Add a mechanism to GC when building .dwps.

.gitignore

@@ -1,2 +1,3 @@
 /target
 /tests/.lit_test_times.txt
+*~

Cargo.toml

@@ -4,3 +4,6 @@ members = [
     "thorin",
     "thorin-bin"
 ]
+
+[patch.crates-io]
+gimli = { git = "https://github.com/gimli-rs/gimli", rev = "1bb1457480938ea412c7b2a99e07534e7846d1df" }

tests/gc-all-dead.test

@@ -0,0 +1,28 @@
+# Test that GC does not remove an entirely dead CU. The skeleton CU in the
+# executable references this CU by DwoId; dropping it would leave a dangling
+# reference that we cannot fix without modifying the executable.
+
+# Assemble the .dwo (split debug info).
+# RUN: llvm-mc --triple=x86_64-unknown-linux --filetype=obj %p/inputs/gc-all-dead-dwo.s -o %t.dwo
+
+# Generate the exec assembly with the .dwo path substituted in, then assemble it.
+# RUN: sed 's|DWO_PATH|%t.dwo|' %p/inputs/gc-all-dead-exec.s | llvm-mc --triple=x86_64-unknown-linux --filetype=obj -o %t.o
+
+# Baseline: both subprograms survive when not garbage-collecting.
+# RUN: thorin -e %t.o -o %t-nogc.dwp
+# RUN: llvm-dwarfdump -v %t-nogc.dwp | FileCheck --check-prefixes=CHECK,NOGC %s
+
+# With garbage collection the dead subprograms are removed but the CU survives.
+# RUN: thorin --gc -e %t.o -o %t-gc.dwp
+# RUN: llvm-dwarfdump -v %t-gc.dwp | FileCheck --check-prefixes=CHECK,GC %s
+
+# The compile unit must always be present.
+# CHECK:        .debug_info.dwo contents:
+# CHECK:        DW_TAG_compile_unit
+
+# Without GC both dead subprograms are present.
+# NOGC:         DW_TAG_subprogram
+# NOGC:         DW_TAG_subprogram
+
+# After GC both subprograms are removed but the CU itself survives.
+# GC-NOT:       DW_TAG_subprogram

tests/gc-compact-gnu-str-index.test

@@ -0,0 +1,67 @@
+# Test that GC compacts the .debug_str_offsets.dwo table when string attributes
+# use DW_FORM_GNU_str_index (DWARF4 split-DWARF extension), by removing entries
+# for strings only referenced by dead DIEs, and that surviving DIEs' string
+# attributes are correctly remapped to the new indices.
+#
+# The .dwo has 4 string offset entries (indices 0-3). The dead subprogram
+# uses index 2 ("dead_func") and the live subprogram uses index 3
+# ("live_func"). After GC the dead entry at index 2 is removed, so the live
+# subprogram's string must be remapped from index 3 to index 2.  Without the
+# DW_FORM_GNU_str_index fix, the old index is copied verbatim and the string
+# resolves incorrectly.
+
+# Assemble the .dwo (split debug info).
+# RUN: llvm-mc --triple=x86_64-unknown-linux --filetype=obj %p/inputs/gc-compact-gnu-str-index-dwo.s -o %t.dwo
+
+# Generate the exec assembly with the .dwo path substituted in, then assemble it.
+# RUN: sed 's|DWO_PATH|%t.dwo|' %p/inputs/gc-compact-gnu-str-index-exec.s | llvm-mc --triple=x86_64-unknown-linux --filetype=obj -o %t.o
+
+# Baseline: all subprograms and strings survive when not garbage-collecting.
+# RUN: thorin -e %t.o -o %t-nogc.dwp
+# RUN: llvm-dwarfdump -debug-info -debug-str-offsets -debug-str %t-nogc.dwp | FileCheck --check-prefixes=CHECK,NOGC %s
+
+# With garbage collection the dead subprogram is removed and str_offsets compacted.
+# RUN: thorin --gc -e %t.o -o %t-gc.dwp
+# RUN: llvm-dwarfdump -debug-info -debug-str-offsets -debug-str %t-gc.dwp | FileCheck --check-prefixes=CHECK,GC %s
+
+# -- debug_info checks --
+# CHECK:        .debug_info.dwo contents:
+# CHECK:        DW_TAG_compile_unit
+# CHECK:          DW_AT_name	("test.c")
+# CHECK:          DW_AT_comp_dir	("/home")
+
+# Without GC, the dead subprogram comes first.
+# NOGC:          DW_TAG_subprogram
+# NOGC:            DW_AT_name	("dead_func")
+
+# The live subprogram must always be present with correct name.
+# CHECK:          DW_TAG_subprogram
+# CHECK:            DW_AT_name	("live_func")
+
+# After GC there must be no further subprograms.
+# GC:          NULL
+# GC-NOT:      DW_TAG_subprogram
+
+# -- debug_str checks (appears before str_offsets in output) --
+# After GC the dead string should not be in .debug_str.dwo.
+# GC:          .debug_str.dwo contents:
+# GC:          "test.c"
+# GC:          "live_func"
+# GC-NOT:      "dead_func"
+
+# -- str_offsets checks --
+# CHECK:        .debug_str_offsets.dwo contents:
+
+# Without GC the table has 4 entries (16 bytes, no header in DWARF4).
+# NOGC:        Contribution size = 16
+# NOGC:        "test.c"
+# NOGC:        "/home"
+# NOGC:        "dead_func"
+# NOGC:        "live_func"
+
+# After GC the table has 3 entries (12 bytes, no header in DWARF4).
+# GC:          Contribution size = 12
+# GC:          "test.c"
+# GC:          "/home"
+# GC:          "live_func"
+# GC-NOT:      "dead_func"

tests/gc-compact-loclists.test

@@ -0,0 +1,74 @@
+# Test that GC compacts the .debug_loclists.dwo offset table by removing
+# entries for location lists only referenced by dead subprograms, and that
+# surviving DIEs' loclistx attributes are correctly renumbered.
+#
+# The .dwo has 3 location lists (loclistx 0, 1, 2). The live subprogram's
+# variable uses loclistx 1. After GC the dead entries at indices 0 and 2
+# are removed, so the surviving variable's loclistx must be renumbered
+# from 1 to 0. Without the fix, the old index is copied verbatim and the
+# location list resolves incorrectly.
+
+# Assemble the .dwo (split debug info).
+# RUN: llvm-mc --triple=x86_64-unknown-linux --filetype=obj %p/inputs/gc-compact-loclists-dwo.s -o %t.dwo
+
+# Generate the exec assembly with the .dwo path substituted in, then assemble it.
+# RUN: sed 's|DWO_PATH|%t.dwo|' %p/inputs/gc-compact-loclists-exec.s | llvm-mc --triple=x86_64-unknown-linux --filetype=obj -o %t.o
+
+# Baseline: all subprograms and location lists survive when not garbage-collecting.
+# RUN: thorin -e %t.o -o %t-nogc.dwp
+# RUN: llvm-dwarfdump -debug-info -debug-loclists %t-nogc.dwp | FileCheck --check-prefixes=CHECK,NOGC %s
+
+# With garbage collection the dead subprograms are removed and loclists compacted.
+# RUN: thorin --gc -e %t.o -o %t-gc.dwp
+# RUN: llvm-dwarfdump -debug-info -debug-loclists %t-gc.dwp | FileCheck --check-prefixes=CHECK,GC %s
+
+# -- debug_info checks --
+# CHECK:        .debug_info.dwo contents:
+# CHECK:        DW_TAG_compile_unit
+
+# Without GC, dead_func1 comes first.
+# NOGC:          DW_TAG_subprogram
+# NOGC:            DW_AT_name	("dead_func1")
+
+# The live subprogram must always be present.
+# CHECK:          DW_TAG_subprogram
+# CHECK:            DW_AT_name	("live_func")
+
+# Its variable with a location list must survive.
+# Without GC the original loclistx index 1 is preserved.
+# NOGC:            DW_TAG_variable
+# NOGC:              DW_AT_location	(indexed (0x1) loclist =
+# NOGC:                 <default>: DW_OP_reg1 RDX)
+
+# After GC the loclistx must be renumbered from 1 to 0.
+# GC:              DW_TAG_variable
+# GC:                DW_AT_location	(indexed (0x0) loclist =
+# GC:                   <default>: DW_OP_reg1 RDX)
+
+# Without GC, dead_func2 is also present.
+# NOGC:          DW_TAG_subprogram
+# NOGC:            DW_AT_name	("dead_func2")
+
+# After GC there must be no further subprograms.
+# GC:          NULL
+# GC-NOT:      DW_TAG_subprogram
+
+# -- loclists checks --
+# CHECK:        .debug_loclists.dwo contents:
+
+# Without GC the offset table has 3 entries.
+# NOGC:        offset_entry_count = 0x00000003
+# NOGC:        offsets: [
+# NOGC:        0x{{[0-9a-f]+}}
+# NOGC-NEXT:   0x{{[0-9a-f]+}}
+# NOGC-NEXT:   0x{{[0-9a-f]+}}
+# NOGC-NEXT:   ]
+
+# After GC the offset table has only 1 entry (compacted).
+# GC:          offset_entry_count = 0x00000001
+# GC:          offsets: [
+# GC:          0x{{[0-9a-f]+}}
+# GC-NEXT:     ]
+
+# The surviving location list data is correct.
+# GC:          <default>: DW_OP_reg1 RDX

tests/gc-compact-rnglists.test

@@ -0,0 +1,69 @@
+# Test that GC compacts the .debug_rnglists.dwo offset table by removing
+# entries for range lists only referenced by dead subprograms, and that
+# surviving DIEs' rnglistx attributes are correctly renumbered.
+#
+# The .dwo has 3 range lists (rnglistx 0, 1, 2). The live subprogram uses
+# rnglistx 1. After GC the dead entries at indices 0 and 2 are removed,
+# so the surviving subprogram's rnglistx must be renumbered from 1 to 0.
+# Without the fix, the old index is copied verbatim and the range list
+# resolves incorrectly.
+
+# Assemble the .dwo (split debug info).
+# RUN: llvm-mc --triple=x86_64-unknown-linux --filetype=obj %p/inputs/gc-compact-rnglists-dwo.s -o %t.dwo
+
+# Generate the exec assembly with the .dwo path substituted in, then assemble it.
+# RUN: sed 's|DWO_PATH|%t.dwo|' %p/inputs/gc-compact-rnglists-exec.s | llvm-mc --triple=x86_64-unknown-linux --filetype=obj -o %t.o
+
+# Baseline: all subprograms and range lists survive when not garbage-collecting.
+# RUN: thorin -e %t.o -o %t-nogc.dwp
+# RUN: llvm-dwarfdump -debug-info -debug-rnglists %t-nogc.dwp | FileCheck --check-prefixes=CHECK,NOGC %s
+
+# With garbage collection the dead subprograms are removed and rnglists compacted.
+# RUN: thorin --gc -e %t.o -o %t-gc.dwp
+# RUN: llvm-dwarfdump -debug-info -debug-rnglists %t-gc.dwp | FileCheck --check-prefixes=CHECK,GC %s
+
+# -- debug_info checks --
+# CHECK:        .debug_info.dwo contents:
+# CHECK:        DW_TAG_compile_unit
+
+# Without GC, dead_func1 comes first.
+# NOGC:          DW_TAG_subprogram
+# NOGC:            DW_AT_name	("dead_func1")
+
+# The live subprogram must always be present.
+# CHECK:          DW_TAG_subprogram
+
+# Without GC the original rnglistx index 1 is preserved.
+# NOGC:            DW_AT_ranges	(indexed (0x1) rangelist =
+# After GC the rnglistx must be renumbered from 1 to 0.
+# GC:              DW_AT_ranges	(indexed (0x0) rangelist =
+
+# CHECK:            DW_AT_name	("live_func")
+
+# Without GC, dead_func2 is also present.
+# NOGC:          DW_TAG_subprogram
+# NOGC:            DW_AT_name	("dead_func2")
+
+# After GC there must be no further subprograms.
+# GC:          NULL
+# GC-NOT:      DW_TAG_subprogram
+
+# -- rnglists checks --
+# CHECK:        .debug_rnglists.dwo contents:
+
+# Without GC the offset table has 3 entries.
+# NOGC:        offset_entry_count = 0x00000003
+# NOGC:        offsets: [
+# NOGC:        0x{{[0-9a-f]+}}
+# NOGC-NEXT:   0x{{[0-9a-f]+}}
+# NOGC-NEXT:   0x{{[0-9a-f]+}}
+# NOGC-NEXT:   ]
+
+# After GC the offset table has only 1 entry (compacted).
+# GC:          offset_entry_count = 0x00000001
+# GC:          offsets: [
+# GC:          0x{{[0-9a-f]+}}
+# GC-NEXT:     ]
+
+# The surviving range list data for live_func is correct.
+# GC:          [0x{{[0-9a-f]+}}, 0x{{[0-9a-f]+}})