On Fri, 22 Jul 2011, Richard Guenther wrote:
> On Fri, 22 Jul 2011, Richard Guenther wrote:
>
> > On Fri, 22 Jul 2011, Richard Guenther wrote:
> >
> > > On Thu, 21 Jul 2011, Joseph S. Myers wrote:
> > >
> > > > On Thu, 21 Jul 2011, Richard Guenther wrote:
> > > >
> > > > > Patch also handling wider modes and not starting with SImode but
> > > > > the mode of int:
> > > >
> > > > Use of target int for anything not about C ABIs is certainly wrong.
> > > > This
> > > > might be about what operations the target does efficiently, or what
> > > > functions are present in libgcc (both of which would be functions of
> > > > machine modes), but it's not about the choice of C int.
> > >
> > > Ok. Given rths last suggestion I'm testing the following which
> > > checks all integer modes (but never will widen - optabs.c will do
> > > that if it turns out to be profitable).
> >
> > Err, I should refresh the patch before sending it ... here it goes.
>
> Regresses vectorization on i?86 because that defines floathi expanders
> but the vectorizer does not recognize a short -> float conversion
> as that requires different sized vectors (the int -> short truncation
> is also a complication for it). One of the cases that fail to vectorize
> is
>
> #define N 40
> float image[N][N] __attribute__ ((__aligned__(__BIGGEST_ALIGNMENT__)));
> float out[N];
>
> /* Outer-loop vectorization. */
>
> __attribute__ ((noinline)) void
> foo (){
> int i,j;
>
> for (i = 0; i < N; i++) {
> for (j = 0; j < N; j++) {
> image[j][i] = j+i;
> }
> }
> }
>
> where VRP figures out that j+i can be represented as a short.
>
> Now, I can dumb down the VRP code to only consider doing the
> unsigned -> signed demotion and never truncate (unless maybe
> the target does not support floatMODE of the current input mode).
>
> Other suggestions?
The following implements the above - do not truncate if the target
claims to be able to do the unsigned -> float conversion in the
original modes, but unconditionally try to change it to signed -> float.
Bootstrapped and tested on x86_64-unknown-linux-gnu.
Does this look sensible as a start? We can always improve things
incrementally when we discover a case that is worthwhile.
Thanks,
Richard.
2011-07-22 Richard Guenther <rguent...@suse.de>
PR tree-optimization/49715
* tree-vrp.c: Include expr.h and optabs.h.
(range_fits_type_): New function.
(simplify_float_conversion_using_ranges): Likewise.
(simplify_stmt_using_ranges): Call it.
* Makefile.in (tree-vrp.o): Add $(EXPR_H) and $(OPTABS_H) dependencies.
* optabs.c (can_float_p): Export.
* optabs.h (can_float_p): Declare.
* gcc.target/i386/pr49715-1.c: New testcase.
* gcc.target/i386/pr49715-2.c: Likewise.
Index: gcc/tree-vrp.c
===================================================================
*** gcc/tree-vrp.c.orig 2011-07-21 17:46:16.000000000 +0200
--- gcc/tree-vrp.c 2011-07-22 14:41:46.000000000 +0200
*************** along with GCC; see the file COPYING3.
*** 40,45 ****
--- 40,47 ----
#include "tree-ssa-propagate.h"
#include "tree-chrec.h"
#include "gimple-fold.h"
+ #include "expr.h"
+ #include "optabs.h"
/* Type of value ranges. See value_range_d for a description of these
*************** simplify_conversion_using_ranges (gimple
*** 7448,7453 ****
--- 7450,7548 ----
return true;
}
+ /* Return whether the value range *VR fits in an integer type specified
+ by PRECISION and UNSIGNED_P. */
+
+ static bool
+ range_fits_type_p (value_range_t *vr, unsigned precision, bool unsigned_p)
+ {
+ double_int tem;
+
+ /* We can only handle constant ranges. */
+ if (vr->type != VR_RANGE
+ || TREE_CODE (vr->min) != INTEGER_CST
+ || TREE_CODE (vr->max) != INTEGER_CST)
+ return false;
+
+ tem = double_int_ext (tree_to_double_int (vr->min), precision, unsigned_p);
+ if (!double_int_equal_p (tree_to_double_int (vr->min), tem))
+ return false;
+
+ tem = double_int_ext (tree_to_double_int (vr->max), precision, unsigned_p);
+ if (!double_int_equal_p (tree_to_double_int (vr->max), tem))
+ return false;
+
+ return true;
+ }
+
+ /* Simplify a conversion from integral SSA name to float in STMT. */
+
+ static bool
+ simplify_float_conversion_using_ranges (gimple_stmt_iterator *gsi, gimple
stmt)
+ {
+ tree rhs1 = gimple_assign_rhs1 (stmt);
+ value_range_t *vr = get_value_range (rhs1);
+ enum machine_mode fltmode = TYPE_MODE (TREE_TYPE (gimple_assign_lhs
(stmt)));
+ enum machine_mode mode;
+ tree tem;
+ gimple conv;
+
+ /* We can only handle constant ranges. */
+ if (vr->type != VR_RANGE
+ || TREE_CODE (vr->min) != INTEGER_CST
+ || TREE_CODE (vr->max) != INTEGER_CST)
+ return false;
+
+ /* First check if we can use a signed type in place of an unsigned. */
+ if (TYPE_UNSIGNED (TREE_TYPE (rhs1))
+ && (can_float_p (fltmode, TYPE_MODE (TREE_TYPE (rhs1)), 0)
+ != CODE_FOR_nothing)
+ && range_fits_type_p (vr, GET_MODE_PRECISION
+ (TYPE_MODE (TREE_TYPE (rhs1))), 0))
+ mode = TYPE_MODE (TREE_TYPE (rhs1));
+ /* If we can do the conversion in the current input mode do nothing. */
+ else if (can_float_p (fltmode, TYPE_MODE (TREE_TYPE (rhs1)),
+ TYPE_UNSIGNED (TREE_TYPE (rhs1))))
+ return false;
+ /* Otherwise search for a mode we can use, starting from the narrowest
+ integer mode available. */
+ else
+ {
+ mode = GET_CLASS_NARROWEST_MODE (MODE_INT);
+ do
+ {
+ /* If we cannot do a signed conversion to float from mode
+ or if the value-range does not fit in the signed type
+ try with a wider mode. */
+ if (can_float_p (fltmode, mode, 0) != CODE_FOR_nothing
+ && range_fits_type_p (vr, GET_MODE_PRECISION (mode), 0))
+ break;
+
+ mode = GET_MODE_WIDER_MODE (mode);
+ /* But do not widen the input. Instead leave that to the
+ optabs expansion code. */
+ if (GET_MODE_PRECISION (mode) > TYPE_PRECISION (TREE_TYPE (rhs1)))
+ return false;
+ }
+ while (mode != VOIDmode);
+ if (mode == VOIDmode)
+ return false;
+ }
+
+ /* It works, insert a truncation or sign-change before the
+ float conversion. */
+ tem = create_tmp_var (build_nonstandard_integer_type
+ (GET_MODE_PRECISION (mode), 0), NULL);
+ conv = gimple_build_assign_with_ops (NOP_EXPR, tem, rhs1, NULL_TREE);
+ tem = make_ssa_name (tem, conv);
+ gimple_assign_set_lhs (conv, tem);
+ gsi_insert_before (gsi, conv, GSI_SAME_STMT);
+ gimple_assign_set_rhs1 (stmt, tem);
+ update_stmt (stmt);
+
+ return true;
+ }
+
/* Simplify STMT using ranges if possible. */
static bool
*************** simplify_stmt_using_ranges (gimple_stmt_
*** 7507,7512 ****
--- 7602,7613 ----
return simplify_conversion_using_ranges (stmt);
break;
+ case FLOAT_EXPR:
+ if (TREE_CODE (rhs1) == SSA_NAME
+ && INTEGRAL_TYPE_P (TREE_TYPE (rhs1)))
+ return simplify_float_conversion_using_ranges (gsi, stmt);
+ break;
+
default:
break;
}
Index: gcc/testsuite/gcc.target/i386/pr49715-1.c
===================================================================
*** /dev/null 1970-01-01 00:00:00.000000000 +0000
--- gcc/testsuite/gcc.target/i386/pr49715-1.c 2011-07-21 17:46:39.000000000
+0200
***************
*** 0 ****
--- 1,9 ----
+ /* { dg-do compile } */
+ /* { dg-options "-O2 -msse -mfpmath=sse" } */
+
+ float func(unsigned x)
+ {
+ return (x & 0xfffff) * 0.01f;
+ }
+
+ /* { dg-final { scan-assembler-times "cvtsi2ss" 1 } } */
Index: gcc/testsuite/gcc.target/i386/pr49715-2.c
===================================================================
*** /dev/null 1970-01-01 00:00:00.000000000 +0000
--- gcc/testsuite/gcc.target/i386/pr49715-2.c 2011-07-21 17:46:39.000000000
+0200
***************
*** 0 ****
--- 1,12 ----
+ /* { dg-do compile } */
+ /* { dg-require-effective-target lp64 } */
+ /* { dg-options "-O2" } */
+
+ double func(unsigned long long x)
+ {
+ if (x <= 0x7ffffffffffffffeULL)
+ return (x + 1) * 0.01;
+ return 0.0;
+ }
+
+ /* { dg-final { scan-assembler-times "cvtsi2sdq" 1 } } */
Index: gcc/Makefile.in
===================================================================
*** gcc/Makefile.in.orig 2011-07-19 12:59:09.000000000 +0200
--- gcc/Makefile.in 2011-07-22 11:09:16.000000000 +0200
*************** tree-vrp.o : tree-vrp.c $(CONFIG_H) $(SY
*** 2504,2510 ****
$(TREE_FLOW_H) $(TREE_PASS_H) $(TREE_DUMP_H) $(DIAGNOSTIC_H) $(GGC_H) \
$(BASIC_BLOCK_H) tree-ssa-propagate.h $(FLAGS_H) $(TREE_DUMP_H) \
$(CFGLOOP_H) $(SCEV_H) $(TIMEVAR_H) intl.h tree-pretty-print.h \
! gimple-pretty-print.h gimple-fold.h
tree-cfg.o : tree-cfg.c $(TREE_FLOW_H) $(CONFIG_H) $(SYSTEM_H) \
$(TREE_H) $(TM_P_H) $(EXPR_H) $(GGC_H) $(FLAGS_H) output.h \
$(DIAGNOSTIC_H) $(FUNCTION_H) $(TIMEVAR_H) $(TM_H) coretypes.h \
--- 2504,2510 ----
$(TREE_FLOW_H) $(TREE_PASS_H) $(TREE_DUMP_H) $(DIAGNOSTIC_H) $(GGC_H) \
$(BASIC_BLOCK_H) tree-ssa-propagate.h $(FLAGS_H) $(TREE_DUMP_H) \
$(CFGLOOP_H) $(SCEV_H) $(TIMEVAR_H) intl.h tree-pretty-print.h \
! gimple-pretty-print.h gimple-fold.h $(OPTABS_H) $(EXPR_H)
tree-cfg.o : tree-cfg.c $(TREE_FLOW_H) $(CONFIG_H) $(SYSTEM_H) \
$(TREE_H) $(TM_P_H) $(EXPR_H) $(GGC_H) $(FLAGS_H) output.h \
$(DIAGNOSTIC_H) $(FUNCTION_H) $(TIMEVAR_H) $(TM_H) coretypes.h \
Index: gcc/optabs.c
===================================================================
*** gcc/optabs.c.orig 2011-07-11 17:02:51.000000000 +0200
--- gcc/optabs.c 2011-07-22 11:04:42.000000000 +0200
*************** can_fix_p (enum machine_mode fixmode, en
*** 4626,4632 ****
return CODE_FOR_nothing;
}
! static enum insn_code
can_float_p (enum machine_mode fltmode, enum machine_mode fixmode,
int unsignedp)
{
--- 4626,4632 ----
return CODE_FOR_nothing;
}
! enum insn_code
can_float_p (enum machine_mode fltmode, enum machine_mode fixmode,
int unsignedp)
{
Index: gcc/optabs.h
===================================================================
*** gcc/optabs.h.orig 2011-06-22 16:01:28.000000000 +0200
--- gcc/optabs.h 2011-07-22 11:06:01.000000000 +0200
*************** extern void expand_fixed_convert (rtx, r
*** 849,854 ****
--- 849,857 ----
/* Generate code for a FLOAT_EXPR. */
extern void expand_float (rtx, rtx, int);
+ /* Return the insn_code for a FLOAT_EXPR. */
+ enum insn_code can_float_p (enum machine_mode, enum machine_mode, int);
+
/* Generate code for a FIX_EXPR. */
extern void expand_fix (rtx, rtx, int);
