From d959cb66129dceb7e56ce6fbace85c22aebcfe0f Mon Sep 17 00:00:00 2001
From: Thiago Ize <10453105+ThiagoIze@users.noreply.github.com>
Date: Thu, 29 Oct 2026 13:21:18 -0600
Subject: [PATCH] build: Make simd.h and thread.h compile for ARM64EC with
 clang-cl

clang-cl defines _M_AMD64 and __x86_64__ for ARM64EC, so that data
layouts match x64 code, but it provides no x86 intrinsics for it.

- simd.h: the flush-to-zero and denormals functions use the SSE control
  register macros under __x86_64__. They also test !defined(__arm64ec__)
  now, so on ARM64EC they return false and do nothing, as on arm64.
- thread.h: pause() called _mm_pause() under _M_AMD64. It takes the ARM64
  branch (__dmb and __yield) for _M_ARM64EC now.

This completes #5369, which switched simd.h to the NEON code path for
ARM64EC. x86_64 builds are unchanged.

Co-authored-by: Oscar Anson
Assisted-by: Claude Code / Claude Opus 5.5
Signed-off-by: Thiago Ize <10453105+ThiagoIze@users.noreply.github.com>
---
 src/include/OpenImageIO/simd.h   | 8 ++++----
 src/include/OpenImageIO/thread.h | 5 +++--
 2 files changed, 7 insertions(+), 6 deletions(-)

diff --git a/src/include/OpenImageIO/simd.h b/src/include/OpenImageIO/simd.h
index c6853a0e1f..bc843d2c22 100644
--- a/src/include/OpenImageIO/simd.h
+++ b/src/include/OpenImageIO/simd.h
@@ -3152,7 +3152,7 @@ vfloat16 nmsub (const vfloat16& a, const vfloat16& b, const vfloat16& c); // -a*
 // Try to set the flush_zero_mode CPU flag on x86. Return true if we are
 // able, otherwise false (because it's not available on that platform).
 inline bool set_flush_zero_mode (bool on) {
-#if (defined(__x86_64__) || defined(__i386__)) && !defined(__CUDA_ARCH__)
+#if (defined(__x86_64__) || defined(__i386__)) && !defined(__CUDA_ARCH__) && !defined(__arm64ec__)
     _MM_SET_FLUSH_ZERO_MODE (on ? _MM_FLUSH_ZERO_ON : _MM_FLUSH_ZERO_OFF);
     return true;
 #endif
@@ -3162,7 +3162,7 @@ inline bool set_flush_zero_mode (bool on) {
 // Try to set the denorms_zero_mode CPU flag on x86. Return true if we are
 // able, otherwise false (because it's not available on that platform).
 inline bool set_denorms_zero_mode (bool on) {
-#if (defined(__x86_64__) || defined(__i386__)) && !defined(__CUDA_ARCH__)
+#if (defined(__x86_64__) || defined(__i386__)) && !defined(__CUDA_ARCH__) && !defined(__arm64ec__)
     _MM_SET_DENORMALS_ZERO_MODE (on ? _MM_DENORMALS_ZERO_ON : _MM_DENORMALS_ZERO_OFF);
     return true;
 #endif
@@ -3171,7 +3171,7 @@ inline bool set_denorms_zero_mode (bool on) {
 
 // Get the flush_zero_mode CPU flag on x86.
 inline bool get_flush_zero_mode () {
-#if (defined(__x86_64__) || defined(__i386__)) && !defined(__CUDA_ARCH__)
+#if (defined(__x86_64__) || defined(__i386__)) && !defined(__CUDA_ARCH__) && !defined(__arm64ec__)
     return _MM_GET_FLUSH_ZERO_MODE() == _MM_FLUSH_ZERO_ON;
 #endif
     return false;
@@ -3179,7 +3179,7 @@ inline bool get_flush_zero_mode () {
 
 // Get the denorms_zero_mode CPU flag on x86.
 inline bool get_denorms_zero_mode () {
-#if (defined(__x86_64__) || defined(__i386__)) && !defined(__CUDA_ARCH__)
+#if (defined(__x86_64__) || defined(__i386__)) && !defined(__CUDA_ARCH__) && !defined(__arm64ec__)
     return _MM_GET_DENORMALS_ZERO_MODE() == _MM_DENORMALS_ZERO_ON;
 #endif
     return false;
diff --git a/src/include/OpenImageIO/thread.h b/src/include/OpenImageIO/thread.h
index a0881222b9..851782b050 100644
--- a/src/include/OpenImageIO/thread.h
+++ b/src/include/OpenImageIO/thread.h
@@ -154,9 +154,10 @@ pause(int delay) noexcept
 #elif defined(_MSC_VER)
 // A reimplementation of winnt.h YieldProcessor,
 // to avoid including windows headers.
-#    if defined(_M_AMD64)
+#    if defined(_M_AMD64) && !defined(_M_ARM64EC)
         _mm_pause();
-#    elif defined(_M_ARM64) || defined(_M_HYBRID_X86_ARM64)
+#    elif defined(_M_ARM64) || defined(_M_ARM64EC) \
+        || defined(_M_HYBRID_X86_ARM64)
         __dmb(_ARM64_BARRIER_ISHST);
         __yield();
 #    elif defined(_M_ARM)