Workaround an alignment issue for 32-bit MinGW targets

Long story short, 32-bit Windows only guarantees 8-byte alignment from malloc,
while MinGW generates a default operator new that assumes 16-byte alignment
from malloc. Forcing 32-byte alignment for affected structs will make MinGW use
the aligned operator new, ensuring correct (if excessive) alignment.

common/opthelpers.h

@@ -28,6 +28,24 @@
 #define NOINLINE
 #endif
 
+#if defined(__MINGW32__) && defined(__i386__)
+/* 32-bit MinGW targets have a bug where __STDCPP_DEFAULT_NEW_ALIGNMENT__
+ * reports 16, despite the default operator new calling standard malloc which
+ * only guarantees 8-byte alignment. As a result, structs that need and specify
+ * 16-byte alignment only get 8-byte alignment. Explicitly specifying 32-byte
+ * alignment forces the over-aligned operator new to be called, giving the
+ * correct (if larger than necessary) alignment.
+ *
+ * Technically this bug affects 32-bit GCC more generally, but typically only
+ * with fairly old glibc versions as newer versions do guarantee the 16-byte
+ * alignment as specified. MinGW is reliant on msvcrt.dll's malloc however,
+ * which can't be updated to give that guarantee.
+ */
+#define SIMDALIGN alignas(32)
+#else
+#define SIMDALIGN
+#endif
+
 /* Unlike the likely attribute, ASSUME requires the condition to be true or
  * else it invokes undefined behavior. It's essentially an assert without
  * actually checking the condition at run-time, allowing for stronger

common/phase_shifter.h

@@ -21,7 +21,7 @@
  * signal delay (FilterSize/2) to properly align.
  */
 template<std::size_t FilterSize>
-struct PhaseShifterT {
+struct SIMDALIGN PhaseShifterT {
     static_assert(FilterSize >= 16, "FilterSize needs to be at least 16");
     static_assert((FilterSize&(FilterSize-1)) == 0, "FilterSize needs to be power-of-two");
 

core/bformatdec.h

@@ -13,11 +13,12 @@
 #include "devformat.h"
 #include "filters/splitter.h"
 #include "front_stablizer.h"
+#include "opthelpers.h"
 
 
 using ChannelDec = std::array<float,MaxAmbiChannels>;
 
-class BFormatDec {
+class SIMDALIGN BFormatDec {
     static constexpr size_t sHFBand{0};
     static constexpr size_t sLFBand{1};
     static constexpr size_t sNumBands{2};

core/bsinc_tables.cpp

@@ -14,6 +14,7 @@
 #include "alnumeric.h"
 #include "alspan.h"
 #include "bsinc_defs.h"
+#include "opthelpers.h"
 #include "resampler_limits.h"
 
 
@@ -153,7 +154,7 @@ constexpr BSincHeader bsinc24_hdr{60, 23};
 
 
 template<const BSincHeader &hdr>
-struct BSincFilterArray {
+struct SIMDALIGN BSincFilterArray {
     alignas(16) std::array<float, hdr.total_size> mTable{};
 
     BSincFilterArray()

core/cubic_tables.h

@@ -5,9 +5,10 @@
 #include <cstddef>
 
 #include "cubic_defs.h"
+#include "opthelpers.h"
 
 
-struct CubicTable {
+struct SIMDALIGN CubicTable {
     std::array<CubicCoefficients,CubicPhaseCount> mTable{};
 };
 

core/device.h

@@ -179,7 +179,7 @@ enum class DeviceState : std::uint8_t {
     Playing
 };
 
-struct DeviceBase {
+struct SIMDALIGN DeviceBase {
     std::atomic<bool> Connected{true};
     const DeviceType Type{};
 

core/effects/base.h

@@ -8,6 +8,7 @@
 #include "alspan.h"
 #include "core/bufferline.h"
 #include "intrusive_ptr.h"
+#include "opthelpers.h"
 
 struct BufferStorage;
 struct ContextBase;
@@ -193,7 +194,7 @@ struct EffectTarget {
     RealMixParams *RealOut;
 };
 
-struct EffectState : public al::intrusive_ref<EffectState> {
+struct SIMDALIGN EffectState : public al::intrusive_ref<EffectState> {
     al::span<FloatBufferLine> mOutTarget;
 
 

core/hrtf.h

@@ -11,7 +11,6 @@
 
 #include "almalloc.h"
 #include "alspan.h"
-#include "atomic.h"
 #include "ambidefs.h"
 #include "bufferline.h"
 #include "flexarray.h"

core/mastering.cpp

@@ -19,7 +19,7 @@
 /* These structures assume BufferLineSize is a power of 2. */
 static_assert((BufferLineSize & (BufferLineSize-1)) == 0, "BufferLineSize is not a power of 2");
 
-struct SlidingHold {
+struct SIMDALIGN SlidingHold {
     alignas(16) FloatBufferLine mValues;
     std::array<uint,BufferLineSize> mExpiries;
     uint mLowerIndex;

core/mastering.h

@@ -4,10 +4,10 @@
 #include <array>
 #include <memory>
 
-#include "almalloc.h"
 #include "alnumeric.h"
 #include "alspan.h"
 #include "bufferline.h"
+#include "opthelpers.h"
 #include "vector.h"
 
 struct SlidingHold;
@@ -25,7 +25,7 @@ using uint = unsigned int;
  *
  *   http://c4dm.eecs.qmul.ac.uk/audioengineering/compressors/
  */
-class Compressor {
+class SIMDALIGN Compressor {
     size_t mNumChans{0u};
 
     struct AutoFlags {

core/uhjfilter.h

@@ -7,6 +7,7 @@
 
 #include "alspan.h"
 #include "bufferline.h"
+#include "opthelpers.h"
 
 
 inline constexpr std::size_t UhjLength256{256};
@@ -36,7 +37,7 @@ struct UhjAllPassFilter {
 };
 
 
-struct UhjEncoderBase {
+struct SIMDALIGN UhjEncoderBase {
     UhjEncoderBase() = default;
     UhjEncoderBase(const UhjEncoderBase&) = delete;
     UhjEncoderBase(UhjEncoderBase&&) = delete;
@@ -120,7 +121,7 @@ struct UhjEncoderIIR final : public UhjEncoderBase {
 };
 
 
-struct DecoderBase {
+struct SIMDALIGN DecoderBase {
     static constexpr std::size_t sMaxPadding{256};
 
     /* For 2-channel UHJ, shelf filters should use these LF responses. */

core/voice.h

@@ -10,7 +10,6 @@
 #include <optional>
 #include <string>
 
-#include "almalloc.h"
 #include "alspan.h"
 #include "bufferline.h"
 #include "buffer_storage.h"
@@ -20,6 +19,7 @@
 #include "filters/splitter.h"
 #include "mixer/defs.h"
 #include "mixer/hrtfdefs.h"
+#include "opthelpers.h"
 #include "resampler_limits.h"
 #include "uhjfilter.h"
 #include "vector.h"
@@ -180,7 +180,7 @@ enum : uint {
     VoiceFlagCount
 };
 
-struct Voice {
+struct SIMDALIGN Voice {
     enum State {
         Stopped,
         Playing,