This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: Patch to fix bit-testing performance regression
Roger Sayle <roger@eyesopen.com> writes:
> My only thought was whether this should be placed before or after the
> other transformation of BIT_AND_EXPR in dojump.c which changes the mode
> of comparisons with integer constants. If you moved your transformation
> earlier "(X >> 7) & 1" could be expanded as a test of a signed char's
> sign bit, "(char)X < 0", for example.
Good point. I should have noticed this myself, sorry. It doesn't make
any difference for MIPS, what with it being a SLOW_BYTE_ACCESS target,
but it could clearly make a difference elsewhere.
The patch below moves the do_jump hunk to the place you suggested.
Bootstrapped & regression tested on i686-pc-linux-gnu. OK to install?
Richard
* Makefile.in (dojump.o): Depend on $(GGC_H) and dojump.h.
(GTFILES): Add $(srcdir)/dojump.h.
(gt-dojump.h): New dependency.
* dojump.c (and_reg, and_test, shift_test): New static variables.
(prefer_and_bit_test): New function.
(do_jump): Use it to choose between (X & (1 << C)) and (X >> C) & 1.
Index: Makefile.in
===================================================================
RCS file: /cvs/gcc/gcc/gcc/Makefile.in,v
retrieving revision 1.1262
diff -c -p -F^\([(a-zA-Z0-9_]\|#define\) -r1.1262 Makefile.in
*** Makefile.in 16 Mar 2004 21:09:22 -0000 1.1262
--- Makefile.in 18 Mar 2004 16:45:33 -0000
*************** expr.o : expr.c $(CONFIG_H) $(SYSTEM_H)
*** 1597,1603 ****
except.h reload.h $(GGC_H) langhooks.h intl.h $(TM_P_H) real.h $(TARGET_H)
dojump.o : dojump.c $(CONFIG_H) $(SYSTEM_H) coretypes.h $(TM_H) $(RTL_H) $(TREE_H) \
flags.h function.h $(EXPR_H) $(OPTABS_H) $(INSN_ATTR_H) insn-config.h \
! langhooks.h
builtins.o : builtins.c $(CONFIG_H) $(SYSTEM_H) coretypes.h $(TM_H) $(RTL_H) $(TREE_H)\
flags.h $(TARGET_H) function.h $(REGS_H) $(EXPR_H) $(OPTABS_H) insn-config.h \
$(RECOG_H) output.h typeclass.h hard-reg-set.h toplev.h hard-reg-set.h \
--- 1597,1603 ----
except.h reload.h $(GGC_H) langhooks.h intl.h $(TM_P_H) real.h $(TARGET_H)
dojump.o : dojump.c $(CONFIG_H) $(SYSTEM_H) coretypes.h $(TM_H) $(RTL_H) $(TREE_H) \
flags.h function.h $(EXPR_H) $(OPTABS_H) $(INSN_ATTR_H) insn-config.h \
! langhooks.h $(GGC_H) gt-dojump.h
builtins.o : builtins.c $(CONFIG_H) $(SYSTEM_H) coretypes.h $(TM_H) $(RTL_H) $(TREE_H)\
flags.h $(TARGET_H) function.h $(REGS_H) $(EXPR_H) $(OPTABS_H) insn-config.h \
$(RECOG_H) output.h typeclass.h hard-reg-set.h toplev.h hard-reg-set.h \
*************** GTFILES = $(srcdir)/input.h $(srcdir)/co
*** 2086,2091 ****
--- 2086,2092 ----
$(srcdir)/c-common.h $(srcdir)/c-tree.h \
$(srcdir)/alias.c $(srcdir)/bitmap.c $(srcdir)/cselib.c $(srcdir)/cgraph.c \
$(srcdir)/dbxout.c $(srcdir)/dwarf2out.c $(srcdir)/dwarf2asm.c \
+ $(srcdir)/dojump.c \
$(srcdir)/emit-rtl.c $(srcdir)/except.c $(srcdir)/explow.c $(srcdir)/expr.c \
$(srcdir)/fold-const.c $(srcdir)/function.c \
$(srcdir)/gcse.c $(srcdir)/integrate.c $(srcdir)/lists.c $(srcdir)/optabs.c \
*************** gt-cgraph.h gt-coverage.h gtype-desc.h g
*** 2105,2111 ****
gt-function.h gt-integrate.h gt-stmt.h gt-tree.h gt-varasm.h \
gt-emit-rtl.h gt-explow.h gt-stor-layout.h gt-regclass.h \
gt-lists.h gt-alias.h gt-cselib.h gt-fold-const.h gt-gcse.h \
! gt-expr.h gt-sdbout.h gt-optabs.h gt-bitmap.h \
gt-dwarf2out.h gt-ra-build.h gt-reg-stack.h gt-dwarf2asm.h \
gt-dbxout.h gt-c-common.h gt-c-decl.h gt-c-parse.h \
gt-c-pragma.h gtype-c.h gt-input.h gt-cfglayout.h \
--- 2106,2112 ----
gt-function.h gt-integrate.h gt-stmt.h gt-tree.h gt-varasm.h \
gt-emit-rtl.h gt-explow.h gt-stor-layout.h gt-regclass.h \
gt-lists.h gt-alias.h gt-cselib.h gt-fold-const.h gt-gcse.h \
! gt-expr.h gt-sdbout.h gt-optabs.h gt-bitmap.h gt-dojump.h \
gt-dwarf2out.h gt-ra-build.h gt-reg-stack.h gt-dwarf2asm.h \
gt-dbxout.h gt-c-common.h gt-c-decl.h gt-c-parse.h \
gt-c-pragma.h gtype-c.h gt-input.h gt-cfglayout.h \
Index: dojump.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/dojump.c,v
retrieving revision 1.11
diff -c -p -F^\([(a-zA-Z0-9_]\|#define\) -r1.11 dojump.c
*** dojump.c 14 Mar 2004 22:26:04 -0000 1.11
--- dojump.c 18 Mar 2004 16:45:33 -0000
*************** 02111-1307, USA. */
*** 33,39 ****
--- 33,41 ----
#include "expr.h"
#include "optabs.h"
#include "langhooks.h"
+ #include "ggc.h"
+ static bool prefer_and_bit_test (enum machine_mode, int);
static void do_jump_by_parts_greater (tree, int, rtx, rtx);
static void do_jump_by_parts_equality (tree, rtx, rtx);
static void do_compare_and_jump (tree, enum rtx_code, enum rtx_code, rtx,
*************** jumpif (tree exp, rtx label)
*** 101,106 ****
--- 103,147 ----
do_jump (exp, NULL_RTX, label);
}
+ /* Used internally by prefer_and_bit_test. */
+
+ static GTY(()) rtx and_reg;
+ static GTY(()) rtx and_test;
+ static GTY(()) rtx shift_test;
+
+ /* Compare the relative costs of "(X & (1 << BITNUM))" and "(X >> BITNUM) & 1",
+ where X is an arbitrary register of mode MODE. Return true if the former
+ is preferred. */
+
+ static bool
+ prefer_and_bit_test (enum machine_mode mode, int bitnum)
+ {
+ if (and_test == 0)
+ {
+ /* Set up rtxes for the two variations. Use NULL as a placeholder
+ for the BITNUM-based constants. */
+ and_reg = gen_rtx_REG (mode, FIRST_PSEUDO_REGISTER);
+ and_test = gen_rtx_AND (mode, and_reg, NULL);
+ shift_test = gen_rtx_AND (mode, gen_rtx_ASHIFTRT (mode, and_reg, NULL),
+ const1_rtx);
+ }
+ else
+ {
+ /* Change the mode of the previously-created rtxes. */
+ PUT_MODE (and_reg, mode);
+ PUT_MODE (and_test, mode);
+ PUT_MODE (shift_test, mode);
+ PUT_MODE (XEXP (shift_test, 0), mode);
+ }
+
+ /* Fill in the integers. */
+ XEXP (and_test, 0) = GEN_INT ((unsigned HOST_WIDE_INT) 1 << bitnum);
+ XEXP (XEXP (shift_test, 0), 1) = GEN_INT (bitnum);
+
+ return (rtx_cost (and_test, IF_THEN_ELSE)
+ <= rtx_cost (shift_test, IF_THEN_ELSE));
+ }
+
/* Generate code to evaluate EXP and jump to IF_FALSE_LABEL if
the result is zero, or IF_TRUE_LABEL if the result is one.
Either of IF_FALSE_LABEL and IF_TRUE_LABEL may be zero,
*************** do_jump (tree exp, rtx if_false_label, r
*** 206,211 ****
--- 247,275 ----
break;
case BIT_AND_EXPR:
+ /* fold_single_bit_test() converts (X & (1 << C)) into (X >> C) & 1.
+ See if the former is preferred for jump tests and restore it
+ if so. */
+ if (TREE_CODE (TREE_OPERAND (exp, 0)) == RSHIFT_EXPR
+ && integer_onep (TREE_OPERAND (exp, 1)))
+ {
+ tree arg = TREE_OPERAND (TREE_OPERAND (exp, 0), 0);
+ tree shift = TREE_OPERAND (TREE_OPERAND (exp, 0), 1);
+ tree one = TREE_OPERAND (exp, 1);
+ tree argtype = TREE_TYPE (arg);
+ if (TREE_CODE (shift) == INTEGER_CST
+ && compare_tree_int (shift, 0) > 0
+ && compare_tree_int (shift, HOST_BITS_PER_WIDE_INT) < 0
+ && prefer_and_bit_test (TYPE_MODE (argtype),
+ TREE_INT_CST_LOW (shift)))
+ {
+ do_jump (build (BIT_AND_EXPR, argtype, arg,
+ fold (build (LSHIFT_EXPR, argtype, one, shift))),
+ if_false_label, if_true_label);
+ break;
+ }
+ }
+
/* If we are AND'ing with a small constant, do this comparison in the
smallest type that fits. If the machine doesn't have comparisons
that small, it will be converted back to the wider comparison.
*************** do_compare_and_jump (tree exp, enum rtx_
*** 999,1001 ****
--- 1063,1067 ----
? expr_size (TREE_OPERAND (exp, 0)) : NULL_RTX),
if_false_label, if_true_label);
}
+
+ #include "gt-dojump.h"