/* The Blackfin code generation auxiliary output file.
- Copyright (C) 2005, 2006 Free Software Foundation, Inc.
+ Copyright (C) 2005, 2006, 2007, 2008 Free Software Foundation, Inc.
Contributed by Analog Devices.
This file is part of GCC.
GCC is free software; you can redistribute it and/or modify it
under the terms of the GNU General Public License as published
- by the Free Software Foundation; either version 2, or (at your
+ by the Free Software Foundation; either version 3, or (at your
option) any later version.
GCC is distributed in the hope that it will be useful, but WITHOUT
License for more details.
You should have received a copy of the GNU General Public License
- along with GCC; see the file COPYING. If not, write to
- the Free Software Foundation, 51 Franklin Street, Fifth Floor,
- Boston, MA 02110-1301, USA. */
+ along with GCC; see the file COPYING3. If not see
+ <http://www.gnu.org/licenses/>. */
#include "config.h"
#include "system.h"
#include "langhooks.h"
#include "bfin-protos.h"
#include "tm-preds.h"
+#include "tm-constrs.h"
#include "gt-bfin.h"
#include "basic-block.h"
+#include "cfglayout.h"
#include "timevar.h"
+#include "df.h"
/* A C structure for machine-specific, per-function data.
This is added to the cfun structure. */
reorganization. */
static int bfin_flag_var_tracking;
+/* -mcpu support */
+bfin_cpu_t bfin_cpu_type = BFIN_CPU_UNKNOWN;
+
+/* -msi-revision support. There are three special values:
+ -1 -msi-revision=none.
+ 0xffff -msi-revision=any. */
+int bfin_si_revision;
+
+/* The workarounds enabled */
+unsigned int bfin_workarounds = 0;
+
+struct bfin_cpu
+{
+ const char *name;
+ bfin_cpu_t type;
+ int si_revision;
+ unsigned int workarounds;
+};
+
+struct bfin_cpu bfin_cpus[] =
+{
+ {"bf522", BFIN_CPU_BF522, 0x0001,
+ WA_SPECULATIVE_LOADS | WA_RETS},
+ {"bf522", BFIN_CPU_BF522, 0x0000,
+ WA_SPECULATIVE_LOADS | WA_RETS},
+
+ {"bf523", BFIN_CPU_BF523, 0x0001,
+ WA_SPECULATIVE_LOADS | WA_RETS},
+ {"bf523", BFIN_CPU_BF523, 0x0000,
+ WA_SPECULATIVE_LOADS | WA_RETS},
+
+ {"bf524", BFIN_CPU_BF524, 0x0001,
+ WA_SPECULATIVE_LOADS | WA_RETS},
+ {"bf524", BFIN_CPU_BF524, 0x0000,
+ WA_SPECULATIVE_LOADS | WA_RETS},
+
+ {"bf525", BFIN_CPU_BF525, 0x0001,
+ WA_SPECULATIVE_LOADS | WA_RETS},
+ {"bf525", BFIN_CPU_BF525, 0x0000,
+ WA_SPECULATIVE_LOADS | WA_RETS},
+
+ {"bf526", BFIN_CPU_BF526, 0x0001,
+ WA_SPECULATIVE_LOADS | WA_RETS},
+ {"bf526", BFIN_CPU_BF526, 0x0000,
+ WA_SPECULATIVE_LOADS | WA_RETS},
+
+ {"bf527", BFIN_CPU_BF527, 0x0001,
+ WA_SPECULATIVE_LOADS | WA_RETS},
+ {"bf527", BFIN_CPU_BF527, 0x0000,
+ WA_SPECULATIVE_LOADS | WA_RETS},
+
+ {"bf531", BFIN_CPU_BF531, 0x0005,
+ WA_SPECULATIVE_LOADS | WA_RETS},
+ {"bf531", BFIN_CPU_BF531, 0x0004,
+ WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS},
+ {"bf531", BFIN_CPU_BF531, 0x0003,
+ WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS},
+
+ {"bf532", BFIN_CPU_BF532, 0x0005,
+ WA_SPECULATIVE_LOADS | WA_RETS},
+ {"bf532", BFIN_CPU_BF532, 0x0004,
+ WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS},
+ {"bf532", BFIN_CPU_BF532, 0x0003,
+ WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS},
+
+ {"bf533", BFIN_CPU_BF533, 0x0005,
+ WA_SPECULATIVE_LOADS | WA_RETS},
+ {"bf533", BFIN_CPU_BF533, 0x0004,
+ WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS},
+ {"bf533", BFIN_CPU_BF533, 0x0003,
+ WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS},
+
+ {"bf534", BFIN_CPU_BF534, 0x0003,
+ WA_SPECULATIVE_LOADS | WA_RETS},
+ {"bf534", BFIN_CPU_BF534, 0x0002,
+ WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS},
+ {"bf534", BFIN_CPU_BF534, 0x0001,
+ WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS},
+
+ {"bf536", BFIN_CPU_BF536, 0x0003,
+ WA_SPECULATIVE_LOADS | WA_RETS},
+ {"bf536", BFIN_CPU_BF536, 0x0002,
+ WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS},
+ {"bf536", BFIN_CPU_BF536, 0x0001,
+ WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS},
+
+ {"bf537", BFIN_CPU_BF537, 0x0003,
+ WA_SPECULATIVE_LOADS | WA_RETS},
+ {"bf537", BFIN_CPU_BF537, 0x0002,
+ WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS},
+ {"bf537", BFIN_CPU_BF537, 0x0001,
+ WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS},
+
+ {"bf538", BFIN_CPU_BF538, 0x0004,
+ WA_SPECULATIVE_LOADS | WA_RETS},
+ {"bf538", BFIN_CPU_BF538, 0x0003,
+ WA_SPECULATIVE_LOADS | WA_RETS},
+ {"bf538", BFIN_CPU_BF538, 0x0002,
+ WA_SPECULATIVE_LOADS | WA_RETS},
+
+ {"bf539", BFIN_CPU_BF539, 0x0004,
+ WA_SPECULATIVE_LOADS | WA_RETS},
+ {"bf539", BFIN_CPU_BF539, 0x0003,
+ WA_SPECULATIVE_LOADS | WA_RETS},
+ {"bf539", BFIN_CPU_BF539, 0x0002,
+ WA_SPECULATIVE_LOADS | WA_RETS},
+
+ {"bf542", BFIN_CPU_BF542, 0x0001,
+ WA_SPECULATIVE_LOADS | WA_RETS},
+ {"bf542", BFIN_CPU_BF542, 0x0000,
+ WA_SPECULATIVE_LOADS | WA_RETS},
+
+ {"bf544", BFIN_CPU_BF544, 0x0001,
+ WA_SPECULATIVE_LOADS | WA_RETS},
+ {"bf544", BFIN_CPU_BF544, 0x0000,
+ WA_SPECULATIVE_LOADS | WA_RETS},
+
+ {"bf547", BFIN_CPU_BF547, 0x0001,
+ WA_SPECULATIVE_LOADS | WA_RETS},
+ {"bf547", BFIN_CPU_BF547, 0x0000,
+ WA_SPECULATIVE_LOADS | WA_RETS},
+
+ {"bf548", BFIN_CPU_BF548, 0x0001,
+ WA_SPECULATIVE_LOADS | WA_RETS},
+ {"bf548", BFIN_CPU_BF548, 0x0000,
+ WA_SPECULATIVE_LOADS | WA_RETS},
+
+ {"bf549", BFIN_CPU_BF549, 0x0001,
+ WA_SPECULATIVE_LOADS | WA_RETS},
+ {"bf549", BFIN_CPU_BF549, 0x0000,
+ WA_SPECULATIVE_LOADS | WA_RETS},
+
+ {"bf561", BFIN_CPU_BF561, 0x0005, WA_RETS},
+ {"bf561", BFIN_CPU_BF561, 0x0003,
+ WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS},
+ {"bf561", BFIN_CPU_BF561, 0x0002,
+ WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS},
+
+ {NULL, 0, 0, 0}
+};
+
int splitting_for_sched;
static void
/* Examine machine-dependent attributes of function type FUNTYPE and return its
type. See the definition of E_FUNKIND. */
-static e_funkind funkind (tree funtype)
+static e_funkind
+funkind (const_tree funtype)
{
tree attrs = TYPE_ATTRIBUTES (funtype);
if (lookup_attribute ("interrupt_handler", attrs))
legitimize_pic_address (rtx orig, rtx reg, rtx picreg)
{
rtx addr = orig;
- rtx new = orig;
+ rtx new_rtx = orig;
if (GET_CODE (addr) == SYMBOL_REF || GET_CODE (addr) == LABEL_REF)
{
if (reg == 0)
{
- gcc_assert (!no_new_pseudos);
+ gcc_assert (can_create_pseudo_p ());
reg = gen_reg_rtx (Pmode);
}
tmp = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, addr), unspec);
- new = gen_const_mem (Pmode, gen_rtx_PLUS (Pmode, picreg, tmp));
+ new_rtx = gen_const_mem (Pmode, gen_rtx_PLUS (Pmode, picreg, tmp));
- emit_move_insn (reg, new);
+ emit_move_insn (reg, new_rtx);
if (picreg == pic_offset_table_rtx)
- current_function_uses_pic_offset_table = 1;
+ crtl->uses_pic_offset_table = 1;
return reg;
}
if (reg == 0)
{
- gcc_assert (!no_new_pseudos);
+ gcc_assert (can_create_pseudo_p ());
reg = gen_reg_rtx (Pmode);
}
return gen_rtx_PLUS (Pmode, base, addr);
}
- return new;
+ return new_rtx;
}
\f
/* Stack frame layout. */
-/* Compute the number of DREGS to save with a push_multiple operation.
- This could include registers that aren't modified in the function,
- since push_multiple only takes a range of registers.
- If IS_INTHANDLER, then everything that is live must be saved, even
- if normally call-clobbered. */
-
-static int
-n_dregs_to_save (bool is_inthandler)
+/* For a given REGNO, determine whether it must be saved in the function
+ prologue. IS_INTHANDLER specifies whether we're generating a normal
+ prologue or an interrupt/exception one. */
+static bool
+must_save_p (bool is_inthandler, unsigned regno)
{
- unsigned i;
-
- for (i = REG_R0; i <= REG_R7; i++)
+ if (D_REGNO_P (regno))
{
- if (regs_ever_live[i] && (is_inthandler || ! call_used_regs[i]))
- return REG_R7 - i + 1;
-
- if (current_function_calls_eh_return)
+ bool is_eh_return_reg = false;
+ if (crtl->calls_eh_return)
{
unsigned j;
for (j = 0; ; j++)
unsigned test = EH_RETURN_DATA_REGNO (j);
if (test == INVALID_REGNUM)
break;
- if (test == i)
- return REG_R7 - i + 1;
+ if (test == regno)
+ is_eh_return_reg = true;
}
}
+ return (is_eh_return_reg
+ || (df_regs_ever_live_p (regno)
+ && !fixed_regs[regno]
+ && (is_inthandler || !call_used_regs[regno])));
}
- return 0;
+ else if (P_REGNO_P (regno))
+ {
+ return ((df_regs_ever_live_p (regno)
+ && !fixed_regs[regno]
+ && (is_inthandler || !call_used_regs[regno]))
+ || (!TARGET_FDPIC
+ && regno == PIC_OFFSET_TABLE_REGNUM
+ && (crtl->uses_pic_offset_table
+ || (TARGET_ID_SHARED_LIBRARY && !current_function_is_leaf))));
+ }
+ else
+ return ((is_inthandler || !call_used_regs[regno])
+ && (df_regs_ever_live_p (regno)
+ || (!leaf_function_p () && call_used_regs[regno])));
+
+}
+
+/* Compute the number of DREGS to save with a push_multiple operation.
+ This could include registers that aren't modified in the function,
+ since push_multiple only takes a range of registers.
+ If IS_INTHANDLER, then everything that is live must be saved, even
+ if normally call-clobbered.
+ If CONSECUTIVE, return the number of registers we can save in one
+ instruction with a push/pop multiple instruction. */
+
+static int
+n_dregs_to_save (bool is_inthandler, bool consecutive)
+{
+ int count = 0;
+ unsigned i;
+
+ for (i = REG_R7 + 1; i-- != REG_R0;)
+ {
+ if (must_save_p (is_inthandler, i))
+ count++;
+ else if (consecutive)
+ return count;
+ }
+ return count;
}
/* Like n_dregs_to_save, but compute number of PREGS to save. */
static int
-n_pregs_to_save (bool is_inthandler)
+n_pregs_to_save (bool is_inthandler, bool consecutive)
{
+ int count = 0;
unsigned i;
- for (i = REG_P0; i <= REG_P5; i++)
- if ((regs_ever_live[i] && (is_inthandler || ! call_used_regs[i]))
- || (!TARGET_FDPIC
- && i == PIC_OFFSET_TABLE_REGNUM
- && (current_function_uses_pic_offset_table
- || (TARGET_ID_SHARED_LIBRARY && ! current_function_is_leaf))))
- return REG_P5 - i + 1;
- return 0;
+ for (i = REG_P5 + 1; i-- != REG_P0;)
+ if (must_save_p (is_inthandler, i))
+ count++;
+ else if (consecutive)
+ return count;
+ return count;
}
/* Determine if we are going to save the frame pointer in the prologue. */
static bool
must_save_fp_p (void)
{
- return frame_pointer_needed || regs_ever_live[REG_FP];
+ return frame_pointer_needed || df_regs_ever_live_p (REG_FP);
}
static bool
{
/* EH return puts a new return address into the frame using an
address relative to the frame pointer. */
- if (current_function_calls_eh_return)
+ if (crtl->calls_eh_return)
return true;
return frame_pointer_needed;
}
static void
expand_prologue_reg_save (rtx spreg, int saveall, bool is_inthandler)
{
- int ndregs = saveall ? 8 : n_dregs_to_save (is_inthandler);
- int npregs = saveall ? 6 : n_pregs_to_save (is_inthandler);
- int dregno = REG_R7 + 1 - ndregs;
- int pregno = REG_P5 + 1 - npregs;
- int total = ndregs + npregs;
- int i;
- rtx pat, insn, val;
+ rtx predec1 = gen_rtx_PRE_DEC (SImode, spreg);
+ rtx predec = gen_rtx_MEM (SImode, predec1);
+ int ndregs = saveall ? 8 : n_dregs_to_save (is_inthandler, false);
+ int npregs = saveall ? 6 : n_pregs_to_save (is_inthandler, false);
+ int ndregs_consec = saveall ? 8 : n_dregs_to_save (is_inthandler, true);
+ int npregs_consec = saveall ? 6 : n_pregs_to_save (is_inthandler, true);
+ int dregno, pregno;
+ int total_consec = ndregs_consec + npregs_consec;
+ int i, d_to_save;
+
+ if (saveall || is_inthandler)
+ {
+ rtx insn = emit_move_insn (predec, gen_rtx_REG (SImode, REG_ASTAT));
+ RTX_FRAME_RELATED_P (insn) = 1;
+ }
- if (total == 0)
- return;
+ if (total_consec != 0)
+ {
+ rtx insn;
+ rtx val = GEN_INT (-total_consec * 4);
+ rtx pat = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (total_consec + 2));
+
+ XVECEXP (pat, 0, 0) = gen_rtx_UNSPEC (VOIDmode, gen_rtvec (1, val),
+ UNSPEC_PUSH_MULTIPLE);
+ XVECEXP (pat, 0, total_consec + 1) = gen_rtx_SET (VOIDmode, spreg,
+ gen_rtx_PLUS (Pmode,
+ spreg,
+ val));
+ RTX_FRAME_RELATED_P (XVECEXP (pat, 0, total_consec + 1)) = 1;
+ d_to_save = ndregs_consec;
+ dregno = REG_R7 + 1 - ndregs_consec;
+ pregno = REG_P5 + 1 - npregs_consec;
+ for (i = 0; i < total_consec; i++)
+ {
+ rtx memref = gen_rtx_MEM (word_mode,
+ gen_rtx_PLUS (Pmode, spreg,
+ GEN_INT (- i * 4 - 4)));
+ rtx subpat;
+ if (d_to_save > 0)
+ {
+ subpat = gen_rtx_SET (VOIDmode, memref, gen_rtx_REG (word_mode,
+ dregno++));
+ d_to_save--;
+ }
+ else
+ {
+ subpat = gen_rtx_SET (VOIDmode, memref, gen_rtx_REG (word_mode,
+ pregno++));
+ }
+ XVECEXP (pat, 0, i + 1) = subpat;
+ RTX_FRAME_RELATED_P (subpat) = 1;
+ }
+ insn = emit_insn (pat);
+ RTX_FRAME_RELATED_P (insn) = 1;
+ }
- val = GEN_INT (-total * 4);
- pat = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (total + 2));
- XVECEXP (pat, 0, 0) = gen_rtx_UNSPEC (VOIDmode, gen_rtvec (1, val),
- UNSPEC_PUSH_MULTIPLE);
- XVECEXP (pat, 0, total + 1) = gen_rtx_SET (VOIDmode, spreg,
- gen_rtx_PLUS (Pmode, spreg,
- val));
- RTX_FRAME_RELATED_P (XVECEXP (pat, 0, total + 1)) = 1;
- for (i = 0; i < total; i++)
- {
- rtx memref = gen_rtx_MEM (word_mode,
- gen_rtx_PLUS (Pmode, spreg,
- GEN_INT (- i * 4 - 4)));
- rtx subpat;
- if (ndregs > 0)
+ for (dregno = REG_R0; ndregs != ndregs_consec; dregno++)
+ {
+ if (must_save_p (is_inthandler, dregno))
{
- subpat = gen_rtx_SET (VOIDmode, memref, gen_rtx_REG (word_mode,
- dregno++));
+ rtx insn = emit_move_insn (predec, gen_rtx_REG (word_mode, dregno));
+ RTX_FRAME_RELATED_P (insn) = 1;
ndregs--;
}
- else
+ }
+ for (pregno = REG_P0; npregs != npregs_consec; pregno++)
+ {
+ if (must_save_p (is_inthandler, pregno))
{
- subpat = gen_rtx_SET (VOIDmode, memref, gen_rtx_REG (word_mode,
- pregno++));
- npregs++;
+ rtx insn = emit_move_insn (predec, gen_rtx_REG (word_mode, pregno));
+ RTX_FRAME_RELATED_P (insn) = 1;
+ npregs--;
}
- XVECEXP (pat, 0, i + 1) = subpat;
- RTX_FRAME_RELATED_P (subpat) = 1;
}
- insn = emit_insn (pat);
- RTX_FRAME_RELATED_P (insn) = 1;
+ for (i = REG_P7 + 1; i < REG_CC; i++)
+ if (saveall
+ || (is_inthandler
+ && (df_regs_ever_live_p (i)
+ || (!leaf_function_p () && call_used_regs[i]))))
+ {
+ rtx insn;
+ if (i == REG_A0 || i == REG_A1)
+ insn = emit_move_insn (gen_rtx_MEM (PDImode, predec1),
+ gen_rtx_REG (PDImode, i));
+ else
+ insn = emit_move_insn (predec, gen_rtx_REG (SImode, i));
+ RTX_FRAME_RELATED_P (insn) = 1;
+ }
}
/* Emit code to restore registers in the epilogue. SAVEALL is nonzero if we
static void
expand_epilogue_reg_restore (rtx spreg, bool saveall, bool is_inthandler)
{
- int ndregs = saveall ? 8 : n_dregs_to_save (is_inthandler);
- int npregs = saveall ? 6 : n_pregs_to_save (is_inthandler);
- int total = ndregs + npregs;
- int i, regno;
- rtx pat, insn;
+ rtx postinc1 = gen_rtx_POST_INC (SImode, spreg);
+ rtx postinc = gen_rtx_MEM (SImode, postinc1);
- if (total == 0)
- return;
+ int ndregs = saveall ? 8 : n_dregs_to_save (is_inthandler, false);
+ int npregs = saveall ? 6 : n_pregs_to_save (is_inthandler, false);
+ int ndregs_consec = saveall ? 8 : n_dregs_to_save (is_inthandler, true);
+ int npregs_consec = saveall ? 6 : n_pregs_to_save (is_inthandler, true);
+ int total_consec = ndregs_consec + npregs_consec;
+ int i, regno;
+ rtx insn;
- pat = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (total + 1));
- XVECEXP (pat, 0, 0) = gen_rtx_SET (VOIDmode, spreg,
- gen_rtx_PLUS (Pmode, spreg,
- GEN_INT (total * 4)));
+ /* A slightly crude technique to stop flow from trying to delete "dead"
+ insns. */
+ MEM_VOLATILE_P (postinc) = 1;
- if (npregs > 0)
- regno = REG_P5 + 1;
- else
- regno = REG_R7 + 1;
+ for (i = REG_CC - 1; i > REG_P7; i--)
+ if (saveall
+ || (is_inthandler
+ && (df_regs_ever_live_p (i)
+ || (!leaf_function_p () && call_used_regs[i]))))
+ {
+ if (i == REG_A0 || i == REG_A1)
+ {
+ rtx mem = gen_rtx_MEM (PDImode, postinc1);
+ MEM_VOLATILE_P (mem) = 1;
+ emit_move_insn (gen_rtx_REG (PDImode, i), mem);
+ }
+ else
+ emit_move_insn (gen_rtx_REG (SImode, i), postinc);
+ }
- for (i = 0; i < total; i++)
+ regno = REG_P5 - npregs_consec;
+ for (; npregs != npregs_consec; regno--)
+ {
+ if (must_save_p (is_inthandler, regno))
+ {
+ emit_move_insn (gen_rtx_REG (word_mode, regno), postinc);
+ npregs--;
+ }
+ }
+ regno = REG_R7 - ndregs_consec;
+ for (; ndregs != ndregs_consec; regno--)
{
- rtx addr = (i > 0
- ? gen_rtx_PLUS (Pmode, spreg, GEN_INT (i * 4))
- : spreg);
- rtx memref = gen_rtx_MEM (word_mode, addr);
+ if (must_save_p (is_inthandler, regno))
+ {
+ emit_move_insn (gen_rtx_REG (word_mode, regno), postinc);
+ ndregs--;
+ }
+ }
- regno--;
- XVECEXP (pat, 0, i + 1)
- = gen_rtx_SET (VOIDmode, gen_rtx_REG (word_mode, regno), memref);
+ if (total_consec != 0)
+ {
+ rtx pat = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (total_consec + 1));
+ XVECEXP (pat, 0, 0)
+ = gen_rtx_SET (VOIDmode, spreg,
+ gen_rtx_PLUS (Pmode, spreg,
+ GEN_INT (total_consec * 4)));
+
+ if (npregs_consec > 0)
+ regno = REG_P5 + 1;
+ else
+ regno = REG_R7 + 1;
- if (npregs > 0)
+ for (i = 0; i < total_consec; i++)
{
- if (--npregs == 0)
- regno = REG_R7 + 1;
+ rtx addr = (i > 0
+ ? gen_rtx_PLUS (Pmode, spreg, GEN_INT (i * 4))
+ : spreg);
+ rtx memref = gen_rtx_MEM (word_mode, addr);
+
+ regno--;
+ XVECEXP (pat, 0, i + 1)
+ = gen_rtx_SET (VOIDmode, gen_rtx_REG (word_mode, regno), memref);
+
+ if (npregs_consec > 0)
+ {
+ if (--npregs_consec == 0)
+ regno = REG_R7 + 1;
+ }
}
- }
- insn = emit_insn (pat);
- RTX_FRAME_RELATED_P (insn) = 1;
+ insn = emit_insn (pat);
+ RTX_FRAME_RELATED_P (insn) = 1;
+ }
+ if (saveall || is_inthandler)
+ emit_move_insn (gen_rtx_REG (SImode, REG_ASTAT), postinc);
}
/* Perform any needed actions needed for a function that is receiving a
tree attrs = TYPE_ATTRIBUTES (TREE_TYPE (current_function_decl));
bool all = (lookup_attribute ("saveall", attrs) != NULL_TREE
|| (is_inthandler && !current_function_is_leaf));
- int ndregs = all ? 8 : n_dregs_to_save (is_inthandler);
- int npregs = all ? 6 : n_pregs_to_save (is_inthandler);
+ int ndregs = all ? 8 : n_dregs_to_save (is_inthandler, false);
+ int npregs = all ? 6 : n_pregs_to_save (is_inthandler, false);
int n = ndregs + npregs;
+ int i;
if (all || stack_frame_needed_p ())
/* We use a LINK instruction in this case. */
n++;
}
+ if (fkind != SUBROUTINE || all)
+ /* Increment once for ASTAT. */
+ n++;
+
if (fkind != SUBROUTINE)
{
- int i;
-
- /* Increment once for ASTAT. */
- n++;
-
/* RETE/X/N. */
if (lookup_attribute ("nesting", attrs))
n++;
-
- for (i = REG_P7 + 1; i < REG_CC; i++)
- if (all
- || regs_ever_live[i]
- || (!leaf_function_p () && call_used_regs[i]))
- n += i == REG_A0 || i == REG_A1 ? 2 : 1;
}
+
+ for (i = REG_P7 + 1; i < REG_CC; i++)
+ if (all
+ || (fkind != SUBROUTINE
+ && (df_regs_ever_live_p (i)
+ || (!leaf_function_p () && call_used_regs[i]))))
+ n += i == REG_A0 || i == REG_A1 ? 2 : 1;
+
return n;
}
if (to == STACK_POINTER_REGNUM)
{
- if (current_function_outgoing_args_size >= FIXED_STACK_AREA)
- offset += current_function_outgoing_args_size;
- else if (current_function_outgoing_args_size)
+ if (crtl->outgoing_args_size >= FIXED_STACK_AREA)
+ offset += crtl->outgoing_args_size;
+ else if (crtl->outgoing_args_size)
offset += FIXED_STACK_AREA;
offset += get_frame_size ();
RTX_FRAME_RELATED_P (insn) = 1;
}
-/* Generate efficient code to add a value to a P register. We can use
- P1 as a scratch register. Set RTX_FRAME_RELATED_P on the generated
- insns if FRAME is nonzero. */
+/* Generate efficient code to add a value to a P register.
+ Set RTX_FRAME_RELATED_P on the generated insns if FRAME is nonzero.
+ EPILOGUE_P is zero if this function is called for prologue,
+ otherwise it's nonzero. And it's less than zero if this is for
+ sibcall epilogue. */
static void
-add_to_reg (rtx reg, HOST_WIDE_INT value, int frame)
+add_to_reg (rtx reg, HOST_WIDE_INT value, int frame, int epilogue_p)
{
if (value == 0)
return;
/* Choose whether to use a sequence using a temporary register, or
- a sequence with multiple adds. We can add a signed 7 bit value
+ a sequence with multiple adds. We can add a signed 7-bit value
in one instruction. */
if (value > 120 || value < -120)
{
- rtx tmpreg = gen_rtx_REG (SImode, REG_P1);
+ rtx tmpreg;
+ rtx tmpreg2;
rtx insn;
+ tmpreg2 = NULL_RTX;
+
+ /* For prologue or normal epilogue, P1 can be safely used
+ as the temporary register. For sibcall epilogue, we try to find
+ a call used P register, which will be restored in epilogue.
+ If we cannot find such a P register, we have to use one I register
+ to help us. */
+
+ if (epilogue_p >= 0)
+ tmpreg = gen_rtx_REG (SImode, REG_P1);
+ else
+ {
+ int i;
+ for (i = REG_P0; i <= REG_P5; i++)
+ if ((df_regs_ever_live_p (i) && ! call_used_regs[i])
+ || (!TARGET_FDPIC
+ && i == PIC_OFFSET_TABLE_REGNUM
+ && (crtl->uses_pic_offset_table
+ || (TARGET_ID_SHARED_LIBRARY
+ && ! current_function_is_leaf))))
+ break;
+ if (i <= REG_P5)
+ tmpreg = gen_rtx_REG (SImode, i);
+ else
+ {
+ tmpreg = gen_rtx_REG (SImode, REG_P1);
+ tmpreg2 = gen_rtx_REG (SImode, REG_I0);
+ emit_move_insn (tmpreg2, tmpreg);
+ }
+ }
+
if (frame)
frame_related_constant_load (tmpreg, value, TRUE);
else
insn = emit_insn (gen_addsi3 (reg, reg, tmpreg));
if (frame)
RTX_FRAME_RELATED_P (insn) = 1;
+
+ if (tmpreg2 != NULL_RTX)
+ emit_move_insn (tmpreg, tmpreg2);
}
else
do
static HOST_WIDE_INT
arg_area_size (void)
{
- if (current_function_outgoing_args_size)
+ if (crtl->outgoing_args_size)
{
- if (current_function_outgoing_args_size >= FIXED_STACK_AREA)
- return current_function_outgoing_args_size;
+ if (crtl->outgoing_args_size >= FIXED_STACK_AREA)
+ return crtl->outgoing_args_size;
else
return FIXED_STACK_AREA;
}
rtx insn = emit_insn (pat);
RTX_FRAME_RELATED_P (insn) = 1;
}
- add_to_reg (spreg, -frame_size, 1);
+ add_to_reg (spreg, -frame_size, 1, 0);
}
}
-/* Like do_link, but used for epilogues to deallocate the stack frame. */
+/* Like do_link, but used for epilogues to deallocate the stack frame.
+ EPILOGUE_P is zero if this function is called for prologue,
+ otherwise it's nonzero. And it's less than zero if this is for
+ sibcall epilogue. */
static void
-do_unlink (rtx spreg, HOST_WIDE_INT frame_size, bool all)
+do_unlink (rtx spreg, HOST_WIDE_INT frame_size, bool all, int epilogue_p)
{
frame_size += arg_area_size ();
{
rtx postinc = gen_rtx_MEM (Pmode, gen_rtx_POST_INC (Pmode, spreg));
- add_to_reg (spreg, frame_size, 0);
+ add_to_reg (spreg, frame_size, 0, epilogue_p);
if (must_save_fp_p ())
{
rtx fpreg = gen_rtx_REG (Pmode, REG_FP);
emit_move_insn (fpreg, postinc);
- emit_insn (gen_rtx_USE (VOIDmode, fpreg));
+ emit_use (fpreg);
}
if (! current_function_is_leaf)
{
emit_move_insn (bfin_rets_rtx, postinc);
- emit_insn (gen_rtx_USE (VOIDmode, bfin_rets_rtx));
+ emit_use (bfin_rets_rtx);
}
}
}
SPREG contains (reg:SI REG_SP). */
static void
-expand_interrupt_handler_prologue (rtx spreg, e_funkind fkind)
+expand_interrupt_handler_prologue (rtx spreg, e_funkind fkind, bool all)
{
- int i;
HOST_WIDE_INT frame_size = get_frame_size ();
rtx predec1 = gen_rtx_PRE_DEC (SImode, spreg);
rtx predec = gen_rtx_MEM (SImode, predec1);
rtx insn;
tree attrs = TYPE_ATTRIBUTES (TREE_TYPE (current_function_decl));
- bool all = lookup_attribute ("saveall", attrs) != NULL_TREE;
tree kspisusp = lookup_attribute ("kspisusp", attrs);
if (kspisusp)
RTX_FRAME_RELATED_P (insn) = 1;
}
- insn = emit_move_insn (predec, gen_rtx_REG (SImode, REG_ASTAT));
- RTX_FRAME_RELATED_P (insn) = 1;
-
/* If we're calling other functions, they won't save their call-clobbered
registers, so we must save everything here. */
if (!current_function_is_leaf)
all = true;
expand_prologue_reg_save (spreg, all, true);
- for (i = REG_P7 + 1; i < REG_CC; i++)
- if (all
- || regs_ever_live[i]
- || (!leaf_function_p () && call_used_regs[i]))
- {
- if (i == REG_A0 || i == REG_A1)
- insn = emit_move_insn (gen_rtx_MEM (PDImode, predec1),
- gen_rtx_REG (PDImode, i));
- else
- insn = emit_move_insn (predec, gen_rtx_REG (SImode, i));
- RTX_FRAME_RELATED_P (insn) = 1;
- }
-
if (lookup_attribute ("nesting", attrs))
{
rtx srcreg = gen_rtx_REG (Pmode, (fkind == EXCPT_HANDLER ? REG_RETX
rtx insn;
insn = emit_move_insn (r0reg, gen_rtx_REG (SImode, REG_SEQSTAT));
- REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_MAYBE_DEAD, const0_rtx,
- NULL_RTX);
insn = emit_insn (gen_ashrsi3 (r0reg, r0reg, GEN_INT (26)));
- REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_MAYBE_DEAD, const0_rtx,
- NULL_RTX);
insn = emit_insn (gen_ashlsi3 (r0reg, r0reg, GEN_INT (26)));
- REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_MAYBE_DEAD, const0_rtx,
- NULL_RTX);
insn = emit_move_insn (r1reg, spreg);
- REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_MAYBE_DEAD, const0_rtx,
- NULL_RTX);
insn = emit_move_insn (r2reg, gen_rtx_REG (Pmode, REG_FP));
- REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_MAYBE_DEAD, const0_rtx,
- NULL_RTX);
insn = emit_insn (gen_addsi3 (r2reg, r2reg, GEN_INT (8)));
- REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_MAYBE_DEAD, const0_rtx,
- NULL_RTX);
}
}
SPREG contains (reg:SI REG_SP). */
static void
-expand_interrupt_handler_epilogue (rtx spreg, e_funkind fkind)
+expand_interrupt_handler_epilogue (rtx spreg, e_funkind fkind, bool all)
{
- int i;
+ tree attrs = TYPE_ATTRIBUTES (TREE_TYPE (current_function_decl));
rtx postinc1 = gen_rtx_POST_INC (SImode, spreg);
rtx postinc = gen_rtx_MEM (SImode, postinc1);
- tree attrs = TYPE_ATTRIBUTES (TREE_TYPE (current_function_decl));
- bool all = lookup_attribute ("saveall", attrs) != NULL_TREE;
/* A slightly crude technique to stop flow from trying to delete "dead"
insns. */
MEM_VOLATILE_P (postinc) = 1;
- do_unlink (spreg, get_frame_size (), all);
+ do_unlink (spreg, get_frame_size (), all, 1);
if (lookup_attribute ("nesting", attrs))
{
if (!current_function_is_leaf)
all = true;
- for (i = REG_CC - 1; i > REG_P7; i--)
- if (all
- || regs_ever_live[i]
- || (!leaf_function_p () && call_used_regs[i]))
- {
- if (i == REG_A0 || i == REG_A1)
- {
- rtx mem = gen_rtx_MEM (PDImode, postinc1);
- MEM_VOLATILE_P (mem) = 1;
- emit_move_insn (gen_rtx_REG (PDImode, i), mem);
- }
- else
- emit_move_insn (gen_rtx_REG (SImode, i), postinc);
- }
-
expand_epilogue_reg_restore (spreg, all, true);
- emit_move_insn (gen_rtx_REG (SImode, REG_ASTAT), postinc);
-
/* Deallocate any space we left on the stack in case we needed to save the
argument registers. */
if (fkind == EXCPT_HANDLER)
struct cgraph_local_info *i = NULL;
rtx addr, insn;
- if (flag_unit_at_a_time)
- i = cgraph_local_info (current_function_decl);
+ i = cgraph_local_info (current_function_decl);
/* Functions local to the translation unit don't need to reload the
pic reg, since the caller always passes a usable one. */
gen_rtx_UNSPEC (Pmode, gen_rtvec (1, const0_rtx),
UNSPEC_LIBRARY_OFFSET));
insn = emit_insn (gen_movsi (dest, gen_rtx_MEM (Pmode, addr)));
- REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_MAYBE_DEAD, const0_rtx, NULL);
return dest;
}
void
bfin_expand_prologue (void)
{
- rtx insn;
HOST_WIDE_INT frame_size = get_frame_size ();
rtx spreg = gen_rtx_REG (Pmode, REG_SP);
e_funkind fkind = funkind (TREE_TYPE (current_function_decl));
rtx pic_reg_loaded = NULL_RTX;
+ tree attrs = TYPE_ATTRIBUTES (TREE_TYPE (current_function_decl));
+ bool all = lookup_attribute ("saveall", attrs) != NULL_TREE;
if (fkind != SUBROUTINE)
{
- expand_interrupt_handler_prologue (spreg, fkind);
+ expand_interrupt_handler_prologue (spreg, fkind, all);
return;
}
- if (current_function_limit_stack
- || TARGET_STACK_CHECK_L1)
+ if (crtl->limit_stack
+ || (TARGET_STACK_CHECK_L1
+ && !DECL_NO_LIMIT_STACK (current_function_decl)))
{
HOST_WIDE_INT offset
= bfin_initial_elimination_offset (ARG_POINTER_REGNUM,
STACK_POINTER_REGNUM);
- rtx lim = current_function_limit_stack ? stack_limit_rtx : NULL_RTX;
+ rtx lim = crtl->limit_stack ? stack_limit_rtx : NULL_RTX;
rtx p2reg = gen_rtx_REG (Pmode, REG_P2);
if (!lim)
{
- rtx p1reg = gen_rtx_REG (Pmode, REG_P1);
emit_move_insn (p2reg, gen_int_mode (0xFFB00000, SImode));
emit_move_insn (p2reg, gen_rtx_MEM (Pmode, p2reg));
lim = p2reg;
{
if (lim != p2reg)
emit_move_insn (p2reg, lim);
- add_to_reg (p2reg, offset, 0);
+ add_to_reg (p2reg, offset, 0, 0);
lim = p2reg;
}
emit_insn (gen_compare_lt (bfin_cc_rtx, spreg, lim));
emit_insn (gen_trapifcc ());
}
- expand_prologue_reg_save (spreg, 0, false);
+ expand_prologue_reg_save (spreg, all, false);
do_link (spreg, frame_size, false);
if (TARGET_ID_SHARED_LIBRARY
&& !TARGET_SEP_DATA
- && (current_function_uses_pic_offset_table
+ && (crtl->uses_pic_offset_table
|| !current_function_is_leaf))
bfin_load_pic_reg (pic_offset_table_rtx);
}
/* Generate RTL for the epilogue of the current function. NEED_RETURN is zero
if this is for a sibcall. EH_RETURN is nonzero if we're expanding an
- eh_return pattern. */
+ eh_return pattern. SIBCALL_P is true if this is a sibcall epilogue,
+ false otherwise. */
void
-bfin_expand_epilogue (int need_return, int eh_return)
+bfin_expand_epilogue (int need_return, int eh_return, bool sibcall_p)
{
rtx spreg = gen_rtx_REG (Pmode, REG_SP);
e_funkind fkind = funkind (TREE_TYPE (current_function_decl));
+ int e = sibcall_p ? -1 : 1;
+ tree attrs = TYPE_ATTRIBUTES (TREE_TYPE (current_function_decl));
+ bool all = lookup_attribute ("saveall", attrs) != NULL_TREE;
if (fkind != SUBROUTINE)
{
- expand_interrupt_handler_epilogue (spreg, fkind);
+ expand_interrupt_handler_epilogue (spreg, fkind, all);
return;
}
- do_unlink (spreg, get_frame_size (), false);
+ do_unlink (spreg, get_frame_size (), false, e);
- expand_epilogue_reg_restore (spreg, false, false);
+ expand_epilogue_reg_restore (spreg, all, false);
/* Omit the return insn if this is for a sibcall. */
if (! need_return)
call-clobbered. */
if (funkind (TREE_TYPE (current_function_decl)) != SUBROUTINE
- && !regs_ever_live[new_reg])
+ && !df_regs_ever_live_p (new_reg))
return 0;
return 1;
static rtx
bfin_delegitimize_address (rtx orig_x)
{
- rtx x = orig_x, y;
+ rtx x = orig_x;
if (GET_CODE (x) != MEM)
return orig_x;
/* This predicate is used to compute the length of a load/store insn.
OP is a MEM rtx, we return nonzero if its addressing mode requires a
- 32 bit instruction. */
+ 32-bit instruction. */
int
effective_address_32bit_p (rtx op, enum machine_mode mode)
return 0;
}
+ if (GET_CODE (XEXP (op, 1)) == UNSPEC)
+ return 1;
+
offset = INTVAL (XEXP (op, 1));
- /* All byte loads use a 16 bit offset. */
+ /* All byte loads use a 16-bit offset. */
if (GET_MODE_SIZE (mode) == 1)
return 1;
All addressing modes are equally cheap on the Blackfin. */
static int
-bfin_address_cost (rtx addr ATTRIBUTE_UNUSED)
+bfin_address_cost (rtx addr ATTRIBUTE_UNUSED, bool speed ATTRIBUTE_UNUSED)
{
return 1;
}
case REG:
if (code == 'h')
{
- gcc_assert (REGNO (x) < 32);
- fprintf (file, "%s", short_reg_names[REGNO (x)]);
- /*fprintf (file, "\n%d\n ", REGNO (x));*/
- break;
+ if (REGNO (x) < 32)
+ fprintf (file, "%s", short_reg_names[REGNO (x)]);
+ else
+ output_operand_lossage ("invalid operand for code '%c'", code);
}
else if (code == 'd')
{
- gcc_assert (REGNO (x) < 32);
- fprintf (file, "%s", high_reg_names[REGNO (x)]);
- break;
+ if (REGNO (x) < 32)
+ fprintf (file, "%s", high_reg_names[REGNO (x)]);
+ else
+ output_operand_lossage ("invalid operand for code '%c'", code);
}
else if (code == 'w')
{
- gcc_assert (REGNO (x) == REG_A0 || REGNO (x) == REG_A1);
- fprintf (file, "%s.w", reg_names[REGNO (x)]);
+ if (REGNO (x) == REG_A0 || REGNO (x) == REG_A1)
+ fprintf (file, "%s.w", reg_names[REGNO (x)]);
+ else
+ output_operand_lossage ("invalid operand for code '%c'", code);
}
else if (code == 'x')
{
- gcc_assert (REGNO (x) == REG_A0 || REGNO (x) == REG_A1);
- fprintf (file, "%s.x", reg_names[REGNO (x)]);
+ if (REGNO (x) == REG_A0 || REGNO (x) == REG_A1)
+ fprintf (file, "%s.x", reg_names[REGNO (x)]);
+ else
+ output_operand_lossage ("invalid operand for code '%c'", code);
+ }
+ else if (code == 'v')
+ {
+ if (REGNO (x) == REG_A0)
+ fprintf (file, "AV0");
+ else if (REGNO (x) == REG_A1)
+ fprintf (file, "AV1");
+ else
+ output_operand_lossage ("invalid operand for code '%c'", code);
}
else if (code == 'D')
{
- fprintf (file, "%s", dregs_pair_names[REGNO (x)]);
+ if (D_REGNO_P (REGNO (x)))
+ fprintf (file, "%s", dregs_pair_names[REGNO (x)]);
+ else
+ output_operand_lossage ("invalid operand for code '%c'", code);
}
else if (code == 'H')
{
- gcc_assert (mode == DImode || mode == DFmode);
- gcc_assert (REG_P (x));
- fprintf (file, "%s", reg_names[REGNO (x) + 1]);
+ if ((mode == DImode || mode == DFmode) && REG_P (x))
+ fprintf (file, "%s", reg_names[REGNO (x) + 1]);
+ else
+ output_operand_lossage ("invalid operand for code '%c'", code);
}
else if (code == 'T')
{
- gcc_assert (D_REGNO_P (REGNO (x)));
- fprintf (file, "%s", byte_reg_names[REGNO (x)]);
+ if (D_REGNO_P (REGNO (x)))
+ fprintf (file, "%s", byte_reg_names[REGNO (x)]);
+ else
+ output_operand_lossage ("invalid operand for code '%c'", code);
}
else
fprintf (file, "%s", reg_names[REGNO (x)]);
case MACFLAG_M:
fputs ("(M)", file);
break;
+ case MACFLAG_IS_M:
+ fputs ("(IS,M)", file);
+ break;
case MACFLAG_ISS2:
fputs ("(ISS2)", file);
break;
x = GEN_INT ((INTVAL (x) >> 16) & 0xffff);
else if (code == 'h')
x = GEN_INT (INTVAL (x) & 0xffff);
+ else if (code == 'N')
+ x = GEN_INT (-INTVAL (x));
else if (code == 'X')
x = GEN_INT (exact_log2 (0xffffffff & INTVAL (x)));
else if (code == 'Y')
For args passed entirely in registers or entirely in memory, zero.
Refer VDSP C Compiler manual, our ABI.
- First 3 words are in registers. So, if a an argument is larger
+ First 3 words are in registers. So, if an argument is larger
than the registers available, it will span the register and
stack. */
static bool
bfin_pass_by_reference (CUMULATIVE_ARGS *cum ATTRIBUTE_UNUSED,
enum machine_mode mode ATTRIBUTE_UNUSED,
- tree type, bool named ATTRIBUTE_UNUSED)
+ const_tree type, bool named ATTRIBUTE_UNUSED)
{
return type && TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST;
}
/* Decide whether a type should be returned in memory (true)
or in a register (false). This is called by the macro
- RETURN_IN_MEMORY. */
+ TARGET_RETURN_IN_MEMORY. */
-int
-bfin_return_in_memory (tree type)
+static bool
+bfin_return_in_memory (const_tree type, const_tree fntype ATTRIBUTE_UNUSED)
{
int size = int_size_in_bytes (type);
return size > 2 * UNITS_PER_WORD || size == -1;
bfin_function_ok_for_sibcall (tree decl ATTRIBUTE_UNUSED,
tree exp ATTRIBUTE_UNUSED)
{
+ struct cgraph_local_info *this_func, *called_func;
e_funkind fkind = funkind (TREE_TYPE (current_function_decl));
if (fkind != SUBROUTINE)
return false;
not need to reload P5 in the prologue, but the sibcall wil pop P5 in the
sibcall epilogue, and we end up with the wrong value in P5. */
- if (!flag_unit_at_a_time || decl == NULL)
+ if (!decl)
/* Not enough information. */
return false;
-
- {
- struct cgraph_local_info *this_func, *called_func;
- rtx addr, insn;
- this_func = cgraph_local_info (current_function_decl);
- called_func = cgraph_local_info (decl);
- return !called_func->local || this_func->local;
- }
+ this_func = cgraph_local_info (current_function_decl);
+ called_func = cgraph_local_info (decl);
+ return !called_func->local || this_func->local;
}
\f
/* Emit RTL insns to initialize the variable parts of a trampoline at
code. CXT is an RTX for the static chain value for the function. */
void
-initialize_trampoline (tramp, fnaddr, cxt)
- rtx tramp, fnaddr, cxt;
+initialize_trampoline (rtx tramp, rtx fnaddr, rtx cxt)
{
rtx t1 = copy_to_reg (fnaddr);
rtx t2 = copy_to_reg (cxt);
if (TARGET_FDPIC)
{
+ int caller_has_l1_text, callee_has_l1_text;
+
+ caller_has_l1_text = callee_has_l1_text = 0;
+
+ if (lookup_attribute ("l1_text",
+ DECL_ATTRIBUTES (cfun->decl)) != NULL_TREE)
+ caller_has_l1_text = 1;
+
+ if (GET_CODE (callee) == SYMBOL_REF
+ && SYMBOL_REF_DECL (callee) && DECL_P (SYMBOL_REF_DECL (callee))
+ && lookup_attribute
+ ("l1_text",
+ DECL_ATTRIBUTES (SYMBOL_REF_DECL (callee))) != NULL_TREE)
+ callee_has_l1_text = 1;
+
if (GET_CODE (callee) != SYMBOL_REF
- || bfin_longcall_p (callee, INTVAL (cookie)))
+ || bfin_longcall_p (callee, INTVAL (cookie))
+ || (GET_CODE (callee) == SYMBOL_REF
+ && !SYMBOL_REF_LOCAL_P (callee)
+ && TARGET_INLINE_PLT)
+ || caller_has_l1_text != callee_has_l1_text
+ || (caller_has_l1_text && callee_has_l1_text
+ && (GET_CODE (callee) != SYMBOL_REF
+ || !SYMBOL_REF_LOCAL_P (callee))))
{
rtx addr = callee;
if (! address_operand (addr, Pmode))
hard_regno_mode_ok (int regno, enum machine_mode mode)
{
/* Allow only dregs to store value of mode HI or QI */
- enum reg_class class = REGNO_REG_CLASS (regno);
+ enum reg_class rclass = REGNO_REG_CLASS (regno);
if (mode == CCmode)
return 0;
if (mode == V2HImode)
return D_REGNO_P (regno);
- if (class == CCREGS)
+ if (rclass == CCREGS)
return mode == BImode;
if (mode == PDImode || mode == V2PDImode)
return regno == REG_A0 || regno == REG_A1;
- /* Allow all normal 32 bit regs, except REG_M3, in case regclass ever comes
+ /* Allow all normal 32-bit regs, except REG_M3, in case regclass ever comes
up with a bad register class (such as ALL_REGS) for DImode. */
if (mode == DImode)
return regno < REG_M3;
int
bfin_memory_move_cost (enum machine_mode mode ATTRIBUTE_UNUSED,
- enum reg_class class,
+ enum reg_class rclass,
int in ATTRIBUTE_UNUSED)
{
/* Make memory accesses slightly more expensive than any register-register
move. Also, penalize non-DP registers, since they need secondary
reloads to load and store. */
- if (! reg_class_subset_p (class, DPREGS))
+ if (! reg_class_subset_p (rclass, DPREGS))
return 10;
return 8;
}
/* Inform reload about cases where moving X with a mode MODE to a register in
- CLASS requires an extra scratch register. Return the class needed for the
+ RCLASS requires an extra scratch register. Return the class needed for the
scratch register. */
static enum reg_class
-bfin_secondary_reload (bool in_p, rtx x, enum reg_class class,
- enum machine_mode mode, secondary_reload_info *sri)
+bfin_secondary_reload (bool in_p, rtx x, enum reg_class rclass,
+ enum machine_mode mode, secondary_reload_info *sri)
{
/* If we have HImode or QImode, we can only use DREGS as secondary registers;
in most other cases we can also use PREGS. */
if (fp_plus_const_operand (x, mode))
{
rtx op2 = XEXP (x, 1);
- int large_constant_p = ! CONST_7BIT_IMM_P (INTVAL (op2));
+ int large_constant_p = ! satisfies_constraint_Ks7 (op2);
- if (class == PREGS || class == PREGS_CLOBBERED)
+ if (rclass == PREGS || rclass == PREGS_CLOBBERED)
return NO_REGS;
/* If destination is a DREG, we can do this without a scratch register
if the constant is valid for an add instruction. */
- if ((class == DREGS || class == DPREGS)
+ if ((rclass == DREGS || rclass == DPREGS)
&& ! large_constant_p)
return NO_REGS;
/* Reloading to anything other than a DREG? Use a PREG scratch
/* Data can usually be moved freely between registers of most classes.
AREGS are an exception; they can only move to or from another register
in AREGS or one in DREGS. They can also be assigned the constant 0. */
- if (x_class == AREGS)
- return class == DREGS || class == AREGS ? NO_REGS : DREGS;
+ if (x_class == AREGS || x_class == EVEN_AREGS || x_class == ODD_AREGS)
+ return (rclass == DREGS || rclass == AREGS || rclass == EVEN_AREGS
+ || rclass == ODD_AREGS
+ ? NO_REGS : DREGS);
- if (class == AREGS)
+ if (rclass == AREGS || rclass == EVEN_AREGS || rclass == ODD_AREGS)
{
+ if (code == MEM)
+ {
+ sri->icode = in_p ? CODE_FOR_reload_inpdi : CODE_FOR_reload_outpdi;
+ return NO_REGS;
+ }
+
if (x != const0_rtx && x_class != DREGS)
- return DREGS;
+ {
+ return DREGS;
+ }
else
return NO_REGS;
}
/* CCREGS can only be moved from/to DREGS. */
- if (class == CCREGS && x_class != DREGS)
+ if (rclass == CCREGS && x_class != DREGS)
return DREGS;
- if (x_class == CCREGS && class != DREGS)
+ if (x_class == CCREGS && rclass != DREGS)
return DREGS;
/* All registers other than AREGS can load arbitrary constants. The only
case that remains is MEM. */
if (code == MEM)
- if (! reg_class_subset_p (class, default_class))
+ if (! reg_class_subset_p (rclass, default_class))
return default_class;
+
return NO_REGS;
}
\f
bfin_lib_id_given = 1;
return true;
+ case OPT_mcpu_:
+ {
+ const char *p, *q;
+ int i;
+
+ i = 0;
+ while ((p = bfin_cpus[i].name) != NULL)
+ {
+ if (strncmp (arg, p, strlen (p)) == 0)
+ break;
+ i++;
+ }
+
+ if (p == NULL)
+ {
+ error ("-mcpu=%s is not valid", arg);
+ return false;
+ }
+
+ bfin_cpu_type = bfin_cpus[i].type;
+
+ q = arg + strlen (p);
+
+ if (*q == '\0')
+ {
+ bfin_si_revision = bfin_cpus[i].si_revision;
+ bfin_workarounds |= bfin_cpus[i].workarounds;
+ }
+ else if (strcmp (q, "-none") == 0)
+ bfin_si_revision = -1;
+ else if (strcmp (q, "-any") == 0)
+ {
+ bfin_si_revision = 0xffff;
+ while (bfin_cpus[i].type == bfin_cpu_type)
+ {
+ bfin_workarounds |= bfin_cpus[i].workarounds;
+ i++;
+ }
+ }
+ else
+ {
+ unsigned int si_major, si_minor;
+ int rev_len, n;
+
+ rev_len = strlen (q);
+
+ if (sscanf (q, "-%u.%u%n", &si_major, &si_minor, &n) != 2
+ || n != rev_len
+ || si_major > 0xff || si_minor > 0xff)
+ {
+ invalid_silicon_revision:
+ error ("-mcpu=%s has invalid silicon revision", arg);
+ return false;
+ }
+
+ bfin_si_revision = (si_major << 8) | si_minor;
+
+ while (bfin_cpus[i].type == bfin_cpu_type
+ && bfin_cpus[i].si_revision != bfin_si_revision)
+ i++;
+
+ if (bfin_cpus[i].type != bfin_cpu_type)
+ goto invalid_silicon_revision;
+
+ bfin_workarounds |= bfin_cpus[i].workarounds;
+ }
+
+ return true;
+ }
+
default:
return true;
}
{
struct machine_function *f;
- f = ggc_alloc_cleared (sizeof (struct machine_function));
+ f = GGC_CNEW (struct machine_function);
return f;
}
void
override_options (void)
{
+ /* If processor type is not specified, enable all workarounds. */
+ if (bfin_cpu_type == BFIN_CPU_UNKNOWN)
+ {
+ int i;
+
+ for (i = 0; bfin_cpus[i].name != NULL; i++)
+ bfin_workarounds |= bfin_cpus[i].workarounds;
+
+ bfin_si_revision = 0xffff;
+ }
+
+ if (bfin_csync_anomaly == 1)
+ bfin_workarounds |= WA_SPECULATIVE_SYNCS;
+ else if (bfin_csync_anomaly == 0)
+ bfin_workarounds &= ~WA_SPECULATIVE_SYNCS;
+
+ if (bfin_specld_anomaly == 1)
+ bfin_workarounds |= WA_SPECULATIVE_LOADS;
+ else if (bfin_specld_anomaly == 0)
+ bfin_workarounds &= ~WA_SPECULATIVE_LOADS;
+
if (TARGET_OMIT_LEAF_FRAME_POINTER)
flag_omit_frame_pointer = 1;
if (bfin_lib_id_given && ! TARGET_ID_SHARED_LIBRARY)
error ("-mshared-library-id= specified without -mid-shared-library");
- if (TARGET_ID_SHARED_LIBRARY && flag_pic == 0)
- flag_pic = 1;
-
if (stack_limit_rtx && TARGET_STACK_CHECK_L1)
error ("Can't use multiple stack checking methods together.");
if (TARGET_SEP_DATA)
target_flags |= MASK_ID_SHARED_LIBRARY | MASK_LEAF_ID_SHARED_LIBRARY;
+ if (TARGET_ID_SHARED_LIBRARY && flag_pic == 0)
+ flag_pic = 1;
+
/* There is no single unaligned SI op for PIC code. Sometimes we
need to use ".4byte" and sometimes we need to use ".picptr".
See bfin_assemble_integer for details. */
if (flag_pic && !TARGET_FDPIC && !TARGET_ID_SHARED_LIBRARY)
flag_pic = 0;
+ if (TARGET_MULTICORE && bfin_cpu_type != BFIN_CPU_BF561)
+ error ("-mmulticore can only be used with BF561");
+
+ if (TARGET_COREA && !TARGET_MULTICORE)
+ error ("-mcorea should be used with -mmulticore");
+
+ if (TARGET_COREB && !TARGET_MULTICORE)
+ error ("-mcoreb should be used with -mmulticore");
+
+ if (TARGET_COREA && TARGET_COREB)
+ error ("-mcorea and -mcoreb can't be used together");
+
flag_schedule_insns = 0;
/* Passes after sched2 can break the helpful TImode annotations that
}
\f
/* Return nonzero iff C has exactly one bit set if it is interpreted
- as a 32 bit constant. */
+ as a 32-bit constant. */
int
log2constp (unsigned HOST_WIDE_INT c)
int num_zero = shiftr_zero (&shifted);
int num_compl_zero = shiftr_zero (&shifted_compl);
unsigned int regno = REGNO (operands[0]);
- enum reg_class class1 = REGNO_REG_CLASS (regno);
/* This case takes care of single-bit set/clear constants, which we could
also implement with BITSET/BITCLR. */
if (D_REGNO_P (regno))
{
- if (CONST_7BIT_IMM_P (tmp))
+ if (tmp >= -64 && tmp <= 63)
{
emit_insn (gen_movsi (operands[0], GEN_INT (tmp)));
emit_insn (gen_movstricthi_high (operands[0], GEN_INT (val & -65536)));
return 0;
if (optimize_size
- && num_compl_zero && CONST_7BIT_IMM_P (shifted_compl))
+ && num_compl_zero && shifted_compl >= -64 && shifted_compl <= 63)
{
/* If optimizing for size, generate a sequence that has more instructions
but is shorter. */
}
static bool
-bfin_rtx_costs (rtx x, int code, int outer_code, int *total)
+bfin_rtx_costs (rtx x, int code, int outer_code, int *total, bool speed)
{
int cost2 = COSTS_N_INSNS (1);
rtx op0, op1;
{
case CONST_INT:
if (outer_code == SET || outer_code == PLUS)
- *total = CONST_7BIT_IMM_P (INTVAL (x)) ? 0 : cost2;
+ *total = satisfies_constraint_Ks7 (x) ? 0 : cost2;
else if (outer_code == AND)
*total = log2constp (~INTVAL (x)) ? 0 : cost2;
else if (outer_code == LE || outer_code == LT || outer_code == EQ)
if (val == 2 || val == 4)
{
*total = cost2;
- *total += rtx_cost (XEXP (op0, 0), outer_code);
- *total += rtx_cost (op1, outer_code);
+ *total += rtx_cost (XEXP (op0, 0), outer_code, speed);
+ *total += rtx_cost (op1, outer_code, speed);
return true;
}
}
*total = cost2;
if (GET_CODE (op0) != REG
&& (GET_CODE (op0) != SUBREG || GET_CODE (SUBREG_REG (op0)) != REG))
- *total += rtx_cost (op0, SET);
+ *total += rtx_cost (op0, SET, speed);
#if 0 /* We'd like to do this for accuracy, but it biases the loop optimizer
towards creating too many induction variables. */
if (!reg_or_7bit_operand (op1, SImode))
- *total += rtx_cost (op1, SET);
+ *total += rtx_cost (op1, SET, speed);
#endif
}
else if (GET_MODE (x) == DImode)
{
*total = 6 * cost2;
if (GET_CODE (op1) != CONST_INT
- || !CONST_7BIT_IMM_P (INTVAL (op1)))
- *total += rtx_cost (op1, PLUS);
+ || !satisfies_constraint_Ks7 (op1))
+ *total += rtx_cost (op1, PLUS, speed);
if (GET_CODE (op0) != REG
&& (GET_CODE (op0) != SUBREG || GET_CODE (SUBREG_REG (op0)) != REG))
- *total += rtx_cost (op0, PLUS);
+ *total += rtx_cost (op0, PLUS, speed);
}
return true;
op1 = XEXP (x, 1);
if (GET_CODE (op0) != REG
&& (GET_CODE (op0) != SUBREG || GET_CODE (SUBREG_REG (op0)) != REG))
- *total += rtx_cost (op0, code);
+ *total += rtx_cost (op0, code, speed);
return true;
if (GET_CODE (op0) != REG
&& (GET_CODE (op0) != SUBREG || GET_CODE (SUBREG_REG (op0)) != REG))
- *total += rtx_cost (op0, code);
+ *total += rtx_cost (op0, code, speed);
if (GET_MODE (x) == DImode)
{
if (code == AND)
{
if (! rhs_andsi3_operand (XEXP (x, 1), SImode))
- *total += rtx_cost (XEXP (x, 1), code);
+ *total += rtx_cost (XEXP (x, 1), code, speed);
}
else
{
if (! regorlog2_operand (XEXP (x, 1), SImode))
- *total += rtx_cost (XEXP (x, 1), code);
+ *total += rtx_cost (XEXP (x, 1), code, speed);
}
return true;
op0 = XEXP (op0, 0);
op1 = XEXP (op1, 0);
}
- else if (optimize_size)
+ else if (!speed)
*total = COSTS_N_INSNS (1);
else
*total = COSTS_N_INSNS (3);
if (GET_CODE (op0) != REG
&& (GET_CODE (op0) != SUBREG || GET_CODE (SUBREG_REG (op0)) != REG))
- *total += rtx_cost (op0, MULT);
+ *total += rtx_cost (op0, MULT, speed);
if (GET_CODE (op1) != REG
&& (GET_CODE (op1) != SUBREG || GET_CODE (SUBREG_REG (op1)) != REG))
- *total += rtx_cost (op1, MULT);
+ *total += rtx_cost (op1, MULT, speed);
}
return true;
return false;
}
}
-
-static void
-bfin_internal_label (FILE *stream, const char *prefix, unsigned long num)
-{
- fprintf (stream, "%s%s$%ld:\n", LOCAL_LABEL_PREFIX, prefix, num);
-}
\f
/* Used for communication between {push,pop}_multiple_operation (which
we use not only as a predicate) and the corresponding output functions. */
static int first_preg_to_save, first_dreg_to_save;
+static int n_regs_to_save;
int
push_multiple_operation (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
lastpreg++;
}
}
+ n_regs_to_save = 8 - first_dreg_to_save + 6 - first_preg_to_save;
return 1;
}
}
first_dreg_to_save = lastdreg;
first_preg_to_save = lastpreg;
+ n_regs_to_save = 8 - first_dreg_to_save + 6 - first_preg_to_save;
return 1;
}
return false;
}
\f
-/* Implement TARGET_SCHED_ISSUE_RATE. */
+/* Compute the alignment for a local variable.
+ TYPE is the data type, and ALIGN is the alignment that
+ the object would ordinarily have. The value of this macro is used
+ instead of that alignment to align the object. */
-static int
-bfin_issue_rate (void)
+int
+bfin_local_alignment (tree type, int align)
+{
+ /* Increasing alignment for (relatively) big types allows the builtin
+ memcpy can use 32 bit loads/stores. */
+ if (TYPE_SIZE (type)
+ && TREE_CODE (TYPE_SIZE (type)) == INTEGER_CST
+ && (TREE_INT_CST_LOW (TYPE_SIZE (type)) > 8
+ || TREE_INT_CST_HIGH (TYPE_SIZE (type))) && align < 32)
+ return 32;
+ return align;
+}
+\f
+/* Implement TARGET_SCHED_ISSUE_RATE. */
+
+static int
+bfin_issue_rate (void)
{
return 3;
}
if (dep_insn_type == TYPE_MOVE || dep_insn_type == TYPE_MCLD)
{
rtx pat = PATTERN (dep_insn);
+ if (GET_CODE (pat) == PARALLEL)
+ pat = XVECEXP (pat, 0, 0);
rtx dest = SET_DEST (pat);
rtx src = SET_SRC (pat);
- if (! ADDRESS_REGNO_P (REGNO (dest)) || ! D_REGNO_P (REGNO (src)))
+ if (! ADDRESS_REGNO_P (REGNO (dest))
+ || ! (MEM_P (src) || D_REGNO_P (REGNO (src))))
return cost;
return cost + (dep_insn_type == TYPE_MOVE ? 4 : 3);
}
/* Maximum size of a loop. */
#define MAX_LOOP_LENGTH 2042
+/* Maximum distance of the LSETUP instruction from the loop start. */
+#define MAX_LSETUP_DISTANCE 30
+
/* We need to keep a vector of loops */
typedef struct loop_info *loop_info;
DEF_VEC_P (loop_info);
/* loop number, for dumps */
int loop_no;
- /* Predecessor block of the loop. This is the one that falls into
- the loop and contains the initialization instruction. */
- basic_block predecessor;
+ /* All edges that jump into and out of the loop. */
+ VEC(edge,gc) *incoming;
+
+ /* We can handle two cases: all incoming edges have the same destination
+ block, or all incoming edges have the same source block. These two
+ members are set to the common source or destination we found, or NULL
+ if different blocks were found. If both are NULL the loop can't be
+ optimized. */
+ basic_block incoming_src;
+ basic_block incoming_dest;
/* First block in the loop. This is the one branched to by the loop_end
insn. */
return false;
}
+/* Estimate the length of INSN conservatively. */
+
+static int
+length_for_loop (rtx insn)
+{
+ int length = 0;
+ if (JUMP_P (insn) && any_condjump_p (insn) && !optimize_size)
+ {
+ if (ENABLE_WA_SPECULATIVE_SYNCS)
+ length = 8;
+ else if (ENABLE_WA_SPECULATIVE_LOADS)
+ length = 6;
+ }
+ else if (LABEL_P (insn))
+ {
+ if (ENABLE_WA_SPECULATIVE_SYNCS)
+ length = 4;
+ }
+
+ if (INSN_P (insn))
+ length += get_attr_length (insn);
+
+ return length;
+}
+
/* Optimize LOOP. */
static void
rtx reg_lc0, reg_lc1, reg_lt0, reg_lt1, reg_lb0, reg_lb1;
rtx iter_reg;
rtx lc_reg, lt_reg, lb_reg;
- rtx seq;
+ rtx seq, seq_end;
int length;
unsigned ix;
int inner_depth = 0;
goto bad_loop;
}
+ if (loop->incoming_src)
+ {
+ /* Make sure the predecessor is before the loop start label, as required by
+ the LSETUP instruction. */
+ length = 0;
+ for (insn = BB_END (loop->incoming_src);
+ insn && insn != loop->start_label;
+ insn = NEXT_INSN (insn))
+ length += length_for_loop (insn);
+
+ if (!insn)
+ {
+ if (dump_file)
+ fprintf (dump_file, ";; loop %d lsetup not before loop_start\n",
+ loop->loop_no);
+ goto bad_loop;
+ }
+
+ if (length > MAX_LSETUP_DISTANCE)
+ {
+ if (dump_file)
+ fprintf (dump_file, ";; loop %d lsetup too far away\n", loop->loop_no);
+ goto bad_loop;
+ }
+ }
+
/* Check if start_label appears before loop_end and calculate the
offset between them. We calculate the length of instructions
conservatively. */
for (insn = loop->start_label;
insn && insn != loop->loop_end;
insn = NEXT_INSN (insn))
- {
- if (JUMP_P (insn) && any_condjump_p (insn) && !optimize_size)
- {
- if (TARGET_CSYNC_ANOMALY)
- length += 8;
- else if (TARGET_SPECLD_ANOMALY)
- length += 6;
- }
- else if (LABEL_P (insn))
- {
- if (TARGET_CSYNC_ANOMALY)
- length += 4;
- }
-
- if (INSN_P (insn))
- length += get_attr_length (insn);
- }
+ length += length_for_loop (insn);
if (!insn)
{
if (JUMP_P (last_insn))
{
- loop_info inner = bb->aux;
+ loop_info inner = (loop_info) bb->aux;
if (inner
&& inner->outer == loop
&& inner->loop_end == last_insn
if (loop->init != NULL_RTX)
emit_insn (loop->init);
- emit_insn(loop->loop_init);
- emit_label (loop->start_label);
+ seq_end = emit_insn (loop->loop_init);
seq = get_insns ();
end_sequence ();
- emit_insn_after (seq, BB_END (loop->predecessor));
- delete_insn (loop->loop_end);
+ if (loop->incoming_src)
+ {
+ rtx prev = BB_END (loop->incoming_src);
+ if (VEC_length (edge, loop->incoming) > 1
+ || !(VEC_last (edge, loop->incoming)->flags & EDGE_FALLTHRU))
+ {
+ gcc_assert (JUMP_P (prev));
+ prev = PREV_INSN (prev);
+ }
+ emit_insn_after (seq, prev);
+ }
+ else
+ {
+ basic_block new_bb;
+ edge e;
+ edge_iterator ei;
+
+ if (loop->head != loop->incoming_dest)
+ {
+ FOR_EACH_EDGE (e, ei, loop->head->preds)
+ {
+ if (e->flags & EDGE_FALLTHRU)
+ {
+ rtx newjump = gen_jump (loop->start_label);
+ emit_insn_before (newjump, BB_HEAD (loop->head));
+ new_bb = create_basic_block (newjump, newjump, loop->head->prev_bb);
+ gcc_assert (new_bb = loop->head->prev_bb);
+ break;
+ }
+ }
+ }
+
+ emit_insn_before (seq, BB_HEAD (loop->head));
+ seq = emit_label_before (gen_label_rtx (), seq);
+ new_bb = create_basic_block (seq, seq_end, loop->head->prev_bb);
+ FOR_EACH_EDGE (e, ei, loop->incoming)
+ {
+ if (!(e->flags & EDGE_FALLTHRU)
+ || e->dest != loop->head)
+ redirect_edge_and_branch_force (e, new_bb);
+ else
+ redirect_edge_succ (e, new_bb);
+ }
+ }
+
+ delete_insn (loop->loop_end);
/* Insert the loop end label before the last instruction of the loop. */
emit_label_before (loop->end_label, loop->last_insn);
return;
-bad_loop:
+ bad_loop:
if (dump_file)
fprintf (dump_file, ";; loop %d is bad\n", loop->loop_no);
loop->tail = tail_bb;
loop->head = BRANCH_EDGE (tail_bb)->dest;
loop->successor = FALLTHRU_EDGE (tail_bb)->dest;
- loop->predecessor = NULL;
loop->loop_end = tail_insn;
loop->last_insn = NULL_RTX;
loop->iter_reg = SET_DEST (XVECEXP (PATTERN (tail_insn), 0, 1));
loop->clobber_loop0 = loop->clobber_loop1 = 0;
loop->outer = NULL;
loop->loops = NULL;
-
+ loop->incoming = VEC_alloc (edge, gc, 2);
loop->init = loop->loop_init = NULL_RTX;
loop->start_label = XEXP (XEXP (SET_SRC (XVECEXP (PATTERN (tail_insn), 0, 0)), 1), 0);
loop->end_label = NULL_RTX;
FOR_EACH_EDGE (e, ei, bb->succs)
{
basic_block succ = EDGE_SUCC (bb, ei.index)->dest;
- if (!REGNO_REG_SET_P (succ->il.rtl->global_live_at_start,
+ if (!REGNO_REG_SET_P (df_get_live_in (succ),
REGNO (loop->iter_reg)))
continue;
if (!VEC_space (basic_block, works, 1))
}
}
+ /* Find the predecessor, and make sure nothing else jumps into this loop. */
if (!loop->bad)
{
- /* Make sure we only have one entry point. */
- if (EDGE_COUNT (loop->head->preds) == 2)
+ int pass, retry;
+ for (dwork = 0; VEC_iterate (basic_block, loop->blocks, dwork, bb); dwork++)
{
- loop->predecessor = EDGE_PRED (loop->head, 0)->src;
- if (loop->predecessor == loop->tail)
- /* We wanted the other predecessor. */
- loop->predecessor = EDGE_PRED (loop->head, 1)->src;
-
- /* We can only place a loop insn on a fall through edge of a
- single exit block. */
- if (EDGE_COUNT (loop->predecessor->succs) != 1
- || !(EDGE_SUCC (loop->predecessor, 0)->flags & EDGE_FALLTHRU)
- /* If loop->predecessor is in loop, loop->head is not really
- the head of the loop. */
- || bfin_bb_in_loop (loop, loop->predecessor))
- loop->predecessor = NULL;
+ edge e;
+ edge_iterator ei;
+ FOR_EACH_EDGE (e, ei, bb->preds)
+ {
+ basic_block pred = e->src;
+
+ if (!bfin_bb_in_loop (loop, pred))
+ {
+ if (dump_file)
+ fprintf (dump_file, ";; Loop %d: incoming edge %d -> %d\n",
+ loop->loop_no, pred->index,
+ e->dest->index);
+ VEC_safe_push (edge, gc, loop->incoming, e);
+ }
+ }
}
- if (loop->predecessor == NULL)
+ for (pass = 0, retry = 1; retry && pass < 2; pass++)
{
- if (dump_file)
- fprintf (dump_file, ";; loop has bad predecessor\n");
- loop->bad = 1;
+ edge e;
+ edge_iterator ei;
+ bool first = true;
+ retry = 0;
+
+ FOR_EACH_EDGE (e, ei, loop->incoming)
+ {
+ if (first)
+ {
+ loop->incoming_src = e->src;
+ loop->incoming_dest = e->dest;
+ first = false;
+ }
+ else
+ {
+ if (e->dest != loop->incoming_dest)
+ loop->incoming_dest = NULL;
+ if (e->src != loop->incoming_src)
+ loop->incoming_src = NULL;
+ }
+ if (loop->incoming_src == NULL && loop->incoming_dest == NULL)
+ {
+ if (pass == 0)
+ {
+ if (dump_file)
+ fprintf (dump_file,
+ ";; retrying loop %d with forwarder blocks\n",
+ loop->loop_no);
+ retry = 1;
+ break;
+ }
+ loop->bad = 1;
+ if (dump_file)
+ fprintf (dump_file,
+ ";; can't find suitable entry for loop %d\n",
+ loop->loop_no);
+ goto out;
+ }
+ }
+ if (retry)
+ {
+ retry = 0;
+ FOR_EACH_EDGE (e, ei, loop->incoming)
+ {
+ if (forwarder_block_p (e->src))
+ {
+ edge e2;
+ edge_iterator ei2;
+
+ if (dump_file)
+ fprintf (dump_file,
+ ";; Adding forwarder block %d to loop %d and retrying\n",
+ e->src->index, loop->loop_no);
+ VEC_safe_push (basic_block, heap, loop->blocks, e->src);
+ bitmap_set_bit (loop->block_bitmap, e->src->index);
+ FOR_EACH_EDGE (e2, ei2, e->src->preds)
+ VEC_safe_push (edge, gc, loop->incoming, e2);
+ VEC_unordered_remove (edge, loop->incoming, ei.index);
+ retry = 1;
+ break;
+ }
+ }
+ }
}
}
-#ifdef ENABLE_CHECKING
- /* Make sure nothing jumps into this loop. This shouldn't happen as we
- wouldn't have generated the counted loop patterns in such a case.
- However, this test must be done after the test above to detect loops
- with invalid headers. */
- if (!loop->bad)
- for (dwork = 0; VEC_iterate (basic_block, loop->blocks, dwork, bb); dwork++)
- {
- edge e;
- edge_iterator ei;
- if (bb == loop->head)
- continue;
- FOR_EACH_EDGE (e, ei, bb->preds)
- {
- basic_block pred = EDGE_PRED (bb, ei.index)->src;
- if (!bfin_bb_in_loop (loop, pred))
- abort ();
- }
- }
-#endif
+ out:
VEC_free (basic_block, heap, works);
}
-static void
-bfin_reorg_loops (FILE *dump_file)
+/* Analyze the structure of the loops in the current function. Use STACK
+ for bitmap allocations. Returns all the valid candidates for hardware
+ loops found in this function. */
+static loop_info
+bfin_discover_loops (bitmap_obstack *stack, FILE *dump_file)
{
- bitmap_obstack stack;
- bitmap tmp_bitmap;
- basic_block bb;
loop_info loops = NULL;
loop_info loop;
+ basic_block bb;
+ bitmap tmp_bitmap;
int nloops = 0;
- bitmap_obstack_initialize (&stack);
-
/* Find all the possible loop tails. This means searching for every
loop_end instruction. For each one found, create a loop_info
structure and add the head block to the work list. */
if (INSN_P (tail) && recog_memoized (tail) == CODE_FOR_loop_end)
{
+ rtx insn;
/* A possible loop end */
+ /* There's a degenerate case we can handle - an empty loop consisting
+ of only a back branch. Handle that by deleting the branch. */
+ insn = BB_HEAD (BRANCH_EDGE (bb)->dest);
+ if (next_real_insn (insn) == tail)
+ {
+ if (dump_file)
+ {
+ fprintf (dump_file, ";; degenerate loop ending at\n");
+ print_rtl_single (dump_file, tail);
+ }
+ delete_insn_and_edges (tail);
+ continue;
+ }
+
loop = XNEW (struct loop_info);
loop->next = loops;
loops = loop;
loop->loop_no = nloops++;
loop->blocks = VEC_alloc (basic_block, heap, 20);
- loop->block_bitmap = BITMAP_ALLOC (&stack);
+ loop->block_bitmap = BITMAP_ALLOC (stack);
bb->aux = loop;
if (dump_file)
}
}
- tmp_bitmap = BITMAP_ALLOC (&stack);
+ tmp_bitmap = BITMAP_ALLOC (stack);
/* Compute loop nestings. */
for (loop = loops; loop; loop = loop->next)
{
}
else
{
+ if (dump_file)
+ fprintf (dump_file,
+ ";; can't find suitable nesting for loops %d and %d\n",
+ loop->loop_no, other->loop_no);
loop->bad = other->bad = 1;
}
}
}
BITMAP_FREE (tmp_bitmap);
+ return loops;
+}
+
+/* Free up the loop structures in LOOPS. */
+static void
+free_loops (loop_info loops)
+{
+ while (loops)
+ {
+ loop_info loop = loops;
+ loops = loop->next;
+ VEC_free (loop_info, heap, loop->loops);
+ VEC_free (basic_block, heap, loop->blocks);
+ BITMAP_FREE (loop->block_bitmap);
+ XDELETE (loop);
+ }
+}
+
+#define BB_AUX_INDEX(BB) ((unsigned)(BB)->aux)
+
+/* The taken-branch edge from the loop end can actually go forward. Since the
+ Blackfin's LSETUP instruction requires that the loop end be after the loop
+ start, try to reorder a loop's basic blocks when we find such a case. */
+static void
+bfin_reorder_loops (loop_info loops, FILE *dump_file)
+{
+ basic_block bb;
+ loop_info loop;
+
+ FOR_EACH_BB (bb)
+ bb->aux = NULL;
+ cfg_layout_initialize (0);
+
+ for (loop = loops; loop; loop = loop->next)
+ {
+ unsigned index;
+ basic_block bb;
+ edge e;
+ edge_iterator ei;
+
+ if (loop->bad)
+ continue;
+
+ /* Recreate an index for basic blocks that represents their order. */
+ for (bb = ENTRY_BLOCK_PTR->next_bb, index = 0;
+ bb != EXIT_BLOCK_PTR;
+ bb = bb->next_bb, index++)
+ bb->aux = (PTR) index;
+
+ if (BB_AUX_INDEX (loop->head) < BB_AUX_INDEX (loop->tail))
+ continue;
+
+ FOR_EACH_EDGE (e, ei, loop->head->succs)
+ {
+ if (bitmap_bit_p (loop->block_bitmap, e->dest->index)
+ && BB_AUX_INDEX (e->dest) < BB_AUX_INDEX (loop->tail))
+ {
+ basic_block start_bb = e->dest;
+ basic_block start_prev_bb = start_bb->prev_bb;
+
+ if (dump_file)
+ fprintf (dump_file, ";; Moving block %d before block %d\n",
+ loop->head->index, start_bb->index);
+ loop->head->prev_bb->next_bb = loop->head->next_bb;
+ loop->head->next_bb->prev_bb = loop->head->prev_bb;
+
+ loop->head->prev_bb = start_prev_bb;
+ loop->head->next_bb = start_bb;
+ start_prev_bb->next_bb = start_bb->prev_bb = loop->head;
+ break;
+ }
+ }
+ loops = loops->next;
+ }
+
+ FOR_EACH_BB (bb)
+ {
+ if (bb->next_bb != EXIT_BLOCK_PTR)
+ bb->aux = bb->next_bb;
+ else
+ bb->aux = NULL;
+ }
+ cfg_layout_finalize ();
+ df_analyze ();
+}
+
+/* Run from machine_dependent_reorg, this pass looks for doloop_end insns
+ and tries to rewrite the RTL of these loops so that proper Blackfin
+ hardware loops are generated. */
+
+static void
+bfin_reorg_loops (FILE *dump_file)
+{
+ loop_info loops = NULL;
+ loop_info loop;
+ basic_block bb;
+ bitmap_obstack stack;
+
+ bitmap_obstack_initialize (&stack);
+
+ if (dump_file)
+ fprintf (dump_file, ";; Find loops, first pass\n\n");
+
+ loops = bfin_discover_loops (&stack, dump_file);
+
+ if (dump_file)
+ bfin_dump_loops (loops);
+
+ bfin_reorder_loops (loops, dump_file);
+ free_loops (loops);
+
+ if (dump_file)
+ fprintf (dump_file, ";; Find loops, second pass\n\n");
+
+ loops = bfin_discover_loops (&stack, dump_file);
if (dump_file)
{
fprintf (dump_file, ";; All loops found:\n\n");
bfin_dump_loops (loops);
}
- /* Free up the loop structures */
- while (loops)
- {
- loop = loops;
- loops = loop->next;
- VEC_free (loop_info, heap, loop->loops);
- VEC_free (basic_block, heap, loop->blocks);
- BITMAP_FREE (loop->block_bitmap);
- XDELETE (loop);
- }
+ free_loops (loops);
if (dump_file)
print_rtl (dump_file, get_insns ());
static bool
gen_one_bundle (rtx slot[3])
{
- rtx bundle;
-
gcc_assert (slot[1] != NULL_RTX);
+ /* Don't add extra NOPs if optimizing for size. */
+ if (optimize_size
+ && (slot[0] == NULL_RTX || slot[2] == NULL_RTX))
+ return false;
+
/* Verify that we really can do the multi-issue. */
if (slot[0])
{
while (t != slot[1])
{
if (GET_CODE (t) != NOTE
- || NOTE_LINE_NUMBER (t) != NOTE_INSN_DELETED)
+ || NOTE_KIND (t) != NOTE_INSN_DELETED)
return false;
t = NEXT_INSN (t);
}
while (t != slot[2])
{
if (GET_CODE (t) != NOTE
- || NOTE_LINE_NUMBER (t) != NOTE_INSN_DELETED)
+ || NOTE_KIND (t) != NOTE_INSN_DELETED)
return false;
t = NEXT_INSN (t);
}
}
if (slot[0] == NULL_RTX)
- slot[0] = emit_insn_before (gen_mnop (), slot[1]);
+ {
+ slot[0] = emit_insn_before (gen_mnop (), slot[1]);
+ df_insn_rescan (slot[0]);
+ }
if (slot[2] == NULL_RTX)
- slot[2] = emit_insn_after (gen_nop (), slot[1]);
+ {
+ slot[2] = emit_insn_after (gen_forced_nop (), slot[1]);
+ df_insn_rescan (slot[2]);
+ }
/* Avoid line number information being printed inside one bundle. */
if (INSN_LOCATOR (slot[1])
/* Terminate them with "|| " instead of ";" in the output. */
PUT_MODE (slot[0], SImode);
PUT_MODE (slot[1], SImode);
-
- /* This is a cheat to avoid emit_insn's special handling of SEQUENCEs.
- Generating a PARALLEL first and changing its code later is the
- easiest way to emit a SEQUENCE insn. */
- bundle = gen_rtx_PARALLEL (VOIDmode, gen_rtvec (3, slot[0], slot[1], slot[2]));
- emit_insn_before (bundle, slot[0]);
- remove_insn (slot[0]);
- remove_insn (slot[1]);
- remove_insn (slot[2]);
- PUT_CODE (bundle, SEQUENCE);
-
+ /* Terminate the bundle, for the benefit of reorder_var_tracking_notes. */
+ PUT_MODE (slot[2], QImode);
return true;
}
{
SET_SRC (pat) = XVECEXP (SET_SRC (pat), 0, 0);
INSN_CODE (slot[0]) = -1;
+ df_insn_rescan (slot[0]);
}
}
n_filled = 0;
}
}
}
+
+/* Ensure that no var tracking notes are emitted in the middle of a
+ three-instruction bundle. */
+
+static void
+reorder_var_tracking_notes (void)
+{
+ basic_block bb;
+ FOR_EACH_BB (bb)
+ {
+ rtx insn, next;
+ rtx queue = NULL_RTX;
+ bool in_bundle = false;
+
+ for (insn = BB_HEAD (bb); insn != BB_END (bb); insn = next)
+ {
+ next = NEXT_INSN (insn);
+
+ if (INSN_P (insn))
+ {
+ /* Emit queued up notes at the last instruction of a bundle. */
+ if (GET_MODE (insn) == QImode)
+ {
+ while (queue)
+ {
+ rtx next_queue = PREV_INSN (queue);
+ PREV_INSN (NEXT_INSN (insn)) = queue;
+ NEXT_INSN (queue) = NEXT_INSN (insn);
+ NEXT_INSN (insn) = queue;
+ PREV_INSN (queue) = insn;
+ queue = next_queue;
+ }
+ in_bundle = false;
+ }
+ else if (GET_MODE (insn) == SImode)
+ in_bundle = true;
+ }
+ else if (NOTE_P (insn) && NOTE_KIND (insn) == NOTE_INSN_VAR_LOCATION)
+ {
+ if (in_bundle)
+ {
+ rtx prev = PREV_INSN (insn);
+ PREV_INSN (next) = prev;
+ NEXT_INSN (prev) = next;
+
+ PREV_INSN (insn) = queue;
+ queue = insn;
+ }
+ }
+ }
+ }
+}
\f
+/* On some silicon revisions, functions shorter than a certain number of cycles
+ can cause unpredictable behaviour. Work around this by adding NOPs as
+ needed. */
+static void
+workaround_rts_anomaly (void)
+{
+ rtx insn, first_insn = NULL_RTX;
+ int cycles = 4;
+
+ if (! ENABLE_WA_RETS)
+ return;
+
+ for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
+ {
+ rtx pat;
+
+ if (BARRIER_P (insn))
+ return;
+
+ if (NOTE_P (insn) || LABEL_P (insn))
+ continue;
+
+ if (first_insn == NULL_RTX)
+ first_insn = insn;
+ pat = PATTERN (insn);
+ if (GET_CODE (pat) == USE || GET_CODE (pat) == CLOBBER
+ || GET_CODE (pat) == ASM_INPUT || GET_CODE (pat) == ADDR_VEC
+ || GET_CODE (pat) == ADDR_DIFF_VEC || asm_noperands (pat) >= 0)
+ continue;
+
+ if (CALL_P (insn))
+ return;
+
+ if (JUMP_P (insn))
+ {
+ if (recog_memoized (insn) == CODE_FOR_return_internal)
+ break;
+
+ /* Nothing to worry about for direct jumps. */
+ if (!any_condjump_p (insn))
+ return;
+ if (cycles <= 1)
+ return;
+ cycles--;
+ }
+ else if (INSN_P (insn))
+ {
+ rtx pat = PATTERN (insn);
+ int this_cycles = 1;
+
+ if (GET_CODE (pat) == PARALLEL)
+ {
+ if (push_multiple_operation (pat, VOIDmode)
+ || pop_multiple_operation (pat, VOIDmode))
+ this_cycles = n_regs_to_save;
+ }
+ else
+ {
+ enum insn_code icode = recog_memoized (insn);
+ if (icode == CODE_FOR_link)
+ this_cycles = 4;
+ else if (icode == CODE_FOR_unlink)
+ this_cycles = 3;
+ else if (icode == CODE_FOR_mulsi3)
+ this_cycles = 5;
+ }
+ if (this_cycles >= cycles)
+ return;
+
+ cycles -= this_cycles;
+ }
+ }
+ while (cycles > 0)
+ {
+ emit_insn_before (gen_nop (), first_insn);
+ cycles--;
+ }
+}
+
/* Return an insn type for INSN that can be used by the caller for anomaly
workarounds. This differs from plain get_attr_type in that it handles
SEQUENCEs. */
{
enum attr_type t;
t = get_attr_type (XVECEXP (pat, 0, 1));
- if (t == TYPE_MCLD && may_trap_p (SET_SRC (XVECEXP (pat, 0, 1))))
+ if (t == TYPE_MCLD
+ && may_trap_p (SET_SRC (PATTERN (XVECEXP (pat, 0, 1)))))
return true;
t = get_attr_type (XVECEXP (pat, 0, 2));
- if (t == TYPE_MCLD && may_trap_p (SET_SRC (XVECEXP (pat, 0, 2))))
+ if (t == TYPE_MCLD
+ && may_trap_p (SET_SRC (PATTERN (XVECEXP (pat, 0, 2)))))
return true;
return false;
}
return may_trap_p (SET_SRC (single_set (insn)));
}
-/* We use the machine specific reorg pass for emitting CSYNC instructions
- after conditional branches as needed.
-
- The Blackfin is unusual in that a code sequence like
- if cc jump label
- r0 = (p0)
- may speculatively perform the load even if the condition isn't true. This
- happens for a branch that is predicted not taken, because the pipeline
- isn't flushed or stalled, so the early stages of the following instructions,
- which perform the memory reference, are allowed to execute before the
- jump condition is evaluated.
- Therefore, we must insert additional instructions in all places where this
- could lead to incorrect behavior. The manual recommends CSYNC, while
- VDSP seems to use NOPs (even though its corresponding compiler option is
- named CSYNC).
+/* This function acts like NEXT_INSN, but is aware of three-insn bundles and
+ skips all subsequent parallel instructions if INSN is the start of such
+ a group. */
+static rtx
+find_next_insn_start (rtx insn)
+{
+ if (GET_MODE (insn) == SImode)
+ {
+ while (GET_MODE (insn) != QImode)
+ insn = NEXT_INSN (insn);
+ }
+ return NEXT_INSN (insn);
+}
- When optimizing for speed, we emit NOPs, which seems faster than a CSYNC.
- When optimizing for size, we turn the branch into a predicted taken one.
- This may be slower due to mispredicts, but saves code size. */
+/* Return INSN if it is of TYPE_MCLD. Alternatively, if INSN is the start of
+ a three-insn bundle, see if one of them is a load and return that if so.
+ Return NULL_RTX if the insn does not contain loads. */
+static rtx
+find_load (rtx insn)
+{
+ if (get_attr_type (insn) == TYPE_MCLD)
+ return insn;
+ if (GET_MODE (insn) != SImode)
+ return NULL_RTX;
+ do {
+ insn = NEXT_INSN (insn);
+ if ((GET_MODE (insn) == SImode || GET_MODE (insn) == QImode)
+ && get_attr_type (insn) == TYPE_MCLD)
+ return insn;
+ } while (GET_MODE (insn) != QImode);
+ return NULL_RTX;
+}
static void
-bfin_reorg (void)
+workaround_speculation (void)
{
- rtx insn, last_condjump = NULL_RTX;
+ rtx insn, next;
+ rtx last_condjump = NULL_RTX;
int cycles_since_jump = INT_MAX;
- /* We are freeing block_for_insn in the toplev to keep compatibility
- with old MDEP_REORGS that are not CFG based. Recompute it now. */
- compute_bb_for_insn ();
-
- if (bfin_flag_schedule_insns2)
- {
- splitting_for_sched = 1;
- split_all_insns (0);
- splitting_for_sched = 0;
-
- update_life_info (NULL, UPDATE_LIFE_GLOBAL_RM_NOTES, PROP_DEATH_NOTES);
-
- timevar_push (TV_SCHED2);
- schedule_insns ();
- timevar_pop (TV_SCHED2);
-
- /* Examine the schedule and insert nops as necessary for 64 bit parallel
- instructions. */
- bfin_gen_bundles ();
- }
-
- /* Doloop optimization */
- if (cfun->machine->has_hardware_loops)
- bfin_reorg_loops (dump_file);
-
- if (! TARGET_SPECLD_ANOMALY && ! TARGET_CSYNC_ANOMALY)
+ if (! ENABLE_WA_SPECULATIVE_LOADS && ! ENABLE_WA_SPECULATIVE_SYNCS)
return;
/* First pass: find predicted-false branches; if something after them
needs nops, insert them or change the branch to predict true. */
- for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
+ for (insn = get_insns (); insn; insn = next)
{
rtx pat;
+ next = find_next_insn_start (insn);
+
if (NOTE_P (insn) || BARRIER_P (insn) || LABEL_P (insn))
continue;
}
else if (INSN_P (insn))
{
+ rtx load_insn = find_load (insn);
enum attr_type type = type_for_anomaly (insn);
int delay_needed = 0;
if (cycles_since_jump < INT_MAX)
cycles_since_jump++;
- if (type == TYPE_MCLD && TARGET_SPECLD_ANOMALY)
+ if (load_insn && ENABLE_WA_SPECULATIVE_LOADS)
{
- if (trapping_loads_p (insn))
+ if (trapping_loads_p (load_insn))
delay_needed = 3;
}
- else if (type == TYPE_SYNC && TARGET_CSYNC_ANOMALY)
+ else if (type == TYPE_SYNC && ENABLE_WA_SPECULATIVE_SYNCS)
delay_needed = 4;
if (delay_needed > cycles_since_jump)
}
/* Second pass: for predicted-true branches, see if anything at the
branch destination needs extra nops. */
- if (! TARGET_CSYNC_ANOMALY)
+ if (! ENABLE_WA_SPECULATIVE_SYNCS)
return;
for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
{
+ int cycles_since_jump;
if (JUMP_P (insn)
&& any_condjump_p (insn)
&& (INSN_CODE (insn) == CODE_FOR_cbranch_predicted_taken
if (cycles_since_jump < INT_MAX)
cycles_since_jump++;
- if (type == TYPE_SYNC && TARGET_CSYNC_ANOMALY)
+ if (type == TYPE_SYNC && ENABLE_WA_SPECULATIVE_SYNCS)
delay_needed = 2;
if (delay_needed > cycles_since_jump)
}
}
}
+}
+
+/* We use the machine specific reorg pass for emitting CSYNC instructions
+ after conditional branches as needed.
+
+ The Blackfin is unusual in that a code sequence like
+ if cc jump label
+ r0 = (p0)
+ may speculatively perform the load even if the condition isn't true. This
+ happens for a branch that is predicted not taken, because the pipeline
+ isn't flushed or stalled, so the early stages of the following instructions,
+ which perform the memory reference, are allowed to execute before the
+ jump condition is evaluated.
+ Therefore, we must insert additional instructions in all places where this
+ could lead to incorrect behavior. The manual recommends CSYNC, while
+ VDSP seems to use NOPs (even though its corresponding compiler option is
+ named CSYNC).
+
+ When optimizing for speed, we emit NOPs, which seems faster than a CSYNC.
+ When optimizing for size, we turn the branch into a predicted taken one.
+ This may be slower due to mispredicts, but saves code size. */
+
+static void
+bfin_reorg (void)
+{
+ /* We are freeing block_for_insn in the toplev to keep compatibility
+ with old MDEP_REORGS that are not CFG based. Recompute it now. */
+ compute_bb_for_insn ();
+
+ if (bfin_flag_schedule_insns2)
+ {
+ splitting_for_sched = 1;
+ split_all_insns ();
+ splitting_for_sched = 0;
+
+ timevar_push (TV_SCHED2);
+ schedule_insns ();
+ timevar_pop (TV_SCHED2);
+
+ /* Examine the schedule and insert nops as necessary for 64-bit parallel
+ instructions. */
+ bfin_gen_bundles ();
+ }
+
+ df_analyze ();
+
+ /* Doloop optimization */
+ if (cfun->machine->has_hardware_loops)
+ bfin_reorg_loops (dump_file);
+
+ workaround_speculation ();
if (bfin_flag_var_tracking)
{
timevar_push (TV_VAR_TRACKING);
variable_tracking_main ();
+ reorder_var_tracking_notes ();
timevar_pop (TV_VAR_TRACKING);
}
+
+ df_finish_pass (false);
+
+ workaround_rts_anomaly ();
}
\f
/* Handle interrupt_handler, exception_handler and nmi_handler function
warning to be generated). */
static int
-bfin_comp_type_attributes (tree type1, tree type2)
+bfin_comp_type_attributes (const_tree type1, const_tree type2)
{
e_funkind kind1, kind2;
return NULL_TREE;
}
+/* Handle a "l1_text" attribute; arguments as in
+ struct attribute_spec.handler. */
+
+static tree
+bfin_handle_l1_text_attribute (tree *node, tree name, tree ARG_UNUSED (args),
+ int ARG_UNUSED (flags), bool *no_add_attrs)
+{
+ tree decl = *node;
+
+ if (TREE_CODE (decl) != FUNCTION_DECL)
+ {
+ error ("`%s' attribute only applies to functions",
+ IDENTIFIER_POINTER (name));
+ *no_add_attrs = true;
+ }
+
+ /* The decl may have already been given a section attribute
+ from a previous declaration. Ensure they match. */
+ else if (DECL_SECTION_NAME (decl) != NULL_TREE
+ && strcmp (TREE_STRING_POINTER (DECL_SECTION_NAME (decl)),
+ ".l1.text") != 0)
+ {
+ error ("section of %q+D conflicts with previous declaration",
+ decl);
+ *no_add_attrs = true;
+ }
+ else
+ DECL_SECTION_NAME (decl) = build_string (9, ".l1.text");
+
+ return NULL_TREE;
+}
+
+/* Handle a "l1_data", "l1_data_A" or "l1_data_B" attribute;
+ arguments as in struct attribute_spec.handler. */
+
+static tree
+bfin_handle_l1_data_attribute (tree *node, tree name, tree ARG_UNUSED (args),
+ int ARG_UNUSED (flags), bool *no_add_attrs)
+{
+ tree decl = *node;
+
+ if (TREE_CODE (decl) != VAR_DECL)
+ {
+ error ("`%s' attribute only applies to variables",
+ IDENTIFIER_POINTER (name));
+ *no_add_attrs = true;
+ }
+ else if (current_function_decl != NULL_TREE
+ && !TREE_STATIC (decl))
+ {
+ error ("`%s' attribute cannot be specified for local variables",
+ IDENTIFIER_POINTER (name));
+ *no_add_attrs = true;
+ }
+ else
+ {
+ const char *section_name;
+
+ if (strcmp (IDENTIFIER_POINTER (name), "l1_data") == 0)
+ section_name = ".l1.data";
+ else if (strcmp (IDENTIFIER_POINTER (name), "l1_data_A") == 0)
+ section_name = ".l1.data.A";
+ else if (strcmp (IDENTIFIER_POINTER (name), "l1_data_B") == 0)
+ section_name = ".l1.data.B";
+ else
+ gcc_unreachable ();
+
+ /* The decl may have already been given a section attribute
+ from a previous declaration. Ensure they match. */
+ if (DECL_SECTION_NAME (decl) != NULL_TREE
+ && strcmp (TREE_STRING_POINTER (DECL_SECTION_NAME (decl)),
+ section_name) != 0)
+ {
+ error ("section of %q+D conflicts with previous declaration",
+ decl);
+ *no_add_attrs = true;
+ }
+ else
+ DECL_SECTION_NAME (decl)
+ = build_string (strlen (section_name) + 1, section_name);
+ }
+
+ return NULL_TREE;
+}
+
/* Table of valid machine attributes. */
const struct attribute_spec bfin_attribute_table[] =
{
{ "saveall", 0, 0, false, true, true, NULL },
{ "longcall", 0, 0, false, true, true, bfin_handle_longcall_attribute },
{ "shortcall", 0, 0, false, true, true, bfin_handle_longcall_attribute },
+ { "l1_text", 0, 0, true, false, false, bfin_handle_l1_text_attribute },
+ { "l1_data", 0, 0, true, false, false, bfin_handle_l1_data_attribute },
+ { "l1_data_A", 0, 0, true, false, false, bfin_handle_l1_data_attribute },
+ { "l1_data_B", 0, 0, true, false, false, bfin_handle_l1_data_attribute },
{ NULL, 0, 0, false, false, false, NULL }
};
\f
{
rtx xops[3];
/* The this parameter is passed as the first argument. */
- rtx this = gen_rtx_REG (Pmode, REG_R0);
+ rtx this_rtx = gen_rtx_REG (Pmode, REG_R0);
/* Adjust the this parameter by a fixed constant. */
if (delta)
{
- xops[1] = this;
+ xops[1] = this_rtx;
if (delta >= -64 && delta <= 63)
{
xops[0] = GEN_INT (delta);
if (vcall_offset)
{
rtx p2tmp = gen_rtx_REG (Pmode, REG_P2);
- rtx tmp = gen_rtx_REG (Pmode, REG_R2);
+ rtx tmp = gen_rtx_REG (Pmode, REG_R3);
xops[1] = tmp;
xops[2] = p2tmp;
output_asm_insn ("%h1 = %h0; %d1 = %d0; %2 = %2 + %1", xops);
xops[0] = gen_rtx_MEM (Pmode, p2tmp);
}
- xops[2] = this;
+ xops[2] = this_rtx;
output_asm_insn ("%1 = %0; %2 = %2 + %1;", xops);
}
{
BFIN_BUILTIN_CSYNC,
BFIN_BUILTIN_SSYNC,
+ BFIN_BUILTIN_ONES,
BFIN_BUILTIN_COMPOSE_2X16,
BFIN_BUILTIN_EXTRACTLO,
BFIN_BUILTIN_EXTRACTHI,
BFIN_BUILTIN_MIN_1X16,
BFIN_BUILTIN_MAX_1X16,
+ BFIN_BUILTIN_SUM_2X16,
BFIN_BUILTIN_DIFFHL_2X16,
BFIN_BUILTIN_DIFFLH_2X16,
BFIN_BUILTIN_SSADD_1X32,
BFIN_BUILTIN_SSSUB_1X32,
BFIN_BUILTIN_NORM_1X32,
+ BFIN_BUILTIN_ROUND_1X32,
BFIN_BUILTIN_NEG_1X32,
+ BFIN_BUILTIN_ABS_1X32,
BFIN_BUILTIN_MIN_1X32,
BFIN_BUILTIN_MAX_1X32,
BFIN_BUILTIN_MULT_1X32,
+ BFIN_BUILTIN_MULT_1X32X32,
+ BFIN_BUILTIN_MULT_1X32X32NS,
BFIN_BUILTIN_MULHISILL,
BFIN_BUILTIN_MULHISILH,
BFIN_BUILTIN_LSHIFT_2X16,
BFIN_BUILTIN_SSASHIFT_1X16,
BFIN_BUILTIN_SSASHIFT_2X16,
+ BFIN_BUILTIN_SSASHIFT_1X32,
BFIN_BUILTIN_CPLX_MUL_16,
BFIN_BUILTIN_CPLX_MAC_16,
BFIN_BUILTIN_CPLX_MSU_16,
+ BFIN_BUILTIN_CPLX_MUL_16_S40,
+ BFIN_BUILTIN_CPLX_MAC_16_S40,
+ BFIN_BUILTIN_CPLX_MSU_16_S40,
+
+ BFIN_BUILTIN_CPLX_SQU,
+
+ BFIN_BUILTIN_LOADBYTES,
+
BFIN_BUILTIN_MAX
};
tree short_ftype_v2hi
= build_function_type_list (short_integer_type_node, V2HI_type_node,
NULL_TREE);
-
+ tree int_ftype_pint
+ = build_function_type_list (integer_type_node,
+ build_pointer_type (integer_type_node),
+ NULL_TREE);
+
/* Add the remaining MMX insns with somewhat more complicated types. */
def_builtin ("__builtin_bfin_csync", void_ftype_void, BFIN_BUILTIN_CSYNC);
def_builtin ("__builtin_bfin_ssync", void_ftype_void, BFIN_BUILTIN_SSYNC);
+ def_builtin ("__builtin_bfin_ones", short_ftype_int, BFIN_BUILTIN_ONES);
+
def_builtin ("__builtin_bfin_compose_2x16", v2hi_ftype_int_int,
BFIN_BUILTIN_COMPOSE_2X16);
def_builtin ("__builtin_bfin_extract_hi", short_ftype_v2hi,
def_builtin ("__builtin_bfin_abs_fr2x16", v2hi_ftype_v2hi,
BFIN_BUILTIN_ABS_2X16);
+ def_builtin ("__builtin_bfin_min_fr1x16", short_ftype_int_int,
+ BFIN_BUILTIN_MIN_1X16);
+ def_builtin ("__builtin_bfin_max_fr1x16", short_ftype_int_int,
+ BFIN_BUILTIN_MAX_1X16);
+
def_builtin ("__builtin_bfin_add_fr1x16", short_ftype_int_int,
BFIN_BUILTIN_SSADD_1X16);
def_builtin ("__builtin_bfin_sub_fr1x16", short_ftype_int_int,
def_builtin ("__builtin_bfin_norm_fr1x16", short_ftype_int,
BFIN_BUILTIN_NORM_1X16);
+ def_builtin ("__builtin_bfin_sum_fr2x16", short_ftype_v2hi,
+ BFIN_BUILTIN_SUM_2X16);
def_builtin ("__builtin_bfin_diff_hl_fr2x16", short_ftype_v2hi,
BFIN_BUILTIN_DIFFHL_2X16);
def_builtin ("__builtin_bfin_diff_lh_fr2x16", short_ftype_v2hi,
def_builtin ("__builtin_bfin_mulhisihh", int_ftype_v2hi_v2hi,
BFIN_BUILTIN_MULHISIHH);
+ def_builtin ("__builtin_bfin_min_fr1x32", int_ftype_int_int,
+ BFIN_BUILTIN_MIN_1X32);
+ def_builtin ("__builtin_bfin_max_fr1x32", int_ftype_int_int,
+ BFIN_BUILTIN_MAX_1X32);
+
def_builtin ("__builtin_bfin_add_fr1x32", int_ftype_int_int,
BFIN_BUILTIN_SSADD_1X32);
def_builtin ("__builtin_bfin_sub_fr1x32", int_ftype_int_int,
BFIN_BUILTIN_SSSUB_1X32);
def_builtin ("__builtin_bfin_negate_fr1x32", int_ftype_int,
BFIN_BUILTIN_NEG_1X32);
+ def_builtin ("__builtin_bfin_abs_fr1x32", int_ftype_int,
+ BFIN_BUILTIN_ABS_1X32);
def_builtin ("__builtin_bfin_norm_fr1x32", short_ftype_int,
BFIN_BUILTIN_NORM_1X32);
+ def_builtin ("__builtin_bfin_round_fr1x32", short_ftype_int,
+ BFIN_BUILTIN_ROUND_1X32);
def_builtin ("__builtin_bfin_mult_fr1x32", int_ftype_short_short,
BFIN_BUILTIN_MULT_1X32);
+ def_builtin ("__builtin_bfin_mult_fr1x32x32", int_ftype_int_int,
+ BFIN_BUILTIN_MULT_1X32X32);
+ def_builtin ("__builtin_bfin_mult_fr1x32x32NS", int_ftype_int_int,
+ BFIN_BUILTIN_MULT_1X32X32NS);
/* Shifts. */
def_builtin ("__builtin_bfin_shl_fr1x16", short_ftype_int_int,
BFIN_BUILTIN_LSHIFT_1X16);
def_builtin ("__builtin_bfin_lshl_fr2x16", v2hi_ftype_v2hi_int,
BFIN_BUILTIN_LSHIFT_2X16);
+ def_builtin ("__builtin_bfin_shl_fr1x32", int_ftype_int_int,
+ BFIN_BUILTIN_SSASHIFT_1X32);
/* Complex numbers. */
+ def_builtin ("__builtin_bfin_cmplx_add", v2hi_ftype_v2hi_v2hi,
+ BFIN_BUILTIN_SSADD_2X16);
+ def_builtin ("__builtin_bfin_cmplx_sub", v2hi_ftype_v2hi_v2hi,
+ BFIN_BUILTIN_SSSUB_2X16);
def_builtin ("__builtin_bfin_cmplx_mul", v2hi_ftype_v2hi_v2hi,
BFIN_BUILTIN_CPLX_MUL_16);
def_builtin ("__builtin_bfin_cmplx_mac", v2hi_ftype_v2hi_v2hi_v2hi,
BFIN_BUILTIN_CPLX_MAC_16);
def_builtin ("__builtin_bfin_cmplx_msu", v2hi_ftype_v2hi_v2hi_v2hi,
BFIN_BUILTIN_CPLX_MSU_16);
+ def_builtin ("__builtin_bfin_cmplx_mul_s40", v2hi_ftype_v2hi_v2hi,
+ BFIN_BUILTIN_CPLX_MUL_16_S40);
+ def_builtin ("__builtin_bfin_cmplx_mac_s40", v2hi_ftype_v2hi_v2hi_v2hi,
+ BFIN_BUILTIN_CPLX_MAC_16_S40);
+ def_builtin ("__builtin_bfin_cmplx_msu_s40", v2hi_ftype_v2hi_v2hi_v2hi,
+ BFIN_BUILTIN_CPLX_MSU_16_S40);
+ def_builtin ("__builtin_bfin_csqu_fr16", v2hi_ftype_v2hi,
+ BFIN_BUILTIN_CPLX_SQU);
+
+ /* "Unaligned" load. */
+ def_builtin ("__builtin_bfin_loadbytes", int_ftype_pint,
+ BFIN_BUILTIN_LOADBYTES);
+
}
{ CODE_FOR_ssashifthi3, "__builtin_bfin_shl_fr1x16", BFIN_BUILTIN_SSASHIFT_1X16, -1 },
{ CODE_FOR_lshiftv2hi3, "__builtin_bfin_lshl_fr2x16", BFIN_BUILTIN_LSHIFT_2X16, -1 },
{ CODE_FOR_lshifthi3, "__builtin_bfin_lshl_fr1x16", BFIN_BUILTIN_LSHIFT_1X16, -1 },
+ { CODE_FOR_ssashiftsi3, "__builtin_bfin_shl_fr1x32", BFIN_BUILTIN_SSASHIFT_1X32, -1 },
{ CODE_FOR_sminhi3, "__builtin_bfin_min_fr1x16", BFIN_BUILTIN_MIN_1X16, -1 },
{ CODE_FOR_smaxhi3, "__builtin_bfin_max_fr1x16", BFIN_BUILTIN_MAX_1X16, -1 },
static const struct builtin_description bdesc_1arg[] =
{
+ { CODE_FOR_loadbytes, "__builtin_bfin_loadbytes", BFIN_BUILTIN_LOADBYTES, 0 },
+
+ { CODE_FOR_ones, "__builtin_bfin_ones", BFIN_BUILTIN_ONES, 0 },
+
{ CODE_FOR_signbitshi2, "__builtin_bfin_norm_fr1x16", BFIN_BUILTIN_NORM_1X16, 0 },
{ CODE_FOR_ssneghi2, "__builtin_bfin_negate_fr1x16", BFIN_BUILTIN_NEG_1X16, 0 },
{ CODE_FOR_abshi2, "__builtin_bfin_abs_fr1x16", BFIN_BUILTIN_ABS_1X16, 0 },
{ CODE_FOR_signbitssi2, "__builtin_bfin_norm_fr1x32", BFIN_BUILTIN_NORM_1X32, 0 },
+ { CODE_FOR_ssroundsi2, "__builtin_bfin_round_fr1x32", BFIN_BUILTIN_ROUND_1X32, 0 },
{ CODE_FOR_ssnegsi2, "__builtin_bfin_negate_fr1x32", BFIN_BUILTIN_NEG_1X32, 0 },
+ { CODE_FOR_ssabssi2, "__builtin_bfin_abs_fr1x32", BFIN_BUILTIN_ABS_1X32, 0 },
{ CODE_FOR_movv2hi_hi_low, "__builtin_bfin_extract_lo", BFIN_BUILTIN_EXTRACTLO, 0 },
{ CODE_FOR_movv2hi_hi_high, "__builtin_bfin_extract_hi", BFIN_BUILTIN_EXTRACTHI, 0 },
{ CODE_FOR_ssnegv2hi2, "__builtin_bfin_negate_fr2x16", BFIN_BUILTIN_NEG_2X16, 0 },
- { CODE_FOR_absv2hi2, "__builtin_bfin_abs_fr2x16", BFIN_BUILTIN_ABS_2X16, 0 }
+ { CODE_FOR_ssabsv2hi2, "__builtin_bfin_abs_fr2x16", BFIN_BUILTIN_ABS_2X16, 0 }
};
/* Errors in the source file can cause expand_expr to return const0_rtx
if this is a normal binary op, or one of the MACFLAG_xxx constants. */
static rtx
-bfin_expand_binop_builtin (enum insn_code icode, tree arglist, rtx target,
+bfin_expand_binop_builtin (enum insn_code icode, tree exp, rtx target,
int macflag)
{
rtx pat;
- tree arg0 = TREE_VALUE (arglist);
- tree arg1 = TREE_VALUE (TREE_CHAIN (arglist));
+ tree arg0 = CALL_EXPR_ARG (exp, 0);
+ tree arg1 = CALL_EXPR_ARG (exp, 1);
rtx op0 = expand_expr (arg0, NULL_RTX, VOIDmode, 0);
rtx op1 = expand_expr (arg1, NULL_RTX, VOIDmode, 0);
enum machine_mode op0mode = GET_MODE (op0);
/* Subroutine of bfin_expand_builtin to take care of unop insns. */
static rtx
-bfin_expand_unop_builtin (enum insn_code icode, tree arglist,
+bfin_expand_unop_builtin (enum insn_code icode, tree exp,
rtx target)
{
rtx pat;
- tree arg0 = TREE_VALUE (arglist);
+ tree arg0 = CALL_EXPR_ARG (exp, 0);
rtx op0 = expand_expr (arg0, NULL_RTX, VOIDmode, 0);
enum machine_mode op0mode = GET_MODE (op0);
enum machine_mode tmode = insn_data[icode].operand[0].mode;
size_t i;
enum insn_code icode;
const struct builtin_description *d;
- tree fndecl = TREE_OPERAND (TREE_OPERAND (exp, 0), 0);
- tree arglist = TREE_OPERAND (exp, 1);
+ tree fndecl = TREE_OPERAND (CALL_EXPR_FN (exp), 0);
unsigned int fcode = DECL_FUNCTION_CODE (fndecl);
tree arg0, arg1, arg2;
- rtx op0, op1, op2, accvec, pat, tmp1, tmp2;
+ rtx op0, op1, op2, accvec, pat, tmp1, tmp2, a0reg, a1reg;
enum machine_mode tmode, mode0;
switch (fcode)
case BFIN_BUILTIN_DIFFHL_2X16:
case BFIN_BUILTIN_DIFFLH_2X16:
- arg0 = TREE_VALUE (arglist);
+ case BFIN_BUILTIN_SUM_2X16:
+ arg0 = CALL_EXPR_ARG (exp, 0);
op0 = expand_expr (arg0, NULL_RTX, VOIDmode, 0);
- icode = (fcode == BFIN_BUILTIN_DIFFHL_2X16
- ? CODE_FOR_subhilov2hi3 : CODE_FOR_sublohiv2hi3);
+ icode = (fcode == BFIN_BUILTIN_DIFFHL_2X16 ? CODE_FOR_subhilov2hi3
+ : fcode == BFIN_BUILTIN_DIFFLH_2X16 ? CODE_FOR_sublohiv2hi3
+ : CODE_FOR_ssaddhilov2hi3);
tmode = insn_data[icode].operand[0].mode;
mode0 = insn_data[icode].operand[1].mode;
emit_insn (pat);
return target;
+ case BFIN_BUILTIN_MULT_1X32X32:
+ case BFIN_BUILTIN_MULT_1X32X32NS:
+ arg0 = CALL_EXPR_ARG (exp, 0);
+ arg1 = CALL_EXPR_ARG (exp, 1);
+ op0 = expand_expr (arg0, NULL_RTX, VOIDmode, 0);
+ op1 = expand_expr (arg1, NULL_RTX, VOIDmode, 0);
+ if (! target
+ || !register_operand (target, SImode))
+ target = gen_reg_rtx (SImode);
+
+ a1reg = gen_rtx_REG (PDImode, REG_A1);
+ a0reg = gen_rtx_REG (PDImode, REG_A0);
+ tmp1 = gen_lowpart (V2HImode, op0);
+ tmp2 = gen_lowpart (V2HImode, op1);
+ emit_insn (gen_flag_macinit1hi (a1reg,
+ gen_lowpart (HImode, op0),
+ gen_lowpart (HImode, op1),
+ GEN_INT (MACFLAG_FU)));
+ emit_insn (gen_lshrpdi3 (a1reg, a1reg, GEN_INT (16)));
+
+ if (fcode == BFIN_BUILTIN_MULT_1X32X32)
+ emit_insn (gen_flag_mul_macv2hi_parts_acconly (a0reg, a1reg, tmp1, tmp2,
+ const1_rtx, const1_rtx,
+ const1_rtx, const0_rtx, a1reg,
+ const0_rtx, GEN_INT (MACFLAG_NONE),
+ GEN_INT (MACFLAG_M)));
+ else
+ {
+ /* For saturating multiplication, there's exactly one special case
+ to be handled: multiplying the smallest negative value with
+ itself. Due to shift correction in fractional multiplies, this
+ can overflow. Iff this happens, OP2 will contain 1, which, when
+ added in 32 bits to the smallest negative, wraps to the largest
+ positive, which is the result we want. */
+ op2 = gen_reg_rtx (V2HImode);
+ emit_insn (gen_packv2hi (op2, tmp1, tmp2, const0_rtx, const0_rtx));
+ emit_insn (gen_movsibi (gen_rtx_REG (BImode, REG_CC),
+ gen_lowpart (SImode, op2)));
+ emit_insn (gen_flag_mul_macv2hi_parts_acconly_andcc0 (a0reg, a1reg, tmp1, tmp2,
+ const1_rtx, const1_rtx,
+ const1_rtx, const0_rtx, a1reg,
+ const0_rtx, GEN_INT (MACFLAG_NONE),
+ GEN_INT (MACFLAG_M)));
+ op2 = gen_reg_rtx (SImode);
+ emit_insn (gen_movbisi (op2, gen_rtx_REG (BImode, REG_CC)));
+ }
+ emit_insn (gen_flag_machi_parts_acconly (a1reg, tmp2, tmp1,
+ const1_rtx, const0_rtx,
+ a1reg, const0_rtx, GEN_INT (MACFLAG_M)));
+ emit_insn (gen_ashrpdi3 (a1reg, a1reg, GEN_INT (15)));
+ emit_insn (gen_sum_of_accumulators (target, a0reg, a0reg, a1reg));
+ if (fcode == BFIN_BUILTIN_MULT_1X32X32NS)
+ emit_insn (gen_addsi3 (target, target, op2));
+ return target;
+
case BFIN_BUILTIN_CPLX_MUL_16:
- arg0 = TREE_VALUE (arglist);
- arg1 = TREE_VALUE (TREE_CHAIN (arglist));
+ case BFIN_BUILTIN_CPLX_MUL_16_S40:
+ arg0 = CALL_EXPR_ARG (exp, 0);
+ arg1 = CALL_EXPR_ARG (exp, 1);
op0 = expand_expr (arg0, NULL_RTX, VOIDmode, 0);
op1 = expand_expr (arg1, NULL_RTX, VOIDmode, 0);
accvec = gen_reg_rtx (V2PDImode);
if (! register_operand (op1, GET_MODE (op1)))
op1 = copy_to_mode_reg (GET_MODE (op1), op1);
- emit_insn (gen_flag_macinit1v2hi_parts (accvec, op0, op1, const0_rtx,
- const0_rtx, const0_rtx,
- const1_rtx, GEN_INT (MACFLAG_NONE)));
+ if (fcode == BFIN_BUILTIN_CPLX_MUL_16)
+ emit_insn (gen_flag_macinit1v2hi_parts (accvec, op0, op1, const0_rtx,
+ const0_rtx, const0_rtx,
+ const1_rtx, GEN_INT (MACFLAG_W32)));
+ else
+ emit_insn (gen_flag_macinit1v2hi_parts (accvec, op0, op1, const0_rtx,
+ const0_rtx, const0_rtx,
+ const1_rtx, GEN_INT (MACFLAG_NONE)));
emit_insn (gen_flag_macv2hi_parts (target, op0, op1, const1_rtx,
const1_rtx, const1_rtx,
const0_rtx, accvec, const1_rtx, const0_rtx,
case BFIN_BUILTIN_CPLX_MAC_16:
case BFIN_BUILTIN_CPLX_MSU_16:
- arg0 = TREE_VALUE (arglist);
- arg1 = TREE_VALUE (TREE_CHAIN (arglist));
- arg2 = TREE_VALUE (TREE_CHAIN (TREE_CHAIN (arglist)));
+ case BFIN_BUILTIN_CPLX_MAC_16_S40:
+ case BFIN_BUILTIN_CPLX_MSU_16_S40:
+ arg0 = CALL_EXPR_ARG (exp, 0);
+ arg1 = CALL_EXPR_ARG (exp, 1);
+ arg2 = CALL_EXPR_ARG (exp, 2);
op0 = expand_expr (arg0, NULL_RTX, VOIDmode, 0);
op1 = expand_expr (arg1, NULL_RTX, VOIDmode, 0);
op2 = expand_expr (arg2, NULL_RTX, VOIDmode, 0);
|| GET_MODE (target) != V2HImode
|| ! (*insn_data[icode].operand[0].predicate) (target, V2HImode))
target = gen_reg_rtx (tmode);
- if (! register_operand (op0, GET_MODE (op0)))
- op0 = copy_to_mode_reg (GET_MODE (op0), op0);
if (! register_operand (op1, GET_MODE (op1)))
op1 = copy_to_mode_reg (GET_MODE (op1), op1);
+ if (! register_operand (op2, GET_MODE (op2)))
+ op2 = copy_to_mode_reg (GET_MODE (op2), op2);
tmp1 = gen_reg_rtx (SImode);
tmp2 = gen_reg_rtx (SImode);
- emit_insn (gen_ashlsi3 (tmp1, gen_lowpart (SImode, op2), GEN_INT (16)));
- emit_move_insn (tmp2, gen_lowpart (SImode, op2));
+ emit_insn (gen_ashlsi3 (tmp1, gen_lowpart (SImode, op0), GEN_INT (16)));
+ emit_move_insn (tmp2, gen_lowpart (SImode, op0));
emit_insn (gen_movstricthi_1 (gen_lowpart (HImode, tmp2), const0_rtx));
emit_insn (gen_load_accumulator_pair (accvec, tmp1, tmp2));
- emit_insn (gen_flag_macv2hi_parts_acconly (accvec, op0, op1, const0_rtx,
- const0_rtx, const0_rtx,
- const1_rtx, accvec, const0_rtx,
- const0_rtx,
- GEN_INT (MACFLAG_W32)));
- tmp1 = (fcode == BFIN_BUILTIN_CPLX_MAC_16 ? const1_rtx : const0_rtx);
- tmp2 = (fcode == BFIN_BUILTIN_CPLX_MAC_16 ? const0_rtx : const1_rtx);
- emit_insn (gen_flag_macv2hi_parts (target, op0, op1, const1_rtx,
+ if (fcode == BFIN_BUILTIN_CPLX_MAC_16
+ || fcode == BFIN_BUILTIN_CPLX_MSU_16)
+ emit_insn (gen_flag_macv2hi_parts_acconly (accvec, op1, op2, const0_rtx,
+ const0_rtx, const0_rtx,
+ const1_rtx, accvec, const0_rtx,
+ const0_rtx,
+ GEN_INT (MACFLAG_W32)));
+ else
+ emit_insn (gen_flag_macv2hi_parts_acconly (accvec, op1, op2, const0_rtx,
+ const0_rtx, const0_rtx,
+ const1_rtx, accvec, const0_rtx,
+ const0_rtx,
+ GEN_INT (MACFLAG_NONE)));
+ if (fcode == BFIN_BUILTIN_CPLX_MAC_16
+ || fcode == BFIN_BUILTIN_CPLX_MAC_16_S40)
+ {
+ tmp1 = const1_rtx;
+ tmp2 = const0_rtx;
+ }
+ else
+ {
+ tmp1 = const0_rtx;
+ tmp2 = const1_rtx;
+ }
+ emit_insn (gen_flag_macv2hi_parts (target, op1, op2, const1_rtx,
const1_rtx, const1_rtx,
const0_rtx, accvec, tmp1, tmp2,
GEN_INT (MACFLAG_NONE), accvec));
return target;
+ case BFIN_BUILTIN_CPLX_SQU:
+ arg0 = CALL_EXPR_ARG (exp, 0);
+ op0 = expand_expr (arg0, NULL_RTX, VOIDmode, 0);
+ accvec = gen_reg_rtx (V2PDImode);
+ icode = CODE_FOR_flag_mulv2hi;
+ tmp1 = gen_reg_rtx (V2HImode);
+ tmp2 = gen_reg_rtx (V2HImode);
+
+ if (! target
+ || GET_MODE (target) != V2HImode
+ || ! (*insn_data[icode].operand[0].predicate) (target, V2HImode))
+ target = gen_reg_rtx (V2HImode);
+ if (! register_operand (op0, GET_MODE (op0)))
+ op0 = copy_to_mode_reg (GET_MODE (op0), op0);
+
+ emit_insn (gen_flag_mulv2hi (tmp1, op0, op0, GEN_INT (MACFLAG_NONE)));
+
+ emit_insn (gen_flag_mulhi_parts (tmp2, op0, op0, const0_rtx,
+ const0_rtx, const1_rtx,
+ GEN_INT (MACFLAG_NONE)));
+
+ emit_insn (gen_ssaddhi3_parts (target, tmp2, tmp2, const1_rtx,
+ const0_rtx, const0_rtx));
+
+ emit_insn (gen_sssubhi3_parts (target, tmp1, tmp1, const0_rtx,
+ const0_rtx, const1_rtx));
+
+ return target;
+
default:
break;
}
for (i = 0, d = bdesc_2arg; i < ARRAY_SIZE (bdesc_2arg); i++, d++)
if (d->code == fcode)
- return bfin_expand_binop_builtin (d->icode, arglist, target,
+ return bfin_expand_binop_builtin (d->icode, exp, target,
d->macflag);
for (i = 0, d = bdesc_1arg; i < ARRAY_SIZE (bdesc_1arg); i++, d++)
if (d->code == fcode)
- return bfin_expand_unop_builtin (d->icode, arglist, target);
+ return bfin_expand_unop_builtin (d->icode, exp, target);
gcc_unreachable ();
}
#undef TARGET_ADDRESS_COST
#define TARGET_ADDRESS_COST bfin_address_cost
-#undef TARGET_ASM_INTERNAL_LABEL
-#define TARGET_ASM_INTERNAL_LABEL bfin_internal_label
-
#undef TARGET_ASM_INTEGER
#define TARGET_ASM_INTEGER bfin_assemble_integer
#undef TARGET_ASM_OUTPUT_MI_THUNK
#define TARGET_ASM_OUTPUT_MI_THUNK bfin_output_mi_thunk
#undef TARGET_ASM_CAN_OUTPUT_MI_THUNK
-#define TARGET_ASM_CAN_OUTPUT_MI_THUNK hook_bool_tree_hwi_hwi_tree_true
+#define TARGET_ASM_CAN_OUTPUT_MI_THUNK hook_bool_const_tree_hwi_hwi_const_tree_true
#undef TARGET_SCHED_ADJUST_COST
#define TARGET_SCHED_ADJUST_COST bfin_adjust_cost
#define TARGET_SCHED_ISSUE_RATE bfin_issue_rate
#undef TARGET_PROMOTE_PROTOTYPES
-#define TARGET_PROMOTE_PROTOTYPES hook_bool_tree_true
+#define TARGET_PROMOTE_PROTOTYPES hook_bool_const_tree_true
#undef TARGET_PROMOTE_FUNCTION_ARGS
-#define TARGET_PROMOTE_FUNCTION_ARGS hook_bool_tree_true
+#define TARGET_PROMOTE_FUNCTION_ARGS hook_bool_const_tree_true
#undef TARGET_PROMOTE_FUNCTION_RETURN
-#define TARGET_PROMOTE_FUNCTION_RETURN hook_bool_tree_true
+#define TARGET_PROMOTE_FUNCTION_RETURN hook_bool_const_tree_true
#undef TARGET_ARG_PARTIAL_BYTES
#define TARGET_ARG_PARTIAL_BYTES bfin_arg_partial_bytes
#undef TARGET_CANNOT_FORCE_CONST_MEM
#define TARGET_CANNOT_FORCE_CONST_MEM bfin_cannot_force_const_mem
+#undef TARGET_RETURN_IN_MEMORY
+#define TARGET_RETURN_IN_MEMORY bfin_return_in_memory
+
struct gcc_target targetm = TARGET_INITIALIZER;