From 5a491cb2d889f03952182d611d2ac6e15c3c5845 Mon Sep 17 00:00:00 2001
From: Leonid Chistov <lchistov@pathscale.com>
Date: Thu, 1 Oct 2015 11:23:50 +0300
Subject: [PATCH] arm64: Support for restore of ARM64 Neon callee-saved
 registers during unwind

---
 src/aarch64/Ginit.c    |  4 +++-
 src/aarch64/Gregs.c    |  7 +++++--
 src/aarch64/Gresume.c  | 21 +++++++++++++++++----
 src/aarch64/unwind_i.h |  2 ++
 4 files changed, 27 insertions(+), 7 deletions(-)

diff --git a/src/aarch64/Ginit.c b/src/aarch64/Ginit.c
index 667c4764..ab3999f3 100644
--- a/src/aarch64/Ginit.c
+++ b/src/aarch64/Ginit.c
@@ -43,8 +43,10 @@ PROTECTED unw_addr_space_t unw_local_addr_space = &local_addr_space;
 static inline void *
 uc_addr (ucontext_t *uc, int reg)
 {
-  if (reg >= UNW_AARCH64_X0 && reg <= UNW_AARCH64_V31)
+  if (reg >= UNW_AARCH64_X0 && reg < UNW_AARCH64_V0)
     return &uc->uc_mcontext.regs[reg];
+  else if (reg >= UNW_AARCH64_V0 && reg <= UNW_AARCH64_V31)
+    return &GET_FPCTX(uc)->vregs[reg - UNW_AARCH64_V0];
   else
     return NULL;
 }
diff --git a/src/aarch64/Gregs.c b/src/aarch64/Gregs.c
index 104fd62a..a8843734 100644
--- a/src/aarch64/Gregs.c
+++ b/src/aarch64/Gregs.c
@@ -110,6 +110,9 @@ HIDDEN int
 tdep_access_fpreg (struct cursor *c, unw_regnum_t reg, unw_fpreg_t *valp,
                    int write)
 {
-  Debug (1, "bad register number %u\n", reg);
-  return -UNW_EBADREG;
+  dwarf_loc_t loc = c->dwarf.loc[reg];
+  if (write)
+    return dwarf_putfp (&c->dwarf, loc, *valp);
+  else
+    return dwarf_getfp (&c->dwarf, loc, valp);
 }
diff --git a/src/aarch64/Gresume.c b/src/aarch64/Gresume.c
index d9acfa7c..65517a25 100644
--- a/src/aarch64/Gresume.c
+++ b/src/aarch64/Gresume.c
@@ -40,7 +40,7 @@ aarch64_local_resume (unw_addr_space_t as, unw_cursor_t *cursor, void *arg)
     {
       /* Since there are no signals involved here we restore EH and non scratch
          registers only.  */
-      unsigned long regs[15];
+      unsigned long regs[24];
       regs[0] = uc->uc_mcontext.regs[0];
       regs[1] = uc->uc_mcontext.regs[1];
       regs[2] = uc->uc_mcontext.regs[2];
@@ -55,7 +55,16 @@ aarch64_local_resume (unw_addr_space_t as, unw_cursor_t *cursor, void *arg)
       regs[11] = uc->uc_mcontext.regs[26];
       regs[12] = uc->uc_mcontext.regs[27];
       regs[13] = uc->uc_mcontext.regs[28];
-      regs[14] = uc->uc_mcontext.regs[30]; /* LR */
+      regs[14] = uc->uc_mcontext.regs[29]; /* FP */
+      regs[15] = uc->uc_mcontext.regs[30]; /* LR */
+      regs[16] = GET_FPCTX(uc)->vregs[8];
+      regs[17] = GET_FPCTX(uc)->vregs[9];
+      regs[18] = GET_FPCTX(uc)->vregs[10];
+      regs[19] = GET_FPCTX(uc)->vregs[11];
+      regs[20] = GET_FPCTX(uc)->vregs[12];
+      regs[21] = GET_FPCTX(uc)->vregs[13];
+      regs[22] = GET_FPCTX(uc)->vregs[14];
+      regs[23] = GET_FPCTX(uc)->vregs[15];
       unsigned long sp = uc->uc_mcontext.sp;
 
       struct regs_overlay {
@@ -72,7 +81,11 @@ aarch64_local_resume (unw_addr_space_t as, unw_cursor_t *cursor, void *arg)
         "ldp x23, x24, [x4,64]\n"
         "ldp x25, x26, [x4,80]\n"
         "ldp x27, x28, [x4,96]\n"
-        "ldr x30, [x4,112]\n"
+        "ldp x29, x30, [x4,112]\n"
+        "ldp d8, d9,   [x4,128]\n"
+        "ldp d10, d11, [x4,144]\n"
+        "ldp d12, d13, [x4,160]\n"
+        "ldp d14, d15, [x4,176]\n"
         "mov sp, x5\n"
         "ret \n"
         :
@@ -147,7 +160,7 @@ establish_machine_state (struct cursor *c)
 
   Debug (8, "copying out cursor state\n");
 
-  for (reg = 0; reg <= UNW_AARCH64_PSTATE; ++reg)
+  for (reg = 0; reg <= UNW_AARCH64_V31; ++reg)
     {
       Debug (16, "copying %s %d\n", unw_regname (reg), reg);
       if (unw_is_fpreg (reg))
diff --git a/src/aarch64/unwind_i.h b/src/aarch64/unwind_i.h
index 79b342cd..3d324c2b 100644
--- a/src/aarch64/unwind_i.h
+++ b/src/aarch64/unwind_i.h
@@ -59,4 +59,6 @@ extern int aarch64_local_resume (unw_addr_space_t as, unw_cursor_t *cursor,
   } while (0)
 #endif
 
+#define GET_FPCTX(uc) ((struct fpsimd_context *)(&uc->uc_mcontext.__reserved))
+
 #endif /* unwind_i_h */