Arm: Fix performance issue with thumb-2 tailcalls

We currently use a padding NOP after a Thumb to Arm interworking veneer (BX pc).
The NOP is never executed but may result in a performance penalty on some cores.

For this reason this patch changes the NOPs after Thumb to Arm veneers into B .-2
and adds a note to this in the source code for future reference.

bfd/ChangeLog:

	* elf32-arm.c (elf32_thumb2_plt_entry, elf32_arm_plt_thumb_stub,
	elf32_arm_stub_long_branch_v4t_thumb_thumb,
	elf32_arm_stub_long_branch_v4t_thumb_arm,
	elf32_arm_stub_short_branch_v4t_thumb_arm,
	elf32_arm_stub_long_branch_v4t_thumb_arm_pic,
	elf32_arm_stub_long_branch_v4t_thumb_thumb_pic,
	elf32_arm_stub_long_branch_v4t_thumb_tls_pic): Change nop to branch to
	previous instruction.

ld/ChangeLog:

	* testsuite/ld-arm/cortex-a8-fix-b-plt.d: Update Testcase.
	* testsuite/ld-arm/cortex-a8-fix-b-rel-arm.d: Likewise.
	* testsuite/ld-arm/cortex-a8-fix-bcc-plt.d: Likewise.
	* testsuite/ld-arm/farcall-cond-thumb-arm.d: Likewise.
	* testsuite/ld-arm/farcall-mixed-app.d: Likewise.
	* testsuite/ld-arm/farcall-mixed-app2.d: Likewise.
	* testsuite/ld-arm/farcall-mixed-lib-v4t.d: Likewise.
	* testsuite/ld-arm/farcall-thumb-arm-pic-veneer.d: Likewise.
	* testsuite/ld-arm/farcall-thumb-arm-short.d: Likewise.
	* testsuite/ld-arm/farcall-thumb-arm.d: Likewise.
	* testsuite/ld-arm/farcall-thumb-thumb-pic-veneer.d: Likewise.
	* testsuite/ld-arm/farcall-thumb-thumb.d: Likewise.
	* testsuite/ld-arm/fix-arm1176-on.d: Likewise.
	* testsuite/ld-arm/ifunc-10.dd: Likewise.
	* testsuite/ld-arm/ifunc-2.dd: Likewise.
	* testsuite/ld-arm/ifunc-4.dd: Likewise.
	* testsuite/ld-arm/ifunc-6.dd: Likewise.
	* testsuite/ld-arm/ifunc-8.dd: Likewise.
	* testsuite/ld-arm/jump-reloc-veneers-long.d: Likewise.
	* testsuite/ld-arm/mixed-app.d: Likewise.
	* testsuite/ld-arm/thumb2-b-interwork.d: Likewise.
	* testsuite/ld-arm/tls-longplt.d: Likewise.
	* testsuite/ld-arm/tls-thumb1.d: Likewise.

2 files changed, 27 insertions, 9 deletions

diff --git a/bfd/ChangeLog b/bfd/ChangeLog
index bfaec7b..e566656 100644
--- a/bfd/ChangeLog
+++ b/bfd/ChangeLog
@@ -1,3 +1,14 @@
+2019-08-20  Tamar Christina  <tamar.christina@arm.com>
+
+	* elf32-arm.c (elf32_thumb2_plt_entry, elf32_arm_plt_thumb_stub,
+	elf32_arm_stub_long_branch_v4t_thumb_thumb,
+	elf32_arm_stub_long_branch_v4t_thumb_arm,
+	elf32_arm_stub_short_branch_v4t_thumb_arm,
+	elf32_arm_stub_long_branch_v4t_thumb_arm_pic,
+	elf32_arm_stub_long_branch_v4t_thumb_thumb_pic,
+	elf32_arm_stub_long_branch_v4t_thumb_tls_pic): Change nop to branch to
+	previous instruction.
+
 2019-08-19  Tom Tromey  <tromey@adacore.com>
 
 	* dwarf2.c (_bfd_dwarf2_find_symbol_bias): Create hash table
diff --git a/bfd/elf32-arm.c b/bfd/elf32-arm.c
index f1895df..d1548d6 100644
--- a/bfd/elf32-arm.c
+++ b/bfd/elf32-arm.c
@@ -2329,6 +2329,11 @@ static const unsigned long dl_tlsdesc_lazy_trampoline [] =
   0x00000018, /* 4:   .word  _GLOBAL_OFFSET_TABLE_ - 2b - 8 */
 };
 
+/* NOTE: [Thumb nop sequence]
+   When adding code that transitions from Thumb to Arm the instruction that
+   should be used for the alignment padding should be 0xe7fd (b .-2) instead of
+   a nop for performance reasons.  */
+
 /* ARM FDPIC PLT entry.  */
 /* The last 5 words contain PLT lazy fragment code and data.  */
 static const bfd_vma elf32_arm_fdpic_plt_entry [] =
@@ -2446,8 +2451,8 @@ static const bfd_vma elf32_thumb2_plt_entry [] =
   0x0c00f240,		/* movw	   ip, #0xNNNN	  */
   0x0c00f2c0,		/* movt	   ip, #0xNNNN	  */
   0xf8dc44fc,		/* add	   ip, pc	  */
-  0xbf00f000		/* ldr.w   pc, [ip]	  */
-			/* nop			  */
+  0xe7fdf000		/* ldr.w   pc, [ip]	  */
+			/* b      .-2		  */
 };
 
 /* The format of the first entry in the procedure linkage table
@@ -2487,7 +2492,7 @@ static const bfd_vma elf32_arm_vxworks_shared_plt_entry[] =
 static const bfd_vma elf32_arm_plt_thumb_stub [] =
 {
   0x4778,		/* bx pc */
-  0x46c0		/* nop   */
+  0xe7fd		/* b .-2 */
 };
 
 /* The entries in a PLT when using a DLL-based target with multiple
@@ -2574,6 +2579,8 @@ typedef struct
   int		       reloc_addend;
 }  insn_sequence;
 
+/* See note [Thumb nop sequence] when adding a veneer.  */
+
 /* Arm/Thumb -> Arm/Thumb long branch stub. On V5T and above, use blx
    to reach the stub if necessary.  */
 static const insn_sequence elf32_arm_stub_long_branch_any_any[] =
@@ -2624,7 +2631,7 @@ static const insn_sequence elf32_arm_stub_long_branch_thumb2_only_pure[] =
 static const insn_sequence elf32_arm_stub_long_branch_v4t_thumb_thumb[] =
 {
   THUMB16_INSN (0x4778),	     /* bx   pc */
-  THUMB16_INSN (0x46c0),	     /* nop */
+  THUMB16_INSN (0xe7fd),	     /* b   .-2 */
   ARM_INSN (0xe59fc000),	     /* ldr  ip, [pc, #0] */
   ARM_INSN (0xe12fff1c),	     /* bx   ip */
   DATA_WORD (0, R_ARM_ABS32, 0),     /* dcd  R_ARM_ABS32(X) */
@@ -2635,7 +2642,7 @@ static const insn_sequence elf32_arm_stub_long_branch_v4t_thumb_thumb[] =
 static const insn_sequence elf32_arm_stub_long_branch_v4t_thumb_arm[] =
 {
   THUMB16_INSN (0x4778),	     /* bx   pc */
-  THUMB16_INSN (0x46c0),	     /* nop   */
+  THUMB16_INSN (0xe7fd),	     /* b   .-2 */
   ARM_INSN (0xe51ff004),	     /* ldr   pc, [pc, #-4] */
   DATA_WORD (0, R_ARM_ABS32, 0),     /* dcd   R_ARM_ABS32(X) */
 };
@@ -2645,7 +2652,7 @@ static const insn_sequence elf32_arm_stub_long_branch_v4t_thumb_arm[] =
 static const insn_sequence elf32_arm_stub_short_branch_v4t_thumb_arm[] =
 {
   THUMB16_INSN (0x4778),	     /* bx   pc */
-  THUMB16_INSN (0x46c0),	     /* nop   */
+  THUMB16_INSN (0xe7fd),	     /* b   .-2 */
   ARM_REL_INSN (0xea000000, -8),     /* b    (X-8) */
 };
 
@@ -2683,7 +2690,7 @@ static const insn_sequence elf32_arm_stub_long_branch_v4t_arm_thumb_pic[] =
 static const insn_sequence elf32_arm_stub_long_branch_v4t_thumb_arm_pic[] =
 {
   THUMB16_INSN (0x4778),	     /* bx   pc */
-  THUMB16_INSN (0x46c0),	     /* nop  */
+  THUMB16_INSN (0xe7fd),	     /* b   .-2 */
   ARM_INSN (0xe59fc000),	     /* ldr  ip, [pc, #0] */
   ARM_INSN (0xe08cf00f),	     /* add  pc, ip, pc */
   DATA_WORD (0, R_ARM_REL32, -4),     /* dcd  R_ARM_REL32(X) */
@@ -2707,7 +2714,7 @@ static const insn_sequence elf32_arm_stub_long_branch_thumb_only_pic[] =
 static const insn_sequence elf32_arm_stub_long_branch_v4t_thumb_thumb_pic[] =
 {
   THUMB16_INSN (0x4778),	     /* bx   pc */
-  THUMB16_INSN (0x46c0),	     /* nop */
+  THUMB16_INSN (0xe7fd),	     /* b   .-2 */
   ARM_INSN (0xe59fc004),	     /* ldr  ip, [pc, #4] */
   ARM_INSN (0xe08fc00c),	     /* add   ip, pc, ip */
   ARM_INSN (0xe12fff1c),	     /* bx   ip */
@@ -2728,7 +2735,7 @@ static const insn_sequence elf32_arm_stub_long_branch_any_tls_pic[] =
 static const insn_sequence elf32_arm_stub_long_branch_v4t_thumb_tls_pic[] =
 {
   THUMB16_INSN (0x4778),	     /* bx   pc */
-  THUMB16_INSN (0x46c0),	     /* nop */
+  THUMB16_INSN (0xe7fd),	     /* b   .-2 */
   ARM_INSN (0xe59f1000),	     /* ldr  r1, [pc, #0] */
   ARM_INSN (0xe081f00f),	     /* add  pc, r1, pc */
   DATA_WORD (0, R_ARM_REL32, -4),    /* dcd  R_ARM_REL32(X) */