[PATCH 55/58] powerpc: Work around gcc miscompilation of __pa() on 64-bit

[Luis Henriques]

3.5.7.21 -stable review patch.  If anyone has any objections, please let me know.

------------------

From: Paul Mackerras <paulus at samba.org>

commit bdbc29c19b2633b1d9c52638fb732bcde7a2031a upstream.

On 64-bit, __pa(&static_var) gets miscompiled by recent versions of
gcc as something like:

        addis 3,2,.LANCHOR1+4611686018427387904 at toc@ha
        addi 3,3,.LANCHOR1+4611686018427387904 at toc@l

This ends up effectively ignoring the offset, since its bottom 32 bits
are zero, and means that the result of __pa() still has 0xC in the top
nibble.  This happens with gcc 4.8.1, at least.

To work around this, for 64-bit we make __pa() use an AND operator,
and for symmetry, we make __va() use an OR operator.  Using an AND
operator rather than a subtraction ends up with slightly shorter code
since it can be done with a single clrldi instruction, whereas it
takes three instructions to form the constant (-PAGE_OFFSET) and add
it on.  (Note that MEMORY_START is always 0 on 64-bit.)

Signed-off-by: Paul Mackerras <paulus at samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh at kernel.crashing.org>
Signed-off-by: Luis Henriques <luis.henriques at canonical.com>
---
 arch/powerpc/Kconfig            |  1 +
 arch/powerpc/include/asm/page.h | 10 ++++++++++
 2 files changed, 11 insertions(+)

diff --git a/arch/powerpc/Kconfig b/arch/powerpc/Kconfig
index 4d8336c..74b6ce4 100644
--- a/arch/powerpc/Kconfig
+++ b/arch/powerpc/Kconfig
@@ -975,6 +975,7 @@ config RELOCATABLE
 	  must live at a different physical address than the primary
 	  kernel.
 
+# This value must have zeroes in the bottom 60 bits otherwise lots will break
 config PAGE_OFFSET
 	hex
 	default "0xc000000000000000"
diff --git a/arch/powerpc/include/asm/page.h b/arch/powerpc/include/asm/page.h
index f072e97..2e6c4e5a 100644
--- a/arch/powerpc/include/asm/page.h
+++ b/arch/powerpc/include/asm/page.h
@@ -211,9 +211,19 @@ extern long long virt_phys_offset;
 #define __va(x) ((void *)(unsigned long)((phys_addr_t)(x) + VIRT_PHYS_OFFSET))
 #define __pa(x) ((unsigned long)(x) - VIRT_PHYS_OFFSET)
 #else
+#ifdef CONFIG_PPC64
+/*
+ * gcc miscompiles (unsigned long)(&static_var) - PAGE_OFFSET
+ * with -mcmodel=medium, so we use & and | instead of - and + on 64-bit.
+ */
+#define __va(x) ((void *)(unsigned long)((phys_addr_t)(x) | PAGE_OFFSET))
+#define __pa(x) ((unsigned long)(x) & 0x0fffffffffffffffUL)
+
+#else /* 32-bit, non book E */
 #define __va(x) ((void *)(unsigned long)((phys_addr_t)(x) + PAGE_OFFSET - MEMORY_START))
 #define __pa(x) ((unsigned long)(x) - PAGE_OFFSET + MEMORY_START)
 #endif
+#endif
 
 /*
  * Unfortunately the PLT is in the BSS in the PPC32 ELF ABI,
-- 
1.8.3.2