fix instructions missing on 32-bit

kalenedrael · kalenedrael · commit 532340fc1e0d · 2025-09-30T01:04:28.000-05:00
diff --git a/include/xsimd/arch/xsimd_avx2.hpp b/include/xsimd/arch/xsimd_avx2.hpp
@@ -590,9 +590,15 @@ namespace xsimd
                 // GCC/Clang/MSVC will turn it into the correct load.
                 else XSIMD_IF_CONSTEXPR(sizeof(T) == 4)
                 {
+#if defined(__x86_64__)
                     uint64_t tmp;
                     memcpy(&tmp, mem, sizeof(tmp));
-                    return _mm256_sub_epi32(_mm256_set1_epi8(0), _mm256_cvtepu8_epi32(_mm_cvtsi64_si128(tmp)));
+                    auto val = _mm_cvtsi64_si128(tmp);
+#else
+                    __m128i val;
+                    memcpy(&val, mem, sizeof(uint64_t));
+#endif
+                    return _mm256_sub_epi32(_mm256_set1_epi8(0), _mm256_cvtepu8_epi32(val));
                 }
                 else XSIMD_IF_CONSTEXPR(sizeof(T) == 8)
                 {
@@ -1005,8 +1011,13 @@ namespace xsimd
                 else XSIMD_IF_CONSTEXPR(sizeof(T) == 4)
                 {
                     auto pack_16 = _mm_packs_epi32(b_lo, b_hi);
-                    auto val = _mm_cvtsi128_si64(_mm_sub_epi8(_mm_set1_epi8(0), _mm_packs_epi16(pack_16, pack_16)));
-                    memcpy(mem, &val, sizeof(val));
+                    auto val = _mm_sub_epi8(_mm_set1_epi8(0), _mm_packs_epi16(pack_16, pack_16));
+#if defined(__x86_64__)
+                    auto val_lo = _mm_cvtsi128_si64(val);
+                    memcpy(mem, &val_lo, sizeof(val_lo));
+#else
+                    memcpy(mem, &val, sizeof(uint64_t));
+#endif
                 }
                 else XSIMD_IF_CONSTEXPR(sizeof(T) == 8)
                 {
diff --git a/include/xsimd/arch/xsimd_sse2.hpp b/include/xsimd/arch/xsimd_sse2.hpp
@@ -1714,8 +1714,12 @@ namespace xsimd
                 else XSIMD_IF_CONSTEXPR(sizeof(T) == 2)
                 {
                     auto val = _mm_sub_epi8(_mm_set1_epi8(0), _mm_packs_epi16(b, b));
-                    // store only lower 64 bits
+#if defined(__x86_64__)
+                    auto val_lo = _mm_cvtsi128_si64(val);
+                    memcpy(mem, &val_lo, sizeof(val_lo));
+#else
                     memcpy(mem, &val, sizeof(uint64_t));
+#endif
                 }
                 else XSIMD_IF_CONSTEXPR(sizeof(T) == 4)
                 {
diff --git a/include/xsimd/arch/xsimd_sse4_1.hpp b/include/xsimd/arch/xsimd_sse4_1.hpp
@@ -136,9 +136,15 @@ namespace xsimd
                 // GCC/Clang/MSVC will turn it into the correct load.
                 else XSIMD_IF_CONSTEXPR(sizeof(T) == 2)
                 {
+#if defined(__x86_64__)
                     uint64_t tmp;
                     memcpy(&tmp, mem, sizeof(tmp));
-                    return _mm_sub_epi16(_mm_set1_epi8(0), _mm_cvtepu8_epi16(_mm_cvtsi64_si128(tmp)));
+                    auto val = _mm_cvtsi64_si128(tmp);
+#else
+                    __m128i val;
+                    memcpy(&val, mem, sizeof(uint64_t));
+#endif
+                    return _mm_sub_epi16(_mm_set1_epi8(0), _mm_cvtepu8_epi16(val));
                 }
                 else XSIMD_IF_CONSTEXPR(sizeof(T) == 4)
                 {

Original file line number	Diff line number	Diff line change
`@@ -590,9 +590,15 @@ namespace xsimd`
`590`	`590`	`// GCC/Clang/MSVC will turn it into the correct load.`
`591`	`591`	`else XSIMD_IF_CONSTEXPR(sizeof(T) == 4)`
`592`	`592`	`{`
	`593`	`+#if defined(__x86_64__)`
`593`	`594`	`uint64_t tmp;`
`594`	`595`	`memcpy(&tmp, mem, sizeof(tmp));`
`595`		`- return _mm256_sub_epi32(_mm256_set1_epi8(0), _mm256_cvtepu8_epi32(_mm_cvtsi64_si128(tmp)));`
	`596`	`+ auto val = _mm_cvtsi64_si128(tmp);`
	`597`	`+#else`
	`598`	`+ __m128i val;`
	`599`	`+ memcpy(&val, mem, sizeof(uint64_t));`
	`600`	`+#endif`
	`601`	`+ return _mm256_sub_epi32(_mm256_set1_epi8(0), _mm256_cvtepu8_epi32(val));`
`596`	`602`	`}`
`597`	`603`	`else XSIMD_IF_CONSTEXPR(sizeof(T) == 8)`
`598`	`604`	`{`
`@@ -1005,8 +1011,13 @@ namespace xsimd`
`1005`	`1011`	`else XSIMD_IF_CONSTEXPR(sizeof(T) == 4)`
`1006`	`1012`	`{`
`1007`	`1013`	`auto pack_16 = _mm_packs_epi32(b_lo, b_hi);`
`1008`		`- auto val = _mm_cvtsi128_si64(_mm_sub_epi8(_mm_set1_epi8(0), _mm_packs_epi16(pack_16, pack_16)));`
`1009`		`- memcpy(mem, &val, sizeof(val));`
	`1014`	`+ auto val = _mm_sub_epi8(_mm_set1_epi8(0), _mm_packs_epi16(pack_16, pack_16));`
	`1015`	`+#if defined(__x86_64__)`
	`1016`	`+ auto val_lo = _mm_cvtsi128_si64(val);`
	`1017`	`+ memcpy(mem, &val_lo, sizeof(val_lo));`
	`1018`	`+#else`
	`1019`	`+ memcpy(mem, &val, sizeof(uint64_t));`
	`1020`	`+#endif`
`1010`	`1021`	`}`
`1011`	`1022`	`else XSIMD_IF_CONSTEXPR(sizeof(T) == 8)`
`1012`	`1023`	`{`
Original file line number	Diff line number	Diff line change
`@@ -1714,8 +1714,12 @@ namespace xsimd`
`1714`	`1714`	`else XSIMD_IF_CONSTEXPR(sizeof(T) == 2)`
`1715`	`1715`	`{`
`1716`	`1716`	`auto val = _mm_sub_epi8(_mm_set1_epi8(0), _mm_packs_epi16(b, b));`
`1717`		`- // store only lower 64 bits`
	`1717`	`+#if defined(__x86_64__)`
	`1718`	`+ auto val_lo = _mm_cvtsi128_si64(val);`
	`1719`	`+ memcpy(mem, &val_lo, sizeof(val_lo));`
	`1720`	`+#else`
`1718`	`1721`	`memcpy(mem, &val, sizeof(uint64_t));`
	`1722`	`+#endif`
`1719`	`1723`	`}`
`1720`	`1724`	`else XSIMD_IF_CONSTEXPR(sizeof(T) == 4)`
`1721`	`1725`	`{`
Original file line number	Diff line number	Diff line change
`@@ -136,9 +136,15 @@ namespace xsimd`
`136`	`136`	`// GCC/Clang/MSVC will turn it into the correct load.`
`137`	`137`	`else XSIMD_IF_CONSTEXPR(sizeof(T) == 2)`
`138`	`138`	`{`
	`139`	`+#if defined(__x86_64__)`
`139`	`140`	`uint64_t tmp;`
`140`	`141`	`memcpy(&tmp, mem, sizeof(tmp));`
`141`		`- return _mm_sub_epi16(_mm_set1_epi8(0), _mm_cvtepu8_epi16(_mm_cvtsi64_si128(tmp)));`
	`142`	`+ auto val = _mm_cvtsi64_si128(tmp);`
	`143`	`+#else`
	`144`	`+ __m128i val;`
	`145`	`+ memcpy(&val, mem, sizeof(uint64_t));`
	`146`	`+#endif`
	`147`	`+ return _mm_sub_epi16(_mm_set1_epi8(0), _mm_cvtepu8_epi16(val));`
`142`	`148`	`}`
`143`	`149`	`else XSIMD_IF_CONSTEXPR(sizeof(T) == 4)`
`144`	`150`	`{`