[lld] [lld][ARM] Don't emit veneers for wraparound branches. (PR #165263)

[Peter Smith via llvm-commits]

================
@@ -0,0 +1,118 @@
+// REQUIRES: arm
+// RUN: rm -rf %t && split-file %s %t && cd %t
+// RUN: llvm-mc -filetype=obj -triple=armv7-none-eabi code.s -o code.o
+// RUN: ld.lld -T unsigned1.ld code.o -o unsigned1.elf
+// RUN: llvm-objdump --triple=armv7 -d unsigned1.elf | FileCheck %s --check-prefix=UNSIGNED1
+// RUN: ld.lld -T unsigned2.ld code.o -o unsigned2.elf
+// RUN: llvm-objdump --triple=armv7 -d unsigned2.elf | FileCheck %s --check-prefix=UNSIGNED2
+// RUN: ld.lld -T signed1.ld code.o -o signed1.elf
+// RUN: llvm-objdump --triple=armv7 -d signed1.elf | FileCheck %s --check-prefix=SIGNED1
+// RUN: ld.lld -T signed2.ld code.o -o signed2.elf
+// RUN: llvm-objdump --triple=armv7 -d signed2.elf | FileCheck %s --check-prefix=SIGNED2
+
+// The aim of this test is to ensure that a BL instruction near one end of the
+// address space can reach a function at the extreme other end, directly, using
+// a branch offset that makes the address wrap round. We check this at both the
+// unsigned wraparound point (one address near 0 and the other near 0xFFFFFFFF)
+// and the signed wraparound point (addresses either side of 0x80000000),
+// crossing the boundary in both directions. In all four cases we expect a
+// direct branch with no veneer.
+
+// UNSIGNED1: Disassembly of section .text.lowaddr:
+// UNSIGNED1: 00010000 <func>:
+// UNSIGNED1:    10000: e12fff1e      bx      lr
+//
+// UNSIGNED1: Disassembly of section .text.highaddr:
+// UNSIGNED1: ffff0000 <_start>:
----------------
smithp35 wrote:

LLD convention is to only put the first address in a sequence of instructions. For example:
```
// UNSIGNED1: ffff0000: eb007ffe      bl      0x10000
// UNSIGNED1-NEXT:      e12fff1e      bx      lr
```

https://github.com/llvm/llvm-project/pull/165263