[AVR][MC] Add ELF flag 'EF_AVR_LINKRELAX_PREPARED' to OBJ files

[benshi001]

Summary

This is in accordance with avr-gcc, even '-mno-relax' is specified
to avr-gcc, this flag will also be added to the output relocatables.

[benshi001]

llvm/test/MC/AVR/elf_header.s

71

This is the newly added ELF flag.

[MaskRay]

I'm concerned of this change. Do you intend to add linker relaxation to lld/ELF's AVR port? Implementing features is indeed fun but I am unsure whether the use case has a sufficient demand of the feature.
RISC-V linker relaxation has greatly increased code complexity and LoongArch folks want to copy that. I don't want to see a third arch which may use a non-recommended way implementing linker relaxation.

[benshi001]

In D144617#4146563, @MaskRay wrote:

I'm concerned of this change. Do you intend to add linker relaxation to lld/ELF's AVR port? Implementing features is indeed fun but I am unsure whether the use case has a sufficient demand of the feature.
RISC-V linker relaxation has greatly increased code complexity and LoongArch folks want to copy that. I don't want to see a third arch which may use a non-recommended way implementing linker relaxation.

I have no plan for adding linker relaxation to lld/ELF's AVR port, at least, you can deny that if you think it introduces complexity to lld.

However, avr-ld indeed has already implemented linker relaxation, my current patch can make use of it, and clang+avr-ld can generate shorter assembly.

[MaskRay]

I think we should understand what this flag does and how it affects assemblers/linkers.
If it does not cause behavior difference, I am unsure we should add it.

[benshi001]

In D144617#4148171, @MaskRay wrote:

I think we should understand what this flag does and how it affects assemblers/linkers.
If it does not cause behavior difference, I am unsure we should add it.

This flag does cause difference behavior of gnu ld. With this flag added and --relax specified to gnu-ld, the gnu-ld will optimize long call (4-byte instruction) to short call (2-byte instruction), otherwise either this flag or --relax is absent, gnu-ld will not do the long call to short call optimization.

[benshi001]

In D144617#4148171, @MaskRay wrote:

I think we should understand what this flag does and how it affects assemblers/linkers.
If it does not cause behavior difference, I am unsure we should add it.

This is how gnu ld perform linker relaxation for AVR
https://sourceware.org/git/?p=binutils-gdb.git;a=blob;f=bfd/elf32-avr.c;h=702719136d09acbc8c98ec49ab8129d0f33fffa8;hb=6777dece58127236db900215857f9070ad63e0bf#l2423

We can see this flag EF_AVR_LINKRELAX_PREPARED does affect
https://sourceware.org/git/?p=binutils-gdb.git;a=blob;f=bfd/elf32-avr.c;h=702719136d09acbc8c98ec49ab8129d0f33fffa8;hb=6777dece58127236db900215857f9070ad63e0bf#l2533

The gnu ld will optimize long call (absolute address) to short call (PC relative address) only if this flag is set. (Of course --relax must also be specified.)

[MaskRay]

In D144617#4149029, @benshi001 wrote:

In D144617#4148171, @MaskRay wrote:

I think we should understand what this flag does and how it affects assemblers/linkers.
If it does not cause behavior difference, I am unsure we should add it.

This is how gnu ld perform linker relaxation for AVR
https://sourceware.org/git/?p=binutils-gdb.git;a=blob;f=bfd/elf32-avr.c;h=702719136d09acbc8c98ec49ab8129d0f33fffa8;hb=6777dece58127236db900215857f9070ad63e0bf#l2423

We can see this flag EF_AVR_LINKRELAX_PREPARED does affect
https://sourceware.org/git/?p=binutils-gdb.git;a=blob;f=bfd/elf32-avr.c;h=702719136d09acbc8c98ec49ab8129d0f33fffa8;hb=6777dece58127236db900215857f9070ad63e0bf#l2533

The gnu ld will optimize long call (absolute address) to short call (PC relative address) only if this flag is set. (Of course --relax must also be specified.)

s/gnu ld/GNU ld/. Change the summary to say that it affects GNU ld.

[benshi001]

In D144617#4149200, @MaskRay wrote:

In D144617#4149029, @benshi001 wrote:

In D144617#4148171, @MaskRay wrote:

I think we should understand what this flag does and how it affects assemblers/linkers.
If it does not cause behavior difference, I am unsure we should add it.

This is how gnu ld perform linker relaxation for AVR
https://sourceware.org/git/?p=binutils-gdb.git;a=blob;f=bfd/elf32-avr.c;h=702719136d09acbc8c98ec49ab8129d0f33fffa8;hb=6777dece58127236db900215857f9070ad63e0bf#l2423

We can see this flag EF_AVR_LINKRELAX_PREPARED does affect
https://sourceware.org/git/?p=binutils-gdb.git;a=blob;f=bfd/elf32-avr.c;h=702719136d09acbc8c98ec49ab8129d0f33fffa8;hb=6777dece58127236db900215857f9070ad63e0bf#l2533

The gnu ld will optimize long call (absolute address) to short call (PC relative address) only if this flag is set. (Of course --relax must also be specified.)

s/gnu ld/GNU ld/. Change the summary to say that it affects GNU ld.

Thanks for your comment. I will update my commit message when committing.

[MaskRay]

The commit has Reviewed By: MaskRay, jacquesguan, aykevl but @jacquesguan/@aykevl haven't signed off on the patch so it should just have my username :)

If you use arc amend (https://llvm.org/docs/Phabricator.html#requesting-a-review-via-the-command-line), Reviewed By: will be added correctly.

lld/test/ELF/avr-flags.s

 ; REQUIRES: avr
 
 ; RUN: llvm-mc -filetype=obj -triple=avr -mcpu=avr5 %s -o %t-v5
 ; RUN: llvm-mc -filetype=obj -triple=avr -mcpu=avrxmega3 %s -o %t-xmega3
 ; RUN: ld.lld %t-v5 -o %t-v5.out
 ; RUN: ld.lld %t-xmega3 -o %t-xmega3.out
 ; RUN: llvm-readobj -h %t-v5.out | FileCheck --check-prefix V5 %s
 ; RUN: llvm-readobj -h %t-xmega3.out | FileCheck --check-prefix XMEGA3 %s
 
 ;; Ensure LLD won't silently mix object files targeting different ISAs.
 ; RUN: not ld.lld %t-v5 %t-xmega3 -o /dev/null 2>&1 | FileCheck --check-prefix ERR %s
 ; ERR: error: {{.*}}: cannot link object files with incompatible target ISA
 
-; V5:  Flags [ (0x5)
+; V5:  Flags [ (0x85)
 ; V5:  EF_AVR_ARCH_AVR5 (0x5)
-; XMEGA3: Flags [ (0x67)
+; V5:  EF_AVR_LINKRELAX_PREPARED (0x80)
+; XMEGA3: Flags [ (0xE7)
 ; XMEGA3: EF_AVR_ARCH_XMEGA3 (0x67)
+; XMEGA3: EF_AVR_LINKRELAX_PREPARED (0x80)

lld/test/ELF/linkerscript/avr5.test

 
 # RELOC:      ldi  r24, 2
 # RELOC-NEXT: ldi  r25, 8
 
 # RUN: llvm-readelf --headers %t/avr5a.out | FileCheck %s --check-prefix=HEAD
 # RUN: llvm-readelf --headers %t/avr5b.out | FileCheck %s --check-prefix=HEAD
 
 # HEAD: Atmel AVR 8-bit microcontroller
-# HEAD: 0x5, EF_AVR_ARCH_AVR5
+# HEAD: 0x85, EF_AVR_ARCH_AVR5, relaxable
 
 # HEAD:      Name     Type      Address   Off     Size
 # HEAD-NEXT:          NULL      00000000  000000  000000
 # HEAD-NEXT: .text    PROGBITS  00000000  001000  000006
 # HEAD-NEXT: .data    PROGBITS  00000800  001800  000004
 
 #--- avr5.s
   .text

llvm/lib/Target/AVR/MCTargetDesc/AVRELFStreamer.cpp

 
 AVRELFStreamer::AVRELFStreamer(MCStreamer &S, const MCSubtargetInfo &STI)
     : AVRTargetStreamer(S) {
 
   MCAssembler &MCA = getStreamer().getAssembler();
   unsigned EFlags = MCA.getELFHeaderEFlags();
 
   EFlags |= getEFlagsForFeatureSet(STI.getFeatureBits());
+  EFlags |= ELF::EF_AVR_LINKRELAX_PREPARED;
 
   MCA.setELFHeaderEFlags(EFlags);
 }
 
 } // end namespace llvm

llvm/test/MC/AVR/elf_header.s

+; RUN: llvm-mc -filetype=obj -triple avr -mcpu=at90s8515 %s -o - \
+; RUN:     | llvm-readobj -h - | FileCheck --check-prefixes=ALL,AVR2 %s
+; RUN: llvm-mc -filetype=obj -triple avr -mcpu=attiny13a %s -o - \
+; RUN:     | llvm-readobj -h - | FileCheck --check-prefixes=ALL,AVR25 %s
+; RUN: llvm-mc -filetype=obj -triple avr -mcpu=attiny167 %s -o - \
+; RUN:     | llvm-readobj -h - | FileCheck --check-prefixes=ALL,AVR35 %s
+; RUN: llvm-mc -filetype=obj -triple avr -mcpu=atmega88 %s -o - \
+; RUN:     | llvm-readobj -h - | FileCheck --check-prefixes=ALL,AVR4 %s
+; RUN: llvm-mc -filetype=obj -triple avr -mcpu=atmega16 %s -o - \
+; RUN:     | llvm-readobj -h - | FileCheck --check-prefixes=ALL,AVR5 %s
+; RUN: llvm-mc -filetype=obj -triple avr -mcpu=atmega128 %s -o - \
+; RUN:     | llvm-readobj -h - | FileCheck --check-prefixes=ALL,AVR51 %s
+; RUN: llvm-mc -filetype=obj -triple avr -mcpu=attiny817 %s -o - \
+; RUN:     | llvm-readobj -h - | FileCheck --check-prefixes=ALL,XM3 %s
+; RUN: llvm-mc -filetype=obj -triple avr -mcpu=atxmega256a3u %s -o - \
+; RUN:     | llvm-readobj -h - | FileCheck --check-prefixes=ALL,XM6 %s
+; RUN: llvm-mc -filetype=obj -triple avr -mcpu=atxmega256a3u %s -o - \
+; RUN:     | llvm-readobj -h - | FileCheck --check-prefixes=ALL,XM6 %s
+; RUN: llvm-mc -filetype=obj -triple avr -mcpu=attiny10 %s -o - \
+; RUN:     | llvm-readobj -h - | FileCheck --check-prefixes=ALL,TINY %s
+
+; ALL:       ElfHeader {
+; ALL-NEXT:    Ident {
+; ALL-NEXT:      Magic: (7F 45 4C 46)
+; ALL-NEXT:      Class: 32-bit (0x1)
+; ALL-NEXT:      DataEncoding: LittleEndian (0x1)
+; ALL-NEXT:      FileVersion: 1
+; ALL-NEXT:      OS/ABI: SystemV (0x0)
+; ALL-NEXT:      ABIVersion: 0
+; ALL-NEXT:      Unused: (00 00 00 00 00 00 00)
+; ALL-NEXT:    }
+; ALL-NEXT:    Type: Relocatable (0x1)
+; ALL-NEXT:    Machine: EM_AVR (0x53)
+; ALL-NEXT:    Version: 1
+; ALL-NEXT:    Entry: 0x0
+; ALL-NEXT:    ProgramHeaderOffset: 0x0
+; ALL-NEXT:    SectionHeaderOffset: 0x5C
+
+; AVR2:        Flags [ (0x82)
+; AVR2-NEXT:     EF_AVR_ARCH_AVR2 (0x2)
+
+; AVR25:       Flags [ (0x99)
+; AVR25-NEXT:    EF_AVR_ARCH_AVR25 (0x19)
+
+; AVR35:       Flags [ (0xA3)
+; AVR35-NEXT:    EF_AVR_ARCH_AVR35 (0x23)
+
+; AVR4:        Flags [ (0x84)
+; AVR4-NEXT:     EF_AVR_ARCH_AVR4 (0x4)
+
+; AVR5:        Flags [ (0x85)
+; AVR5-NEXT:     EF_AVR_ARCH_AVR5 (0x5)
+
+; AVR51:       Flags [ (0xB3)
+; AVR51-NEXT:    EF_AVR_ARCH_AVR51 (0x33)
+
+; XM3:         Flags [ (0xE7)
+; XM3-NEXT:      EF_AVR_ARCH_XMEGA3 (0x67)
+
+; XM6:         Flags [ (0xEA)
+; XM6-NEXT:      EF_AVR_ARCH_XMEGA6 (0x6A)
+
+; TINY:        Flags [ (0xE4)
+; TINY-NEXT:     EF_AVR_ARCH_AVRTINY (0x64)
+
+; ALL:           EF_AVR_LINKRELAX_PREPARED (0x80)
+; ALL-NEXT:    ]
+; ALL-NEXT:    HeaderSize: 52
+; ALL-NEXT:    ProgramHeaderEntrySize: 0
+; ALL-NEXT:    ProgramHeaderCount: 0
+; ALL-NEXT:    SectionHeaderEntrySize: 40
+; ALL-NEXT:    SectionHeaderCount: 4
+; ALL-NEXT:    StringTableSectionIndex: 1
+; ALL-NEXT:  }