[ELF] - Provide a hint about --no-check-sections

[grimar]

Summary

This adds a hint about --no-check-sections option to the
sections overlapping error message.

One known legal use case for this option is OVERLAYS,
which are not yet supported in LLD, though with --no-check-sections
it would be possible to write linker script equal by functionality to
OVERLAY syntax.

[ruiu]

You can either disable error checking or ignore errors, but "ignoring this check" doesn't make much sense.

In what situation this hint is useful? Showing a hint suggests people do that, but IIUC, overlapping sections is almost always an error, and it is pretty rare that you need --no-check-sections option. It may give a wrong impression that the error can be resolved by passing that option.

[grimar]

In D45080#1053141, @ruiu wrote:

You can either disable error checking or ignore errors, but "ignoring this check" doesn't make much sense.

In what situation this hint is useful? Showing a hint suggests people do that, but IIUC, overlapping sections is almost always an error, and it is pretty rare that you need --no-check-sections option. It may give a wrong impression that the error can be resolved by passing that option.

In this message, I am saying "if you need overlays". That suggests that user *need* it. It is for cases like PR36768. LLD does not support OVERLAYS, though it is a possible legal thing to do.

[grimar]

• Addressed review comment (updated error message wording).

[ruiu]

In this message, I am saying "if you need overlays". That suggests that user *need* it. It is for cases like PR36768. LLD does not support OVERLAYS, though it is a possible legal thing to do.

As I said, if you show something, it suggests users do it even if the hint message comes with a footnote. You cannot really trust all users to understand what "if you need overlays" means. I think you can easily imagine that a user seeing that error message blindly adds that option to "fix" the problem. Unlike you, users are not linker experts, and our hint messages must be relevant to most users, or otherwise it could do harm than good.

Say, do you think that a half of the users who saw this hint message actually need --no-check-sections? I honestly don't think so. I cannot even imagine that 10% users who saw this error message have needed --no-check-sections because it is hard to imagine that you want to intentionally create overlapped sections.

[grimar]

In D45080#1054585, @ruiu wrote:

In this message, I am saying "if you need overlays". That suggests that user *need* it. It is for cases like PR36768. LLD does not support OVERLAYS, though it is a possible legal thing to do.

As I said, if you show something, it suggests users do it even if the hint message comes with a footnote. You cannot really trust all users to understand what "if you need overlays" means. I think you can easily imagine that a user seeing that error message blindly adds that option to "fix" the problem. Unlike you, users are not linker experts, and our hint messages must be relevant to most users, or otherwise it could do harm than good.

Say, do you think that a half of the users who saw this hint message actually need --no-check-sections? I honestly don't think so. I cannot even imagine that 10% users who saw this error message have needed --no-check-sections because it is hard to imagine that you want to intentionally create overlapped sections.

I generally find it difficult to imagine a user who sees this error honestly. If you write straightforward linker scripts without moving location counter backwards,
I guess such an error is either impossible or means we have a bug in linker. If you don't then most probably you write some kernel script and should know what you do :)
What is the difference with the "pass '-Wl,-z,notext' to allow text relocations in the output" hint we provide, btw? It is also can be an error, though we hint here.

I think the main use case we really want to think about is a case when somebody tries to switch from bfd to LLD. In this case, providing such a hint would help this user
to resolve the issue. It is user-friendly. First of all, the issue happens because we are incompatible with tools that users used for decades and forcing them to search
in the web or read help about all options we support to make things just work again is not a great experience.

There is a discussion in PR36768's comments about the new option for allowing overlays.
With that, "--no-check-sections" hint for all kind of overlaps will not make sense.
We might add a just hint for VA overlaps specific to OVERLAY feature only then.

[grimar]

We will not need this when/if we land OVERLAYs (D44780).

ELF/Writer.cpp

std::sort(Sections.begin(), Sections.end(),
             });
 
   // Finding overlap is easy given a vector is sorted by start position.
   // If an element starts before the end of the previous element, they overlap.
   for (size_t I = 1, End = Sections.size(); I < End; ++I) {
     SectionOffset A = Sections[I - 1];
     SectionOffset B = Sections[I];
     if (B.Offset < A.Offset + A.Sec->Size)
-      errorOrWarn(
-          "section " + A.Sec->Name + " " + Name + " range overlaps with " +
-          B.Sec->Name + "\n>>> " + A.Sec->Name + " range is " +
-          rangeToString(A.Offset, A.Sec->Size) + "\n>>> " + B.Sec->Name +
-          " range is " + rangeToString(B.Offset, B.Sec->Size));
+      errorOrWarn("section " + A.Sec->Name + " " + Name +
+                  " range overlaps with " + B.Sec->Name +
+                  "; use --no-check-sections to disable this check if you need "
+                  "overlays\n>>> " +
+                  A.Sec->Name + " range is " +
+                  rangeToString(A.Offset, A.Sec->Size) + "\n>>> " +
+                  B.Sec->Name + " range is " +
+                  rangeToString(B.Offset, B.Sec->Size));
   }
 }
 
 // Check for overlapping sections
 //
 // In this function we check that none of the output sections have overlapping
 // file offsets. For SHF_ALLOC sections we also check that the load address
 // ranges and the virtual address ranges don't overlap

test/ELF/linkerscript/overlapping-sections.s

 # REQUIRES: x86
 # RUN: llvm-mc -filetype=obj -triple=x86_64-unknown-linux %s -o %t.o
 
 # RUN: echo "SECTIONS { \
 # RUN:   .sec1 0x8000 : AT(0x8000) { sec1_start = .; *(.first_sec) sec1_end = .;} \
 # RUN:   .sec2 0x8800 : AT(0x8080) { sec2_start = .; *(.second_sec) sec2_end = .;} \
 # RUN: }" > %t-lma.script
 # RUN: not ld.lld -o %t.so --script %t-lma.script %t.o -shared 2>&1 | FileCheck %s -check-prefix LMA-OVERLAP-ERR
-# LMA-OVERLAP-ERR:      error: section .sec1 load address range overlaps with .sec2
+# LMA-OVERLAP-ERR:      error: section .sec1 load address range overlaps with .sec2; use --no-check-sections to disable this check if you need overlays
 # LMA-OVERLAP-ERR-NEXT: >>> .sec1 range is [0x8000, 0x80FF]
 # LMA-OVERLAP-ERR-NEXT: >>> .sec2 range is [0x8080, 0x817F]
 
 # Check that we create the expected binary with --noinhibit-exec or --no-check-sections:
 # RUN: ld.lld -o %t.so --script %t-lma.script %t.o -shared --noinhibit-exec
 # RUN: ld.lld -o %t.so --script %t-lma.script %t.o -shared --no-check-sections -fatal-warnings
 # RUN: ld.lld -o %t.so --script %t-lma.script %t.o -shared --check-sections --no-check-sections -fatal-warnings
 

 
 # Now try a script where the virtual memory addresses overlap but ensure that the
 # load addresses don't:
 # RUN: echo "SECTIONS { \
 # RUN:   .sec1 0x8000 : AT(0x8000) { sec1_start = .; *(.first_sec) sec1_end = .;} \
 # RUN:   .sec2 0x8020 : AT(0x8800) { sec2_start = .; *(.second_sec) sec2_end = .;} \
 # RUN: }" > %t-vaddr.script
 # RUN: not ld.lld -o %t.so --script %t-vaddr.script %t.o -shared 2>&1 | FileCheck %s -check-prefix VADDR-OVERLAP-ERR
-# VADDR-OVERLAP-ERR:      error: section .sec1 virtual address range overlaps with .sec2
+# VADDR-OVERLAP-ERR:      error: section .sec1 virtual address range overlaps with .sec2; use --no-check-sections to disable this check if you need overlays
 # VADDR-OVERLAP-ERR-NEXT: >>> .sec1 range is [0x8000, 0x80FF]
 # VADDR-OVERLAP-ERR-NEXT: >>> .sec2 range is [0x8020, 0x811F]
 
 # Check that the expected binary was created with --noinhibit-exec:
 # RUN: ld.lld -o %t.so --script %t-vaddr.script %t.o -shared --noinhibit-exec
 # RUN: llvm-readobj -sections -program-headers -elf-output-style=GNU %t.so | FileCheck %s -check-prefix BAD-VADDR
 # BAD-VADDR-LABEL: Section Headers:
 # BAD-VADDR: .sec1             PROGBITS        0000000000008000 002000 000100 00  WA  0   0  1

 
 # RUN: echo "SECTIONS { \
 # RUN:   .sec1 0x8000 : { sec1_start = .; *(.first_sec) sec1_end = .;} \
 # RUN:   .sec2 0x8040 : { sec2_start = .; *(.second_sec) sec2_end = .;} \
 # RUN: }" > %t-both-overlap.script
 
 # RUN: not ld.lld -o %t.so --script %t-both-overlap.script %t.o -shared 2>&1 | FileCheck %s -check-prefix BOTH-OVERLAP-ERR
 
-# BOTH-OVERLAP-ERR:      error: section .sec1 file range overlaps with .sec2
+# BOTH-OVERLAP-ERR:      error: section .sec1 file range overlaps with .sec2; use --no-check-sections to disable this check if you need overlays
 # BOTH-OVERLAP-ERR-NEXT: >>> .sec1 range is [0x2000, 0x20FF]
 # BOTH-OVERLAP-ERR-NEXT: >>> .sec2 range is [0x2040, 0x213F]
-# BOTH-OVERLAP-ERR:      error: section .sec1 virtual address range overlaps with .sec2
+# BOTH-OVERLAP-ERR:      error: section .sec1 virtual address range overlaps with .sec2; use --no-check-sections to disable this check if you need overlays
 # BOTH-OVERLAP-ERR-NEXT: >>> .sec1 range is [0x8000, 0x80FF]
 # BOTH-OVERLAP-ERR-NEXT: >>> .sec2 range is [0x8040, 0x813F]
-# BOTH-OVERLAP-ERR:      error: section .sec1 load address range overlaps with .sec2
+# BOTH-OVERLAP-ERR:      error: section .sec1 load address range overlaps with .sec2; use --no-check-sections to disable this check if you need overlays
 # BOTH-OVERLAP-ERR-NEXT: >>> .sec1 range is [0x8000, 0x80FF]
 # BOTH-OVERLAP-ERR-NEXT: >>> .sec2 range is [0x8040, 0x813F]
 
 # RUN: ld.lld -o %t.so --script %t-both-overlap.script %t.o -shared --noinhibit-exec
 # Note: In case everything overlaps we create a binary with overlapping file
 # offsets. ld.bfd seems to place .sec1 to file offset 18000 and .sec2
 # at 18100 so that only virtual addr and LMA overlap
 # However, in order to create such a broken binary the user has to ignore a