Fix stub_test for Mips64

Force ENTRY macro to load GP. Instead .cpload (used for o32 abi,
not working for n64) use .cpsetup to set up GP and save previous
GP value.

Fix expectations in GetSet32Static.

Change-Id: Ie97e32fa2fee7a6e8bae6bc36e78976b8be62b79

4 files changed, 57 insertions, 24 deletions

diff --git a/runtime/arch/mips64/asm_support_mips64.S b/runtime/arch/mips64/asm_support_mips64.S
index 10976bb70d..26137777e3 100644
--- a/runtime/arch/mips64/asm_support_mips64.S
+++ b/runtime/arch/mips64/asm_support_mips64.S
@@ -27,7 +27,8 @@
 #define rSELF $s1
 
 
-    //  Declare a function called name, sets up $gp.
+    // Declare a function called name, sets up $gp.
+    // This macro modifies t8.
 .macro ENTRY name
     .type \name, %function
     .global \name
@@ -35,10 +36,11 @@
     .balign 16
 \name:
     .cfi_startproc
+    // Set up $gp and store the previous $gp value to $t8. It will be pushed to the
+    // stack after the frame has been constructed.
+    .cpsetup $t9, $t8, \name
     // Ensure we get a sane starting CFA.
     .cfi_def_cfa $sp,0
-    // Load $gp. We expect that ".set noreorder" is in effect.
-    .cpload $t9
     // Declare a local convenience label to be branched to when $gp is already set up.
 .L\name\()_gp_set:
 .endm
diff --git a/runtime/arch/mips64/jni_entrypoints_mips64.S b/runtime/arch/mips64/jni_entrypoints_mips64.S
index 10856668a2..70d7d9742c 100644
--- a/runtime/arch/mips64/jni_entrypoints_mips64.S
+++ b/runtime/arch/mips64/jni_entrypoints_mips64.S
@@ -44,8 +44,11 @@ ENTRY art_jni_dlsym_lookup_stub
     .cfi_rel_offset 5, 8
     sd     $a0, 0($sp)
     .cfi_rel_offset 4, 0
-    jal    artFindNativeMethod  # (Thread*)
     move   $a0, $s1             # pass Thread::Current()
+    jal    artFindNativeMethod  # (Thread*)
+    .cpreturn                   # Restore gp from t8 in branch delay slot. gp is not used
+                                # anymore, and t8 may be clobbered in artFindNativeMethod.
+
     ld     $a0, 0($sp)          # restore registers from stack
     .cfi_restore 4
     ld     $a1, 8($sp)
diff --git a/runtime/arch/mips64/quick_entrypoints_mips64.S b/runtime/arch/mips64/quick_entrypoints_mips64.S
index 8330d0c253..ff79b5d77c 100644
--- a/runtime/arch/mips64/quick_entrypoints_mips64.S
+++ b/runtime/arch/mips64/quick_entrypoints_mips64.S
@@ -27,6 +27,19 @@
     .extern artDeliverPendingExceptionFromCode
 
     /*
+     * Macro that sets up $gp and stores the previous $gp value to $t8.
+     * This macro modifies v1 and t8.
+     */
+.macro SETUP_GP
+    move $v1, $ra
+    bal 1f
+    nop
+1:
+    .cpsetup $ra, $t8, 1b
+    move $ra, $v1
+.endm
+
+    /*
      * Macro that sets up the callee save frame to conform with
      * Runtime::CreateCalleeSaveMethod(kSaveAll)
      * callee-save: padding + $f24-$f31 + $s0-$s7 + $gp + $ra + $s8 = 19 total + 1x8 bytes padding
@@ -44,8 +57,8 @@
     .cfi_rel_offset 31, 152
     sd     $s8, 144($sp)
     .cfi_rel_offset 30, 144
-    sd     $gp, 136($sp)
-    .cfi_rel_offset 28, 136
+    sd     $t8, 136($sp)           # t8 holds caller's gp, now save it to the stack.
+    .cfi_rel_offset 28, 136        # Value from gp is pushed, so set the cfi offset accordingly.
     sd     $s7, 128($sp)
     .cfi_rel_offset 23, 128
     sd     $s6, 120($sp)
@@ -102,8 +115,8 @@
     .cfi_rel_offset 31, 72
     sd     $s8, 64($sp)
     .cfi_rel_offset 30, 64
-    sd     $gp, 56($sp)
-    .cfi_rel_offset 28, 56
+    sd     $t8, 56($sp)            # t8 holds caller's gp, now save it to the stack.
+    .cfi_rel_offset 28, 56         # Value from gp is pushed, so set the cfi offset accordingly.
     sd     $s7, 48($sp)
     .cfi_rel_offset 23, 48
     sd     $s6, 40($sp)
@@ -130,7 +143,7 @@
     .cfi_restore 31
     ld     $s8, 64($sp)
     .cfi_restore 30
-    ld     $gp, 56($sp)
+    ld     $t8, 56($sp)            # Restore gp back to it's temp storage.
     .cfi_restore 28
     ld     $s7, 48($sp)
     .cfi_restore 23
@@ -146,6 +159,7 @@
     .cfi_restore 18
     daddiu $sp, $sp, 80
     .cfi_adjust_cfa_offset -80
+    .cpreturn
 .endm
 
 .macro RESTORE_REFS_ONLY_CALLEE_SAVE_FRAME_AND_RETURN
@@ -153,7 +167,7 @@
     .cfi_restore 31
     ld     $s8, 64($sp)
     .cfi_restore 30
-    ld     $gp, 56($sp)
+    ld     $t8, 56($sp)            # Restore gp back to it's temp storage.
     .cfi_restore 28
     ld     $s7, 48($sp)
     .cfi_restore 23
@@ -167,6 +181,7 @@
     .cfi_restore 19
     ld     $s2, 8($sp)
     .cfi_restore 18
+    .cpreturn
     jalr   $zero, $ra
     daddiu $sp, $sp, 80
     .cfi_adjust_cfa_offset -80
@@ -188,8 +203,8 @@
     .cfi_rel_offset 31, 200
     sd     $s8, 192($sp)
     .cfi_rel_offset 30, 192
-    sd     $gp, 184($sp)
-    .cfi_rel_offset 28, 184
+    sd     $t8, 184($sp)           # t8 holds caller's gp, now save it to the stack.
+    .cfi_rel_offset 28, 184        # Value from gp is pushed, so set the cfi offset accordingly.
     sd     $s7, 176($sp)
     .cfi_rel_offset 23, 176
     sd     $s6, 168($sp)
@@ -257,7 +272,7 @@
     .cfi_restore 31
     ld     $s8, 192($sp)
     .cfi_restore 30
-    ld     $gp, 184($sp)
+    ld     $t8, 184($sp)           # Restore gp back to it's temp storage.
     .cfi_restore 28
     ld     $s7, 176($sp)
     .cfi_restore 23
@@ -296,6 +311,7 @@
     l.d    $f13, 24($sp)
     l.d    $f12, 16($sp)
 
+    .cpreturn
     daddiu $sp, $sp, 208
     .cfi_adjust_cfa_offset -208
 .endm
@@ -306,6 +322,7 @@
      * exception is Thread::Current()->exception_
      */
 .macro DELIVER_PENDING_EXCEPTION
+    SETUP_GP
     SETUP_SAVE_ALL_CALLEE_SAVE_FRAME     # save callee saves for throw
     dla     $t9, artDeliverPendingExceptionFromCode
     jalr    $zero, $t9                   # artDeliverPendingExceptionFromCode(Thread*)
@@ -347,7 +364,7 @@
      * On entry $a0 is uint32_t* gprs_ and $a1 is uint32_t* fprs_
      * FIXME: just guessing about the shape of the jmpbuf.  Where will pc be?
      */
-ENTRY art_quick_do_long_jump
+ENTRY_NO_GP art_quick_do_long_jump
     l.d     $f0, 0($a1)
     l.d     $f1, 8($a1)
     l.d     $f2, 16($a1)
@@ -604,7 +621,7 @@ INVOKE_TRAMPOLINE art_quick_invoke_virtual_trampoline_with_access_check, artInvo
      *   a4 = JValue* result
      *   a5 = shorty
      */
-ENTRY art_quick_invoke_stub
+ENTRY_NO_GP art_quick_invoke_stub
     # push a4, a5, s0(rSUSPEND), s1(rSELF), s8, ra onto the stack
     daddiu $sp, $sp, -48
     .cfi_adjust_cfa_offset 48
@@ -706,7 +723,7 @@ END art_quick_invoke_stub
      *   a4 = JValue* result
      *   a5 = shorty
      */
-ENTRY art_quick_invoke_static_stub
+ENTRY_NO_GP art_quick_invoke_static_stub
 
     # push a4, a5, s0(rSUSPEND), s1(rSELF), s8, ra, onto the stack
     daddiu $sp, $sp, -48
@@ -850,7 +867,8 @@ ENTRY art_quick_check_cast
     sd     $a1, 8($sp)
     sd     $a0, 0($sp)
     jal    artIsAssignableFromCode
-    nop
+    .cpreturn                       # Restore gp from t8 in branch delay slot.
+                                    # t8 may be clobbered in artIsAssignableFromCode.
     beq    $v0, $zero, .Lthrow_class_cast_exception
     ld     $ra, 24($sp)
     jalr   $zero, $ra
@@ -862,6 +880,7 @@ ENTRY art_quick_check_cast
     ld     $a0, 0($sp)
     daddiu $sp, $sp, 32
     .cfi_adjust_cfa_offset -32
+    SETUP_GP
     SETUP_SAVE_ALL_CALLEE_SAVE_FRAME
     dla  $t9, artThrowClassCastException
     jalr $zero, $t9                 # artThrowClassCastException (Class*, Class*, Thread*)
@@ -907,13 +926,13 @@ ENTRY art_quick_aput_obj
     daddu $t1, $t1, $t0
     sb   $t0, ($t1)
     jalr $zero, $ra
-    nop
+    .cpreturn                       # Restore gp from t8 in branch delay slot.
 .Ldo_aput_null:
     dsll  $a1, $a1, 2
     daddu $t0, $a0, $a1
     sw   $a2, MIRROR_OBJECT_ARRAY_DATA_OFFSET($t0)
     jalr $zero, $ra
-    nop
+    .cpreturn                       # Restore gp from t8 in branch delay slot.
 .Lcheck_assignability:
     daddiu $sp, $sp, -64
     .cfi_adjust_cfa_offset 64
@@ -926,7 +945,8 @@ ENTRY art_quick_aput_obj
     move   $a1, $t1
     move   $a0, $t0
     jal    artIsAssignableFromCode  # (Class*, Class*)
-    nop
+    .cpreturn                       # Restore gp from t8 in branch delay slot.
+                                    # t8 may be clobbered in artIsAssignableFromCode.
     ld     $ra, 56($sp)
     ld     $t9, 24($sp)
     ld     $a2, 16($sp)
@@ -934,6 +954,7 @@ ENTRY art_quick_aput_obj
     ld     $a0, 0($sp)
     daddiu $sp, $sp, 64
     .cfi_adjust_cfa_offset -64
+    SETUP_GP
     bne    $v0, $zero, .Ldo_aput
     nop
     SETUP_SAVE_ALL_CALLEE_SAVE_FRAME
@@ -1311,7 +1332,7 @@ ENTRY art_quick_test_suspend
     bne    $a0, $zero, 1f
     daddiu rSUSPEND, $zero, SUSPEND_CHECK_INTERVAL   # reset rSUSPEND to SUSPEND_CHECK_INTERVAL
     jalr   $zero, $ra
-    nop
+    .cpreturn                                 # Restore gp from t8 in branch delay slot.
 1:
     SETUP_REFS_ONLY_CALLEE_SAVE_FRAME         # save callee saves for stack crawl
     jal    artTestSuspendFromCode             # (Thread*)
@@ -1350,6 +1371,7 @@ ENTRY art_quick_imt_conflict_trampoline
     dsll    $t0, 2                 # convert target method offset to bytes
     daddu   $a0, $t0               # get address of target method
     dla     $t9, art_quick_invoke_interface_trampoline
+    .cpreturn
     jalr    $zero, $t9
     lwu     $a0, MIRROR_OBJECT_ARRAY_DATA_OFFSET($a0)  # load the target method
 END art_quick_imt_conflict_trampoline
@@ -1478,8 +1500,7 @@ END art_quick_instrumentation_entry
     .global art_quick_instrumentation_exit
 art_quick_instrumentation_exit:
     .cfi_startproc
-    daddiu   $t9, $ra, 4       # put current address into $t9 to rebuild $gp
-    .cpload  $t9
+    SETUP_GP
     move     $ra, $zero        # link register is to here, so clobber with 0 for later checks
     SETUP_REFS_ONLY_CALLEE_SAVE_FRAME
     move     $t0, $sp          # remember bottom of caller's frame
@@ -1491,8 +1512,11 @@ art_quick_instrumentation_exit:
     mov.d    $f15, $f0         # pass fpr result
     move     $a2, $v0          # pass gpr result
     move     $a1, $t0          # pass $sp
-    jal      artInstrumentationMethodExitFromCode  # (Thread*, SP, gpr_res, fpr_res)
     move     $a0, rSELF        # pass Thread::Current
+    jal      artInstrumentationMethodExitFromCode  # (Thread*, SP, gpr_res, fpr_res)
+    .cpreturn                  # Restore gp from t8 in branch delay slot. gp is not used anymore,
+                               # and t8 may be clobbered in artInstrumentationMethodExitFromCode.
+
     move     $t9, $v0          # set aside returned link register
     move     $ra, $v1          # set link register for deoptimization
     ld       $v0, 0($sp)       # restore return values
diff --git a/runtime/arch/stub_test.cc b/runtime/arch/stub_test.cc
index de7804f759..6291164cc4 100644
--- a/runtime/arch/stub_test.cc
+++ b/runtime/arch/stub_test.cc
@@ -1563,7 +1563,11 @@ static void GetSet32Static(ArtField* f, Thread* self, mirror::ArtMethod* referre
                                            self,
                                            referrer);
 
+#if defined(__mips__) && defined(__LP64__)
+    EXPECT_EQ(static_cast<uint32_t>(res), values[i]) << "Iteration " << i;
+#else
     EXPECT_EQ(res, values[i]) << "Iteration " << i;
+#endif
   }
 #else
   UNUSED(f, self, referrer, test);