1 files changed, 67 insertions, 0 deletions
diff --git a/compiler-rt/lib/sanitizer_common/sanitizer_coverage_win_sections.cpp b/compiler-rt/lib/sanitizer_common/sanitizer_coverage_win_sections.cpp
new file mode 100644
index 000000000000..e7d6563393cf
--- /dev/null
+++ b/compiler-rt/lib/sanitizer_common/sanitizer_coverage_win_sections.cpp
@@ -0,0 +1,67 @@
+//===-- sanitizer_coverage_win_sections.cpp -------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This file defines delimiters for Sanitizer Coverage's section. It contains
+// Windows specific tricks to coax the linker into giving us the start and stop
+// addresses of a section, as ELF linkers can do, to get the size of certain
+// arrays. According to https://msdn.microsoft.com/en-us/library/7977wcck.aspx
+// sections with the same name before "$" are sorted alphabetically by the
+// string that comes after "$" and merged into one section. We take advantage
+// of this by putting data we want the size of into the middle (M) of a section,
+// by using the letter "M" after "$". We get the start of this data (ie:
+// __start_section_name) by making the start variable come at the start of the
+// section (using the letter A after "$"). We do the same to get the end of the
+// data by using the letter "Z" after "$" to make the end variable come after
+// the data. Note that because of our technique the address of the start
+// variable is actually the address of data that comes before our middle
+// section. We also need to prevent the linker from adding any padding. Each
+// technique we use for this is explained in the comments below.
+//===----------------------------------------------------------------------===//
+
+#include "sanitizer_platform.h"
+#if SANITIZER_WINDOWS
+#include <stdint.h>
+
+extern "C" {
+// Use uint64_t so the linker won't need to add any padding if it tries to word
+// align the start of the 8-bit counters array. The array will always start 8
+// bytes after __start_sancov_cntrs.
+#pragma section(".SCOV$CA", read, write)
+__declspec(allocate(".SCOV$CA")) uint64_t __start___sancov_cntrs = 0;
+
+// Even though we said not to align __stop__sancov_cntrs (using the "align"
+// declspec), MSVC's linker may try to align the section, .SCOV$CZ, containing
+// it. This can cause a mismatch between the number of PCs and counters since
+// each PCTable element is 8 bytes (unlike counters which are 1 byte) so no
+// padding would be added to align .SCOVP$Z, However, if .SCOV$CZ section is 1
+// byte, the linker won't try to align it on an 8-byte boundary, so use a
+// uint8_t for __stop_sancov_cntrs.
+#pragma section(".SCOV$CZ", read, write)
+__declspec(allocate(".SCOV$CZ")) __declspec(align(1)) uint8_t
+    __stop___sancov_cntrs = 0;
+
+#pragma section(".SCOV$GA", read, write)
+__declspec(allocate(".SCOV$GA")) uint64_t __start___sancov_guards = 0;
+#pragma section(".SCOV$GZ", read, write)
+__declspec(allocate(".SCOV$GZ")) __declspec(align(1)) uint8_t
+    __stop___sancov_guards = 0;
+
+// The guard array and counter array should both be merged into the .data
+// section to reduce the number of PE sections. However, because PCTable is
+// constant it should be merged with the .rdata section.
+#pragma comment(linker, "/MERGE:.SCOV=.data")
+
+#pragma section(".SCOVP$A", read)
+__declspec(allocate(".SCOVP$A")) uint64_t __start___sancov_pcs = 0;
+#pragma section(".SCOVP$Z", read)
+__declspec(allocate(".SCOVP$Z")) __declspec(align(1)) uint8_t
+    __stop___sancov_pcs = 0;
+
+#pragma comment(linker, "/MERGE:.SCOVP=.rdata")
+}
+#endif  // SANITIZER_WINDOWS