diff mbox

[AArch64] Improve stack adjustment

Message ID VI1PR0802MB2621AAE5FE2CC2B7378E62D683D30@VI1PR0802MB2621.eurprd08.prod.outlook.com
State New
Headers show

Commit Message

Wilco Dijkstra Oct. 18, 2016, 6:10 p.m. UTC
James Greenhalgh wrote:
On Mon, Oct 17, 2016 at 12:38:36PM +0000, Wilco Dijkstra wrote:

>> +  /* We need two add/sub instructions, each one perform part of the
>> +     addition/subtraction, but don't this if the addend can be loaded into
>> +     register by single instruction, in that case we prefer a move to scratch
>> +     register following by addition.  */

> This sentence is missing some words.

Sorry, badly edited old comment. I decided to just rewrite it, so here is the new version:

+  /* Emit 2 additions/subtractions if the adjustment is less than 24 bits.
+     Only do this if mdelta is not a 16-bit move as adjusting using a move
+     is better.  */


> > +  if (mdelta < 0x1000000 && !aarch64_move_imm (mdelta, mode))

> Could you explain this change? The comment makes it seem like delta would
> still be correct. Probably the comment needs to say "followed by
> addition/subtraction" rather than what is currently written?

aarch64_move_imm (mdelta, mode) is not always the same as aarch64_move_imm
(delta, mode). We later emit a move using mdelta (so that both prolog and epilog use
positive adjustments). Therefore we must check mdelta here, not delta.

> > +  if (emit_move_imm)
> > +    aarch64_internal_mov_immediate (scratch_rtx, GEN_INT (mdelta), true, mode);
> > +  insn = emit_insn (delta < 0 ? gen_sub2_insn (this_rtx, scratch_rtx)
> > +                             : gen_add2_insn (this_rtx, scratch_rtx));

> What is contained in scratch_rtx here if we didn't set it up with
> aarch64_internal_move_immediate? Are you not adding junk values in the
> !emit_move_imm case?

This function should only be called with !emit_move_imm if scratchreg is known to contain
mdelta. The prolog initializes the scratch register, the liveness check is done in the epilog:

+  aarch64_add_sp (IP0_REGNUM, initial_adjust, df_regs_ever_live_p (IP0_REGNUM));

Wilco


ChangeLog:
2016-10-18  Wilco Dijkstra  <wdijkstr@arm.com>

gcc/
        * config/aarch64/aarch64.c (aarch64_add_constant_internal):
        Add extra argument to allow emitting the move immediate.
        Use add/sub with positive immediate.
        (aarch64_add_constant): Add inline function.
        (aarch64_add_sp): Likewise.
        (aarch64_sub_sp): Likewise.
        (aarch64_expand_prologue): Call aarch64_sub_sp.
        (aarch64_expand_epilogue): Call aarch64_add_sp.
        Decide when to leave out move.
        (aarch64_output_mi_thunk): Call aarch64_add_constant.

testsuite/
        * gcc.target/aarch64/test_frame_17.c: New test.
--

Comments

James Greenhalgh Oct. 21, 2016, 1:10 p.m. UTC | #1
On Tue, Oct 18, 2016 at 07:10:07PM +0100, Wilco Dijkstra wrote:
> James Greenhalgh wrote:
> On Mon, Oct 17, 2016 at 12:38:36PM +0000, Wilco Dijkstra wrote:
> 
> >> +  /* We need two add/sub instructions, each one perform part of the
> >> +     addition/subtraction, but don't this if the addend can be loaded into
> >> +     register by single instruction, in that case we prefer a move to scratch
> >> +     register following by addition.  */
> 
> > This sentence is missing some words.
> 
> Sorry, badly edited old comment. I decided to just rewrite it, so here is the new version:
> 
> +  /* Emit 2 additions/subtractions if the adjustment is less than 24 bits.
> +     Only do this if mdelta is not a 16-bit move as adjusting using a move
> +     is better.  */
> 
> 
> > > +  if (mdelta < 0x1000000 && !aarch64_move_imm (mdelta, mode))
> 
> > Could you explain this change? The comment makes it seem like delta would
> > still be correct. Probably the comment needs to say "followed by
> > addition/subtraction" rather than what is currently written?
> 
> aarch64_move_imm (mdelta, mode) is not always the same as aarch64_move_imm
> (delta, mode). We later emit a move using mdelta (so that both prolog and epilog use
> positive adjustments). Therefore we must check mdelta here, not delta.
> 
> > > +  if (emit_move_imm)
> > > +    aarch64_internal_mov_immediate (scratch_rtx, GEN_INT (mdelta), true, mode);
> > > +  insn = emit_insn (delta < 0 ? gen_sub2_insn (this_rtx, scratch_rtx)
> > > +                             : gen_add2_insn (this_rtx, scratch_rtx));
> 
> > What is contained in scratch_rtx here if we didn't set it up with
> > aarch64_internal_move_immediate? Are you not adding junk values in the
> > !emit_move_imm case?
> 
> This function should only be called with !emit_move_imm if scratchreg is known to contain
> mdelta. The prolog initializes the scratch register, the liveness check is done in the epilog:
> 
> +  aarch64_add_sp (IP0_REGNUM, initial_adjust, df_regs_ever_live_p (IP0_REGNUM));

OK. I see where I went wrong here. I thought you were constructing a fresh
scratch register, and only initialising it if emit_move_imm - now I see that
scratch_rtx is just wrapping scratchreg.

This is OK.

Thanks,
James
diff mbox

Patch

diff --git a/gcc/config/aarch64/aarch64.c b/gcc/config/aarch64/aarch64.c
index 899ce219763a1b62e20de49e279b3fe26ee4bc05..582a9a752f788b57c6e762005c35b460eb45bec8 100644
--- a/gcc/config/aarch64/aarch64.c
+++ b/gcc/config/aarch64/aarch64.c
@@ -1956,25 +1956,30 @@  aarch64_expand_mov_immediate (rtx dest, rtx imm)
 }
 
 /* Add DELTA to REGNUM in mode MODE.  SCRATCHREG can be used to held
-   intermediate value if necessary.
+   intermediate value if necessary.  FRAME_RELATED_P should be true if
+   the RTX_FRAME_RELATED flag should be set and CFA adjustments added
+   to the generated instructions.  If SCRATCHREG is known to hold
+   abs (delta), EMIT_MOVE_IMM can be set to false to avoid emitting the
+   immediate again.
 
-   This function is sometimes used to adjust the stack pointer, so we must
-   ensure that it can never cause transient stack deallocation by writing an
-   invalid value into REGNUM.  */
+   Since this function may be used to adjust the stack pointer, we must
+   ensure that it cannot cause transient stack deallocation (for example
+   by first incrementing SP and then decrementing when adjusting by a
+   large immediate).  */
 
 static void
-aarch64_add_constant (machine_mode mode, int regnum, int scratchreg,
-		      HOST_WIDE_INT delta, bool frame_related_p)
+aarch64_add_constant_internal (machine_mode mode, int regnum, int scratchreg,
+			       HOST_WIDE_INT delta, bool frame_related_p,
+			       bool emit_move_imm)
 {
   HOST_WIDE_INT mdelta = abs_hwi (delta);
   rtx this_rtx = gen_rtx_REG (mode, regnum);
   rtx_insn *insn;
 
-  /* Do nothing if mdelta is zero.  */
   if (!mdelta)
     return;
 
-  /* We only need single instruction if the offset fit into add/sub.  */
+  /* Single instruction adjustment.  */
   if (aarch64_uimm12_shift (mdelta))
     {
       insn = emit_insn (gen_add2_insn (this_rtx, GEN_INT (delta)));
@@ -1982,11 +1987,10 @@  aarch64_add_constant (machine_mode mode, int regnum, int scratchreg,
       return;
     }
 
-  /* We need two add/sub instructions, each one performing part of the
-     calculation.  Don't do this if the addend can be loaded into register with
-     a single instruction, in that case we prefer a move to a scratch register
-     following by an addition.  */
-  if (mdelta < 0x1000000 && !aarch64_move_imm (delta, mode))
+  /* Emit 2 additions/subtractions if the adjustment is less than 24 bits.
+     Only do this if mdelta is not a 16-bit move as adjusting using a move
+     is better.  */
+  if (mdelta < 0x1000000 && !aarch64_move_imm (mdelta, mode))
     {
       HOST_WIDE_INT low_off = mdelta & 0xfff;
 
@@ -1998,10 +2002,12 @@  aarch64_add_constant (machine_mode mode, int regnum, int scratchreg,
       return;
     }
 
-  /* Otherwise use generic function to handle all other situations.  */
+  /* Emit a move immediate if required and an addition/subtraction.  */
   rtx scratch_rtx = gen_rtx_REG (mode, scratchreg);
-  aarch64_internal_mov_immediate (scratch_rtx, GEN_INT (delta), true, mode);
-  insn = emit_insn (gen_add2_insn (this_rtx, scratch_rtx));
+  if (emit_move_imm)
+    aarch64_internal_mov_immediate (scratch_rtx, GEN_INT (mdelta), true, mode);
+  insn = emit_insn (delta < 0 ? gen_sub2_insn (this_rtx, scratch_rtx)
+			      : gen_add2_insn (this_rtx, scratch_rtx));
   if (frame_related_p)
     {
       RTX_FRAME_RELATED_P (insn) = frame_related_p;
@@ -2010,6 +2016,27 @@  aarch64_add_constant (machine_mode mode, int regnum, int scratchreg,
     }
 }
 
+static inline void
+aarch64_add_constant (machine_mode mode, int regnum, int scratchreg,
+		      HOST_WIDE_INT delta)
+{
+  aarch64_add_constant_internal (mode, regnum, scratchreg, delta, false, true);
+}
+
+static inline void
+aarch64_add_sp (int scratchreg, HOST_WIDE_INT delta, bool emit_move_imm)
+{
+  aarch64_add_constant_internal (Pmode, SP_REGNUM, scratchreg, delta,
+				 true, emit_move_imm);
+}
+
+static inline void
+aarch64_sub_sp (int scratchreg, HOST_WIDE_INT delta, bool frame_related_p)
+{
+  aarch64_add_constant_internal (Pmode, SP_REGNUM, scratchreg, -delta,
+				 frame_related_p, true);
+}
+
 static bool
 aarch64_function_ok_for_sibcall (tree decl ATTRIBUTE_UNUSED,
 				 tree exp ATTRIBUTE_UNUSED)
@@ -3223,7 +3250,7 @@  aarch64_expand_prologue (void)
 	aarch64_emit_probe_stack_range (STACK_CHECK_PROTECT, frame_size);
     }
 
-  aarch64_add_constant (Pmode, SP_REGNUM, IP0_REGNUM, -initial_adjust, true);
+  aarch64_sub_sp (IP0_REGNUM, initial_adjust, true);
 
   if (callee_adjust != 0)
     aarch64_push_regs (reg1, reg2, callee_adjust);
@@ -3244,8 +3271,7 @@  aarch64_expand_prologue (void)
 			     callee_adjust != 0 || frame_pointer_needed);
   aarch64_save_callee_saves (DFmode, callee_offset, V0_REGNUM, V31_REGNUM,
 			     callee_adjust != 0 || frame_pointer_needed);
-  aarch64_add_constant (Pmode, SP_REGNUM, IP1_REGNUM, -final_adjust,
-			!frame_pointer_needed);
+  aarch64_sub_sp (IP1_REGNUM, final_adjust, !frame_pointer_needed);
 }
 
 /* Return TRUE if we can use a simple_return insn.
@@ -3311,7 +3337,7 @@  aarch64_expand_epilogue (bool for_sibcall)
       RTX_FRAME_RELATED_P (insn) = callee_adjust == 0;
     }
   else
-    aarch64_add_constant (Pmode, SP_REGNUM, IP1_REGNUM, final_adjust, true);
+    aarch64_add_sp (IP1_REGNUM, final_adjust, df_regs_ever_live_p (IP1_REGNUM));
 
   aarch64_restore_callee_saves (DImode, callee_offset, R0_REGNUM, R30_REGNUM,
 				callee_adjust != 0, &cfi_ops);
@@ -3334,7 +3360,7 @@  aarch64_expand_epilogue (bool for_sibcall)
       cfi_ops = NULL;
     }
 
-  aarch64_add_constant (Pmode, SP_REGNUM, IP0_REGNUM, initial_adjust, true);
+  aarch64_add_sp (IP0_REGNUM, initial_adjust, df_regs_ever_live_p (IP0_REGNUM));
 
   if (cfi_ops)
     {
@@ -3392,7 +3418,7 @@  aarch64_output_mi_thunk (FILE *file, tree thunk ATTRIBUTE_UNUSED,
   emit_note (NOTE_INSN_PROLOGUE_END);
 
   if (vcall_offset == 0)
-    aarch64_add_constant (Pmode, this_regno, IP1_REGNUM, delta, false);
+    aarch64_add_constant (Pmode, this_regno, IP1_REGNUM, delta);
   else
     {
       gcc_assert ((vcall_offset & (POINTER_BYTES - 1)) == 0);
@@ -3408,7 +3434,7 @@  aarch64_output_mi_thunk (FILE *file, tree thunk ATTRIBUTE_UNUSED,
 	    addr = gen_rtx_PRE_MODIFY (Pmode, this_rtx,
 				       plus_constant (Pmode, this_rtx, delta));
 	  else
-	    aarch64_add_constant (Pmode, this_regno, IP1_REGNUM, delta, false);
+	    aarch64_add_constant (Pmode, this_regno, IP1_REGNUM, delta);
 	}
 
       if (Pmode == ptr_mode)
diff --git a/gcc/testsuite/gcc.target/aarch64/test_frame_17.c b/gcc/testsuite/gcc.target/aarch64/test_frame_17.c
new file mode 100644
index 0000000000000000000000000000000000000000..c214431999b60cce8a75204876a8c73ec6304128
--- /dev/null
+++ b/gcc/testsuite/gcc.target/aarch64/test_frame_17.c
@@ -0,0 +1,21 @@ 
+/* { dg-do compile } */
+/* { dg-options "-O2 --save-temps" } */
+
+/* Test reuse of stack adjustment temporaries.  */
+
+void foo ();
+
+int reuse_mov (int i)
+{
+  int arr[1025];
+  return arr[i];
+}
+
+int no_reuse_mov (int i)
+{
+  int arr[1025];
+  foo ();
+  return arr[i];
+}
+
+/* { dg-final { scan-assembler-times "mov\tx16, \[0-9\]+" 3 } } */