This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [RFC] Fold floating-point comparisons


On Fri, 28 May 2004, Paolo Bonzini wrote:
> gcc/ChangeLog:
> 2004-05-27  Paolo Bonzini  <bonzini@gnu.org>
> 	    Roger Sayle  <roger@eyesopen.com>
>
>	PR rtl-optimization/15649
>	Add LTGT_EXPR and improve pretty-printing of unordered
>	comparisons.
>	* c-common.c (c_common_truthvalue_conversion):
>	Handle LTGT_EXPR.
>	* c-typeck.c (build_binary_op): Likewise.
>	* dojump.c (do_jump): Likewise.
>	* expr.c (expand_expr_real_1, do_store_flag): Likewise.
>	* predict.c (tree_predict_by_opcode): Likewise.
>	* real.c (real_compare): Likewise.
>	* tree-cfg.c (verify_expr): Likewise.
>	* tree-inline.c (estimate_num_insns_1): Likewise.
>	* tree-pretty-print.c (dump_generic_node): Likewise.
>	Handle ORDERED_EXPR, UNORDERED_EXPR.
>	(op_symbol): Print unordered comparisons differently
>	than ordered ones.
>	* tree.def (LTGT_EXPR): New '<' tree code.
>	* doc/c-tree.texi (Expressions): Document floating-point
>	comparison nodes.
>
>	Fold comparisons between floating point values.
>	* fold-const.c (enum comparison_code): New, from
>	#define'd constants.  Define compcodes for unordered
>	comparisons and for invalid transformations.
>	(invert_truthvalue): Fold the resulting trees whenever
>	useful.
>	(invert_tree_comparison): Add "honor_nans" parameter.
>	(fold_truthop): Revamp to work on floating-point types too.
>	(comparison_to_compcode): Support unordered comparisons.
>	Use new enum comparison_code.
>	(compcode_to_comparison): Likewise.
>	(combine_compcodes): New function.
>	(invert_truthvalue): Let invert_tree_comparison decide
>	whether it is valid to fold the comparison.  Fold ORDERED
>	and UNORDERED even if flag_unsafe_math_optimizations is off,
>	and the remaining even if flag_unsafe_math_optimizations
>	is off but we are under -fno-trapping-math.
>	(fold_relational_const): Integer modes do not honor NaNs.
>
>gcc/testsuite/ChangeLog:
>2004-05-27  Paolo Bonzini  <bonzini@gnu.org>
>
>	* gcc.dg/compare-fp-1.c, gcc.dg/compare-fp-2.c,
>	gcc.dg/compare-fp-3.c, gcc.dg/compare-fp-4.c,
>	gcc.dg/pr15649-1.c: New.


This is OK for mainline.  The remaining issues can always be sorted
out as follow-up patches, and this is no worse than what's currently
there.

Roger
--


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]