[llvm-bugs] [Bug 32291] New: [PowerPC] codegen problem in loop, results in run-time SIGSEGV in application

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=32291

            Bug ID: 32291
           Summary: [PowerPC] codegen problem in loop, results in run-time
                    SIGSEGV in application
           Product: new-bugs
           Version: 4.0
          Hardware: Other
                OS: Linux
            Status: NEW
          Severity: normal
          Priority: P
         Component: new bugs
          Assignee: unassignedbugs at nondot.org
          Reporter: eric.schweitz at pgroup.com
                CC: llvm-bugs at lists.llvm.org

Created attachment 18103
  --> https://bugs.llvm.org/attachment.cgi?id=18103&action=edit
reduced test case

In the (bugpoint reduced) test case, I've marked a basic block with the label
"L.bug_here", which is where I found the SIGSEGV was happening in the original
program.

Here is the LLVM IR:
L.bug_here:                                  ; preds = %L.bug_here,
%L.LB2_5666.prol.preheader
  %z00go = phi i64 [ 0, %L.LB2_5666.prol.preheader ], [ %696, %L.bug_here ]
  %z00gp = phi i64 [ %z00fr, %L.LB2_5666.prol.preheader ], [ %z00gq,
%L.bug_here ]
  %695 = load i8*, i8** %445, align 8, !dbg !1187, !tbaa !1157
  %696 = add nuw nsw i64 %z00go, 1, !dbg !1187
  %697 = mul i64 %z00gn, %696, !dbg !1187
  %698 = add i64 %694, %697, !dbg !1187
  %699 = mul i64 %698, %z00gl, !dbg !1187
  %700 = getelementptr i8, i8* %695, i64 %699, !dbg !1187
  %701 = bitcast i8* %700 to i64*, !dbg !1187
  %702 = load i64, i64* %701, align 8, !dbg !1187, !tbaa !1157
  %z00gr = shl i64 %z00go, 32, !dbg !1187
  %703 = ashr exact i64 %z00gr, 32, !dbg !1187
  %704 = getelementptr double, double* %693, i64 %703, !dbg !1187
  %705 = bitcast double* %704 to i64*, !dbg !1187
  store i64 %702, i64* %705, align 8, !dbg !1187, !tbaa !1240
  %z00gq = add i64 %z00gp, -1, !dbg !1187
  %z00gs = icmp eq i64 %z00gq, 0, !dbg !1187
  br i1 %z00gs, label %L.LB2_5666.prol.loopexit.unr-lcssa, label %L.bug_here,
!dbg !1187, !llvm.loop !1242

Using llc v3.9, this basic block does NOT suffer a SIGSEGV and appears as:

.LBB1_83:                               # %L.bug_here
                                        #   Parent Loop BB1_26 Depth=1
                                        #     Parent Loop BB1_42 Depth=2
                                        # =>    This Inner Loop Header: Depth=3
        .loc    1 107 1                 # z000002b9:107:1
        ld 11, 0(20)
        sldi 0, 28, 32
        sradi 12, 10, 29
        addi 7, 7, 1
        add 10, 10, 0
        ldx 11, 11, 8
        add 8, 8, 9
        stdx 11, 3, 12
        bdnz .LBB1_83

Using llc v4.0, this basic block appears as:

.LBB1_88:                               # %L.bug_here
                                        #   Parent Loop BB1_29 Depth=1
                                        #     Parent Loop BB1_47 Depth=2
                                        # =>    This Inner Loop Header: Depth=3
        .loc    1 107 1 is_stmt 1       # z000002b9:107:1
        ld 11, 0(25)
        addi 7, 7, 1
        ldx 11, 11, 9
        add 9, 9, 10
        stdu 11, 8(8)
        bdnz .LBB1_88

In this latter asm, r8 is loaded with an unmapped memory address and the
program takes a signal on the stdu insn.

The command used to compile the .ll file was simply:

% llc -o simplified.s simplified.ll

PS: It appears that the master branch generates a code pattern much like v3.9.

-- 
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/20170315/b3026bcd/attachment.html>