* [PATCH] FMA on trees
@ 2010-11-02 13:35 Richard Guenther
2010-11-02 15:40 ` Richard Henderson
0 siblings, 1 reply; 12+ messages in thread
From: Richard Guenther @ 2010-11-02 13:35 UTC (permalink / raw)
To: gcc-patches; +Cc: rth
This accumulates the patches from RTH and me and implements the suggestion
made by Joseph for -ffp-contract. The patch passes bootstrap & regtest
on x86_64-unknown-linux-gnu with the following fallout:
FAIL: gcc.target/i386/fma3-builtin-2.c scan-assembler-times vfmadd...ps 1
FAIL: gcc.target/i386/fma3-builtin-2.c scan-assembler-times vfmadd...pd 1
FAIL: gcc.target/i386/fma3-builtin-2.c scan-assembler-times vfmsub...ps 1
FAIL: gcc.target/i386/fma3-builtin-2.c scan-assembler-times vfmsub...pd 1
FAIL: gcc.target/i386/fma3-builtin-2.c scan-assembler-times vfnmadd...ps 1
FAIL: gcc.target/i386/fma3-builtin-2.c scan-assembler-times vfnmadd...pd 1
FAIL: gcc.target/i386/fma3-builtin-2.c scan-assembler-times vfnmsub...ps 1
FAIL: gcc.target/i386/fma3-builtin-2.c scan-assembler-times vfnmsub...pd 1
FAIL: gcc.target/i386/fma4-builtin-2.c scan-assembler-times vfmaddps 1
FAIL: gcc.target/i386/fma4-builtin-2.c scan-assembler-times vfmaddpd 1
FAIL: gcc.target/i386/fma4-builtin-2.c scan-assembler-times vfmsubps 1
FAIL: gcc.target/i386/fma4-builtin-2.c scan-assembler-times vfmsubpd 1
FAIL: gcc.target/i386/fma4-builtin-2.c scan-assembler-times vfnmaddps 1
FAIL: gcc.target/i386/fma4-builtin-2.c scan-assembler-times vfnmaddpd 1
FAIL: gcc.target/i386/fma4-builtin-2.c scan-assembler-times vfnmsubps 1
FAIL: gcc.target/i386/fma4-builtin-2.c scan-assembler-times vfnmsubpd 1
FAIL: gcc.target/i386/fma4-vector-2.c scan-assembler vfmaddps
FAIL: gcc.target/i386/fma4-vector-2.c scan-assembler vfmsubps
without investigating this probably means that folding __builtin_fma
to FMA_EXPR breaks vectorization here. I'll fix that and re-submit.
Still posted to make the stage1 deadline.
Richard.
2010-10-22 Richard Guenther <rguenther@suse.de>
Richard Henderson <rth@redhat.com>
* tree.def (FMA_EXPR): New tree code.
* expr.c (expand_expr_real_2): Add FMA_EXPR expansion code.
* gimple.c (gimple_rhs_class_table): FMA_EXPR is a GIMPLE_TERNARY_RHS.
* tree-cfg.c (verify_gimple_assign_ternary): Verify FMA_EXPR types.
* tree-inline.c (estimate_operator_cost): Handle FMA_EXPR.
* gimple-pretty-print.c (dump_ternary_rhs): Likewise.
* tree-ssa-math-opts.c (convert_mult_to_fma): New function.
(execute_optimize_widening_mul): Call it. Reorganize to allow
dead stmt removal. Move TODO flags ...
(pass_optimize_widening_mul): ... here.
* flag-types.h (enum fp_contract_mode): New enum.
* common.opt (flag_fp_contract_mode): New variable.
(-ffp-contract): New option.
* opts.c (common_handle_option): Handle it.
(set_unsafe_math_optimizations_flags): Enable -ffp-contract=fast.
(fast_math_flags_set_p): Check flag_fp_contract_mode.
* doc/invoke.texi (-ffp-contract): Document.
(-funsafe-math-optimizations): Adjust.
* builtins.c (fold_builtin_fma): New function.
(fold_builtin_3): Call it for fma.
* config/i386/sse.md (fms<mode>4, fnma<mode>, fnms<mode>4):
New expanders.
* doc/md.texi (fms<mode>4, fnma<mode>, fnms<mode>4): Document new
named patterns.
* genopinit.c (optabs): Initialize fms_optab, fnma_optab and fnms_optab.
* optabs.h (enum optab_index): Add OTI_fms, OTI_fnma and OTI_fnms.
(fms_optab, fnma_optab, fnms_optab): New defines.
* gimplify.c (gimplify_expr): Handle binary truth expressions
explicitly. Handle FMA_EXPR.
* gcc.target/i386/fma4-vector-2.c: New testcase.
Index: gcc/tree.def
===================================================================
*** gcc/tree.def.orig 2010-10-22 17:06:23.000000000 +0200
--- gcc/tree.def 2010-11-02 12:08:20.000000000 +0100
*************** DEFTREECODE (WIDEN_MULT_PLUS_EXPR, "wide
*** 1092,1097 ****
--- 1092,1103 ----
is subtracted from t3. */
DEFTREECODE (WIDEN_MULT_MINUS_EXPR, "widen_mult_plus_expr", tcc_expression, 3)
+ /* Fused multiply-add.
+ All operands and the result are of the same type. No intermediate
+ rounding is performed after multiplying operand one with operand two
+ before adding operand three. */
+ DEFTREECODE (FMA_EXPR, "fma_expr", tcc_expression, 3)
+
/* Whole vector left/right shift in bits.
Operand 0 is a vector to be shifted.
Operand 1 is an integer shift amount in bits. */
Index: gcc/expr.c
===================================================================
*** gcc/expr.c.orig 2010-10-22 17:06:23.000000000 +0200
--- gcc/expr.c 2010-11-02 12:18:39.000000000 +0100
*************** expand_expr_real_2 (sepops ops, rtx targ
*** 7254,7260 ****
int ignore;
bool reduce_bit_field;
location_t loc = ops->location;
! tree treeop0, treeop1;
#define REDUCE_BIT_FIELD(expr) (reduce_bit_field \
? reduce_to_bit_field_precision ((expr), \
target, \
--- 7254,7260 ----
int ignore;
bool reduce_bit_field;
location_t loc = ops->location;
! tree treeop0, treeop1, treeop2;
#define REDUCE_BIT_FIELD(expr) (reduce_bit_field \
? reduce_to_bit_field_precision ((expr), \
target, \
*************** expand_expr_real_2 (sepops ops, rtx targ
*** 7267,7272 ****
--- 7267,7273 ----
treeop0 = ops->op0;
treeop1 = ops->op1;
+ treeop2 = ops->op2;
/* We should be called only on simple (binary or unary) expressions,
exactly those that are valid in gimple expressions that aren't
*************** expand_expr_real_2 (sepops ops, rtx targ
*** 7624,7630 ****
case WIDEN_MULT_PLUS_EXPR:
case WIDEN_MULT_MINUS_EXPR:
expand_operands (treeop0, treeop1, NULL_RTX, &op0, &op1, EXPAND_NORMAL);
! op2 = expand_normal (ops->op2);
target = expand_widen_pattern_expr (ops, op0, op1, op2,
target, unsignedp);
return target;
--- 7625,7631 ----
case WIDEN_MULT_PLUS_EXPR:
case WIDEN_MULT_MINUS_EXPR:
expand_operands (treeop0, treeop1, NULL_RTX, &op0, &op1, EXPAND_NORMAL);
! op2 = expand_normal (treeop2);
target = expand_widen_pattern_expr (ops, op0, op1, op2,
target, unsignedp);
return target;
*************** expand_expr_real_2 (sepops ops, rtx targ
*** 7711,7716 ****
--- 7712,7757 ----
expand_operands (treeop0, treeop1, subtarget, &op0, &op1, EXPAND_NORMAL);
return REDUCE_BIT_FIELD (expand_mult (mode, op0, op1, target, unsignedp));
+ case FMA_EXPR:
+ {
+ optab opt = fma_optab;
+ gimple def0, def2;
+
+ def0 = get_def_for_expr (treeop0, NEGATE_EXPR);
+ def2 = get_def_for_expr (treeop2, NEGATE_EXPR);
+
+ op0 = op2 = NULL;
+
+ if (def0 && def2
+ && optab_handler (fnms_optab, mode) != CODE_FOR_nothing)
+ {
+ opt = fnms_optab;
+ op0 = expand_normal (gimple_assign_rhs1 (def0));
+ op2 = expand_normal (gimple_assign_rhs1 (def2));
+ }
+ else if (def0
+ && optab_handler (fnma_optab, mode) != CODE_FOR_nothing)
+ {
+ opt = fnma_optab;
+ op0 = expand_normal (gimple_assign_rhs1 (def0));
+ }
+ else if (def2
+ && optab_handler (fms_optab, mode) != CODE_FOR_nothing)
+ {
+ opt = fms_optab;
+ op2 = expand_normal (gimple_assign_rhs1 (def2));
+ }
+
+ if (op0 == NULL)
+ op0 = expand_expr (treeop0, subtarget, VOIDmode, EXPAND_NORMAL);
+ if (op2 == NULL)
+ op2 = expand_normal (treeop2);
+ op1 = expand_normal (treeop1);
+
+ return expand_ternary_op (TYPE_MODE (type), opt,
+ op0, op1, op2, target, 0);
+ }
+
case MULT_EXPR:
/* If this is a fixed-point operation, then we cannot use the code
below because "expand_mult" doesn't support sat/no-sat fixed-point
Index: gcc/gimple.c
===================================================================
*** gcc/gimple.c.orig 2010-11-02 11:16:39.000000000 +0100
--- gcc/gimple.c 2010-11-02 12:08:20.000000000 +0100
*************** get_gimple_rhs_num_ops (enum tree_code c
*** 2530,2536 ****
|| (SYM) == TRUTH_XOR_EXPR) ? GIMPLE_BINARY_RHS \
: (SYM) == TRUTH_NOT_EXPR ? GIMPLE_UNARY_RHS \
: ((SYM) == WIDEN_MULT_PLUS_EXPR \
! || (SYM) == WIDEN_MULT_MINUS_EXPR) ? GIMPLE_TERNARY_RHS \
: ((SYM) == COND_EXPR \
|| (SYM) == CONSTRUCTOR \
|| (SYM) == OBJ_TYPE_REF \
--- 2530,2537 ----
|| (SYM) == TRUTH_XOR_EXPR) ? GIMPLE_BINARY_RHS \
: (SYM) == TRUTH_NOT_EXPR ? GIMPLE_UNARY_RHS \
: ((SYM) == WIDEN_MULT_PLUS_EXPR \
! || (SYM) == WIDEN_MULT_MINUS_EXPR \
! || (SYM) == FMA_EXPR) ? GIMPLE_TERNARY_RHS \
: ((SYM) == COND_EXPR \
|| (SYM) == CONSTRUCTOR \
|| (SYM) == OBJ_TYPE_REF \
Index: gcc/tree-cfg.c
===================================================================
*** gcc/tree-cfg.c.orig 2010-10-22 17:06:23.000000000 +0200
--- gcc/tree-cfg.c 2010-11-02 12:08:20.000000000 +0100
*************** verify_gimple_assign_ternary (gimple stm
*** 3748,3753 ****
--- 3748,3767 ----
}
break;
+ case FMA_EXPR:
+ if (!useless_type_conversion_p (lhs_type, rhs1_type)
+ || !useless_type_conversion_p (lhs_type, rhs2_type)
+ || !useless_type_conversion_p (lhs_type, rhs3_type))
+ {
+ error ("type mismatch in fused multiply-add expression");
+ debug_generic_expr (lhs_type);
+ debug_generic_expr (rhs1_type);
+ debug_generic_expr (rhs2_type);
+ debug_generic_expr (rhs3_type);
+ return true;
+ }
+ break;
+
default:
gcc_unreachable ();
}
Index: gcc/tree-inline.c
===================================================================
*** gcc/tree-inline.c.orig 2010-10-22 17:06:23.000000000 +0200
--- gcc/tree-inline.c 2010-11-02 12:08:20.000000000 +0100
*************** estimate_operator_cost (enum tree_code c
*** 3284,3289 ****
--- 3284,3290 ----
case POINTER_PLUS_EXPR:
case MINUS_EXPR:
case MULT_EXPR:
+ case FMA_EXPR:
case ADDR_SPACE_CONVERT_EXPR:
case FIXED_CONVERT_EXPR:
Index: gcc/gimple-pretty-print.c
===================================================================
*** gcc/gimple-pretty-print.c.orig 2010-10-22 17:06:23.000000000 +0200
--- gcc/gimple-pretty-print.c 2010-11-02 12:08:20.000000000 +0100
*************** dump_ternary_rhs (pretty_printer *buffer
*** 400,405 ****
--- 400,413 ----
pp_character (buffer, '>');
break;
+ case FMA_EXPR:
+ dump_generic_node (buffer, gimple_assign_rhs1 (gs), spc, flags, false);
+ pp_string (buffer, " * ");
+ dump_generic_node (buffer, gimple_assign_rhs2 (gs), spc, flags, false);
+ pp_string (buffer, " + ");
+ dump_generic_node (buffer, gimple_assign_rhs3 (gs), spc, flags, false);
+ break;
+
default:
gcc_unreachable ();
}
Index: gcc/tree-ssa-math-opts.c
===================================================================
*** gcc/tree-ssa-math-opts.c.orig 2010-10-22 17:06:23.000000000 +0200
--- gcc/tree-ssa-math-opts.c 2010-11-02 12:33:53.000000000 +0100
*************** convert_plusminus_to_widen (gimple_stmt_
*** 1494,1499 ****
--- 1494,1616 ----
return true;
}
+ /* Combine the multiplication at MUL_STMT with uses in additions and
+ subtractions to form fused multiply-add operations. Returns true
+ if successful and MUL_STMT should be removed. */
+
+ static bool
+ convert_mult_to_fma (gimple mul_stmt)
+ {
+ tree mul_result = gimple_assign_lhs (mul_stmt);
+ tree type = TREE_TYPE (mul_result);
+ gimple use_stmt, fma_stmt;
+ use_operand_p use_p;
+ imm_use_iterator imm_iter;
+
+ if (FLOAT_TYPE_P (type)
+ && flag_fp_contract_mode == FP_CONTRACT_OFF)
+ return false;
+
+ /* We don't want to do bitfield reduction ops. */
+ if (INTEGRAL_TYPE_P (type)
+ && (TYPE_PRECISION (type)
+ != GET_MODE_PRECISION (TYPE_MODE (type))))
+ return false;
+
+ /* If the target doesn't support it, don't generate it. We assume that
+ if fma isn't available then fms, fnma or fnms are not either. */
+ if (optab_handler (fma_optab, TYPE_MODE (type)) == CODE_FOR_nothing)
+ return false;
+
+ /* Make sure that the multiplication statement becomes dead after
+ the transformation, thus that all uses are transformed to FMAs.
+ This means we assume that an FMA operation has the same cost
+ as an addition. */
+ FOR_EACH_IMM_USE_FAST (use_p, imm_iter, mul_result)
+ {
+ enum tree_code use_code;
+
+ use_stmt = USE_STMT (use_p);
+
+ if (!is_gimple_assign (use_stmt))
+ return false;
+ use_code = gimple_assign_rhs_code (use_stmt);
+ /* ??? We need to handle NEGATE_EXPR to eventually form fnms. */
+ if (use_code != PLUS_EXPR
+ && use_code != MINUS_EXPR)
+ return false;
+
+ /* For now restrict this operations to single basic blocks. In theory
+ we would want to support sinking the multiplication in
+ m = a*b;
+ if ()
+ ma = m + c;
+ else
+ d = m;
+ to form a fma in the then block and sink the multiplication to the
+ else block. */
+ if (gimple_bb (use_stmt) != gimple_bb (mul_stmt))
+ return false;
+
+ /* We can't handle a * b + a * b. */
+ if (gimple_assign_rhs1 (use_stmt) == gimple_assign_rhs2 (use_stmt))
+ return false;
+
+ /* If the target doesn't support a * b - c then drop the ball. */
+ if (gimple_assign_rhs1 (use_stmt) == mul_result
+ && use_code == MINUS_EXPR
+ && optab_handler (fms_optab, TYPE_MODE (type)) == CODE_FOR_nothing)
+ return false;
+
+ /* If the target doesn't support -a * b + c then drop the ball. */
+ if (gimple_assign_rhs2 (use_stmt) == mul_result
+ && use_code == MINUS_EXPR
+ && optab_handler (fnma_optab, TYPE_MODE (type)) == CODE_FOR_nothing)
+ return false;
+
+ /* We don't yet generate -a * b - c below yet. */
+ }
+
+ FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, mul_result)
+ {
+ tree addop, mulop1;
+ gimple_stmt_iterator gsi = gsi_for_stmt (use_stmt);
+
+ mulop1 = gimple_assign_rhs1 (mul_stmt);
+ if (gimple_assign_rhs1 (use_stmt) == mul_result)
+ {
+ addop = gimple_assign_rhs2 (use_stmt);
+ /* a * b - c -> a * b + (-c) */
+ if (gimple_assign_rhs_code (use_stmt) == MINUS_EXPR)
+ addop = force_gimple_operand_gsi (&gsi,
+ build1 (NEGATE_EXPR,
+ type, addop),
+ true, NULL_TREE, true,
+ GSI_SAME_STMT);
+ }
+ else
+ {
+ addop = gimple_assign_rhs1 (use_stmt);
+ /* a - b * c -> (-b) * c + a */
+ if (gimple_assign_rhs_code (use_stmt) == MINUS_EXPR)
+ mulop1 = force_gimple_operand_gsi (&gsi,
+ build1 (NEGATE_EXPR,
+ type, mulop1),
+ true, NULL_TREE, true,
+ GSI_SAME_STMT);
+ }
+
+ fma_stmt = gimple_build_assign_with_ops3 (FMA_EXPR,
+ gimple_assign_lhs (use_stmt),
+ mulop1,
+ gimple_assign_rhs2 (mul_stmt),
+ addop);
+ gsi_replace (&gsi, fma_stmt, true);
+ }
+
+ return true;
+ }
+
/* Find integer multiplications where the operands are extended from
smaller types, and replace the MULT_EXPR with a WIDEN_MULT_EXPR
where appropriate. */
*************** convert_plusminus_to_widen (gimple_stmt_
*** 1501,1531 ****
static unsigned int
execute_optimize_widening_mul (void)
{
- bool changed = false;
basic_block bb;
FOR_EACH_BB (bb)
{
gimple_stmt_iterator gsi;
! for (gsi = gsi_after_labels (bb); !gsi_end_p (gsi); gsi_next (&gsi))
{
gimple stmt = gsi_stmt (gsi);
enum tree_code code;
! if (!is_gimple_assign (stmt))
! continue;
!
! code = gimple_assign_rhs_code (stmt);
! if (code == MULT_EXPR)
! changed |= convert_mult_to_widen (stmt);
! else if (code == PLUS_EXPR || code == MINUS_EXPR)
! changed |= convert_plusminus_to_widen (&gsi, stmt, code);
}
}
! return (changed ? TODO_dump_func | TODO_update_ssa | TODO_verify_ssa
! | TODO_verify_stmts : 0);
}
static bool
--- 1618,1662 ----
static unsigned int
execute_optimize_widening_mul (void)
{
basic_block bb;
FOR_EACH_BB (bb)
{
gimple_stmt_iterator gsi;
! for (gsi = gsi_after_labels (bb); !gsi_end_p (gsi);)
{
gimple stmt = gsi_stmt (gsi);
enum tree_code code;
! if (is_gimple_assign (stmt))
! {
! code = gimple_assign_rhs_code (stmt);
! switch (code)
! {
! case MULT_EXPR:
! if (!convert_mult_to_widen (stmt)
! && convert_mult_to_fma (stmt))
! {
! gsi_remove (&gsi, true);
! release_defs (stmt);
! continue;
! }
! break;
!
! case PLUS_EXPR:
! case MINUS_EXPR:
! convert_plusminus_to_widen (&gsi, stmt, code);
! break;
!
! default:;
! }
! }
! gsi_next (&gsi);
}
}
! return 0;
}
static bool
*************** struct gimple_opt_pass pass_optimize_wid
*** 1549,1554 ****
0, /* properties_provided */
0, /* properties_destroyed */
0, /* todo_flags_start */
! 0 /* todo_flags_finish */
}
};
--- 1680,1688 ----
0, /* properties_provided */
0, /* properties_destroyed */
0, /* todo_flags_start */
! TODO_verify_ssa
! | TODO_verify_stmts
! | TODO_dump_func
! | TODO_update_ssa /* todo_flags_finish */
}
};
Index: gcc/testsuite/gcc.target/i386/fma4-vector-2.c
===================================================================
*** /dev/null 1970-01-01 00:00:00.000000000 +0000
--- gcc/testsuite/gcc.target/i386/fma4-vector-2.c 2010-11-02 12:08:20.000000000 +0100
***************
*** 0 ****
--- 1,21 ----
+ /* { dg-do compile } */
+ /* { dg-require-effective-target lp64 } */
+ /* { dg-options "-O2 -mfma4 -ftree-vectorize -mtune=generic" } */
+
+ float r[256], s[256];
+ float x[256];
+ float y[256];
+ float z[256];
+
+ void foo (void)
+ {
+ int i;
+ for (i = 0; i < 256; ++i)
+ {
+ r[i] = x[i] * y[i] - z[i];
+ s[i] = x[i] * y[i] + z[i];
+ }
+ }
+
+ /* { dg-final { scan-assembler "vfmaddps" } } */
+ /* { dg-final { scan-assembler "vfmsubps" } } */
Index: gcc/common.opt
===================================================================
*** gcc/common.opt.orig 2010-11-02 11:16:39.000000000 +0100
--- gcc/common.opt 2010-11-02 12:45:41.000000000 +0100
*************** bool flag_warn_unused_result = false
*** 58,63 ****
--- 58,67 ----
Variable
int *param_values
+ ; Floating-point contraction mode, fast by default.
+ Variable
+ enum fp_contract_mode flag_fp_contract_mode = FP_CONTRACT_FAST
+
###
Driver
*************** fforward-propagate
*** 857,862 ****
--- 861,870 ----
Common Report Var(flag_forward_propagate) Optimization
Perform a forward propagation pass on RTL
+ ffp-contract=
+ Common Joined RejectNegative
+ -ffp-contract=[off|on|fast] Perform floating-point expression contraction.
+
; Nonzero means don't put addresses of constant functions in registers.
; Used for compiling the Unix kernel, where strange substitutions are
; done on the assembly output.
Index: gcc/doc/invoke.texi
===================================================================
*** gcc/doc/invoke.texi.orig 2010-11-02 11:16:10.000000000 +0100
--- gcc/doc/invoke.texi 2010-11-02 12:14:37.000000000 +0100
*************** Objective-C and Objective-C++ Dialects}.
*** 342,348 ****
-fdelayed-branch -fdelete-null-pointer-checks -fdse -fdse @gol
-fearly-inlining -fipa-sra -fexpensive-optimizations -ffast-math @gol
-ffinite-math-only -ffloat-store -fexcess-precision=@var{style} @gol
! -fforward-propagate -ffunction-sections @gol
-fgcse -fgcse-after-reload -fgcse-las -fgcse-lm -fgraphite-identity @gol
-fgcse-sm -fif-conversion -fif-conversion2 -findirect-inlining @gol
-finline-functions -finline-functions-called-once -finline-limit=@var{n} @gol
--- 342,348 ----
-fdelayed-branch -fdelete-null-pointer-checks -fdse -fdse @gol
-fearly-inlining -fipa-sra -fexpensive-optimizations -ffast-math @gol
-ffinite-math-only -ffloat-store -fexcess-precision=@var{style} @gol
! -fforward-propagate -ffp-contract=@var{style} -ffunction-sections @gol
-fgcse -fgcse-after-reload -fgcse-las -fgcse-lm -fgraphite-identity @gol
-fgcse-sm -fif-conversion -fif-conversion2 -findirect-inlining @gol
-finline-functions -finline-functions-called-once -finline-limit=@var{n} @gol
*************** loop unrolling.
*** 5980,5985 ****
--- 5980,5997 ----
This option is enabled by default at optimization levels @option{-O},
@option{-O2}, @option{-O3}, @option{-Os}.
+ @item -ffp-contract=@var{style}
+ @opindex ffp-contract
+ @option{-ffp-contract=off} disables floating-point expression contraction.
+ @option{-ffp-contract=fast} enables floating-point expression contraction
+ such as forming of fused multiply-add operations if the target has
+ native support for them.
+ @option{-ffp-contract=on} enables floating-point expression contraction
+ if allowed by the language standard. This is currently not implemented
+ and treated equal to @option{-ffp-contract=off}.
+
+ The default is @option{-ffp-contract=fast}.
+
@item -fomit-frame-pointer
@opindex fomit-frame-pointer
Don't keep the frame pointer in a register for functions that
*************** an exact implementation of IEEE or ISO r
*** 7816,7822 ****
math functions. It may, however, yield faster code for programs
that do not require the guarantees of these specifications.
Enables @option{-fno-signed-zeros}, @option{-fno-trapping-math},
! @option{-fassociative-math} and @option{-freciprocal-math}.
The default is @option{-fno-unsafe-math-optimizations}.
--- 7828,7835 ----
math functions. It may, however, yield faster code for programs
that do not require the guarantees of these specifications.
Enables @option{-fno-signed-zeros}, @option{-fno-trapping-math},
! @option{-fassociative-math}, @option{-freciprocal-math} and
! @option{-ffp-contract=fast}.
The default is @option{-fno-unsafe-math-optimizations}.
Index: gcc/opts.c
===================================================================
*** gcc/opts.c.orig 2010-10-22 17:06:23.000000000 +0200
--- gcc/opts.c 2010-11-02 12:45:34.000000000 +0100
*************** common_handle_option (struct gcc_options
*** 1901,1906 ****
--- 1901,1918 ----
return false;
break;
+ case OPT_ffp_contract_:
+ if (!strcmp (arg, "on"))
+ /* Not implemented, fall back to conservative FP_CONTRACT_OFF. */
+ flag_fp_contract_mode = FP_CONTRACT_OFF;
+ else if (!strcmp (arg, "off"))
+ flag_fp_contract_mode = FP_CONTRACT_OFF;
+ else if (!strcmp (arg, "fast"))
+ flag_fp_contract_mode = FP_CONTRACT_FAST;
+ else
+ error ("unknown floating point contraction style \"%s\"", arg);
+ break;
+
case OPT_fexcess_precision_:
if (!strcmp (arg, "fast"))
flag_excess_precision_cmdline = EXCESS_PRECISION_FAST;
*************** set_unsafe_math_optimizations_flags (int
*** 2289,2294 ****
--- 2301,2307 ----
flag_signed_zeros = !set;
flag_associative_math = set;
flag_reciprocal_math = set;
+ flag_fp_contract_mode = set ? FP_CONTRACT_FAST : FP_CONTRACT_OFF;
}
/* Return true iff flags are set as if -ffast-math. */
*************** fast_math_flags_set_p (void)
*** 2299,2305 ****
&& flag_unsafe_math_optimizations
&& flag_finite_math_only
&& !flag_signed_zeros
! && !flag_errno_math);
}
/* Return true iff flags are set as if -ffast-math but using the flags stored
--- 2312,2319 ----
&& flag_unsafe_math_optimizations
&& flag_finite_math_only
&& !flag_signed_zeros
! && !flag_errno_math
! && flag_fp_contract_mode == FP_CONTRACT_FAST);
}
/* Return true iff flags are set as if -ffast-math but using the flags stored
Index: gcc/builtins.c
===================================================================
*** gcc/builtins.c.orig 2010-11-02 11:16:39.000000000 +0100
--- gcc/builtins.c 2010-11-02 12:19:44.000000000 +0100
*************** fold_builtin_abs (location_t loc, tree a
*** 9262,9267 ****
--- 9262,9288 ----
return fold_build1_loc (loc, ABS_EXPR, type, arg);
}
+ /* Fold a call to fma, fmaf, or fmal with arguments ARG[012]. */
+
+ static tree
+ fold_builtin_fma (location_t loc, tree arg0, tree arg1, tree arg2, tree type)
+ {
+ if (validate_arg (arg0, REAL_TYPE)
+ && validate_arg(arg1, REAL_TYPE)
+ && validate_arg(arg2, REAL_TYPE))
+ {
+ if (TREE_CODE (arg0) == REAL_CST
+ && TREE_CODE (arg1) == REAL_CST
+ && TREE_CODE (arg2) == REAL_CST)
+ return do_mpfr_arg3 (arg0, arg1, arg2, type, mpfr_fma);
+
+ /* ??? Only expand to FMA_EXPR if it's directly supported. */
+ if (optab_handler (fma_optab, TYPE_MODE (type)) != CODE_FOR_nothing)
+ return fold_build3_loc (loc, FMA_EXPR, type, arg0, arg1, arg2);
+ }
+ return NULL_TREE;
+ }
+
/* Fold a call to builtin fmin or fmax. */
static tree
*************** fold_builtin_3 (location_t loc, tree fnd
*** 10536,10545 ****
return fold_builtin_sincos (loc, arg0, arg1, arg2);
CASE_FLT_FN (BUILT_IN_FMA):
! if (validate_arg (arg0, REAL_TYPE)
! && validate_arg(arg1, REAL_TYPE)
! && validate_arg(arg2, REAL_TYPE))
! return do_mpfr_arg3 (arg0, arg1, arg2, type, mpfr_fma);
break;
CASE_FLT_FN (BUILT_IN_REMQUO):
--- 10557,10563 ----
return fold_builtin_sincos (loc, arg0, arg1, arg2);
CASE_FLT_FN (BUILT_IN_FMA):
! return fold_builtin_fma (loc, arg0, arg1, arg2, type);
break;
CASE_FLT_FN (BUILT_IN_REMQUO):
Index: gcc/config/i386/sse.md
===================================================================
*** gcc/config/i386/sse.md.orig 2010-11-02 11:16:36.000000000 +0100
--- gcc/config/i386/sse.md 2010-11-02 12:18:39.000000000 +0100
***************
*** 1859,1865 ****
;; Intrinsic FMA operations.
! ;; The standard name for fma is only available with SSE math enabled.
(define_expand "fma<mode>4"
[(set (match_operand:FMAMODE 0 "register_operand")
(fma:FMAMODE
--- 1859,1865 ----
;; Intrinsic FMA operations.
! ;; The standard names for fma is only available with SSE math enabled.
(define_expand "fma<mode>4"
[(set (match_operand:FMAMODE 0 "register_operand")
(fma:FMAMODE
***************
*** 1869,1874 ****
--- 1869,1901 ----
"(TARGET_FMA || TARGET_FMA4) && TARGET_SSE_MATH"
"")
+ (define_expand "fms<mode>4"
+ [(set (match_operand:FMAMODE 0 "register_operand")
+ (fma:FMAMODE
+ (match_operand:FMAMODE 1 "nonimmediate_operand")
+ (match_operand:FMAMODE 2 "nonimmediate_operand")
+ (neg:FMAMODE (match_operand:FMAMODE 3 "nonimmediate_operand"))))]
+ "(TARGET_FMA || TARGET_FMA4) && TARGET_SSE_MATH"
+ "")
+
+ (define_expand "fnma<mode>4"
+ [(set (match_operand:FMAMODE 0 "register_operand")
+ (fma:FMAMODE
+ (neg:FMAMODE (match_operand:FMAMODE 1 "nonimmediate_operand"))
+ (match_operand:FMAMODE 2 "nonimmediate_operand")
+ (match_operand:FMAMODE 3 "nonimmediate_operand")))]
+ "(TARGET_FMA || TARGET_FMA4) && TARGET_SSE_MATH"
+ "")
+
+ (define_expand "fnms<mode>4"
+ [(set (match_operand:FMAMODE 0 "register_operand")
+ (fma:FMAMODE
+ (neg:FMAMODE (match_operand:FMAMODE 1 "nonimmediate_operand"))
+ (match_operand:FMAMODE 2 "nonimmediate_operand")
+ (neg:FMAMODE (match_operand:FMAMODE 3 "nonimmediate_operand"))))]
+ "(TARGET_FMA || TARGET_FMA4) && TARGET_SSE_MATH"
+ "")
+
;; The builtin for fma4intrin.h is not constrained by SSE math enabled.
(define_expand "fma4i_fmadd_<mode>"
[(set (match_operand:FMAMODE 0 "register_operand")
Index: gcc/doc/md.texi
===================================================================
*** gcc/doc/md.texi.orig 2010-10-18 11:11:47.000000000 +0200
--- gcc/doc/md.texi 2010-11-02 12:18:39.000000000 +0100
*************** pattern is used to implement the @code{f
*** 3958,3963 ****
--- 3958,3993 ----
multiply followed by the add if the machine does not perform a
rounding step between the operations.
+ @cindex @code{fms@var{m}4} instruction pattern
+ @item @samp{fms@var{m}4}
+ Like @code{fma@var{m}4}, except operand 3 subtracted from the
+ product instead of added to the product. This is represented
+ in the rtl as
+
+ @smallexample
+ (fma:@var{m} @var{op1} @var{op2} (neg:@var{m} @var{op3}))
+ @end smallexample
+
+ @cindex @code{fnma@var{m}4} instruction pattern
+ @item @samp{fnma@var{m}4}
+ Like @code{fma@var{m}4} except that the intermediate product
+ is negated before being added to operand 3. This is represented
+ in the rtl as
+
+ @smallexample
+ (fma:@var{m} (neg:@var{m} @var{op1}) @var{op2} @var{op3})
+ @end smallexample
+
+ @cindex @code{fnms@var{m}4} instruction pattern
+ @item @samp{fnms@var{m}4}
+ Like @code{fms@var{m}4} except that the intermediate product
+ is negated before subtracting operand 3. This is represented
+ in the rtl as
+
+ @smallexample
+ (fma:@var{m} (neg:@var{m} @var{op1}) @var{op2} (neg:@var{m} @var{op3}))
+ @end smallexample
+
@cindex @code{min@var{m}3} instruction pattern
@cindex @code{max@var{m}3} instruction pattern
@item @samp{smin@var{m}3}, @samp{smax@var{m}3}
Index: gcc/flag-types.h
===================================================================
*** gcc/flag-types.h.orig 2010-10-12 14:57:13.000000000 +0200
--- gcc/flag-types.h 2010-11-02 12:12:45.000000000 +0100
*************** enum warn_strict_overflow_code
*** 152,155 ****
--- 152,162 ----
WARN_STRICT_OVERFLOW_MAGNITUDE = 5
};
+ /* Floating-point contraction mode. */
+ enum fp_contract_mode {
+ FP_CONTRACT_OFF = 0,
+ FP_CONTRACT_ON = 1,
+ FP_CONTRACT_FAST = 2
+ };
+
#endif /* ! GCC_FLAG_TYPES_H */
Index: gcc/genopinit.c
===================================================================
*** gcc/genopinit.c.orig 2010-10-18 11:11:51.000000000 +0200
--- gcc/genopinit.c 2010-11-02 12:18:39.000000000 +0100
*************** static const char * const optabs[] =
*** 160,165 ****
--- 160,168 ----
"set_optab_handler (floor_optab, $A, CODE_FOR_$(floor$a2$))",
"set_convert_optab_handler (lfloor_optab, $B, $A, CODE_FOR_$(lfloor$F$a$I$b2$))",
"set_optab_handler (fma_optab, $A, CODE_FOR_$(fma$a4$))",
+ "set_optab_handler (fms_optab, $A, CODE_FOR_$(fms$a4$))",
+ "set_optab_handler (fnma_optab, $A, CODE_FOR_$(fnma$a4$))",
+ "set_optab_handler (fnms_optab, $A, CODE_FOR_$(fnms$a4$))",
"set_optab_handler (ceil_optab, $A, CODE_FOR_$(ceil$a2$))",
"set_convert_optab_handler (lceil_optab, $B, $A, CODE_FOR_$(lceil$F$a$I$b2$))",
"set_optab_handler (round_optab, $A, CODE_FOR_$(round$a2$))",
Index: gcc/gimplify.c
===================================================================
*** gcc/gimplify.c.orig 2010-11-02 11:16:39.000000000 +0100
--- gcc/gimplify.c 2010-11-02 12:19:44.000000000 +0100
*************** gimplify_expr (tree *expr_p, gimple_seq
*** 7170,7175 ****
--- 7170,7185 ----
ret = gimplify_omp_atomic (expr_p, pre_p);
break;
+ case TRUTH_AND_EXPR:
+ case TRUTH_OR_EXPR:
+ case TRUTH_XOR_EXPR:
+ /* Classified as tcc_expression. */
+ goto expr_2;
+
+ case FMA_EXPR:
+ /* Classified as tcc_expression. */
+ goto expr_3;
+
case POINTER_PLUS_EXPR:
/* Convert ((type *)A)+offset into &A->field_of_type_and_offset.
The second is gimple immediate saving a need for extra statement.
*************** gimplify_expr (tree *expr_p, gimple_seq
*** 7249,7264 ****
break;
}
case tcc_declaration:
case tcc_constant:
ret = GS_ALL_DONE;
goto dont_recalculate;
default:
! gcc_assert (TREE_CODE (*expr_p) == TRUTH_AND_EXPR
! || TREE_CODE (*expr_p) == TRUTH_OR_EXPR
! || TREE_CODE (*expr_p) == TRUTH_XOR_EXPR);
! goto expr_2;
}
recalculate_side_effects (*expr_p);
--- 7259,7286 ----
break;
}
+ expr_3:
+ {
+ enum gimplify_status r0, r1, r2;
+
+ r0 = gimplify_expr (&TREE_OPERAND (*expr_p, 0), pre_p,
+ post_p, is_gimple_val, fb_rvalue);
+ r1 = gimplify_expr (&TREE_OPERAND (*expr_p, 1), pre_p,
+ post_p, is_gimple_val, fb_rvalue);
+ r2 = gimplify_expr (&TREE_OPERAND (*expr_p, 2), pre_p,
+ post_p, is_gimple_val, fb_rvalue);
+
+ ret = MIN (MIN (r0, r1), r2);
+ break;
+ }
+
case tcc_declaration:
case tcc_constant:
ret = GS_ALL_DONE;
goto dont_recalculate;
default:
! gcc_unreachable ();
}
recalculate_side_effects (*expr_p);
Index: gcc/optabs.h
===================================================================
*** gcc/optabs.h.orig 2010-10-18 11:11:51.000000000 +0200
--- gcc/optabs.h 2010-11-02 12:18:39.000000000 +0100
*************** enum optab_index
*** 192,197 ****
--- 192,200 ----
OTI_atan2,
/* Floating multiply/add */
OTI_fma,
+ OTI_fms,
+ OTI_fnma,
+ OTI_fnms,
/* Move instruction. */
OTI_mov,
*************** enum optab_index
*** 435,440 ****
--- 438,446 ----
#define pow_optab (&optab_table[OTI_pow])
#define atan2_optab (&optab_table[OTI_atan2])
#define fma_optab (&optab_table[OTI_fma])
+ #define fms_optab (&optab_table[OTI_fms])
+ #define fnma_optab (&optab_table[OTI_fnma])
+ #define fnms_optab (&optab_table[OTI_fnms])
#define mov_optab (&optab_table[OTI_mov])
#define movstrict_optab (&optab_table[OTI_movstrict])
^ permalink raw reply [flat|nested] 12+ messages in thread* Re: [PATCH] FMA on trees 2010-11-02 13:35 [PATCH] FMA on trees Richard Guenther @ 2010-11-02 15:40 ` Richard Henderson 2010-11-03 13:13 ` Richard Guenther 0 siblings, 1 reply; 12+ messages in thread From: Richard Henderson @ 2010-11-02 15:40 UTC (permalink / raw) To: Richard Guenther; +Cc: gcc-patches On 11/02/2010 06:24 AM, Richard Guenther wrote: > *************** set_unsafe_math_optimizations_flags (int > *** 2289,2294 **** > --- 2301,2307 ---- > flag_signed_zeros = !set; > flag_associative_math = set; > flag_reciprocal_math = set; > + flag_fp_contract_mode = set ? FP_CONTRACT_FAST : FP_CONTRACT_OFF; Off? Default is FAST anyway. I can't think what to set here for -fno-fast-math, actually. > + fold_builtin_fma (location_t loc, tree arg0, tree arg1, tree arg2, tree type) > + { > + if (validate_arg (arg0, REAL_TYPE) > + && validate_arg(arg1, REAL_TYPE) > + && validate_arg(arg2, REAL_TYPE)) > + { > + if (TREE_CODE (arg0) == REAL_CST > + && TREE_CODE (arg1) == REAL_CST > + && TREE_CODE (arg2) == REAL_CST) > + return do_mpfr_arg3 (arg0, arg1, arg2, type, mpfr_fma); > + > + /* ??? Only expand to FMA_EXPR if it's directly supported. */ > + if (optab_handler (fma_optab, TYPE_MODE (type)) != CODE_FOR_nothing) > + return fold_build3_loc (loc, FMA_EXPR, type, arg0, arg1, arg2); I know this is my bit, but I just remembered that this means we need a new entry in fold_ternary_loc to fold FMA_EXPR. Finally, if !flag_signed_zeros (part of fast-math), we probably want to fold -fma(a,b,c) -> fma(-a,b,c). r~ ^ permalink raw reply [flat|nested] 12+ messages in thread
* Re: [PATCH] FMA on trees 2010-11-02 15:40 ` Richard Henderson @ 2010-11-03 13:13 ` Richard Guenther 2010-11-03 15:30 ` Richard Guenther 0 siblings, 1 reply; 12+ messages in thread From: Richard Guenther @ 2010-11-03 13:13 UTC (permalink / raw) To: Richard Henderson; +Cc: gcc-patches On Tue, 2 Nov 2010, Richard Henderson wrote: > On 11/02/2010 06:24 AM, Richard Guenther wrote: > > *************** set_unsafe_math_optimizations_flags (int > > *** 2289,2294 **** > > --- 2301,2307 ---- > > flag_signed_zeros = !set; > > flag_associative_math = set; > > flag_reciprocal_math = set; > > + flag_fp_contract_mode = set ? FP_CONTRACT_FAST : FP_CONTRACT_OFF; > > Off? Default is FAST anyway. I can't think what to set here > for -fno-fast-math, actually. Yeah, I wondered about that, too. I've now just removed -ffp-contract from unsafe-math handling completely. We can revisit that if we change the default and/or implement ON. > > + fold_builtin_fma (location_t loc, tree arg0, tree arg1, tree arg2, tree type) > > + { > > + if (validate_arg (arg0, REAL_TYPE) > > + && validate_arg(arg1, REAL_TYPE) > > + && validate_arg(arg2, REAL_TYPE)) > > + { > > + if (TREE_CODE (arg0) == REAL_CST > > + && TREE_CODE (arg1) == REAL_CST > > + && TREE_CODE (arg2) == REAL_CST) > > + return do_mpfr_arg3 (arg0, arg1, arg2, type, mpfr_fma); > > + > > + /* ??? Only expand to FMA_EXPR if it's directly supported. */ > > + if (optab_handler (fma_optab, TYPE_MODE (type)) != CODE_FOR_nothing) > > + return fold_build3_loc (loc, FMA_EXPR, type, arg0, arg1, arg2); > > I know this is my bit, but I just remembered that this means we > need a new entry in fold_ternary_loc to fold FMA_EXPR. Done. > Finally, if !flag_signed_zeros (part of fast-math), we probably > want to fold -fma(a,b,c) -> fma(-a,b,c). Hm, that doesn't sound correct to me ;) The following adds vectorization capabilities as well which fixes the seen regressions. Bootstrap and regtest pending. Richard. 2010-10-22 Richard Guenther <rguenther@suse.de> Richard Henderson <rth@redhat.com> * tree.def (FMA_EXPR): New tree code. * expr.c (expand_expr_real_2): Add FMA_EXPR expansion code. * gimple.c (gimple_rhs_class_table): FMA_EXPR is a GIMPLE_TERNARY_RHS. * tree-cfg.c (verify_gimple_assign_ternary): Verify FMA_EXPR types. * tree-inline.c (estimate_operator_cost): Handle FMA_EXPR. * gimple-pretty-print.c (dump_ternary_rhs): Likewise. * tree-ssa-math-opts.c (convert_mult_to_fma): New function. (execute_optimize_widening_mul): Call it. Reorganize to allow dead stmt removal. Move TODO flags ... (pass_optimize_widening_mul): ... here. * flag-types.h (enum fp_contract_mode): New enum. * common.opt (flag_fp_contract_mode): New variable. (-ffp-contract): New option. * opts.c (common_handle_option): Handle it. * doc/invoke.texi (-ffp-contract): Document. * tree.h (fold_fma): Declare. * builtins.c (fold_fma): New function. (fold_builtin_fma): Likewise. (fold_builtin_3): Call it for fma. * fold-const.c (fold_ternary_loc): Fold FMA_EXPR. * optabs.c (optab_for_tree_code): Handle FMA_EXPR. * config/i386/sse.md (fms<mode>4, fnma<mode>, fnms<mode>4): New expanders. * doc/md.texi (fms<mode>4, fnma<mode>, fnms<mode>4): Document new named patterns. * genopinit.c (optabs): Initialize fms_optab, fnma_optab and fnms_optab. * optabs.h (enum optab_index): Add OTI_fms, OTI_fnma and OTI_fnms. (fms_optab, fnma_optab, fnms_optab): New defines. * gimplify.c (gimplify_expr): Handle binary truth expressions explicitly. Handle FMA_EXPR. * tree-vect-stmts.c (vectorizable_operation): Handle ternary operations. * gcc.target/i386/fma4-vector-2.c: New testcase. Index: gcc/tree.def =================================================================== *** gcc/tree.def.orig 2010-11-02 16:37:35.000000000 +0100 --- gcc/tree.def 2010-11-03 13:21:19.000000000 +0100 *************** DEFTREECODE (WIDEN_MULT_PLUS_EXPR, "wide *** 1092,1097 **** --- 1092,1103 ---- is subtracted from t3. */ DEFTREECODE (WIDEN_MULT_MINUS_EXPR, "widen_mult_plus_expr", tcc_expression, 3) + /* Fused multiply-add. + All operands and the result are of the same type. No intermediate + rounding is performed after multiplying operand one with operand two + before adding operand three. */ + DEFTREECODE (FMA_EXPR, "fma_expr", tcc_expression, 3) + /* Whole vector left/right shift in bits. Operand 0 is a vector to be shifted. Operand 1 is an integer shift amount in bits. */ Index: gcc/expr.c =================================================================== *** gcc/expr.c.orig 2010-11-02 16:37:35.000000000 +0100 --- gcc/expr.c 2010-11-03 13:21:19.000000000 +0100 *************** expand_expr_real_2 (sepops ops, rtx targ *** 7254,7260 **** int ignore; bool reduce_bit_field; location_t loc = ops->location; ! tree treeop0, treeop1; #define REDUCE_BIT_FIELD(expr) (reduce_bit_field \ ? reduce_to_bit_field_precision ((expr), \ target, \ --- 7254,7260 ---- int ignore; bool reduce_bit_field; location_t loc = ops->location; ! tree treeop0, treeop1, treeop2; #define REDUCE_BIT_FIELD(expr) (reduce_bit_field \ ? reduce_to_bit_field_precision ((expr), \ target, \ *************** expand_expr_real_2 (sepops ops, rtx targ *** 7267,7272 **** --- 7267,7273 ---- treeop0 = ops->op0; treeop1 = ops->op1; + treeop2 = ops->op2; /* We should be called only on simple (binary or unary) expressions, exactly those that are valid in gimple expressions that aren't *************** expand_expr_real_2 (sepops ops, rtx targ *** 7624,7630 **** case WIDEN_MULT_PLUS_EXPR: case WIDEN_MULT_MINUS_EXPR: expand_operands (treeop0, treeop1, NULL_RTX, &op0, &op1, EXPAND_NORMAL); ! op2 = expand_normal (ops->op2); target = expand_widen_pattern_expr (ops, op0, op1, op2, target, unsignedp); return target; --- 7625,7631 ---- case WIDEN_MULT_PLUS_EXPR: case WIDEN_MULT_MINUS_EXPR: expand_operands (treeop0, treeop1, NULL_RTX, &op0, &op1, EXPAND_NORMAL); ! op2 = expand_normal (treeop2); target = expand_widen_pattern_expr (ops, op0, op1, op2, target, unsignedp); return target; *************** expand_expr_real_2 (sepops ops, rtx targ *** 7711,7716 **** --- 7712,7757 ---- expand_operands (treeop0, treeop1, subtarget, &op0, &op1, EXPAND_NORMAL); return REDUCE_BIT_FIELD (expand_mult (mode, op0, op1, target, unsignedp)); + case FMA_EXPR: + { + optab opt = fma_optab; + gimple def0, def2; + + def0 = get_def_for_expr (treeop0, NEGATE_EXPR); + def2 = get_def_for_expr (treeop2, NEGATE_EXPR); + + op0 = op2 = NULL; + + if (def0 && def2 + && optab_handler (fnms_optab, mode) != CODE_FOR_nothing) + { + opt = fnms_optab; + op0 = expand_normal (gimple_assign_rhs1 (def0)); + op2 = expand_normal (gimple_assign_rhs1 (def2)); + } + else if (def0 + && optab_handler (fnma_optab, mode) != CODE_FOR_nothing) + { + opt = fnma_optab; + op0 = expand_normal (gimple_assign_rhs1 (def0)); + } + else if (def2 + && optab_handler (fms_optab, mode) != CODE_FOR_nothing) + { + opt = fms_optab; + op2 = expand_normal (gimple_assign_rhs1 (def2)); + } + + if (op0 == NULL) + op0 = expand_expr (treeop0, subtarget, VOIDmode, EXPAND_NORMAL); + if (op2 == NULL) + op2 = expand_normal (treeop2); + op1 = expand_normal (treeop1); + + return expand_ternary_op (TYPE_MODE (type), opt, + op0, op1, op2, target, 0); + } + case MULT_EXPR: /* If this is a fixed-point operation, then we cannot use the code below because "expand_mult" doesn't support sat/no-sat fixed-point Index: gcc/gimple.c =================================================================== *** gcc/gimple.c.orig 2010-11-03 10:55:43.000000000 +0100 --- gcc/gimple.c 2010-11-03 13:21:19.000000000 +0100 *************** get_gimple_rhs_num_ops (enum tree_code c *** 2529,2535 **** || (SYM) == TRUTH_XOR_EXPR) ? GIMPLE_BINARY_RHS \ : (SYM) == TRUTH_NOT_EXPR ? GIMPLE_UNARY_RHS \ : ((SYM) == WIDEN_MULT_PLUS_EXPR \ ! || (SYM) == WIDEN_MULT_MINUS_EXPR) ? GIMPLE_TERNARY_RHS \ : ((SYM) == COND_EXPR \ || (SYM) == CONSTRUCTOR \ || (SYM) == OBJ_TYPE_REF \ --- 2529,2536 ---- || (SYM) == TRUTH_XOR_EXPR) ? GIMPLE_BINARY_RHS \ : (SYM) == TRUTH_NOT_EXPR ? GIMPLE_UNARY_RHS \ : ((SYM) == WIDEN_MULT_PLUS_EXPR \ ! || (SYM) == WIDEN_MULT_MINUS_EXPR \ ! || (SYM) == FMA_EXPR) ? GIMPLE_TERNARY_RHS \ : ((SYM) == COND_EXPR \ || (SYM) == CONSTRUCTOR \ || (SYM) == OBJ_TYPE_REF \ Index: gcc/tree-cfg.c =================================================================== *** gcc/tree-cfg.c.orig 2010-11-02 16:37:53.000000000 +0100 --- gcc/tree-cfg.c 2010-11-03 13:21:19.000000000 +0100 *************** verify_gimple_assign_ternary (gimple stm *** 3748,3753 **** --- 3748,3767 ---- } break; + case FMA_EXPR: + if (!useless_type_conversion_p (lhs_type, rhs1_type) + || !useless_type_conversion_p (lhs_type, rhs2_type) + || !useless_type_conversion_p (lhs_type, rhs3_type)) + { + error ("type mismatch in fused multiply-add expression"); + debug_generic_expr (lhs_type); + debug_generic_expr (rhs1_type); + debug_generic_expr (rhs2_type); + debug_generic_expr (rhs3_type); + return true; + } + break; + default: gcc_unreachable (); } Index: gcc/tree-inline.c =================================================================== *** gcc/tree-inline.c.orig 2010-11-03 10:55:43.000000000 +0100 --- gcc/tree-inline.c 2010-11-03 13:21:19.000000000 +0100 *************** estimate_operator_cost (enum tree_code c *** 3283,3288 **** --- 3283,3289 ---- case POINTER_PLUS_EXPR: case MINUS_EXPR: case MULT_EXPR: + case FMA_EXPR: case ADDR_SPACE_CONVERT_EXPR: case FIXED_CONVERT_EXPR: Index: gcc/gimple-pretty-print.c =================================================================== *** gcc/gimple-pretty-print.c.orig 2010-11-02 16:37:35.000000000 +0100 --- gcc/gimple-pretty-print.c 2010-11-03 13:21:19.000000000 +0100 *************** dump_ternary_rhs (pretty_printer *buffer *** 400,405 **** --- 400,413 ---- pp_character (buffer, '>'); break; + case FMA_EXPR: + dump_generic_node (buffer, gimple_assign_rhs1 (gs), spc, flags, false); + pp_string (buffer, " * "); + dump_generic_node (buffer, gimple_assign_rhs2 (gs), spc, flags, false); + pp_string (buffer, " + "); + dump_generic_node (buffer, gimple_assign_rhs3 (gs), spc, flags, false); + break; + default: gcc_unreachable (); } Index: gcc/tree-ssa-math-opts.c =================================================================== *** gcc/tree-ssa-math-opts.c.orig 2010-11-02 16:37:35.000000000 +0100 --- gcc/tree-ssa-math-opts.c 2010-11-03 13:21:19.000000000 +0100 *************** convert_plusminus_to_widen (gimple_stmt_ *** 1494,1499 **** --- 1494,1616 ---- return true; } + /* Combine the multiplication at MUL_STMT with uses in additions and + subtractions to form fused multiply-add operations. Returns true + if successful and MUL_STMT should be removed. */ + + static bool + convert_mult_to_fma (gimple mul_stmt) + { + tree mul_result = gimple_assign_lhs (mul_stmt); + tree type = TREE_TYPE (mul_result); + gimple use_stmt, fma_stmt; + use_operand_p use_p; + imm_use_iterator imm_iter; + + if (FLOAT_TYPE_P (type) + && flag_fp_contract_mode == FP_CONTRACT_OFF) + return false; + + /* We don't want to do bitfield reduction ops. */ + if (INTEGRAL_TYPE_P (type) + && (TYPE_PRECISION (type) + != GET_MODE_PRECISION (TYPE_MODE (type)))) + return false; + + /* If the target doesn't support it, don't generate it. We assume that + if fma isn't available then fms, fnma or fnms are not either. */ + if (optab_handler (fma_optab, TYPE_MODE (type)) == CODE_FOR_nothing) + return false; + + /* Make sure that the multiplication statement becomes dead after + the transformation, thus that all uses are transformed to FMAs. + This means we assume that an FMA operation has the same cost + as an addition. */ + FOR_EACH_IMM_USE_FAST (use_p, imm_iter, mul_result) + { + enum tree_code use_code; + + use_stmt = USE_STMT (use_p); + + if (!is_gimple_assign (use_stmt)) + return false; + use_code = gimple_assign_rhs_code (use_stmt); + /* ??? We need to handle NEGATE_EXPR to eventually form fnms. */ + if (use_code != PLUS_EXPR + && use_code != MINUS_EXPR) + return false; + + /* For now restrict this operations to single basic blocks. In theory + we would want to support sinking the multiplication in + m = a*b; + if () + ma = m + c; + else + d = m; + to form a fma in the then block and sink the multiplication to the + else block. */ + if (gimple_bb (use_stmt) != gimple_bb (mul_stmt)) + return false; + + /* We can't handle a * b + a * b. */ + if (gimple_assign_rhs1 (use_stmt) == gimple_assign_rhs2 (use_stmt)) + return false; + + /* If the target doesn't support a * b - c then drop the ball. */ + if (gimple_assign_rhs1 (use_stmt) == mul_result + && use_code == MINUS_EXPR + && optab_handler (fms_optab, TYPE_MODE (type)) == CODE_FOR_nothing) + return false; + + /* If the target doesn't support -a * b + c then drop the ball. */ + if (gimple_assign_rhs2 (use_stmt) == mul_result + && use_code == MINUS_EXPR + && optab_handler (fnma_optab, TYPE_MODE (type)) == CODE_FOR_nothing) + return false; + + /* We don't yet generate -a * b - c below yet. */ + } + + FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, mul_result) + { + tree addop, mulop1; + gimple_stmt_iterator gsi = gsi_for_stmt (use_stmt); + + mulop1 = gimple_assign_rhs1 (mul_stmt); + if (gimple_assign_rhs1 (use_stmt) == mul_result) + { + addop = gimple_assign_rhs2 (use_stmt); + /* a * b - c -> a * b + (-c) */ + if (gimple_assign_rhs_code (use_stmt) == MINUS_EXPR) + addop = force_gimple_operand_gsi (&gsi, + build1 (NEGATE_EXPR, + type, addop), + true, NULL_TREE, true, + GSI_SAME_STMT); + } + else + { + addop = gimple_assign_rhs1 (use_stmt); + /* a - b * c -> (-b) * c + a */ + if (gimple_assign_rhs_code (use_stmt) == MINUS_EXPR) + mulop1 = force_gimple_operand_gsi (&gsi, + build1 (NEGATE_EXPR, + type, mulop1), + true, NULL_TREE, true, + GSI_SAME_STMT); + } + + fma_stmt = gimple_build_assign_with_ops3 (FMA_EXPR, + gimple_assign_lhs (use_stmt), + mulop1, + gimple_assign_rhs2 (mul_stmt), + addop); + gsi_replace (&gsi, fma_stmt, true); + } + + return true; + } + /* Find integer multiplications where the operands are extended from smaller types, and replace the MULT_EXPR with a WIDEN_MULT_EXPR where appropriate. */ *************** convert_plusminus_to_widen (gimple_stmt_ *** 1501,1531 **** static unsigned int execute_optimize_widening_mul (void) { - bool changed = false; basic_block bb; FOR_EACH_BB (bb) { gimple_stmt_iterator gsi; ! for (gsi = gsi_after_labels (bb); !gsi_end_p (gsi); gsi_next (&gsi)) { gimple stmt = gsi_stmt (gsi); enum tree_code code; ! if (!is_gimple_assign (stmt)) ! continue; ! ! code = gimple_assign_rhs_code (stmt); ! if (code == MULT_EXPR) ! changed |= convert_mult_to_widen (stmt); ! else if (code == PLUS_EXPR || code == MINUS_EXPR) ! changed |= convert_plusminus_to_widen (&gsi, stmt, code); } } ! return (changed ? TODO_dump_func | TODO_update_ssa | TODO_verify_ssa ! | TODO_verify_stmts : 0); } static bool --- 1618,1662 ---- static unsigned int execute_optimize_widening_mul (void) { basic_block bb; FOR_EACH_BB (bb) { gimple_stmt_iterator gsi; ! for (gsi = gsi_after_labels (bb); !gsi_end_p (gsi);) { gimple stmt = gsi_stmt (gsi); enum tree_code code; ! if (is_gimple_assign (stmt)) ! { ! code = gimple_assign_rhs_code (stmt); ! switch (code) ! { ! case MULT_EXPR: ! if (!convert_mult_to_widen (stmt) ! && convert_mult_to_fma (stmt)) ! { ! gsi_remove (&gsi, true); ! release_defs (stmt); ! continue; ! } ! break; ! ! case PLUS_EXPR: ! case MINUS_EXPR: ! convert_plusminus_to_widen (&gsi, stmt, code); ! break; ! ! default:; ! } ! } ! gsi_next (&gsi); } } ! return 0; } static bool *************** struct gimple_opt_pass pass_optimize_wid *** 1549,1554 **** 0, /* properties_provided */ 0, /* properties_destroyed */ 0, /* todo_flags_start */ ! 0 /* todo_flags_finish */ } }; --- 1680,1688 ---- 0, /* properties_provided */ 0, /* properties_destroyed */ 0, /* todo_flags_start */ ! TODO_verify_ssa ! | TODO_verify_stmts ! | TODO_dump_func ! | TODO_update_ssa /* todo_flags_finish */ } }; Index: gcc/testsuite/gcc.target/i386/fma4-vector-2.c =================================================================== *** /dev/null 1970-01-01 00:00:00.000000000 +0000 --- gcc/testsuite/gcc.target/i386/fma4-vector-2.c 2010-11-03 13:21:19.000000000 +0100 *************** *** 0 **** --- 1,21 ---- + /* { dg-do compile } */ + /* { dg-require-effective-target lp64 } */ + /* { dg-options "-O2 -mfma4 -ftree-vectorize -mtune=generic" } */ + + float r[256], s[256]; + float x[256]; + float y[256]; + float z[256]; + + void foo (void) + { + int i; + for (i = 0; i < 256; ++i) + { + r[i] = x[i] * y[i] - z[i]; + s[i] = x[i] * y[i] + z[i]; + } + } + + /* { dg-final { scan-assembler "vfmaddps" } } */ + /* { dg-final { scan-assembler "vfmsubps" } } */ Index: gcc/common.opt =================================================================== *** gcc/common.opt.orig 2010-11-02 16:37:35.000000000 +0100 --- gcc/common.opt 2010-11-03 13:21:19.000000000 +0100 *************** bool flag_warn_unused_result = false *** 58,63 **** --- 58,67 ---- Variable int *param_values + ; Floating-point contraction mode, fast by default. + Variable + enum fp_contract_mode flag_fp_contract_mode = FP_CONTRACT_FAST + ### Driver *************** fforward-propagate *** 857,862 **** --- 861,870 ---- Common Report Var(flag_forward_propagate) Optimization Perform a forward propagation pass on RTL + ffp-contract= + Common Joined RejectNegative + -ffp-contract=[off|on|fast] Perform floating-point expression contraction. + ; Nonzero means don't put addresses of constant functions in registers. ; Used for compiling the Unix kernel, where strange substitutions are ; done on the assembly output. Index: gcc/doc/invoke.texi =================================================================== *** gcc/doc/invoke.texi.orig 2010-11-02 16:37:50.000000000 +0100 --- gcc/doc/invoke.texi 2010-11-03 13:21:19.000000000 +0100 *************** Objective-C and Objective-C++ Dialects}. *** 342,348 **** -fdelayed-branch -fdelete-null-pointer-checks -fdse -fdse @gol -fearly-inlining -fipa-sra -fexpensive-optimizations -ffast-math @gol -ffinite-math-only -ffloat-store -fexcess-precision=@var{style} @gol ! -fforward-propagate -ffunction-sections @gol -fgcse -fgcse-after-reload -fgcse-las -fgcse-lm -fgraphite-identity @gol -fgcse-sm -fif-conversion -fif-conversion2 -findirect-inlining @gol -finline-functions -finline-functions-called-once -finline-limit=@var{n} @gol --- 342,348 ---- -fdelayed-branch -fdelete-null-pointer-checks -fdse -fdse @gol -fearly-inlining -fipa-sra -fexpensive-optimizations -ffast-math @gol -ffinite-math-only -ffloat-store -fexcess-precision=@var{style} @gol ! -fforward-propagate -ffp-contract=@var{style} -ffunction-sections @gol -fgcse -fgcse-after-reload -fgcse-las -fgcse-lm -fgraphite-identity @gol -fgcse-sm -fif-conversion -fif-conversion2 -findirect-inlining @gol -finline-functions -finline-functions-called-once -finline-limit=@var{n} @gol *************** loop unrolling. *** 5980,5985 **** --- 5980,5997 ---- This option is enabled by default at optimization levels @option{-O}, @option{-O2}, @option{-O3}, @option{-Os}. + @item -ffp-contract=@var{style} + @opindex ffp-contract + @option{-ffp-contract=off} disables floating-point expression contraction. + @option{-ffp-contract=fast} enables floating-point expression contraction + such as forming of fused multiply-add operations if the target has + native support for them. + @option{-ffp-contract=on} enables floating-point expression contraction + if allowed by the language standard. This is currently not implemented + and treated equal to @option{-ffp-contract=off}. + + The default is @option{-ffp-contract=fast}. + @item -fomit-frame-pointer @opindex fomit-frame-pointer Don't keep the frame pointer in a register for functions that *************** an exact implementation of IEEE or ISO r *** 7816,7822 **** math functions. It may, however, yield faster code for programs that do not require the guarantees of these specifications. Enables @option{-fno-signed-zeros}, @option{-fno-trapping-math}, ! @option{-fassociative-math} and @option{-freciprocal-math}. The default is @option{-fno-unsafe-math-optimizations}. --- 7828,7835 ---- math functions. It may, however, yield faster code for programs that do not require the guarantees of these specifications. Enables @option{-fno-signed-zeros}, @option{-fno-trapping-math}, ! @option{-fassociative-math}, @option{-freciprocal-math} and ! @option{-ffp-contract=fast}. The default is @option{-fno-unsafe-math-optimizations}. Index: gcc/opts.c =================================================================== *** gcc/opts.c.orig 2010-11-02 16:37:35.000000000 +0100 --- gcc/opts.c 2010-11-03 13:43:06.000000000 +0100 *************** common_handle_option (struct gcc_options *** 1901,1906 **** --- 1901,1918 ---- return false; break; + case OPT_ffp_contract_: + if (!strcmp (arg, "on")) + /* Not implemented, fall back to conservative FP_CONTRACT_OFF. */ + flag_fp_contract_mode = FP_CONTRACT_OFF; + else if (!strcmp (arg, "off")) + flag_fp_contract_mode = FP_CONTRACT_OFF; + else if (!strcmp (arg, "fast")) + flag_fp_contract_mode = FP_CONTRACT_FAST; + else + error ("unknown floating point contraction style \"%s\"", arg); + break; + case OPT_fexcess_precision_: if (!strcmp (arg, "fast")) flag_excess_precision_cmdline = EXCESS_PRECISION_FAST; Index: gcc/builtins.c =================================================================== *** gcc/builtins.c.orig 2010-11-03 10:55:43.000000000 +0100 --- gcc/builtins.c 2010-11-03 13:34:38.000000000 +0100 *************** fold_builtin_abs (location_t loc, tree a *** 9266,9271 **** --- 9266,9304 ---- return fold_build1_loc (loc, ABS_EXPR, type, arg); } + /* Fold a fma operation with arguments ARG[012]. */ + + tree + fold_fma (location_t loc, tree type, tree arg0, tree arg1, tree arg2) + { + if (TREE_CODE (arg0) == REAL_CST + && TREE_CODE (arg1) == REAL_CST + && TREE_CODE (arg2) == REAL_CST) + return do_mpfr_arg3 (arg0, arg1, arg2, type, mpfr_fma); + + return NULL_TREE; + } + + /* Fold a call to fma, fmaf, or fmal with arguments ARG[012]. */ + + static tree + fold_builtin_fma (location_t loc, tree arg0, tree arg1, tree arg2, tree type) + { + if (validate_arg (arg0, REAL_TYPE) + && validate_arg(arg1, REAL_TYPE) + && validate_arg(arg2, REAL_TYPE)) + { + tree tem = fold_fma (loc, type, arg0, arg1, arg2); + if (tem) + return tem; + + /* ??? Only expand to FMA_EXPR if it's directly supported. */ + if (optab_handler (fma_optab, TYPE_MODE (type)) != CODE_FOR_nothing) + return fold_build3_loc (loc, FMA_EXPR, type, arg0, arg1, arg2); + } + return NULL_TREE; + } + /* Fold a call to builtin fmin or fmax. */ static tree *************** fold_builtin_3 (location_t loc, tree fnd *** 10540,10549 **** return fold_builtin_sincos (loc, arg0, arg1, arg2); CASE_FLT_FN (BUILT_IN_FMA): ! if (validate_arg (arg0, REAL_TYPE) ! && validate_arg(arg1, REAL_TYPE) ! && validate_arg(arg2, REAL_TYPE)) ! return do_mpfr_arg3 (arg0, arg1, arg2, type, mpfr_fma); break; CASE_FLT_FN (BUILT_IN_REMQUO): --- 10573,10579 ---- return fold_builtin_sincos (loc, arg0, arg1, arg2); CASE_FLT_FN (BUILT_IN_FMA): ! return fold_builtin_fma (loc, arg0, arg1, arg2, type); break; CASE_FLT_FN (BUILT_IN_REMQUO): Index: gcc/config/i386/sse.md =================================================================== *** gcc/config/i386/sse.md.orig 2010-11-02 16:37:35.000000000 +0100 --- gcc/config/i386/sse.md 2010-11-03 13:21:19.000000000 +0100 *************** *** 1859,1865 **** ;; Intrinsic FMA operations. ! ;; The standard name for fma is only available with SSE math enabled. (define_expand "fma<mode>4" [(set (match_operand:FMAMODE 0 "register_operand") (fma:FMAMODE --- 1859,1865 ---- ;; Intrinsic FMA operations. ! ;; The standard names for fma is only available with SSE math enabled. (define_expand "fma<mode>4" [(set (match_operand:FMAMODE 0 "register_operand") (fma:FMAMODE *************** *** 1869,1874 **** --- 1869,1901 ---- "(TARGET_FMA || TARGET_FMA4) && TARGET_SSE_MATH" "") + (define_expand "fms<mode>4" + [(set (match_operand:FMAMODE 0 "register_operand") + (fma:FMAMODE + (match_operand:FMAMODE 1 "nonimmediate_operand") + (match_operand:FMAMODE 2 "nonimmediate_operand") + (neg:FMAMODE (match_operand:FMAMODE 3 "nonimmediate_operand"))))] + "(TARGET_FMA || TARGET_FMA4) && TARGET_SSE_MATH" + "") + + (define_expand "fnma<mode>4" + [(set (match_operand:FMAMODE 0 "register_operand") + (fma:FMAMODE + (neg:FMAMODE (match_operand:FMAMODE 1 "nonimmediate_operand")) + (match_operand:FMAMODE 2 "nonimmediate_operand") + (match_operand:FMAMODE 3 "nonimmediate_operand")))] + "(TARGET_FMA || TARGET_FMA4) && TARGET_SSE_MATH" + "") + + (define_expand "fnms<mode>4" + [(set (match_operand:FMAMODE 0 "register_operand") + (fma:FMAMODE + (neg:FMAMODE (match_operand:FMAMODE 1 "nonimmediate_operand")) + (match_operand:FMAMODE 2 "nonimmediate_operand") + (neg:FMAMODE (match_operand:FMAMODE 3 "nonimmediate_operand"))))] + "(TARGET_FMA || TARGET_FMA4) && TARGET_SSE_MATH" + "") + ;; The builtin for fma4intrin.h is not constrained by SSE math enabled. (define_expand "fma4i_fmadd_<mode>" [(set (match_operand:FMAMODE 0 "register_operand") Index: gcc/doc/md.texi =================================================================== *** gcc/doc/md.texi.orig 2010-11-02 16:37:35.000000000 +0100 --- gcc/doc/md.texi 2010-11-03 13:21:19.000000000 +0100 *************** pattern is used to implement the @code{f *** 3958,3963 **** --- 3958,3993 ---- multiply followed by the add if the machine does not perform a rounding step between the operations. + @cindex @code{fms@var{m}4} instruction pattern + @item @samp{fms@var{m}4} + Like @code{fma@var{m}4}, except operand 3 subtracted from the + product instead of added to the product. This is represented + in the rtl as + + @smallexample + (fma:@var{m} @var{op1} @var{op2} (neg:@var{m} @var{op3})) + @end smallexample + + @cindex @code{fnma@var{m}4} instruction pattern + @item @samp{fnma@var{m}4} + Like @code{fma@var{m}4} except that the intermediate product + is negated before being added to operand 3. This is represented + in the rtl as + + @smallexample + (fma:@var{m} (neg:@var{m} @var{op1}) @var{op2} @var{op3}) + @end smallexample + + @cindex @code{fnms@var{m}4} instruction pattern + @item @samp{fnms@var{m}4} + Like @code{fms@var{m}4} except that the intermediate product + is negated before subtracting operand 3. This is represented + in the rtl as + + @smallexample + (fma:@var{m} (neg:@var{m} @var{op1}) @var{op2} (neg:@var{m} @var{op3})) + @end smallexample + @cindex @code{min@var{m}3} instruction pattern @cindex @code{max@var{m}3} instruction pattern @item @samp{smin@var{m}3}, @samp{smax@var{m}3} Index: gcc/flag-types.h =================================================================== *** gcc/flag-types.h.orig 2010-11-02 16:37:35.000000000 +0100 --- gcc/flag-types.h 2010-11-03 13:21:19.000000000 +0100 *************** enum warn_strict_overflow_code *** 152,155 **** --- 152,162 ---- WARN_STRICT_OVERFLOW_MAGNITUDE = 5 }; + /* Floating-point contraction mode. */ + enum fp_contract_mode { + FP_CONTRACT_OFF = 0, + FP_CONTRACT_ON = 1, + FP_CONTRACT_FAST = 2 + }; + #endif /* ! GCC_FLAG_TYPES_H */ Index: gcc/genopinit.c =================================================================== *** gcc/genopinit.c.orig 2010-11-02 16:37:35.000000000 +0100 --- gcc/genopinit.c 2010-11-03 13:21:19.000000000 +0100 *************** static const char * const optabs[] = *** 160,165 **** --- 160,168 ---- "set_optab_handler (floor_optab, $A, CODE_FOR_$(floor$a2$))", "set_convert_optab_handler (lfloor_optab, $B, $A, CODE_FOR_$(lfloor$F$a$I$b2$))", "set_optab_handler (fma_optab, $A, CODE_FOR_$(fma$a4$))", + "set_optab_handler (fms_optab, $A, CODE_FOR_$(fms$a4$))", + "set_optab_handler (fnma_optab, $A, CODE_FOR_$(fnma$a4$))", + "set_optab_handler (fnms_optab, $A, CODE_FOR_$(fnms$a4$))", "set_optab_handler (ceil_optab, $A, CODE_FOR_$(ceil$a2$))", "set_convert_optab_handler (lceil_optab, $B, $A, CODE_FOR_$(lceil$F$a$I$b2$))", "set_optab_handler (round_optab, $A, CODE_FOR_$(round$a2$))", Index: gcc/gimplify.c =================================================================== *** gcc/gimplify.c.orig 2010-11-03 10:55:43.000000000 +0100 --- gcc/gimplify.c 2010-11-03 13:21:19.000000000 +0100 *************** gimplify_expr (tree *expr_p, gimple_seq *** 7170,7175 **** --- 7170,7185 ---- ret = gimplify_omp_atomic (expr_p, pre_p); break; + case TRUTH_AND_EXPR: + case TRUTH_OR_EXPR: + case TRUTH_XOR_EXPR: + /* Classified as tcc_expression. */ + goto expr_2; + + case FMA_EXPR: + /* Classified as tcc_expression. */ + goto expr_3; + case POINTER_PLUS_EXPR: /* Convert ((type *)A)+offset into &A->field_of_type_and_offset. The second is gimple immediate saving a need for extra statement. *************** gimplify_expr (tree *expr_p, gimple_seq *** 7249,7264 **** break; } case tcc_declaration: case tcc_constant: ret = GS_ALL_DONE; goto dont_recalculate; default: ! gcc_assert (TREE_CODE (*expr_p) == TRUTH_AND_EXPR ! || TREE_CODE (*expr_p) == TRUTH_OR_EXPR ! || TREE_CODE (*expr_p) == TRUTH_XOR_EXPR); ! goto expr_2; } recalculate_side_effects (*expr_p); --- 7259,7286 ---- break; } + expr_3: + { + enum gimplify_status r0, r1, r2; + + r0 = gimplify_expr (&TREE_OPERAND (*expr_p, 0), pre_p, + post_p, is_gimple_val, fb_rvalue); + r1 = gimplify_expr (&TREE_OPERAND (*expr_p, 1), pre_p, + post_p, is_gimple_val, fb_rvalue); + r2 = gimplify_expr (&TREE_OPERAND (*expr_p, 2), pre_p, + post_p, is_gimple_val, fb_rvalue); + + ret = MIN (MIN (r0, r1), r2); + break; + } + case tcc_declaration: case tcc_constant: ret = GS_ALL_DONE; goto dont_recalculate; default: ! gcc_unreachable (); } recalculate_side_effects (*expr_p); Index: gcc/optabs.h =================================================================== *** gcc/optabs.h.orig 2010-11-02 16:37:35.000000000 +0100 --- gcc/optabs.h 2010-11-03 13:21:19.000000000 +0100 *************** enum optab_index *** 192,197 **** --- 192,200 ---- OTI_atan2, /* Floating multiply/add */ OTI_fma, + OTI_fms, + OTI_fnma, + OTI_fnms, /* Move instruction. */ OTI_mov, *************** enum optab_index *** 435,440 **** --- 438,446 ---- #define pow_optab (&optab_table[OTI_pow]) #define atan2_optab (&optab_table[OTI_atan2]) #define fma_optab (&optab_table[OTI_fma]) + #define fms_optab (&optab_table[OTI_fms]) + #define fnma_optab (&optab_table[OTI_fnma]) + #define fnms_optab (&optab_table[OTI_fnms]) #define mov_optab (&optab_table[OTI_mov]) #define movstrict_optab (&optab_table[OTI_movstrict]) Index: gcc/fold-const.c =================================================================== *** gcc/fold-const.c.orig 2010-11-03 10:55:43.000000000 +0100 --- gcc/fold-const.c 2010-11-03 13:50:07.000000000 +0100 *************** contains_label_p (tree st) *** 13281,13290 **** tree fold_ternary_loc (location_t loc, enum tree_code code, tree type, ! tree op0, tree op1, tree op2) { tree tem; ! tree arg0 = NULL_TREE, arg1 = NULL_TREE; enum tree_code_class kind = TREE_CODE_CLASS (code); gcc_assert (IS_EXPR_CODE_CLASS (kind) --- 13281,13290 ---- tree fold_ternary_loc (location_t loc, enum tree_code code, tree type, ! tree op0, tree op1, tree op2) { tree tem; ! tree arg0 = NULL_TREE, arg1 = NULL_TREE, arg2 = NULL_TREE; enum tree_code_class kind = TREE_CODE_CLASS (code); gcc_assert (IS_EXPR_CODE_CLASS (kind) *************** fold_ternary_loc (location_t loc, enum t *** 13312,13317 **** --- 13312,13323 ---- STRIP_NOPS (arg1); } + if (op2) + { + arg2 = op2; + STRIP_NOPS (arg2); + } + switch (code) { case COMPONENT_REF: *************** fold_ternary_loc (location_t loc, enum t *** 13610,13615 **** --- 13616,13632 ---- return NULL_TREE; + case FMA_EXPR: + /* For integers we can decompose the FMA if possible. */ + if (TREE_CODE (arg0) == INTEGER_CST + && TREE_CODE (arg1) == INTEGER_CST) + return fold_build2_loc (loc, PLUS_EXPR, type, + const_binop (MULT_EXPR, arg0, arg1), arg2); + if (integer_zerop (arg2)) + return fold_build2_loc (loc, MULT_EXPR, type, arg0, arg1); + + return fold_fma (loc, type, arg0, arg1, arg2); + default: return NULL_TREE; } /* switch (code) */ Index: gcc/optabs.c =================================================================== *** gcc/optabs.c.orig 2010-08-18 16:29:14.000000000 +0200 --- gcc/optabs.c 2010-11-03 13:55:11.000000000 +0100 *************** optab_for_tree_code (enum tree_code code *** 374,379 **** --- 374,382 ---- : (TYPE_SATURATING (type) ? ssmsub_widen_optab : smsub_widen_optab)); + case FMA_EXPR: + return fma_optab; + case REDUC_MAX_EXPR: return TYPE_UNSIGNED (type) ? reduc_umax_optab : reduc_smax_optab; Index: gcc/tree-vect-stmts.c =================================================================== *** gcc/tree-vect-stmts.c.orig 2010-11-03 10:55:43.000000000 +0100 --- gcc/tree-vect-stmts.c 2010-11-03 14:00:55.000000000 +0100 *************** vectorizable_shift (gimple stmt, gimple_ *** 2343,2349 **** /* Function vectorizable_operation. ! Check if STMT performs a binary or unary operation that can be vectorized. If VEC_STMT is also passed, vectorize the STMT: create a vectorized stmt to replace it, put it in VEC_STMT, and insert it at BSI. Return FALSE if not a vectorizable STMT, TRUE otherwise. */ --- 2343,2350 ---- /* Function vectorizable_operation. ! Check if STMT performs a binary, unary or ternary operation that can ! be vectorized. If VEC_STMT is also passed, vectorize the STMT: create a vectorized stmt to replace it, put it in VEC_STMT, and insert it at BSI. Return FALSE if not a vectorizable STMT, TRUE otherwise. */ *************** vectorizable_operation (gimple stmt, gim *** 2354,2360 **** { tree vec_dest; tree scalar_dest; ! tree op0, op1 = NULL; stmt_vec_info stmt_info = vinfo_for_stmt (stmt); tree vectype; loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); --- 2355,2361 ---- { tree vec_dest; tree scalar_dest; ! tree op0, op1 = NULL_TREE, op2 = NULL_TREE; stmt_vec_info stmt_info = vinfo_for_stmt (stmt); tree vectype; loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); *************** vectorizable_operation (gimple stmt, gim *** 2366,2372 **** int icode; tree def; gimple def_stmt; ! enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; gimple new_stmt = NULL; stmt_vec_info prev_stmt_info; int nunits_in; --- 2367,2374 ---- int icode; tree def; gimple def_stmt; ! enum vect_def_type dt[3] ! = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; gimple new_stmt = NULL; stmt_vec_info prev_stmt_info; int nunits_in; *************** vectorizable_operation (gimple stmt, gim *** 2374,2381 **** tree vectype_out; int ncopies; int j, i; ! VEC(tree,heap) *vec_oprnds0 = NULL, *vec_oprnds1 = NULL; ! tree vop0, vop1; bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); int vf; --- 2376,2383 ---- tree vectype_out; int ncopies; int j, i; ! VEC(tree,heap) *vec_oprnds0 = NULL, *vec_oprnds1 = NULL, *vec_oprnds2 = NULL; ! tree vop0, vop1, vop2; bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); int vf; *************** vectorizable_operation (gimple stmt, gim *** 2401,2410 **** /* Support only unary or binary operations. */ op_type = TREE_CODE_LENGTH (code); ! if (op_type != unary_op && op_type != binary_op) { if (vect_print_dump_info (REPORT_DETAILS)) ! fprintf (vect_dump, "num. args = %d (not unary/binary op).", op_type); return false; } --- 2403,2413 ---- /* Support only unary or binary operations. */ op_type = TREE_CODE_LENGTH (code); ! if (op_type != unary_op && op_type != binary_op && op_type != ternary_op) { if (vect_print_dump_info (REPORT_DETAILS)) ! fprintf (vect_dump, "num. args = %d (not unary/binary/ternary op).", ! op_type); return false; } *************** vectorizable_operation (gimple stmt, gim *** 2441,2447 **** if (nunits_out != nunits_in) return false; ! if (op_type == binary_op) { op1 = gimple_assign_rhs2 (stmt); if (!vect_is_simple_use (op1, loop_vinfo, bb_vinfo, &def_stmt, &def, --- 2444,2450 ---- if (nunits_out != nunits_in) return false; ! if (op_type == binary_op || op_type == ternary_op) { op1 = gimple_assign_rhs2 (stmt); if (!vect_is_simple_use (op1, loop_vinfo, bb_vinfo, &def_stmt, &def, *************** vectorizable_operation (gimple stmt, gim *** 2452,2457 **** --- 2455,2471 ---- return false; } } + if (op_type == ternary_op) + { + op2 = gimple_assign_rhs3 (stmt); + if (!vect_is_simple_use (op2, loop_vinfo, bb_vinfo, &def_stmt, &def, + &dt[2])) + { + if (vect_print_dump_info (REPORT_DETAILS)) + fprintf (vect_dump, "use not simple."); + return false; + } + } if (loop_vinfo) vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); *************** vectorizable_operation (gimple stmt, gim *** 2473,2479 **** || code == RROTATE_EXPR) return false; ! optab = optab_for_tree_code (code, vectype, optab_default); /* Supportable by target? */ if (!optab) --- 2487,2493 ---- || code == RROTATE_EXPR) return false; ! optab = optab_for_tree_code (code, vectype, optab_default); /* Supportable by target? */ if (!optab) *************** vectorizable_operation (gimple stmt, gim *** 2534,2541 **** if (!slp_node) { vec_oprnds0 = VEC_alloc (tree, heap, 1); ! if (op_type == binary_op) vec_oprnds1 = VEC_alloc (tree, heap, 1); } /* In case the vectorization factor (VF) is bigger than the number --- 2548,2557 ---- if (!slp_node) { vec_oprnds0 = VEC_alloc (tree, heap, 1); ! if (op_type == binary_op || op_type == ternary_op) vec_oprnds1 = VEC_alloc (tree, heap, 1); + if (op_type == ternary_op) + vec_oprnds2 = VEC_alloc (tree, heap, 1); } /* In case the vectorization factor (VF) is bigger than the number *************** vectorizable_operation (gimple stmt, gim *** 2597,2618 **** /* Handle uses. */ if (j == 0) { ! if (op_type == binary_op) vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, slp_node); else vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, slp_node); } else ! vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); /* Arguments are ready. Create the new vector stmt. */ FOR_EACH_VEC_ELT (tree, vec_oprnds0, i, vop0) { ! vop1 = ((op_type == binary_op) ! ? VEC_index (tree, vec_oprnds1, i) : NULL); ! new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1); new_temp = make_ssa_name (vec_dest, new_stmt); gimple_assign_set_lhs (new_stmt, new_temp); vect_finish_stmt_generation (stmt, new_stmt, gsi); --- 2613,2652 ---- /* Handle uses. */ if (j == 0) { ! if (op_type == binary_op || op_type == ternary_op) vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, slp_node); else vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, slp_node); + if (op_type == ternary_op) + { + vec_oprnds2 = VEC_alloc (tree, heap, 1); + VEC_quick_push (tree, vec_oprnds2, + vect_get_vec_def_for_operand (op2, stmt, NULL)); + } } else ! { ! vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); ! if (op_type == ternary_op) ! { ! tree vec_oprnd = VEC_pop (tree, vec_oprnds2); ! VEC_quick_push (tree, vec_oprnds2, ! vect_get_vec_def_for_stmt_copy (dt[2], ! vec_oprnd)); ! } ! } /* Arguments are ready. Create the new vector stmt. */ FOR_EACH_VEC_ELT (tree, vec_oprnds0, i, vop0) { ! vop1 = ((op_type == binary_op || op_type == ternary_op) ! ? VEC_index (tree, vec_oprnds1, i) : NULL_TREE); ! vop2 = ((op_type == ternary_op) ! ? VEC_index (tree, vec_oprnds2, i) : NULL_TREE); ! new_stmt = gimple_build_assign_with_ops3 (code, vec_dest, ! vop0, vop1, vop2); new_temp = make_ssa_name (vec_dest, new_stmt); gimple_assign_set_lhs (new_stmt, new_temp); vect_finish_stmt_generation (stmt, new_stmt, gsi); *************** vectorizable_operation (gimple stmt, gim *** 2633,2638 **** --- 2667,2674 ---- VEC_free (tree, heap, vec_oprnds0); if (vec_oprnds1) VEC_free (tree, heap, vec_oprnds1); + if (vec_oprnds2) + VEC_free (tree, heap, vec_oprnds2); return true; } Index: gcc/tree.h =================================================================== *** gcc/tree.h.orig 2010-11-02 11:16:39.000000000 +0100 --- gcc/tree.h 2010-11-03 13:34:12.000000000 +0100 *************** extern void fold_defer_overflow_warnings *** 4954,4959 **** --- 4954,4960 ---- extern void fold_undefer_overflow_warnings (bool, const_gimple, int); extern void fold_undefer_and_ignore_overflow_warnings (void); extern bool fold_deferring_overflow_warnings_p (void); + extern tree fold_fma (location_t, tree, tree, tree, tree); enum operand_equal_flag { ^ permalink raw reply [flat|nested] 12+ messages in thread
* Re: [PATCH] FMA on trees 2010-11-03 13:13 ` Richard Guenther @ 2010-11-03 15:30 ` Richard Guenther 2010-11-08 19:23 ` Steve Ellcey 0 siblings, 1 reply; 12+ messages in thread From: Richard Guenther @ 2010-11-03 15:30 UTC (permalink / raw) To: Richard Henderson; +Cc: gcc-patches On Wed, 3 Nov 2010, Richard Guenther wrote: > On Tue, 2 Nov 2010, Richard Henderson wrote: > > > On 11/02/2010 06:24 AM, Richard Guenther wrote: > > > *************** set_unsafe_math_optimizations_flags (int > > > *** 2289,2294 **** > > > --- 2301,2307 ---- > > > flag_signed_zeros = !set; > > > flag_associative_math = set; > > > flag_reciprocal_math = set; > > > + flag_fp_contract_mode = set ? FP_CONTRACT_FAST : FP_CONTRACT_OFF; > > > > Off? Default is FAST anyway. I can't think what to set here > > for -fno-fast-math, actually. > > Yeah, I wondered about that, too. I've now just removed -ffp-contract > from unsafe-math handling completely. We can revisit that if we > change the default and/or implement ON. > > > > + fold_builtin_fma (location_t loc, tree arg0, tree arg1, tree arg2, tree type) > > > + { > > > + if (validate_arg (arg0, REAL_TYPE) > > > + && validate_arg(arg1, REAL_TYPE) > > > + && validate_arg(arg2, REAL_TYPE)) > > > + { > > > + if (TREE_CODE (arg0) == REAL_CST > > > + && TREE_CODE (arg1) == REAL_CST > > > + && TREE_CODE (arg2) == REAL_CST) > > > + return do_mpfr_arg3 (arg0, arg1, arg2, type, mpfr_fma); > > > + > > > + /* ??? Only expand to FMA_EXPR if it's directly supported. */ > > > + if (optab_handler (fma_optab, TYPE_MODE (type)) != CODE_FOR_nothing) > > > + return fold_build3_loc (loc, FMA_EXPR, type, arg0, arg1, arg2); > > > > I know this is my bit, but I just remembered that this means we > > need a new entry in fold_ternary_loc to fold FMA_EXPR. > > Done. > > > Finally, if !flag_signed_zeros (part of fast-math), we probably > > want to fold -fma(a,b,c) -> fma(-a,b,c). > > Hm, that doesn't sound correct to me ;) > > The following adds vectorization capabilities as well which fixes the > seen regressions. > > Bootstrap and regtest pending. Bootstrapped and tested on x86_64-unknown-linux-gnu, I'm going to commit it tomorrow if there are no objections. Everything else can be done as a followup I think (not that I have anything in mind). Richard. ^ permalink raw reply [flat|nested] 12+ messages in thread
* Re: [PATCH] FMA on trees 2010-11-03 15:30 ` Richard Guenther @ 2010-11-08 19:23 ` Steve Ellcey 2010-11-08 19:27 ` -mfused-madd vs -ffp-contract deprication? Richard Henderson 0 siblings, 1 reply; 12+ messages in thread From: Steve Ellcey @ 2010-11-08 19:23 UTC (permalink / raw) To: Richard Guenther; +Cc: gcc-patches This patch seems to have broken -mno-fused-madd on IA64. I don't know if it has broken it on other platforms too. Also I now get about 30 failures on IA64 for gcc.dg/torture/builtin-math-2.c. FAIL: gcc.dg/torture/builtin-math-2.c -O0 scan-tree-dump-times original "fma " 3 (30 or so more reports for fmaf and fmal and at various options) It looks like the expand file contains FMA_EXPR instead of the builtin_fma calls, as mentioned in the powerpc bug report for pr46353. I tried various combinations of -ffp-contract and -fno-fused-madd on this test but none had any effect. Steve Ellcey sje@cup.hp.com ^ permalink raw reply [flat|nested] 12+ messages in thread
* -mfused-madd vs -ffp-contract deprication? 2010-11-08 19:23 ` Steve Ellcey @ 2010-11-08 19:27 ` Richard Henderson 2010-11-08 23:14 ` Joseph S. Myers 2010-11-09 9:53 ` Richard Guenther 0 siblings, 2 replies; 12+ messages in thread From: Richard Henderson @ 2010-11-08 19:27 UTC (permalink / raw) To: joseph; +Cc: Steve Ellcey, Richard Guenther, gcc-patches On 11/08/2010 11:11 AM, Steve Ellcey wrote: > This patch seems to have broken -mno-fused-madd on IA64. I don't > know if it has broken it on other platforms too. Probably. I wonder about making -mfused-madd a generic alias for -ffp-contract, and removing it from the backends. With an appropriate warning, we ought to be able to remove -mfused-madd in gcc 4.7. r~ ^ permalink raw reply [flat|nested] 12+ messages in thread
* Re: -mfused-madd vs -ffp-contract deprication? 2010-11-08 19:27 ` -mfused-madd vs -ffp-contract deprication? Richard Henderson @ 2010-11-08 23:14 ` Joseph S. Myers 2010-11-09 1:41 ` Richard Henderson 2010-11-09 9:53 ` Richard Guenther 1 sibling, 1 reply; 12+ messages in thread From: Joseph S. Myers @ 2010-11-08 23:14 UTC (permalink / raw) To: Richard Henderson; +Cc: Steve Ellcey, Richard Guenther, gcc-patches On Mon, 8 Nov 2010, Richard Henderson wrote: > On 11/08/2010 11:11 AM, Steve Ellcey wrote: > > This patch seems to have broken -mno-fused-madd on IA64. I don't > > know if it has broken it on other platforms too. > > Probably. > > I wonder about making -mfused-madd a generic alias for -ffp-contract, > and removing it from the backends. With an appropriate warning, we > ought to be able to remove -mfused-madd in gcc 4.7. As I indicated in <http://gcc.gnu.org/ml/gcc-patches/2010-10/msg01884.html>, this may depend on the actual intended semantics of existing -mfused-madd options (which I have not attempted to discern). -- Joseph S. Myers joseph@codesourcery.com ^ permalink raw reply [flat|nested] 12+ messages in thread
* Re: -mfused-madd vs -ffp-contract deprication? 2010-11-08 23:14 ` Joseph S. Myers @ 2010-11-09 1:41 ` Richard Henderson 2010-11-09 9:09 ` Richard Sandiford 0 siblings, 1 reply; 12+ messages in thread From: Richard Henderson @ 2010-11-09 1:41 UTC (permalink / raw) To: Joseph S. Myers Cc: Steve Ellcey, Richard Guenther, gcc-patches, sterling, rdsandiford On 11/08/2010 02:58 PM, Joseph S. Myers wrote: > <http://gcc.gnu.org/ml/gcc-patches/2010-10/msg01884.html>: > Existing target options may sometimes be ambiguous as to whether they are > meant to disable instructions that might not be available, or to disable > automatic conversion It's currently used in: xtensa i386 s390 rs6000 mips ia64 Given that MIPS doesn't implement true FMA, MIPS probably should match the (plus (mult a b) c) pattern unconditionally (modulo ISA) and be done with it. The rest of the targets either have the instructions in the base ISA (xtensa, s390, rs6000, ia64) or additionally have some other macro that controls availability (i386). So, if we fix MIPS all remaining users will only control contraction. Only xtensa has not been converted to the named fma patterns, though many still contain additional (plus (mult)) patterns for FUSED_MADD. Given that we have a pass that performs that operation at the tree level, it would seem that all we need to do is delete the obsolete patterns. r~ ^ permalink raw reply [flat|nested] 12+ messages in thread
* Re: -mfused-madd vs -ffp-contract deprication? 2010-11-09 1:41 ` Richard Henderson @ 2010-11-09 9:09 ` Richard Sandiford 2010-11-09 15:49 ` Richard Henderson 0 siblings, 1 reply; 12+ messages in thread From: Richard Sandiford @ 2010-11-09 9:09 UTC (permalink / raw) To: Richard Henderson Cc: Joseph S. Myers, Steve Ellcey, Richard Guenther, gcc-patches, sterling Richard Henderson <rth@redhat.com> writes: > On 11/08/2010 02:58 PM, Joseph S. Myers wrote: >> <http://gcc.gnu.org/ml/gcc-patches/2010-10/msg01884.html>: >> Existing target options may sometimes be ambiguous as to whether they are >> meant to disable instructions that might not be available, or to disable >> automatic conversion > > It's currently used in: > > xtensa > i386 > s390 > rs6000 > mips > ia64 > > Given that MIPS doesn't implement true FMA, MIPS probably should > match the (plus (mult a b) c) pattern unconditionally (modulo ISA) > and be done with it. I think they should still be conditional on TARGET_FUSED_MADD because of the R8000 case. If anyone is still using bleeding edge GCC on R8000, we might as well continue to provide a way for them to disable the fusing. Richard ^ permalink raw reply [flat|nested] 12+ messages in thread
* Re: -mfused-madd vs -ffp-contract deprication? 2010-11-09 9:09 ` Richard Sandiford @ 2010-11-09 15:49 ` Richard Henderson 0 siblings, 0 replies; 12+ messages in thread From: Richard Henderson @ 2010-11-09 15:49 UTC (permalink / raw) To: Joseph S. Myers, Steve Ellcey, Richard Guenther, gcc-patches, sterling, rdsandiford On 11/09/2010 12:48 AM, Richard Sandiford wrote: > I think they should still be conditional on TARGET_FUSED_MADD > because of the R8000 case. If anyone is still using bleeding edge > GCC on R8000, we might as well continue to provide a way for them > to disable the fusing. Fair enough, I suppose. Though the plan is to move away from -mfused-add to -ffp-contract, so a tiny bit of global substitution will be needed. r~ ^ permalink raw reply [flat|nested] 12+ messages in thread
* Re: -mfused-madd vs -ffp-contract deprication? 2010-11-08 19:27 ` -mfused-madd vs -ffp-contract deprication? Richard Henderson 2010-11-08 23:14 ` Joseph S. Myers @ 2010-11-09 9:53 ` Richard Guenther 2010-11-09 17:14 ` Richard Henderson 1 sibling, 1 reply; 12+ messages in thread From: Richard Guenther @ 2010-11-09 9:53 UTC (permalink / raw) To: Richard Henderson; +Cc: joseph, Steve Ellcey, gcc-patches On Mon, 8 Nov 2010, Richard Henderson wrote: > On 11/08/2010 11:11 AM, Steve Ellcey wrote: > > This patch seems to have broken -mno-fused-madd on IA64. I don't > > know if it has broken it on other platforms too. > > Probably. > > I wonder about making -mfused-madd a generic alias for -ffp-contract, > and removing it from the backends. With an appropriate warning, we > ought to be able to remove -mfused-madd in gcc 4.7. Hm, I suspect that we fold __builtin_fma to FMA_EXPR too eagerly as for example ia64 has ;; Official C99 versions of the fmaf family of operations. (define_insn "fmasf4" [(set (match_operand:SF 0 "fr_register_operand" "=f") (fma:SF (match_operand:SF 1 "fr_reg_or_fp01_operand" "fG") (match_operand:SF 2 "fr_reg_or_fp01_operand" "fG") (match_operand:SF 3 "fr_reg_or_signed_fp01_operand" "fZ")))] "" "fma.s %0 = %F1, %F2, %F3" [(set_attr "itanium_class" "fmac")]) which matches the named expander. Thus the scan for the builtin calls are failing (but they are failing spuriously I think). Generally named insns may accidentially match the new named expanders even if they do not implement IEEE fma. Richard. ^ permalink raw reply [flat|nested] 12+ messages in thread
* Re: -mfused-madd vs -ffp-contract deprication? 2010-11-09 9:53 ` Richard Guenther @ 2010-11-09 17:14 ` Richard Henderson 0 siblings, 0 replies; 12+ messages in thread From: Richard Henderson @ 2010-11-09 17:14 UTC (permalink / raw) To: Richard Guenther; +Cc: joseph, Steve Ellcey, gcc-patches On 11/09/2010 01:52 AM, Richard Guenther wrote: > Hm, I suspect that we fold __builtin_fma to FMA_EXPR too eagerly Huh? They're semantically identical. > which matches the named expander. Thus the scan for the builtin > calls are failing (but they are failing spuriously I think). Pointer? r~ ^ permalink raw reply [flat|nested] 12+ messages in thread
end of thread, other threads:[~2010-11-09 17:02 UTC | newest] Thread overview: 12+ messages (download: mbox.gz / follow: Atom feed) -- links below jump to the message on this page -- 2010-11-02 13:35 [PATCH] FMA on trees Richard Guenther 2010-11-02 15:40 ` Richard Henderson 2010-11-03 13:13 ` Richard Guenther 2010-11-03 15:30 ` Richard Guenther 2010-11-08 19:23 ` Steve Ellcey 2010-11-08 19:27 ` -mfused-madd vs -ffp-contract deprication? Richard Henderson 2010-11-08 23:14 ` Joseph S. Myers 2010-11-09 1:41 ` Richard Henderson 2010-11-09 9:09 ` Richard Sandiford 2010-11-09 15:49 ` Richard Henderson 2010-11-09 9:53 ` Richard Guenther 2010-11-09 17:14 ` Richard Henderson
This is a public inbox, see mirroring instructions for how to clone and mirror all data and code used for this inbox; as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).