fold_truth_not_expr (loc0, arg0); >>> + tem = fold_unary_loc (loc0, VECTOR_TYPE_P (type) >>> + ? BIT_NOT_EXPR >>> + : TRUTH_NOT_EXPR, >>> + TREE_TYPE (arg0), arg0); >> >> >> since you don't restrict it here either .... >> >>> if (tem && COMPARISON_CLASS_P (tem)) >>> { >>> tem = fold_cond_expr_with_comparison (loc, type, tem, op2, >>> op1); >>> if (tem) >>> return tem; >>> } >>> } >>> >>> - /* ??? Fixup the code below for VEC_COND_EXPR. */ >>> - if (code == VEC_COND_EXPR) >>> - return NULL_TREE; >>> - >>> /* If the second operand is simpler than the third, swap them >>> since that produces better jump optimization results. */ >>> if (truth_value_p (TREE_CODE (arg0)) >>> && tree_swap_operands_p (op1, op2, false)) >>> { >>> location_t loc0 = expr_location_or (arg0, loc); >>> /* See if this can be inverted. If it can't, possibly because >>> it was a floating-point inequality comparison, don't do >>> anything. */ >>> - tem = fold_truth_not_expr (loc0, arg0); >>> + tem = fold_unary_loc (loc0, VECTOR_TYPE_P (type) >>> + ? BIT_NOT_EXPR >>> + : TRUTH_NOT_EXPR, >>> + TREE_TYPE (arg0), arg0); >> >> >> Btw, splitting this out to a helper would be nice. > > > Called fold_truth_not_expr? Oups ;-) > fold_truth_invert? fold_invert_truth? Well, fold_invert_truthvalue maybe? Inverting a truthvalue is well-defined for vectors and non-vectors and is more appropriate spelling for GIMPLE where we don't have any TRUTH_NOT_EXPR anymore. Richard. > > >> The rest of the patch looks ok to me. >> >> Thanks, >> Richard. >> >>> if (tem) >>> return fold_build3_loc (loc, code, type, tem, op2, op1); >>> } >>> >>> /* Convert A ? 1 : 0 to simply A. */ >>> - if (integer_onep (op1) >>> + if ((code == VEC_COND_EXPR ? integer_all_onesp (op1) >>> + : (integer_onep (op1) >>> + && !VECTOR_TYPE_P (type))) >>> && integer_zerop (op2) >>> /* If we try to convert OP0 to our type, the >>> call to fold will try to move the conversion inside >>> a COND, which will recurse. In that case, the COND_EXPR >>> is probably the best choice, so leave it alone. */ >>> && type == TREE_TYPE (arg0)) >>> return pedantic_non_lvalue_loc (loc, arg0); >>> >>> /* Convert A ? 0 : 1 to !A. This prefers the use of NOT_EXPR >>> over COND_EXPR in cases such as floating point comparisons. */ >>> if (integer_zerop (op1) >>> - && integer_onep (op2) >>> + && (code == VEC_COND_EXPR ? integer_all_onesp (op2) >>> + : (integer_onep (op2) >>> + && !VECTOR_TYPE_P (type))) >>> && truth_value_p (TREE_CODE (arg0))) >>> return pedantic_non_lvalue_loc (loc, >>> fold_convert_loc (loc, type, >>> invert_truthvalue_loc (loc, >>> >>> arg0))); >>> >>> /* A < 0 ? <sign bit of A> : 0 is simply (A & <sign bit of A>). >>> */ >>> if (TREE_CODE (arg0) == LT_EXPR >>> && integer_zerop (TREE_OPERAND (arg0, 1)) >>> && integer_zerop (op2) >>> @@ -14187,60 +14194,73 @@ fold_ternary_loc (location_t loc, enum t >>> && TREE_CODE (TREE_OPERAND (arg0, 0)) == BIT_AND_EXPR >>> && operand_equal_p (TREE_OPERAND (TREE_OPERAND (arg0, 0), 1), >>> arg1, OEP_ONLY_CONST)) >>> return pedantic_non_lvalue_loc (loc, >>> fold_convert_loc (loc, type, >>> TREE_OPERAND (arg0, >>> 0))); >>> >>> /* Convert A ? B : 0 into A && B if A and B are truth values. */ >>> if (integer_zerop (op2) >>> && truth_value_p (TREE_CODE (arg0)) >>> - && truth_value_p (TREE_CODE (arg1))) >>> - return fold_build2_loc (loc, TRUTH_ANDIF_EXPR, type, >>> - fold_convert_loc (loc, type, arg0), >>> - arg1); >>> + && truth_value_p (TREE_CODE (arg1)) >>> + && (code == VEC_COND_EXPR || !VECTOR_TYPE_P (type))) >>> + return fold_build2_loc (loc, code == VEC_COND_EXPR ? BIT_AND_EXPR >>> + : >>> TRUTH_ANDIF_EXPR, >>> + type, fold_convert_loc (loc, type, arg0), >>> arg1); >>> >>> /* Convert A ? B : 1 into !A || B if A and B are truth values. */ >>> - if (integer_onep (op2) >>> + if (code == VEC_COND_EXPR ? integer_all_onesp (op2) : integer_onep >>> (op2) >>> && truth_value_p (TREE_CODE (arg0)) >>> - && truth_value_p (TREE_CODE (arg1))) >>> + && truth_value_p (TREE_CODE (arg1)) >>> + && (code == VEC_COND_EXPR || !VECTOR_TYPE_P (type))) >>> { >>> location_t loc0 = expr_location_or (arg0, loc); >>> /* Only perform transformation if ARG0 is easily inverted. */ >>> - tem = fold_truth_not_expr (loc0, arg0); >>> + tem = fold_unary_loc (loc0, VECTOR_TYPE_P (type) >>> + ? BIT_NOT_EXPR >>> + : TRUTH_NOT_EXPR, >>> + TREE_TYPE (arg0), arg0); >>> if (tem) >>> - return fold_build2_loc (loc, TRUTH_ORIF_EXPR, type, >>> - fold_convert_loc (loc, type, tem), >>> - arg1); >>> + return fold_build2_loc (loc, code == VEC_COND_EXPR >>> + ? BIT_IOR_EXPR >>> + : TRUTH_ORIF_EXPR, >>> + type, fold_convert_loc (loc, type, >>> tem), >>> + arg1); >>> } >>> >>> /* Convert A ? 0 : B into !A && B if A and B are truth values. */ >>> if (integer_zerop (arg1) >>> && truth_value_p (TREE_CODE (arg0)) >>> - && truth_value_p (TREE_CODE (op2))) >>> + && truth_value_p (TREE_CODE (op2)) >>> + && (code == VEC_COND_EXPR || !VECTOR_TYPE_P (type))) >>> { >>> location_t loc0 = expr_location_or (arg0, loc); >>> /* Only perform transformation if ARG0 is easily inverted. */ >>> - tem = fold_truth_not_expr (loc0, arg0); >>> + tem = fold_unary_loc (loc0, VECTOR_TYPE_P (type) >>> + ? BIT_NOT_EXPR >>> + : TRUTH_NOT_EXPR, >>> + TREE_TYPE (arg0), arg0); >>> if (tem) >>> - return fold_build2_loc (loc, TRUTH_ANDIF_EXPR, type, >>> - fold_convert_loc (loc, type, tem), >>> - op2); >>> + return fold_build2_loc (loc, code == VEC_COND_EXPR >>> + ? BIT_AND_EXPR : >>> TRUTH_ANDIF_EXPR, >>> + type, fold_convert_loc (loc, type, >>> tem), >>> + op2); >>> } >>> >>> /* Convert A ? 1 : B into A || B if A and B are truth values. */ >>> - if (integer_onep (arg1) >>> + if (code == VEC_COND_EXPR ? integer_all_onesp (arg1) : >>> integer_onep >>> (arg1) >>> && truth_value_p (TREE_CODE (arg0)) >>> - && truth_value_p (TREE_CODE (op2))) >>> - return fold_build2_loc (loc, TRUTH_ORIF_EXPR, type, >>> - fold_convert_loc (loc, type, arg0), >>> - op2); >>> + && truth_value_p (TREE_CODE (op2)) >>> + && (code == VEC_COND_EXPR || !VECTOR_TYPE_P (type))) >>> + return fold_build2_loc (loc, code == VEC_COND_EXPR >>> + ? BIT_IOR_EXPR : TRUTH_ORIF_EXPR, >>> + type, fold_convert_loc (loc, type, arg0), >>> op2); >>> >>> return NULL_TREE; >>> >>> case CALL_EXPR: >>> /* CALL_EXPRs used to be ternary exprs. Catch any mistaken uses >>> of fold_ternary on them. */ >>> gcc_unreachable (); >>> >>> case BIT_FIELD_REF: >>> if ((TREE_CODE (arg0) == VECTOR_CST >>> Index: gcc/testsuite/gcc.dg/binop-xor3.c >>> =================================================================== >>> --- gcc/testsuite/gcc.dg/binop-xor3.c (revision 198796) >>> +++ gcc/testsuite/gcc.dg/binop-xor3.c (working copy) >>> @@ -1,11 +1,11 @@ >>> /* { dg-do compile } */ >>> /* { dg-options "-O2 -fdump-tree-optimized" } */ >>> >>> int >>> foo (int a, int b) >>> { >>> return ((a && !b) || (!a && b)); >>> } >>> >>> -/* { dg-final { scan-tree-dump-times "\\\^" 1 "optimized" { xfail *-*-* >>> } } >>> } */ >>> +/* { dg-final { scan-tree-dump-times "\\\^" 1 "optimized" } } */ >>> /* { dg-final { cleanup-tree-dump "optimized" } } */ >>> Index: gcc/testsuite/g++.dg/ext/vector22.C >>> =================================================================== >>> --- gcc/testsuite/g++.dg/ext/vector22.C (revision 0) >>> +++ gcc/testsuite/g++.dg/ext/vector22.C (revision 0) >>> @@ -0,0 +1,20 @@ >>> +/* { dg-do compile } */ >>> +/* { dg-options "-O -fdump-tree-gimple" } */ >>> + >>> +typedef unsigned vec __attribute__((vector_size(4*sizeof(int)))); >>> + >>> +void f(vec*a,vec*b){ >>> + *a=(*a)?-1:(*b<10); >>> + *b=(*b)?(*a<10):0; >>> +} >>> +void g(vec*a,vec*b){ >>> + *a=(*a)?(*a<*a):-1; >>> + *b=(*b)?-1:(*b<*b); >>> +} >>> +void h(vec*a){ >>> + *a=(~*a==5); >>> +} >>> + >>> +/* { dg-final { scan-tree-dump-not "~" "gimple" } } */ >>> +/* { dg-final { scan-tree-dump-not "VEC_COND_EXPR" "gimple" } } */ >>> +/* { dg-final { cleanup-tree-dump "gimple" } } */ >>> >>> Property changes on: gcc/testsuite/g++.dg/ext/vector22.C >>> ___________________________________________________________________ >>> Added: svn:keywords >>> + Author Date Id Revision URL >>> Added: svn:eol-style >>> + native >>> >>> Index: gcc/tree.h >>> =================================================================== >>> --- gcc/tree.h (revision 198796) >>> +++ gcc/tree.h (working copy) >>> @@ -5758,21 +5758,20 @@ extern int operand_equal_p (const_tree, >>> extern int multiple_of_p (tree, const_tree, const_tree); >>> #define omit_one_operand(T1,T2,T3)\ >>> omit_one_operand_loc (UNKNOWN_LOCATION, T1, T2, T3) >>> extern tree omit_one_operand_loc (location_t, tree, tree, tree); >>> #define omit_two_operands(T1,T2,T3,T4)\ >>> omit_two_operands_loc (UNKNOWN_LOCATION, T1, T2, T3, T4) >>> extern tree omit_two_operands_loc (location_t, tree, tree, tree, tree); >>> #define invert_truthvalue(T)\ >>> invert_truthvalue_loc(UNKNOWN_LOCATION, T) >>> extern tree invert_truthvalue_loc (location_t, tree); >>> -extern tree fold_truth_not_expr (location_t, tree); >>> extern tree fold_unary_to_constant (enum tree_code, tree, tree); >>> extern tree fold_binary_to_constant (enum tree_code, tree, tree, tree); >>> extern tree fold_read_from_constant_string (tree); >>> extern tree int_const_binop (enum tree_code, const_tree, const_tree); >>> #define build_fold_addr_expr(T)\ >>> build_fold_addr_expr_loc (UNKNOWN_LOCATION, (T)) >>> extern tree build_fold_addr_expr_loc (location_t, tree); >>> #define build_fold_addr_expr_with_type(T,TYPE)\ >>> build_fold_addr_expr_with_type_loc (UNKNOWN_LOCATION, (T), TYPE) >>> extern tree build_fold_addr_expr_with_type_loc (location_t, tree, tree); > > > -- > Marc Glisse