xtensa: don't use a12 in __xtensa_copy_user in call0 ABI

a12 is callee-saved register in xtensa call0 ABI, so a function must not
change it. The main unaligned copy loop of __xtensa_copy_user uses all
low-numbered registers, so a register must be spilled to avoid using a12
as a loop counter.

Signed-off-by: Max Filippov <jcmvbkbc@gmail.com>

diff --git a/arch/xtensa/lib/usercopy.S b/arch/xtensa/lib/usercopy.S
index a0aa4047f94aec5779014c5aaef043f97e76d22c..16128c094c621ed8cdc9c6f8854a9aac611c966e 100644 (file)
--- a/arch/xtensa/lib/usercopy.S
+++ b/arch/xtensa/lib/usercopy.S
@@ -60,7 +60,12 @@
        .text
 ENTRY(__xtensa_copy_user)
 
-       abi_entry_default
+#if !XCHAL_HAVE_LOOPS && defined(__XTENSA_CALL0_ABI__)
+#define STACK_SIZE 4
+#else
+#define STACK_SIZE 0
+#endif
+       abi_entry(STACK_SIZE)
        # a2/ dst, a3/ src, a4/ len
        mov     a5, a2          # copy dst so that a2 is return value
        mov     a11, a4         # preserve original len for error case
@@ -75,7 +80,7 @@ ENTRY(__xtensa_copy_user)
        __ssa8  a3              # set shift amount from byte offset
        bnez    a4, .Lsrcunaligned
        movi    a2, 0           # return success for len==0
-       abi_ret_default
+       abi_ret(STACK_SIZE)
 
 /*
  * Destination is unaligned
@@ -127,7 +132,7 @@ EX(10f)     s8i     a6, a5, 0
 #endif /* !XCHAL_HAVE_LOOPS */
 .Lbytecopydone:
        movi    a2, 0           # return success for len bytes copied
-       abi_ret_default
+       abi_ret(STACK_SIZE)
 
 /*
  * Destination and source are word-aligned.
@@ -187,7 +192,7 @@ EX(10f)     l8ui    a6, a3,  0
 EX(10f)        s8i     a6, a5,  0
 .L5:
        movi    a2, 0           # return success for len bytes copied
-       abi_ret_default
+       abi_ret(STACK_SIZE)
 
 /*
  * Destination is aligned, Source is unaligned
@@ -205,8 +210,14 @@ EX(10f)    l32i    a6, a3, 0       # load first word
        loopnez a7, .Loop2done
 #else /* !XCHAL_HAVE_LOOPS */
        beqz    a7, .Loop2done
+#if defined(__XTENSA_CALL0_ABI__)
+       s32i    a10, a1, 0
+       slli    a10, a7, 4
+       add     a10, a10, a3    # a10 = end of last 16B source chunk
+#else
        slli    a12, a7, 4
        add     a12, a12, a3    # a12 = end of last 16B source chunk
+#endif
 #endif /* !XCHAL_HAVE_LOOPS */
 .Loop2:
 EX(10f)        l32i    a7, a3,  4
@@ -224,7 +235,12 @@ EX(10f)    s32i    a8, a5,  8
 EX(10f)        s32i    a9, a5, 12
        addi    a5, a5, 16
 #if !XCHAL_HAVE_LOOPS
+#if defined(__XTENSA_CALL0_ABI__)
+       blt     a3, a10, .Loop2
+       l32i    a10, a1, 0
+#else
        blt     a3, a12, .Loop2
+#endif
 #endif /* !XCHAL_HAVE_LOOPS */
 .Loop2done:
        bbci.l  a4, 3, .L12
@@ -264,7 +280,7 @@ EX(10f)     l8ui    a6, a3,  0
 EX(10f)        s8i     a6, a5,  0
 .L15:
        movi    a2, 0           # return success for len bytes copied
-       abi_ret_default
+       abi_ret(STACK_SIZE)
 
 ENDPROC(__xtensa_copy_user)
 
@@ -281,4 +297,4 @@ ENDPROC(__xtensa_copy_user)
 10:
        sub     a2, a5, a2      /* a2 <-- bytes copied */
        sub     a2, a11, a2     /* a2 <-- bytes not copied */
-       abi_ret_default
+       abi_ret(STACK_SIZE)