[llvm-bugs] [Bug 37240] New: -g causes different code to be generated due to PostRA scheduling differences

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=37240

            Bug ID: 37240
           Summary: -g causes different code to be generated due to PostRA
                    scheduling differences
           Product: libraries
           Version: 6.0
          Hardware: All
                OS: All
            Status: NEW
          Severity: normal
          Priority: P
         Component: Common Code Generator Code
          Assignee: unassignedbugs at nondot.org
          Reporter: gberry at codeaurora.org
                CC: llvm-bugs at lists.llvm.org

When -g (or just -gline-tables-only) is passed to clang, CFI_INSTRUCTIONs are
added to the function prologue by the Prologue/Epilogue insertion pass.  These
instructions act as scheduling region barriers when the PostRA scheduler is
run.  This can lead to different schedules for the prologue block, which can
further lead to differences in tail merging/duplication.

A simple example the illustrates the problem:

target triple = "aarch64-linux-gnu"

@X1 = global i32 0, align 4
@X2 = global i32 0, align 4
@X3 = global i32 0, align 4
@X4 = global i32 0, align 4

define void @test(i32 %i) #0 {
entry:
  %0 = load i32, i32* @X1, align 4
  %x1 = add i32 %0, 1
  %x2 = add i32 %0, 2
  %x3 = add i32 %0, 3
  %x4 = add i32 %0, 4
  tail call void @foo()
  store i32 %x1, i32* @X1, align 4
  store i32 %x2, i32* @X2, align 4
  store i32 %x3, i32* @X3, align 4
  store i32 %x4, i32* @X4, align 4
  ret void
}

declare void @foo()

attributes #0 = { nounwind }

!llvm.dbg.cu = !{!0}
!llvm.module.flags = !{!3, !4, !5}
!0 = distinct !DICompileUnit(language: DW_LANG_C99, file: !1, producer: "clang
version 7.0.0 (trunk 330790) (llvm/trunk 330787)", isOptimized: true,
runtimeVersion: 0, emissionKind: LineTablesOnly, enums: !2)
!1 = !DIFile(filename: "test.c", directory: "")
!2 = !{}
!3 = !{i32 2, !"Dwarf Version", i32 4}
!4 = !{i32 2, !"Debug Info Version", i32 3}
!5 = !{i32 1, !"wchar_size", i32 4}


when compiled as is (i.e. with debug info), produces the following code:

        stp     x23, x22, [sp, #-48]!   // 8-byte Folded Spill
        stp     x21, x20, [sp, #16]     // 8-byte Folded Spill
        stp     x19, x30, [sp, #32]     // 8-byte Folded Spill
        .cfi_def_cfa_offset 48
        .cfi_offset w30, -8
        .cfi_offset w19, -16
        .cfi_offset w20, -24
        .cfi_offset w21, -32
        .cfi_offset w22, -40
        .cfi_offset w23, -48
        adrp    x19, X1
        ldr     w8, [x19, :lo12:X1]
        add     w20, w8, #1             // =1
        add     w21, w8, #2             // =2
        add     w22, w8, #3             // =3
        add     w23, w8, #4             // =4
        bl      foo
        adrp    x8, X2
        str     w20, [x19, :lo12:X1]
        str     w21, [x8, :lo12:X2]
        ldp     x19, x30, [sp, #32]     // 8-byte Folded Reload
        ldp     x21, x20, [sp, #16]     // 8-byte Folded Reload
        adrp    x9, X3
        adrp    x10, X4
        str     w22, [x9, :lo12:X3]
        str     w23, [x10, :lo12:X4]
        ldp     x23, x22, [sp], #48     // 8-byte Folded Reload
        ret


but when the debug metadata is commented out, it produces the following code
instead:

        stp     x23, x22, [sp, #-48]!   // 8-byte Folded Spill
        stp     x19, x30, [sp, #32]     // 8-byte Folded Spill
        adrp    x19, X1
        ldr     w8, [x19, :lo12:X1]
        stp     x21, x20, [sp, #16]     // 8-byte Folded Spill
        add     w20, w8, #1             // =1
        add     w21, w8, #2             // =2
        add     w22, w8, #3             // =3
        add     w23, w8, #4             // =4
        bl      foo
        adrp    x8, X2
        str     w20, [x19, :lo12:X1]
        str     w21, [x8, :lo12:X2]
        ldp     x19, x30, [sp, #32]     // 8-byte Folded Reload
        ldp     x21, x20, [sp, #16]     // 8-byte Folded Reload
        adrp    x9, X3
        adrp    x10, X4
        str     w22, [x9, :lo12:X3]
        str     w23, [x10, :lo12:X4]
        ldp     x23, x22, [sp], #48     // 8-byte Folded Reload
        ret

Note the different position in the schedule of the first adrp and ldr
instructions.

-- 
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/20180425/151b393b/attachment.html>