Delay cfi restore opcodes until the stack frame is deallocated.
This reduces the number of cfi advance opcodes required.
We perform a similar optimization in the x86_64 epilogue.
* config/aarch64/aarch64.c (aarch64_popwb_single_reg): Remove.
(aarch64_popwb_pair_reg): Remove.
(aarch64_restore_callee_saves): Add CFI_OPS argument; fill it with
the restore ops performed by the insns generated.
(aarch64_expand_epilogue): Attach CFI_OPS to the stack deallocation
insn. Perform the calls_eh_return addition later; do not attempt to
preserve the CFA in that case. Don't use aarch64_set_frame_expr.
---
gcc/config/aarch64/aarch64.c | 177 +++++++++++++------------------------------
1 file changed, 52 insertions(+), 125 deletions(-)
diff --git a/gcc/config/aarch64/aarch64.c b/gcc/config/aarch64/aarch64.c
index c3c871e..9a11e05 100644
--- a/gcc/config/aarch64/aarch64.c
+++ b/gcc/config/aarch64/aarch64.c
@@ -1964,23 +1964,6 @@ aarch64_pushwb_single_reg (enum machine_mode mode,
unsigned regno,
RTX_FRAME_RELATED_P (insn) = 1;
}
-static void
-aarch64_popwb_single_reg (enum machine_mode mode, unsigned regno,
- HOST_WIDE_INT adjustment)
-{
- rtx base_rtx = stack_pointer_rtx;
- rtx insn, reg, mem;
-
- reg = gen_rtx_REG (mode, regno);
- mem = gen_rtx_POST_MODIFY (Pmode, base_rtx,
- plus_constant (Pmode, base_rtx, adjustment));
- mem = gen_rtx_MEM (mode, mem);
-
- insn = emit_move_insn (reg, mem);
- add_reg_note (insn, REG_CFA_RESTORE, reg);
- RTX_FRAME_RELATED_P (insn) = 1;
-}
-
static rtx
aarch64_gen_storewb_pair (enum machine_mode mode, rtx base, rtx reg, rtx reg2,
HOST_WIDE_INT adjustment)
@@ -2011,7 +1994,6 @@ aarch64_pushwb_pair_reg (enum machine_mode mode, unsigned
regno1,
insn = emit_insn (aarch64_gen_storewb_pair (mode, stack_pointer_rtx, reg1,
reg2, adjustment));
RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 2)) = 1;
-
RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 1)) = 1;
RTX_FRAME_RELATED_P (insn) = 1;
}
@@ -2033,29 +2015,6 @@ aarch64_gen_loadwb_pair (enum machine_mode mode, rtx
base, rtx reg, rtx reg2,
}
}
-static void
-aarch64_popwb_pair_reg (enum machine_mode mode, unsigned regno1,
- unsigned regno2, HOST_WIDE_INT adjustment, rtx cfa)
-{
- rtx insn;
- rtx reg1 = gen_rtx_REG (mode, regno1);
- rtx reg2 = gen_rtx_REG (mode, regno2);
-
- insn = emit_insn (aarch64_gen_loadwb_pair (mode, stack_pointer_rtx, reg1,
- reg2, adjustment));
- RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 2)) = 1;
- RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 1)) = 1;
- RTX_FRAME_RELATED_P (insn) = 1;
-
- if (cfa)
- add_reg_note (insn, REG_CFA_ADJUST_CFA,
- (gen_rtx_SET (Pmode, stack_pointer_rtx,
- plus_constant (Pmode, cfa, adjustment))));
-
- add_reg_note (insn, REG_CFA_RESTORE, reg1);
- add_reg_note (insn, REG_CFA_RESTORE, reg2);
-}
-
static rtx
aarch64_gen_store_pair (enum machine_mode mode, rtx mem1, rtx reg1, rtx mem2,
rtx reg2)
@@ -2151,9 +2110,8 @@ aarch64_save_callee_saves (enum machine_mode mode,
HOST_WIDE_INT start_offset,
static void
aarch64_restore_callee_saves (enum machine_mode mode,
HOST_WIDE_INT start_offset, unsigned start,
- unsigned limit, bool skip_wb)
+ unsigned limit, bool skip_wb, rtx *cfi_ops)
{
- rtx insn;
rtx base_rtx = stack_pointer_rtx;
rtx (*gen_mem_ref) (enum machine_mode, rtx) = (frame_pointer_needed
? gen_frame_mem :
gen_rtx_MEM);
@@ -2187,25 +2145,14 @@ aarch64_restore_callee_saves (enum machine_mode mode,
offset = start_offset + cfun->machine->frame.reg_offset[regno2];
mem2 = gen_mem_ref (mode, plus_constant (Pmode, base_rtx, offset));
- insn = emit_insn (aarch64_gen_load_pair (mode, reg, mem, reg2,
- mem2));
- add_reg_note (insn, REG_CFA_RESTORE, reg);
- add_reg_note (insn, REG_CFA_RESTORE, reg2);
+ emit_insn (aarch64_gen_load_pair (mode, reg, mem, reg2, mem2));
- /* The first part of a frame-related parallel insn is
- always assumed to be relevant to the frame
- calculations; subsequent parts, are only
- frame-related if explicitly marked. */
- RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 1)) = 1;
+ *cfi_ops = alloc_reg_note (REG_CFA_RESTORE, reg2, *cfi_ops);
regno = regno2;
}
else
- {
- insn = emit_move_insn (reg, mem);
- add_reg_note (insn, REG_CFA_RESTORE, reg);
- }
-
- RTX_FRAME_RELATED_P (insn) = 1;
+ emit_move_insn (reg, mem);
+ *cfi_ops = alloc_reg_note (REG_CFA_RESTORE, reg, *cfi_ops);
}
}
@@ -2418,7 +2365,6 @@ aarch64_expand_epilogue (bool for_sibcall)
HOST_WIDE_INT frame_size, offset;
HOST_WIDE_INT fp_offset;
rtx insn;
- rtx cfa_reg;
aarch64_layout_frame ();
@@ -2426,8 +2372,6 @@ aarch64_expand_epilogue (bool for_sibcall)
fp_offset = cfun->machine->frame.frame_size
- cfun->machine->frame.hard_fp_offset;
- cfa_reg = frame_pointer_needed ? hard_frame_pointer_rtx : stack_pointer_rtx;
-
/* Store pairs and load pairs have a range only -512 to 504. */
if (offset >= 512)
{
@@ -2459,11 +2403,6 @@ aarch64_expand_epilogue (bool for_sibcall)
hard_frame_pointer_rtx,
GEN_INT (0)));
offset = offset - fp_offset;
- RTX_FRAME_RELATED_P (insn) = 1;
- /* As SP is set to (FP - fp_offset), according to the rules in
- dwarf2cfi.c:dwarf2out_frame_debug_expr, CFA should be calculated
- from the value of SP from now on. */
- cfa_reg = stack_pointer_rtx;
}
if (offset > 0)
@@ -2471,6 +2410,7 @@ aarch64_expand_epilogue (bool for_sibcall)
unsigned reg1 = cfun->machine->frame.wb_candidate1;
unsigned reg2 = cfun->machine->frame.wb_candidate2;
bool skip_wb = true;
+ rtx cfi_ops = NULL;
if (frame_pointer_needed)
fp_offset = 0;
@@ -2481,99 +2421,86 @@ aarch64_expand_epilogue (bool for_sibcall)
skip_wb = false;
aarch64_restore_callee_saves (DImode, fp_offset, R0_REGNUM, R30_REGNUM,
- skip_wb);
+ skip_wb, &cfi_ops);
aarch64_restore_callee_saves (DFmode, fp_offset, V0_REGNUM, V31_REGNUM,
- skip_wb);
+ skip_wb, &cfi_ops);
if (skip_wb)
{
enum machine_mode mode1 = (reg1 <= R30_REGNUM) ? DImode : DFmode;
+ rtx rreg1 = gen_rtx_REG (mode1, reg1);
+ cfi_ops = alloc_reg_note (REG_CFA_RESTORE, rreg1, cfi_ops);
if (reg2 == FIRST_PSEUDO_REGISTER)
- aarch64_popwb_single_reg (mode1, reg1, offset);
+ {
+ rtx mem = plus_constant (Pmode, stack_pointer_rtx, offset);
+ mem = gen_rtx_POST_MODIFY (Pmode, stack_pointer_rtx, mem);
+ mem = gen_rtx_MEM (mode1, mem);
+ insn = emit_move_insn (rreg1, mem);
+ }
else
{
- if (reg1 != HARD_FRAME_POINTER_REGNUM)
- cfa_reg = NULL;
+ rtx rreg2 = gen_rtx_REG (mode1, reg2);
- aarch64_popwb_pair_reg (mode1, reg1, reg2, offset, cfa_reg);
+ cfi_ops = alloc_reg_note (REG_CFA_RESTORE, rreg2, cfi_ops);
+ insn = aarch64_gen_loadwb_pair (mode1, stack_pointer_rtx, rreg1,
+ rreg2, offset);
+ insn = emit_insn (insn);
}
}
else
{
insn = emit_insn (gen_add2_insn (stack_pointer_rtx,
GEN_INT (offset)));
- RTX_FRAME_RELATED_P (insn) = 1;
}
- }
-
- /* Stack adjustment for exception handler. */
- if (crtl->calls_eh_return)
- {
- /* We need to unwind the stack by the offset computed by
- EH_RETURN_STACKADJ_RTX. However, at this point the CFA is
- based on SP. Ideally we would update the SP and define the
- CFA along the lines of:
-
- SP = SP + EH_RETURN_STACKADJ_RTX
- (regnote CFA = SP - EH_RETURN_STACKADJ_RTX)
-
- However the dwarf emitter only understands a constant
- register offset.
-
- The solution chosen here is to use the otherwise unused IP0
- as a temporary register to hold the current SP value. The
- CFA is described using IP0 then SP is modified. */
- rtx ip0 = gen_rtx_REG (DImode, IP0_REGNUM);
-
- insn = emit_move_insn (ip0, stack_pointer_rtx);
- add_reg_note (insn, REG_CFA_DEF_CFA, ip0);
+ /* Reset the CFA to be SP + FRAME_SIZE. */
+ rtx new_cfa = stack_pointer_rtx;
+ if (frame_size > 0)
+ new_cfa = plus_constant (Pmode, new_cfa, frame_size);
+ cfi_ops = alloc_reg_note (REG_CFA_DEF_CFA, new_cfa, cfi_ops);
+ REG_NOTES (insn) = cfi_ops;
RTX_FRAME_RELATED_P (insn) = 1;
-
- emit_insn (gen_add2_insn (stack_pointer_rtx, EH_RETURN_STACKADJ_RTX));
-
- /* Ensure the assignment to IP0 does not get optimized away. */
- emit_use (ip0);
}
- if (frame_size > -1)
+ if (frame_size > 0)
{
if (frame_size >= 0x1000000)
{
rtx op0 = gen_rtx_REG (Pmode, IP0_REGNUM);
emit_move_insn (op0, GEN_INT (frame_size));
- emit_insn (gen_add2_insn (stack_pointer_rtx, op0));
- aarch64_set_frame_expr (gen_rtx_SET
- (Pmode, stack_pointer_rtx,
- plus_constant (Pmode,
- stack_pointer_rtx,
- frame_size)));
+ insn = emit_insn (gen_add2_insn (stack_pointer_rtx, op0));
}
- else if (frame_size > 0)
+ else
{
- if ((frame_size & 0xfff) != 0)
- {
- insn = emit_insn (gen_add2_insn
- (stack_pointer_rtx,
- GEN_INT ((frame_size
- & (HOST_WIDE_INT) 0xfff))));
- RTX_FRAME_RELATED_P (insn) = 1;
- }
- if ((frame_size & 0xfff) != frame_size)
+ int hi_ofs = frame_size & 0xfff000;
+ int lo_ofs = frame_size & 0x000fff;
+
+ if (hi_ofs && lo_ofs)
{
insn = emit_insn (gen_add2_insn
- (stack_pointer_rtx,
- GEN_INT ((frame_size
- & ~ (HOST_WIDE_INT) 0xfff))));
+ (stack_pointer_rtx, GEN_INT (hi_ofs)));
RTX_FRAME_RELATED_P (insn) = 1;
+ frame_size = lo_ofs;
}
+ insn = emit_insn (gen_add2_insn
+ (stack_pointer_rtx, GEN_INT (frame_size)));
}
- aarch64_set_frame_expr (gen_rtx_SET (Pmode, stack_pointer_rtx,
- plus_constant (Pmode,
- stack_pointer_rtx,
- offset)));
+ /* Reset the CFA to be SP + 0. */
+ add_reg_note (insn, REG_CFA_DEF_CFA, stack_pointer_rtx);
+ RTX_FRAME_RELATED_P (insn) = 1;
+ }
+
+ /* Stack adjustment for exception handler. */
+ if (crtl->calls_eh_return)
+ {
+ /* We need to unwind the stack by the offset computed by
+ EH_RETURN_STACKADJ_RTX. We have already reset the CFA
+ to be SP; letting the CFA move during this adjustment
+ is just as correct as retaining the CFA from the body
+ of the function. Therefore, do nothing special. */
+ emit_insn (gen_add2_insn (stack_pointer_rtx, EH_RETURN_STACKADJ_RTX));
}
emit_use (gen_rtx_REG (DImode, LR_REGNUM));
--
1.8.3.1
