[LLVMbugs] [Bug 10857] New: [AVX] incorrect code generated for simple loop

[bugzilla-daemon at llvm.org]

http://llvm.org/bugs/show_bug.cgi?id=10857

           Summary: [AVX] incorrect code generated for simple loop
           Product: new-bugs
           Version: trunk
          Platform: PC
        OS/Version: All
            Status: NEW
          Severity: normal
          Priority: P
         Component: new bugs
        AssignedTo: unassignedbugs at nondot.org
        ReportedBy: matt at pharr.org
                CC: llvmbugs at cs.uiuc.edu


Created an attachment (id=7216)
 --> (http://llvm.org/bugs/attachment.cgi?id=7216)
llvm asm

The attached short test case represents the compilation of the loop:

void result(float *p) {
  for (int i = 0; i < 4; ++i) {
    p[2*i+1] = (float)(10+i);
    p[2*i] = (float)(14+i);
  }
}

(The llvm asm is simple enough that this should be easy to verify by
inspection.)

If I compile it using "llc -O0 -mattr=+avx", incorrect assembly is generated;
the output of the test program indicates that the even array elements aren't
being written to.  Correct code is generated with -O2 -mattr=+avx, or if I
don't use the AVX backend.

Here is the result of disassembling the code generated with -O0, avx.  (Intel
asm syntax.)  Note that the result of the second vcvtsi2ss instruction is never
written to memory, which is aligned with the indication from the test that the
even elements aren't written to.

_result:                                                # Function
        .type   _result, @function
        mov     eax, 0                                  # 0000 _ B8, 00000000
        mov     dword ptr [rsp-0x4], eax                # 0005 _ 89. 44 24, FC
        mov     qword ptr [rsp-0x10], rdi               # 0009 _ 48: 89. 7C 24,
F0
$_001:  mov     eax, dword ptr [rsp-0x4]                # 000E _ 8B. 44 24, FC
# Note: Immediate operand could be made smaller by sign extension
        cmp     eax, 4                                  # 0012 _ 3D, 00000004
        mov     dword ptr [rsp-0x14], eax               # 0017 _ 89. 44 24, EC
        jl      $_003                                   # 001B _ 7C, 11
        jmp     $_004                                   # 001D _ EB, 3E

$_002:  mov     eax, dword ptr [rsp-0x14]               # 001F _ 8B. 44 24, EC
# Note: Immediate operand could be made smaller by sign extension
        add     eax, 1                                  # 0023 _ 05, 00000001
        mov     dword ptr [rsp-0x4], eax                # 0028 _ 89. 44 24, FC
        jmp     $_001                                   # 002C _ EB, E0

$_003:  mov     eax, dword ptr [rsp-0x14]               # 002E _ 8B. 44 24, EC
        mov     ecx, dword ptr [rsp-0x14]               # 0032 _ 8B. 4C 24, EC
        lea     eax, [rcx+rax+0x1]                      # 0036 _ 8D. 44 01, 01
        lea     edx, [rcx+rcx]                          # 003A _ 8D. 14 09
        movsxd  rsi, eax                                # 003D _ 48: 63. F0
        lea     eax, [rcx+0xA]                          # 0040 _ 8D. 41, 0A
        vcvtsi2ss xmm0, xmm0, eax                       # 0043 _ C5 FA: 2A. C0
        mov     rdi, qword ptr [rsp-0x10]               # 0047 _ 48: 8B. 7C 24,
F0
        vmovss  dword ptr [rdi+rsi*4], xmm0             # 004C _ C5 FA: 11. 04
B7
        lea     eax, [rcx+0xE]                          # 0051 _ 8D. 41, 0E
        vcvtsi2ss xmm0, xmm0, eax                       # 0054 _ C5 FA: 2A. C0
        movsxd  rsi, edx                                # 0058 _ 48: 63. F2
        jmp     $_002                                   # 005B _ EB, C2

$_004:  ret                                             # 005D _ C3
        .size   _result, . - _result                    # End of function is
probably here

-- 
Configure bugmail: http://llvm.org/bugs/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are on the CC list for the bug.