[llvm-bugs] [Bug 35529] New: Conversion to TMHH gives wrong result

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=35529

            Bug ID: 35529
           Summary: Conversion to TMHH gives wrong result
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Backend: SystemZ
          Assignee: unassignedbugs at nondot.org
          Reporter: paulsson at linux.vnet.ibm.com
                CC: llvm-bugs at lists.llvm.org

Created attachment 19519
  --> https://bugs.llvm.org/attachment.cgi?id=19519&action=edit
reduced csmith program (pre-processed)

Csmith discovered a program which gave a differing checksum with clang -O0:

*** Analyzing wrong6.c   (seed= 15242346099534488386, size= 153640):
3F15FE04: 16
88680976:  2
Different reduced builds (18):
3F15FE04:  /root/llvm/build/llvm-dev/bin/clang -O3 -march=z10 wrong6.c
-I/root/csmith/runtime -o a.out -w
3F15FE04:  gcc -O0 -march=z10 wrong6.c -I/root/csmith/runtime -o a.out -w
3F15FE04:  gcc -O1 -march=z10 wrong6.c -I/root/csmith/runtime -o a.out -w
3F15FE04:  gcc -O0 -march=z13 wrong6.c -I/root/csmith/runtime -o a.out -w
3F15FE04:  gcc -O1 -march=z13 wrong6.c -I/root/csmith/runtime -o a.out -w
88680976:  /root/llvm/build/llvm-dev/bin/clang -O0 -march=z10 wrong6.c
-I/root/csmith/runtime -o a.out -w
88680976:  /root/llvm/build/llvm-dev/bin/clang -O0 -march=z13 wrong6.c
-I/root/csmith/runtime -o a.out -w
3F15FE04:  /root/llvm/build/llvm-dev/bin/clang -O3 -march=z13 wrong6.c
-I/root/csmith/runtime -o a.out -w


opt -instsimplify corrected the output... The non-simplified condition check
was not correctly handled.

I am not sure exactly what goes wrong, but by commenting out like

  // Check whether the mask is suitable for TMHH, TMHL, TMLH or TMLL.           
  // if (!SystemZ::isImmLL(Mask) && !SystemZ::isImmLH(Mask) &&                  
  //     !SystemZ::isImmHL(Mask) && !SystemZ::isImmHH(Mask))                    
    return 0;

, the correct checksum is printed, which makes it seem that the TMHH conversion
contains the bug.

This if statement is always false (and with -instsimplify it is removed):

static int16_t g_5 = (-9L);
...
int64_t p_3 = g_5;
if (((0 > p_3) >= 3L))
     (*g_1619) = 0;

The store of 0 to g_1619 should never happen (a jump *should* be made to
BB0_17, past BB0_16 which contains the store)

The diff of the .s files is actually very small in this block:

correct                                 bad
.LBB0_15:                               .LBB0_15:                            
        brasl   %r14, platform_main_b           brasl   %r14, platform_main_b
        brasl   %r14, crc32_gentab at PL           brasl   %r14, crc32_gentab at PL
        lghrl   %r0, g_5                        lghrl   %r0, g_5
        srlg    %r1, %r0, 63          |         tmhh    %r0, 32768
        stg     %r0, 184(%r15)                  stg     %r0, 184(%r15)
        cgijl   %r1, 3, .LBB0_17      |         je      .LBB0_17


It seems that to the left both the check for p_3 being negative and then
comparing that to 3, while the tmhh just checks for the sign bit and forgets
the check against 3L..?

bin/clang -O0 -march=z13 w6.i -o a.out -w
./a.out
checksum = 6522DF69  // BAD checksum

bin/clang -O1 -march=z13 w6.i -o a.out -w
./a.out
checksum = B5499A84  // GOOD checksum

-- 
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/20171205/fde6d782/attachment-0001.html>