From 04022a515c76d0acf1ee1147fc4de2cfc12d5c52 Mon Sep 17 00:00:00 2001
From: Scott Graham <scott.github@h4ck3r.net>
Date: Mon, 20 Jan 2025 23:42:03 -0800
Subject: [PATCH] Fixes for building with clang-cl on Windows (#104)

* Avoid unused-variable warning on Windows

When compiling with clang-cl:

.../ir_emit.c(275,8): error: unused variable 'handle' [-Werror,-Wunused-variable]
  275 |         void *handle = NULL;
      |               ^~~~~~

* Avoid redefining __ORDER_LITTLE_ENDIAN__ if already defined

clang-cl defines this on Windows, resulting in an error.

.../ir.h(32,10): error: '__ORDER_LITTLE_ENDIAN__' macro redefined [-Werror,-Wmacro-redefined]
   32 | # define __ORDER_LITTLE_ENDIAN__ 1
      |          ^
<built-in>(39,9): note: previous definition is here
   39 | #define __ORDER_LITTLE_ENDIAN__ 1234

* In ir_ntzl, always use the 64 bit version on Windows

When compiling on Windows with clang-cl, __has_builtin(__builtin_ctzl)
will evaluate to true, but a long on Windows is only 32 bits, resulting
in an incorrect implementation. Move the _WIN64 block ahead of the
general block so that it gets the 64 bit version.

Fixes https://github.com/dstogov/ir/issues/101.
---
 ir.h         | 4 +++-
 ir_emit.c    | 2 +-
 ir_private.h | 9 ++++++---
 3 files changed, 10 insertions(+), 5 deletions(-)

diff --git a/ir.h b/ir.h
index 69a01c21..2670fdfa 100644
--- a/ir.h
+++ b/ir.h
@@ -29,7 +29,9 @@ extern "C" {
 # endif
 /* Only supported is little endian for any arch on Windows,
    so just fake the same for all. */
-# define __ORDER_LITTLE_ENDIAN__ 1
+# ifndef __ORDER_LITTLE_ENDIAN__
+#  define __ORDER_LITTLE_ENDIAN__ 1
+# endif
 # define __BYTE_ORDER__ __ORDER_LITTLE_ENDIAN__
 # ifndef __has_builtin
 #  define __has_builtin(arg) (0)
diff --git a/ir_emit.c b/ir_emit.c
index 83fc242a..367dee72 100644
--- a/ir_emit.c
+++ b/ir_emit.c
@@ -272,10 +272,10 @@ static bool ir_is_same_mem_var(const ir_ctx *ctx, ir_ref r1, int32_t offset)
 
 void *ir_resolve_sym_name(const char *name)
 {
-	void *handle = NULL;
 	void *addr;
 
 #ifndef _WIN32
+	void *handle = NULL;
 # ifdef RTLD_DEFAULT
 	handle = RTLD_DEFAULT;
 # endif
diff --git a/ir_private.h b/ir_private.h
index 4d1e8dd3..f980b86b 100644
--- a/ir_private.h
+++ b/ir_private.h
@@ -137,9 +137,10 @@ IR_ALWAYS_INLINE uint32_t ir_ntz(uint32_t num)
 /* Number of trailing zero bits (0x01 -> 0; 0x40 -> 6; 0x00 -> LEN) */
 IR_ALWAYS_INLINE uint32_t ir_ntzl(uint64_t num)
 {
-#if (defined(__GNUC__) || __has_builtin(__builtin_ctzl))
-	return __builtin_ctzl(num);
-#elif defined(_WIN64)
+  // Note that the _WIN64 case should come before __has_builtin() below so that
+  // clang-cl on Windows will use the uint64_t version, not the "long" uint32_t
+  // version.
+#if defined(_WIN64)
 	unsigned long index;
 
 	if (!_BitScanForward64(&index, num)) {
@@ -148,6 +149,8 @@ IR_ALWAYS_INLINE uint32_t ir_ntzl(uint64_t num)
 	}
 
 	return (uint32_t) index;
+#elif (defined(__GNUC__) || __has_builtin(__builtin_ctzl))
+	return __builtin_ctzl(num);
 #else
 	uint32_t n;