[LLVMbugs] [Bug 23460] New: integrated assembler rejects inlined add w9, w9, #-1 in AtomicIncrement

[bugzilla-daemon at llvm.org]

https://llvm.org/bugs/show_bug.cgi?id=23460

            Bug ID: 23460
           Summary: integrated assembler rejects inlined add w9,w9,#-1 in
                    AtomicIncrement
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: normal
          Priority: P
         Component: Backend: AArch64
          Assignee: unassignedbugs at nondot.org
          Reporter: chh at google.com
                CC: llvmbugs at cs.uiuc.edu
    Classification: Unclassified

This test case shows that clang/llvm -target aarch64-lunix-android
failed to inline assembly code in NoBarrier_AtomicIncrement,
when the increment is -1.

There was no error when NoBarrier_AtomicIncrement is not inlined
or when -no-integrated-as is used.

$ cat a.cc

typedef int int32_t; 
typedef unsigned int Atomic32;
#ifdef NO_INLINE
__attribute__((noinline))
#endif
Atomic32 NoBarrier_AtomicIncrement(volatile Atomic32* ptr,
                                          Atomic32 increment) {
  Atomic32 result;
  int32_t temp;
  __asm__ __volatile__ (
    "0:                                       \n\t"
    "ldxr %w[result], %[ptr]                  \n\t"
    "add %w[result], %w[result], %w[increment]\n\t"
    "stxr %w[temp], %w[result], %[ptr]        \n\t"
    "cbnz %w[temp], 0b                        \n\t"
    : [result]"=&r" (result),
      [temp]"=&r" (temp),
      [ptr]"+Q" (*ptr)
    : [increment]"IJr" (increment)
    : "memory"
  );
  return result;
}
Atomic32* ptr;
void foo() {
  NoBarrier_AtomicIncrement(ptr, 1);
  NoBarrier_AtomicIncrement(ptr, -1);
}



$ clang++ -target aarch64-linux-android a.cc -S -O2 -o /tmp/x.O2.s
$.cc:12:49: error: expected compatible register, symbol or integer in range [0,
4095]
    "ldxr %w[result], %[ptr]                  \n\t" 
                                                ^
<inline asm>:3:14: note: instantiated into assembly here
        add w9, w9, #-1                         
                    ^
1 error generated.  

$ clang++ -target aarch64-linux-android a.cc -S -O2 -no-integrated-as -o
/tmp/x.O2.s

$ clang++ -target aarch64-linux-android a.cc -S -O2 -DNO_INLINE -o /tmp/x.O2.s


This code pattern appears in Android Open Source
external/v8/src/base/atomicops_internals_arm64_gcc.h,
which is now compiled with -no-integrated-as flag.


When compiled with -no-integrated-as:
The instruction
    add w9, w9, #-1
was correctly converted into
    sub w9, w9, #0x1


When compiled with -DNO_INLINE, the embedded assembly code is:
0000000000000000 <_Z25NoBarrier_AtomicIncrementPVjj>:
   0:   885f7c08    ldxr    w8, [x0]
   4:   0b010108    add w8, w8, w1
   8:   88097c08    stxr    w9, w8, [x0]
   c:   35ffffa9    cbnz    w9, 0 <_Z25NoBarrier_AtomicIncrementPVjj>
  10:   2a0803e0    mov w0, w8
  14:   d65f03c0    ret
and -1 was passed in w1.

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20150508/393fa3e5/attachment.html>