From f98495d90ba66f67fe922a4b9229ea787041c418 Mon Sep 17 00:00:00 2001
From: thopre01 <thopre01@138bc75d-0d04-0410-961f-82ee72b054a4>
Date: Thu, 22 Nov 2018 14:46:17 +0000
Subject: [PATCH] PR85434: Prevent spilling of stack protector guard's address
 on ARM

In case of high register pressure in PIC mode, address of the stack
protector's guard can be spilled on ARM targets as shown in PR85434,
thus allowing an attacker to control what the canary would be compared
against. ARM does lack stack_protect_set and stack_protect_test insn
patterns, defining them does not help as the address is expanded
regularly and the patterns only deal with the copy and test of the
guard with the canary.

This problem does not occur for x86 targets because the PIC access and
the test can be done in the same instruction. Aarch64 is exempt too
because PIC access insn pattern are mov of UNSPEC which prevents it from
the second access in the epilogue being CSEd in cse_local pass with the
first access in the prologue.

The approach followed here is to create new "combined" set and test
standard pattern names that take the unexpanded guard and do the set or
test. This allows the target to use an opaque pattern (eg. using UNSPEC)
to hide the individual instructions being generated to the compiler and
split the pattern into generic load, compare and branch instruction
after register allocator, therefore avoiding any spilling. This is here
implemented for the ARM targets. For targets not implementing these new
standard pattern names, the existing stack_protect_set and
stack_protect_test pattern names are used.

To be able to split PIC access after register allocation, the functions
had to be augmented to force a new PIC register load and to control
which register it loads into. This is because sharing the PIC register
between prologue and epilogue could lead to spilling due to CSE again
which an attacker could use to control what the canary gets compared
against.

2018-11-22  Thomas Preud'homme  <thomas.preudhomme@linaro.org>

    gcc/
    PR target/85434
    * target-insns.def (stack_protect_combined_set): Define new standard
    pattern name.
    (stack_protect_combined_test): Likewise.
    * cfgexpand.c (stack_protect_prologue): Try new
    stack_protect_combined_set pattern first.
    * function.c (stack_protect_epilogue): Try new
    stack_protect_combined_test pattern first.
    * config/arm/arm.c (require_pic_register): Add pic_reg and compute_now
    parameters to control which register to use as PIC register and force
    reloading PIC register respectively.  Insert in the stream of insns if
    possible.
    (legitimize_pic_address): Expose above new parameters in prototype and
    adapt recursive calls accordingly.  Use pic_reg if non null instead of
    cached one.
    (arm_load_pic_register): Add pic_reg parameter and use it if non null.
    (arm_legitimize_address): Adapt to new legitimize_pic_address
    prototype.
    (thumb_legitimize_address): Likewise.
    (arm_emit_call_insn): Adapt to require_pic_register prototype change.
    (arm_expand_prologue): Adapt to arm_load_pic_register prototype change.
    (thumb1_expand_prologue): Likewise.
    * config/arm/arm-protos.h (legitimize_pic_address): Adapt to prototype
    change.
    (arm_load_pic_register): Likewise.
    * config/arm/predicated.md (guard_addr_operand): New predicate.
    (guard_operand): New predicate.
    * config/arm/arm.md (movsi expander): Adapt to legitimize_pic_address
    prototype change.
    (builtin_setjmp_receiver expander): Adapt to thumb1_expand_prologue
    prototype change.
    (stack_protect_combined_set): New expander..
    (stack_protect_combined_set_insn): New insn_and_split pattern.
    (stack_protect_set_insn): New insn pattern.
    (stack_protect_combined_test): New expander.
    (stack_protect_combined_test_insn): New insn_and_split pattern.
    (arm_stack_protect_test_insn): New insn pattern.
    * config/arm/thumb1.md (thumb1_stack_protect_test_insn): New insn pattern.
    * config/arm/unspecs.md (UNSPEC_SP_SET): New unspec.
    (UNSPEC_SP_TEST): Likewise.
    * doc/md.texi (stack_protect_combined_set): Document new standard
    pattern name.
    (stack_protect_set): Clarify that the operand for guard's address is
    legal.
    (stack_protect_combined_test): Document new standard pattern name.
    (stack_protect_test): Clarify that the operand for guard's address is
    legal.


git-svn-id: svn+ssh://gcc.gnu.org/svn/gcc/trunk@266379 138bc75d-0d04-0410-961f-82ee72b054a4

Upstream-Status: Backport
CVE: CVE-2018-12886
Signed-off-by: Zhixiong Chi <zhixiong.chi@windriver.com>
---
 gcc/ChangeLog                          |  49 ++++++
 gcc/cfgexpand.c                        |  17 +++
 gcc/config/arm/arm-protos.h            |   4 +-
 gcc/config/arm/arm.c                   |  87 ++++++++---
 gcc/config/arm/arm.md                  | 163 +++++++++++++++++++-
 gcc/config/arm/predicates.md           |  17 +++
 gcc/config/arm/thumb1.md               |  13 ++
 gcc/config/arm/unspecs.md              |   3 +
 gcc/doc/md.texi                        |  55 ++++++-
 gcc/function.c                         |  32 +++-
 gcc/target-insns.def                   |   2 +
 11 files changed, 399 insertions(+), 43 deletions(-)
 create mode 100644 gcc/testsuite/gcc.target/arm/pr85434.c

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index e2ebfd34214..fa41e7112e0 100644
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1537,6 +1537,55 @@
 	* config/arm/neon.md (movv4hf, movv8hf): Refactored to..
 	(mov<mov>): ..this and enable unconditionally.
 
+2018-11-22  Thomas Preud'homme  <thomas.preudhomme@linaro.org>
+
+	* target-insns.def (stack_protect_combined_set): Define new standard
+	pattern name.
+	(stack_protect_combined_test): Likewise.
+	* cfgexpand.c (stack_protect_prologue): Try new
+	stack_protect_combined_set pattern first.
+	* function.c (stack_protect_epilogue): Try new
+	stack_protect_combined_test pattern first.
+	* config/arm/arm.c (require_pic_register): Add pic_reg and compute_now
+	parameters to control which register to use as PIC register and force
+	reloading PIC register respectively.  Insert in the stream of insns if
+	possible.
+	(legitimize_pic_address): Expose above new parameters in prototype and
+	adapt recursive calls accordingly.  Use pic_reg if non null instead of
+	cached one.
+	(arm_load_pic_register): Add pic_reg parameter and use it if non null.
+	(arm_legitimize_address): Adapt to new legitimize_pic_address
+	prototype.
+	(thumb_legitimize_address): Likewise.
+	(arm_emit_call_insn): Adapt to require_pic_register prototype change.
+	(arm_expand_prologue): Adapt to arm_load_pic_register prototype change.
+	(thumb1_expand_prologue): Likewise.
+	* config/arm/arm-protos.h (legitimize_pic_address): Adapt to prototype
+	change.
+	(arm_load_pic_register): Likewise.
+	* config/arm/predicated.md (guard_addr_operand): New predicate.
+	(guard_operand): New predicate.
+	* config/arm/arm.md (movsi expander): Adapt to legitimize_pic_address
+	prototype change.
+	(builtin_setjmp_receiver expander): Adapt to thumb1_expand_prologue
+	prototype change.
+	(stack_protect_combined_set): New expander..
+	(stack_protect_combined_set_insn): New insn_and_split pattern.
+	(stack_protect_set_insn): New insn pattern.
+	(stack_protect_combined_test): New expander.
+	(stack_protect_combined_test_insn): New insn_and_split pattern.
+	(arm_stack_protect_test_insn): New insn pattern.
+	* config/arm/thumb1.md (thumb1_stack_protect_test_insn): New insn pattern.
+	* config/arm/unspecs.md (UNSPEC_SP_SET): New unspec.
+	(UNSPEC_SP_TEST): Likewise.
+	* doc/md.texi (stack_protect_combined_set): Document new standard
+	pattern name.
+	(stack_protect_set): Clarify that the operand for guard's address is
+	legal.
+	(stack_protect_combined_test): Document new standard pattern name.
+	(stack_protect_test): Clarify that the operand for guard's address is
+	legal.
+
 2018-11-22  Uros Bizjak  <ubizjak@gmail.com>
 
 	Backport from mainline
diff --git a/gcc/cfgexpand.c b/gcc/cfgexpand.c
index 8fa392fcd8a..21bdcdaeaa3 100644
--- a/gcc/cfgexpand.c
+++ b/gcc/cfgexpand.c
@@ -6185,6 +6185,23 @@ stack_protect_prologue (void)
   rtx x, y;
 
   x = expand_normal (crtl->stack_protect_guard);
+
+  if (targetm.have_stack_protect_combined_set () && guard_decl)
+    {
+      gcc_assert (DECL_P (guard_decl));
+      y = DECL_RTL (guard_decl);
+
+      /* Allow the target to compute address of Y and copy it to X without
+	 leaking Y into a register.  This combined address + copy pattern
+	 allows the target to prevent spilling of any intermediate results by
+	 splitting it after register allocator.  */
+      if (rtx_insn *insn = targetm.gen_stack_protect_combined_set (x, y))
+	{
+	  emit_insn (insn);
+	  return;
+	}
+    }
+
   if (guard_decl)
     y = expand_normal (guard_decl);
   else
diff --git a/gcc/config/arm/arm-protos.h b/gcc/config/arm/arm-protos.h
index 8d6d2395b84..00f5f16ed02 100644
--- a/gcc/config/arm/arm-protos.h
+++ b/gcc/config/arm/arm-protos.h
@@ -28,7 +28,7 @@ extern enum unwind_info_type arm_except_unwind_info (struct gcc_options *);
 extern int use_return_insn (int, rtx);
 extern bool use_simple_return_p (void);
 extern enum reg_class arm_regno_class (int);
-extern void arm_load_pic_register (unsigned long);
+extern void arm_load_pic_register (unsigned long, rtx);
 extern int arm_volatile_func (void);
 extern void arm_expand_prologue (void);
 extern void arm_expand_epilogue (bool);
@@ -69,7 +69,7 @@ extern int const_ok_for_dimode_op (HOST_WIDE_INT, enum rtx_code);
 extern int arm_split_constant (RTX_CODE, machine_mode, rtx,
 			       HOST_WIDE_INT, rtx, rtx, int);
 extern int legitimate_pic_operand_p (rtx);
-extern rtx legitimize_pic_address (rtx, machine_mode, rtx);
+extern rtx legitimize_pic_address (rtx, machine_mode, rtx, rtx, bool);
 extern rtx legitimize_tls_address (rtx, rtx);
 extern bool arm_legitimate_address_p (machine_mode, rtx, bool);
 extern int arm_legitimate_address_outer_p (machine_mode, rtx, RTX_CODE, int);
diff --git a/gcc/config/arm/arm.c b/gcc/config/arm/arm.c
index 8393f0b87f3..12417de5102 100644
--- a/gcc/config/arm/arm.c
+++ b/gcc/config/arm/arm.c
@@ -7379,21 +7379,34 @@ legitimate_pic_operand_p (rtx x)
   return 1;
 }
 
-/* Record that the current function needs a PIC register.  Initialize
-   cfun->machine->pic_reg if we have not already done so.  */
+/* Record that the current function needs a PIC register.  If PIC_REG is null,
+   a new pseudo is allocated as PIC register, otherwise PIC_REG is used.  In
+   both case cfun->machine->pic_reg is initialized if we have not already done
+   so.  COMPUTE_NOW decide whether and where to set the PIC register.  If true,
+   PIC register is reloaded in the current position of the instruction stream
+   irregardless of whether it was loaded before.  Otherwise, it is only loaded
+   if not already done so (crtl->uses_pic_offset_table is null).  Note that
+   nonnull PIC_REG is only supported iff COMPUTE_NOW is true and null PIC_REG
+   is only supported iff COMPUTE_NOW is false.  */
 
 static void
-require_pic_register (void)
+require_pic_register (rtx pic_reg, bool compute_now)
 {
+  gcc_assert (compute_now == (pic_reg != NULL_RTX));
+
   /* A lot of the logic here is made obscure by the fact that this
      routine gets called as part of the rtx cost estimation process.
      We don't want those calls to affect any assumptions about the real
      function; and further, we can't call entry_of_function() until we
      start the real expansion process.  */
-  if (!crtl->uses_pic_offset_table)
+  if (!crtl->uses_pic_offset_table || compute_now)
     {
-      gcc_assert (can_create_pseudo_p ());
+      gcc_assert (can_create_pseudo_p ()
+		  || (pic_reg != NULL_RTX
+		      && REG_P (pic_reg)
+		      && GET_MODE (pic_reg) == Pmode));
       if (arm_pic_register != INVALID_REGNUM
+	  && !compute_now
 	  && !(TARGET_THUMB1 && arm_pic_register > LAST_LO_REGNUM))
 	{
 	  if (!cfun->machine->pic_reg)
@@ -7409,8 +7422,10 @@ require_pic_register (void)
 	{
 	  rtx_insn *seq, *insn;
 
+	  if (pic_reg == NULL_RTX)
+	    pic_reg = gen_reg_rtx (Pmode);
 	  if (!cfun->machine->pic_reg)
-	    cfun->machine->pic_reg = gen_reg_rtx (Pmode);
+	    cfun->machine->pic_reg = pic_reg;
 
 	  /* Play games to avoid marking the function as needing pic
 	     if we are being called as part of the cost-estimation
@@ -7421,11 +7436,12 @@ require_pic_register (void)
 	      start_sequence ();
 
 	      if (TARGET_THUMB1 && arm_pic_register != INVALID_REGNUM
-		  && arm_pic_register > LAST_LO_REGNUM)
+		  && arm_pic_register > LAST_LO_REGNUM
+		  && !compute_now)
 		emit_move_insn (cfun->machine->pic_reg,
 				gen_rtx_REG (Pmode, arm_pic_register));
 	      else
-		arm_load_pic_register (0UL);
+		arm_load_pic_register (0UL, pic_reg);
 
 	      seq = get_insns ();
 	      end_sequence ();
@@ -7438,16 +7454,33 @@ require_pic_register (void)
 	         we can't yet emit instructions directly in the final
 		 insn stream.  Queue the insns on the entry edge, they will
 		 be committed after everything else is expanded.  */
-	      insert_insn_on_edge (seq,
-				   single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun)));
+	      if (currently_expanding_to_rtl)
+		insert_insn_on_edge (seq,
+				     single_succ_edge
+				     (ENTRY_BLOCK_PTR_FOR_FN (cfun)));
+	      else
+		emit_insn (seq);
 	    }
 	}
     }
 }
 
+/* Legitimize PIC load to ORIG into REG.  If REG is NULL, a new pseudo is
+   created to hold the result of the load.  If not NULL, PIC_REG indicates
+   which register to use as PIC register, otherwise it is decided by register
+   allocator.  COMPUTE_NOW forces the PIC register to be loaded at the current
+   location in the instruction stream, irregardless of whether it was loaded
+   previously.  Note that nonnull PIC_REG is only supported iff COMPUTE_NOW is
+   true and null PIC_REG is only supported iff COMPUTE_NOW is false.
+
+   Returns the register REG into which the PIC load is performed.  */
+
 rtx
-legitimize_pic_address (rtx orig, machine_mode mode, rtx reg)
+legitimize_pic_address (rtx orig, machine_mode mode, rtx reg, rtx pic_reg,
+			bool compute_now)
 {
+  gcc_assert (compute_now == (pic_reg != NULL_RTX));
+
   if (GET_CODE (orig) == SYMBOL_REF
       || GET_CODE (orig) == LABEL_REF)
     {
@@ -7480,9 +7513,12 @@ legitimize_pic_address (rtx orig, machine_mode mode, rtx reg)
 	  rtx mem;
 
 	  /* If this function doesn't have a pic register, create one now.  */
-	  require_pic_register ();
+	  require_pic_register (pic_reg, compute_now);
+
+	  if (pic_reg == NULL_RTX)
+	    pic_reg = cfun->machine->pic_reg;
 
-	  pat = gen_calculate_pic_address (reg, cfun->machine->pic_reg, orig);
+	  pat = gen_calculate_pic_address (reg, pic_reg, orig);
 
 	  /* Make the MEM as close to a constant as possible.  */
 	  mem = SET_SRC (pat);
@@ -7531,9 +7567,11 @@ legitimize_pic_address (rtx orig, machine_mode mode, rtx reg)
 
       gcc_assert (GET_CODE (XEXP (orig, 0)) == PLUS);
 
-      base = legitimize_pic_address (XEXP (XEXP (orig, 0), 0), Pmode, reg);
+      base = legitimize_pic_address (XEXP (XEXP (orig, 0), 0), Pmode, reg,
+				     pic_reg, compute_now);
       offset = legitimize_pic_address (XEXP (XEXP (orig, 0), 1), Pmode,
-				       base == reg ? 0 : reg);
+				       base == reg ? 0 : reg, pic_reg,
+				       compute_now);
 
       if (CONST_INT_P (offset))
 	{
@@ -7633,16 +7671,17 @@ static GTY(()) int pic_labelno;
    low register.  */
 
 void
-arm_load_pic_register (unsigned long saved_regs ATTRIBUTE_UNUSED)
+arm_load_pic_register (unsigned long saved_regs ATTRIBUTE_UNUSED, rtx pic_reg)
 {
-  rtx l1, labelno, pic_tmp, pic_rtx, pic_reg;
+  rtx l1, labelno, pic_tmp, pic_rtx;
 
   if (crtl->uses_pic_offset_table == 0 || TARGET_SINGLE_PIC_BASE)
     return;
 
   gcc_assert (flag_pic);
 
-  pic_reg = cfun->machine->pic_reg;
+  if (pic_reg == NULL_RTX)
+    pic_reg = cfun->machine->pic_reg;
   if (TARGET_VXWORKS_RTP)
     {
       pic_rtx = gen_rtx_SYMBOL_REF (Pmode, VXWORKS_GOTT_BASE);
@@ -8718,7 +8757,8 @@ arm_legitimize_address (rtx x, rtx orig_x, machine_mode mode)
     {
       /* We need to find and carefully transform any SYMBOL and LABEL
 	 references; so go back to the original address expression.  */
-      rtx new_x = legitimize_pic_address (orig_x, mode, NULL_RTX);
+      rtx new_x = legitimize_pic_address (orig_x, mode, NULL_RTX, NULL_RTX,
+					  false /*compute_now*/);
 
       if (new_x != orig_x)
 	x = new_x;
@@ -8786,7 +8826,8 @@ thumb_legitimize_address (rtx x, rtx orig_x, machine_mode mode)
     {
       /* We need to find and carefully transform any SYMBOL and LABEL
 	 references; so go back to the original address expression.  */
-      rtx new_x = legitimize_pic_address (orig_x, mode, NULL_RTX);
+      rtx new_x = legitimize_pic_address (orig_x, mode, NULL_RTX, NULL_RTX,
+					  false /*compute_now*/);
 
       if (new_x != orig_x)
 	x = new_x;
@@ -18074,7 +18115,7 @@ arm_emit_call_insn (rtx pat, rtx addr, bool sibcall)
 	  ? !targetm.binds_local_p (SYMBOL_REF_DECL (addr))
 	  : !SYMBOL_REF_LOCAL_P (addr)))
     {
-      require_pic_register ();
+      require_pic_register (NULL_RTX, false /*compute_now*/);
       use_reg (&CALL_INSN_FUNCTION_USAGE (insn), cfun->machine->pic_reg);
     }
 
@@ -22006,7 +22047,7 @@ arm_expand_prologue (void)
       mask &= THUMB2_WORK_REGS;
       if (!IS_NESTED (func_type))
 	mask |= (1 << IP_REGNUM);
-      arm_load_pic_register (mask);
+      arm_load_pic_register (mask, NULL_RTX);
     }
 
   /* If we are profiling, make sure no instructions are scheduled before
@@ -25237,7 +25278,7 @@ thumb1_expand_prologue (void)
   /* Load the pic register before setting the frame pointer,
      so we can use r7 as a temporary work register.  */
   if (flag_pic && arm_pic_register != INVALID_REGNUM)
-    arm_load_pic_register (live_regs_mask);
+    arm_load_pic_register (live_regs_mask, NULL_RTX);
 
   if (!frame_pointer_needed && CALLER_INTERWORKING_SLOT_SIZE > 0)
     emit_move_insn (gen_rtx_REG (Pmode, ARM_HARD_FRAME_POINTER_REGNUM),
diff --git a/gcc/config/arm/arm.md b/gcc/config/arm/arm.md
index c8dc9474b1b..f6196e93168 100644
--- a/gcc/config/arm/arm.md
+++ b/gcc/config/arm/arm.md
@@ -6021,7 +6021,8 @@
       operands[1] = legitimize_pic_address (operands[1], SImode,
 					    (!can_create_pseudo_p ()
 					     ? operands[0]
-					     : 0));
+					     : NULL_RTX), NULL_RTX,
+					    false /*compute_now*/);
   }
   "
 )
@@ -6309,7 +6310,7 @@
   /* r3 is clobbered by set/longjmp, so we can use it as a scratch
      register.  */
   if (arm_pic_register != INVALID_REGNUM)
-    arm_load_pic_register (1UL << 3);
+    arm_load_pic_register (1UL << 3, NULL_RTX);
   DONE;
 }")
 
@@ -8634,6 +8635,164 @@
    (set_attr "conds" "clob")]
 )
 
+;; Named patterns for stack smashing protection.
+(define_expand "stack_protect_combined_set"
+  [(parallel
+     [(set (match_operand:SI 0 "memory_operand" "")
+	   (unspec:SI [(match_operand:SI 1 "guard_operand" "")]
+		      UNSPEC_SP_SET))
+      (clobber (match_scratch:SI 2 ""))
+      (clobber (match_scratch:SI 3 ""))])]
+  ""
+  ""
+)
+
+;; Use a separate insn from the above expand to be able to have the mem outside
+;; the operand #1 when register allocation comes. This is needed to avoid LRA
+;; try to reload the guard since we need to control how PIC access is done in
+;; the -fpic/-fPIC case (see COMPUTE_NOW parameter when calling
+;; legitimize_pic_address ()).
+(define_insn_and_split "*stack_protect_combined_set_insn"
+  [(set (match_operand:SI 0 "memory_operand" "=m,m")
+	(unspec:SI [(mem:SI (match_operand:SI 1 "guard_addr_operand" "X,X"))]
+		   UNSPEC_SP_SET))
+   (clobber (match_scratch:SI 2 "=&l,&r"))
+   (clobber (match_scratch:SI 3 "=&l,&r"))]
+  ""
+  "#"
+  "reload_completed"
+  [(parallel [(set (match_dup 0) (unspec:SI [(mem:SI (match_dup 2))]
+					    UNSPEC_SP_SET))
+	      (clobber (match_dup 2))])]
+  "
+{
+  if (flag_pic)
+    {
+      /* Forces recomputing of GOT base now.  */
+      legitimize_pic_address (operands[1], SImode, operands[2], operands[3],
+			      true /*compute_now*/);
+    }
+  else
+    {
+      if (address_operand (operands[1], SImode))
+	operands[2] = operands[1];
+      else
+	{
+	  rtx mem = XEXP (force_const_mem (SImode, operands[1]), 0);
+	  emit_move_insn (operands[2], mem);
+	}
+    }
+}"
+  [(set_attr "arch" "t1,32")]
+)
+
+(define_insn "*stack_protect_set_insn"
+  [(set (match_operand:SI 0 "memory_operand" "=m,m")
+	(unspec:SI [(mem:SI (match_operand:SI 1 "register_operand" "+&l,&r"))]
+	 UNSPEC_SP_SET))
+   (clobber (match_dup 1))]
+  ""
+  "@
+   ldr\\t%1, [%1]\;str\\t%1, %0\;movs\t%1,#0
+   ldr\\t%1, [%1]\;str\\t%1, %0\;mov\t%1,#0"
+  [(set_attr "length" "8,12")
+   (set_attr "conds" "clob,nocond")
+   (set_attr "type" "multiple")
+   (set_attr "arch" "t1,32")]
+)
+
+(define_expand "stack_protect_combined_test"
+  [(parallel
+     [(set (pc)
+	   (if_then_else
+		(eq (match_operand:SI 0 "memory_operand" "")
+		    (unspec:SI [(match_operand:SI 1 "guard_operand" "")]
+			       UNSPEC_SP_TEST))
+		(label_ref (match_operand 2))
+		(pc)))
+      (clobber (match_scratch:SI 3 ""))
+      (clobber (match_scratch:SI 4 ""))
+      (clobber (reg:CC CC_REGNUM))])]
+  ""
+  ""
+)
+
+;; Use a separate insn from the above expand to be able to have the mem outside
+;; the operand #1 when register allocation comes. This is needed to avoid LRA
+;; try to reload the guard since we need to control how PIC access is done in
+;; the -fpic/-fPIC case (see COMPUTE_NOW parameter when calling
+;; legitimize_pic_address ()).
+(define_insn_and_split "*stack_protect_combined_test_insn"
+  [(set (pc)
+	(if_then_else
+		(eq (match_operand:SI 0 "memory_operand" "m,m")
+		    (unspec:SI [(mem:SI (match_operand:SI 1 "guard_addr_operand" "X,X"))]
+			       UNSPEC_SP_TEST))
+		(label_ref (match_operand 2))
+		(pc)))
+   (clobber (match_scratch:SI 3 "=&l,&r"))
+   (clobber (match_scratch:SI 4 "=&l,&r"))
+   (clobber (reg:CC CC_REGNUM))]
+  ""
+  "#"
+  "reload_completed"
+  [(const_int 0)]
+{
+  rtx eq;
+
+  if (flag_pic)
+    {
+      /* Forces recomputing of GOT base now.  */
+      legitimize_pic_address (operands[1], SImode, operands[3], operands[4],
+			      true /*compute_now*/);
+    }
+  else
+    {
+      if (address_operand (operands[1], SImode))
+	operands[3] = operands[1];
+      else
+	{
+	  rtx mem = XEXP (force_const_mem (SImode, operands[1]), 0);
+	  emit_move_insn (operands[3], mem);
+	}
+    }
+  if (TARGET_32BIT)
+    {
+      emit_insn (gen_arm_stack_protect_test_insn (operands[4], operands[0],
+						  operands[3]));
+      rtx cc_reg = gen_rtx_REG (CC_Zmode, CC_REGNUM);
+      eq = gen_rtx_EQ (CC_Zmode, cc_reg, const0_rtx);
+      emit_jump_insn (gen_arm_cond_branch (operands[2], eq, cc_reg));
+    }
+  else
+    {
+      emit_insn (gen_thumb1_stack_protect_test_insn (operands[4], operands[0],
+						     operands[3]));
+      eq = gen_rtx_EQ (VOIDmode, operands[4], const0_rtx);
+      emit_jump_insn (gen_cbranchsi4 (eq, operands[4], const0_rtx,
+				      operands[2]));
+    }
+  DONE;
+}
+  [(set_attr "arch" "t1,32")]
+)
+
+(define_insn "arm_stack_protect_test_insn"
+  [(set (reg:CC_Z CC_REGNUM)
+	(compare:CC_Z (unspec:SI [(match_operand:SI 1 "memory_operand" "m,m")
+				  (mem:SI (match_operand:SI 2 "register_operand" "+l,r"))]
+				 UNSPEC_SP_TEST)
+		      (const_int 0)))
+   (clobber (match_operand:SI 0 "register_operand" "=&l,&r"))
+   (clobber (match_dup 2))]
+  "TARGET_32BIT"
+  "ldr\t%0, [%2]\;ldr\t%2, %1\;eors\t%0, %2, %0"
+  [(set_attr "length" "8,12")
+   (set_attr "conds" "set")
+   (set_attr "type" "multiple")
+   (set_attr "arch" "t,32")]
+)
+
 (define_expand "casesi"
   [(match_operand:SI 0 "s_register_operand" "")	; index to jump on
    (match_operand:SI 1 "const_int_operand" "")	; lower bound
diff --git a/gcc/config/arm/predicates.md b/gcc/config/arm/predicates.md
index 7e198f9bce4..69718ee9c7a 100644
--- a/gcc/config/arm/predicates.md
+++ b/gcc/config/arm/predicates.md
@@ -31,6 +31,23 @@
 	      || REGNO_REG_CLASS (REGNO (op)) != NO_REGS));
 })
 
+; Predicate for stack protector guard's address in
+; stack_protect_combined_set_insn and stack_protect_combined_test_insn patterns
+(define_predicate "guard_addr_operand"
+  (match_test "true")
+{
+  return (CONSTANT_ADDRESS_P (op)
+	  || !targetm.cannot_force_const_mem (mode, op));
+})
+
+; Predicate for stack protector guard in stack_protect_combined_set and
+; stack_protect_combined_test patterns
+(define_predicate "guard_operand"
+  (match_code "mem")
+{
+  return guard_addr_operand (XEXP (op, 0), mode);
+})
+
 (define_predicate "imm_for_neon_inv_logic_operand"
   (match_code "const_vector")
 {
diff --git a/gcc/config/arm/thumb1.md b/gcc/config/arm/thumb1.md
index 19dcdbcdd73..cd199c9c529 100644
--- a/gcc/config/arm/thumb1.md
+++ b/gcc/config/arm/thumb1.md
@@ -1962,4 +1962,17 @@
   }"
   [(set_attr "type" "mov_reg")]
 )
+
+(define_insn "thumb1_stack_protect_test_insn"
+  [(set (match_operand:SI 0 "register_operand" "=&l")
+	(unspec:SI [(match_operand:SI 1 "memory_operand" "m")
+		    (mem:SI (match_operand:SI 2 "register_operand" "+l"))]
+	 UNSPEC_SP_TEST))
+   (clobber (match_dup 2))]
+  "TARGET_THUMB1"
+  "ldr\t%0, [%2]\;ldr\t%2, %1\;eors\t%0, %2, %0"
+  [(set_attr "length" "8")
+   (set_attr "conds" "set")
+   (set_attr "type" "multiple")]
+)
 
diff --git a/gcc/config/arm/unspecs.md b/gcc/config/arm/unspecs.md
index 19416736ef9..8f9dbcb08dc 100644
--- a/gcc/config/arm/unspecs.md
+++ b/gcc/config/arm/unspecs.md
@@ -86,6 +86,9 @@
   UNSPEC_PROBE_STACK    ; Probe stack memory reference
   UNSPEC_NONSECURE_MEM	; Represent non-secure memory in ARMv8-M with
 			; security extension
+  UNSPEC_SP_SET		; Represent the setting of stack protector's canary
+  UNSPEC_SP_TEST	; Represent the testing of stack protector's canary
+			; against the guard.
 ])
 
 (define_c_enum "unspec" [
diff --git a/gcc/doc/md.texi b/gcc/doc/md.texi
index 295fc1f1143..895309b2f3c 100644
--- a/gcc/doc/md.texi
+++ b/gcc/doc/md.texi
@@ -7450,22 +7450,61 @@ builtins.
 The get/set patterns have a single output/input operand respectively,
 with @var{mode} intended to be @code{Pmode}.
 
+@cindex @code{stack_protect_combined_set} instruction pattern
+@item @samp{stack_protect_combined_set}
+This pattern, if defined, moves a @code{ptr_mode} value from an address
+whose declaration RTX is given in operand 1 to the memory in operand 0
+without leaving the value in a register afterward.  If several
+instructions are needed by the target to perform the operation (eg. to
+load the address from a GOT entry then load the @code{ptr_mode} value
+and finally store it), it is the backend's responsibility to ensure no
+intermediate result gets spilled.  This is to avoid leaking the value
+some place that an attacker might use to rewrite the stack guard slot
+after having clobbered it.
+
+If this pattern is not defined, then the address declaration is
+expanded first in the standard way and a @code{stack_protect_set}
+pattern is then generated to move the value from that address to the
+address in operand 0.
+
 @cindex @code{stack_protect_set} instruction pattern
 @item @samp{stack_protect_set}
-This pattern, if defined, moves a @code{ptr_mode} value from the memory
-in operand 1 to the memory in operand 0 without leaving the value in
-a register afterward.  This is to avoid leaking the value some place
-that an attacker might use to rewrite the stack guard slot after
-having clobbered it.
+This pattern, if defined, moves a @code{ptr_mode} value from the valid
+memory location in operand 1 to the memory in operand 0 without leaving
+the value in a register afterward.  This is to avoid leaking the value
+some place that an attacker might use to rewrite the stack guard slot
+after having clobbered it.
+
+Note: on targets where the addressing modes do not allow to load
+directly from stack guard address, the address is expanded in a standard
+way first which could cause some spills.
 
 If this pattern is not defined, then a plain move pattern is generated.
 
+@cindex @code{stack_protect_combined_test} instruction pattern
+@item @samp{stack_protect_combined_test}
+This pattern, if defined, compares a @code{ptr_mode} value from an
+address whose declaration RTX is given in operand 1 with the memory in
+operand 0 without leaving the value in a register afterward and
+branches to operand 2 if the values were equal.  If several
+instructions are needed by the target to perform the operation (eg. to
+load the address from a GOT entry then load the @code{ptr_mode} value
+and finally store it), it is the backend's responsibility to ensure no
+intermediate result gets spilled.  This is to avoid leaking the value
+some place that an attacker might use to rewrite the stack guard slot
+after having clobbered it.
+
+If this pattern is not defined, then the address declaration is
+expanded first in the standard way and a @code{stack_protect_test}
+pattern is then generated to compare the value from that address to the
+value at the memory in operand 0.
+
 @cindex @code{stack_protect_test} instruction pattern
 @item @samp{stack_protect_test}
 This pattern, if defined, compares a @code{ptr_mode} value from the
-memory in operand 1 with the memory in operand 0 without leaving the
-value in a register afterward and branches to operand 2 if the values
-were equal.
+valid memory location in operand 1 with the memory in operand 0 without
+leaving the value in a register afterward and branches to operand 2 if
+the values were equal.
 
 If this pattern is not defined, then a plain compare pattern and
 conditional branch pattern is used.
diff --git a/gcc/function.c b/gcc/function.c
index 85a5d9f43f7..69523c1d723 100644
--- a/gcc/function.c
+++ b/gcc/function.c
@@ -4937,18 +4937,34 @@ stack_protect_epilogue (void)
   tree guard_decl = targetm.stack_protect_guard ();
   rtx_code_label *label = gen_label_rtx ();
   rtx x, y;
-  rtx_insn *seq;
+  rtx_insn *seq = NULL;
 
   x = expand_normal (crtl->stack_protect_guard);
-  if (guard_decl)
-    y = expand_normal (guard_decl);
+
+  if (targetm.have_stack_protect_combined_test () && guard_decl)
+    {
+      gcc_assert (DECL_P (guard_decl));
+      y = DECL_RTL (guard_decl);
+      /* Allow the target to compute address of Y and compare it with X without
+	 leaking Y into a register.  This combined address + compare pattern
+	 allows the target to prevent spilling of any intermediate results by
+	 splitting it after register allocator.  */
+      seq = targetm.gen_stack_protect_combined_test (x, y, label);
+    }
   else
-    y = const0_rtx;
+    {
+      if (guard_decl)
+	y = expand_normal (guard_decl);
+      else
+	y = const0_rtx;
+
+      /* Allow the target to compare Y with X without leaking either into
+	 a register.  */
+      if (targetm.have_stack_protect_test ())
+	seq = targetm.gen_stack_protect_test (x, y, label);
+    }
 
-  /* Allow the target to compare Y with X without leaking either into
-     a register.  */
-  if (targetm.have_stack_protect_test ()
-      && ((seq = targetm.gen_stack_protect_test (x, y, label)) != NULL_RTX))
+  if (seq)
     emit_insn (seq);
   else
     emit_cmp_and_jump_insns (x, y, EQ, NULL_RTX, ptr_mode, 1, label);
diff --git a/gcc/target-insns.def b/gcc/target-insns.def
index 9a552c3d11c..d39889b3522 100644
--- a/gcc/target-insns.def
+++ b/gcc/target-insns.def
@@ -96,7 +96,9 @@ DEF_TARGET_INSN (sibcall_value, (rtx x0, rtx x1, rtx opt2, rtx opt3,
 DEF_TARGET_INSN (simple_return, (void))
 DEF_TARGET_INSN (split_stack_prologue, (void))
 DEF_TARGET_INSN (split_stack_space_check, (rtx x0, rtx x1))
+DEF_TARGET_INSN (stack_protect_combined_set, (rtx x0, rtx x1))
 DEF_TARGET_INSN (stack_protect_set, (rtx x0, rtx x1))
+DEF_TARGET_INSN (stack_protect_combined_test, (rtx x0, rtx x1, rtx x2))
 DEF_TARGET_INSN (stack_protect_test, (rtx x0, rtx x1, rtx x2))
 DEF_TARGET_INSN (store_multiple, (rtx x0, rtx x1, rtx x2))
 DEF_TARGET_INSN (tablejump, (rtx x0, rtx x1))
-- 
2.21.0