diff --git a/include/simdjson.h b/include/simdjson.h
index 0e2a08e0..63e90c9a 100644
--- a/include/simdjson.h
+++ b/include/simdjson.h
@@ -35,7 +35,6 @@
  */
 
 #include "simdjson/dom.h"
-#include "simdjson/implementations.h"
 #include "simdjson/builtin.h"
 
 #endif // SIMDJSON_H
diff --git a/include/simdjson/arm64.h b/include/simdjson/arm64.h
index 72deb2a6..2f49c963 100644
--- a/include/simdjson/arm64.h
+++ b/include/simdjson/arm64.h
@@ -3,15 +3,6 @@
 
 #include "simdjson/implementation-base.h"
 
-#ifdef SIMDJSON_FALLBACK_H
-#error "arm64.h must be included before fallback.h"
-#endif
-
-#ifndef SIMDJSON_IMPLEMENTATION_ARM64
-#define SIMDJSON_IMPLEMENTATION_ARM64 (SIMDJSON_IS_ARM64)
-#endif
-#define SIMDJSON_CAN_ALWAYS_RUN_ARM64 SIMDJSON_IMPLEMENTATION_ARM64 && SIMDJSON_IS_ARM64
-
 #if SIMDJSON_IMPLEMENTATION_ARM64
 
 namespace simdjson {
@@ -37,11 +28,9 @@ namespace arm64 {
 #include "simdjson/arm64/stringparsing.h"
 #include "simdjson/arm64/numberparsing.h"
 #include "simdjson/generic/implementation_simdjson_result_base.h"
-#include "simdjson/generic/ondemand.h"
 
 // Inline definitions
 #include "simdjson/generic/implementation_simdjson_result_base-inl.h"
-#include "simdjson/generic/ondemand-inl.h"
 #include "simdjson/arm64/end.h"
 
 #endif // SIMDJSON_IMPLEMENTATION_ARM64
diff --git a/include/simdjson/arm64/implementation.h b/include/simdjson/arm64/implementation.h
index 2e498ff4..7214011a 100644
--- a/include/simdjson/arm64/implementation.h
+++ b/include/simdjson/arm64/implementation.h
@@ -1,7 +1,7 @@
 #ifndef SIMDJSON_ARM64_IMPLEMENTATION_H
 #define SIMDJSON_ARM64_IMPLEMENTATION_H
 
-#include "simdjson.h"
+#include "simdjson/base.h"
 #include "simdjson/internal/isadetection.h"
 
 namespace simdjson {
diff --git a/include/simdjson/arm64/intrinsics.h b/include/simdjson/arm64/intrinsics.h
index ceebaf24..6f6ac4e2 100644
--- a/include/simdjson/arm64/intrinsics.h
+++ b/include/simdjson/arm64/intrinsics.h
@@ -1,8 +1,6 @@
 #ifndef SIMDJSON_ARM64_INTRINSICS_H
 #define SIMDJSON_ARM64_INTRINSICS_H
 
-#include "simdjson.h"
-
 // This should be the correct header whether
 // you use visual studio or other compilers.
 #include <arm_neon.h>
diff --git a/include/simdjson/arm64/simd.h b/include/simdjson/arm64/simd.h
index 64a3a7e1..81764cf9 100644
--- a/include/simdjson/arm64/simd.h
+++ b/include/simdjson/arm64/simd.h
@@ -1,7 +1,7 @@
 #ifndef SIMDJSON_ARM64_SIMD_H
 #define SIMDJSON_ARM64_SIMD_H
 
-#include "simdjson.h"
+#include "simdjson/base.h"
 #include "simdjson/internal/simdprune_tables.h"
 #include "simdjson/arm64/bitmanipulation.h"
 #include <type_traits>
diff --git a/include/simdjson/arm64/stringparsing.h b/include/simdjson/arm64/stringparsing.h
index e1c3b4d9..08c65487 100644
--- a/include/simdjson/arm64/stringparsing.h
+++ b/include/simdjson/arm64/stringparsing.h
@@ -1,7 +1,7 @@
 #ifndef SIMDJSON_ARM64_STRINGPARSING_H
 #define SIMDJSON_ARM64_STRINGPARSING_H
 
-#include "simdjson.h"
+#include "simdjson/base.h"
 #include "simdjson/arm64/simd.h"
 #include "simdjson/arm64/bitmanipulation.h"
 
diff --git a/include/simdjson/builtin.h b/include/simdjson/builtin.h
index 730c9939..47e81501 100644
--- a/include/simdjson/builtin.h
+++ b/include/simdjson/builtin.h
@@ -1,7 +1,7 @@
 #ifndef SIMDJSON_BUILTIN_H
 #define SIMDJSON_BUILTIN_H
 
-#include "simdjson/portability.h"
+#include "simdjson/implementation-base.h"
 
 #ifndef SIMDJSON_BUILTIN_IMPLEMENTATION
 #if SIMDJSON_CAN_ALWAYS_RUN_HASWELL
diff --git a/include/simdjson/fallback.h b/include/simdjson/fallback.h
index 0c85b433..182fa945 100644
--- a/include/simdjson/fallback.h
+++ b/include/simdjson/fallback.h
@@ -3,12 +3,6 @@
 
 #include "simdjson/implementation-base.h"
 
-// Default Fallback to on unless a builtin implementation has already been selected.
-#ifndef SIMDJSON_IMPLEMENTATION_FALLBACK
-#define SIMDJSON_IMPLEMENTATION_FALLBACK 1 // (!SIMDJSON_CAN_ALWAYS_RUN_ARM64 && !SIMDJSON_CAN_ALWAYS_RUN_HASWELL && !SIMDJSON_CAN_ALWAYS_RUN_WESTMERE && !SIMDJSON_CAN_ALWAYS_RUN_PPC64)
-#endif
-#define SIMDJSON_CAN_ALWAYS_RUN_FALLBACK SIMDJSON_IMPLEMENTATION_FALLBACK
-
 #if SIMDJSON_IMPLEMENTATION_FALLBACK
 
 namespace simdjson {
diff --git a/include/simdjson/fallback/bitmanipulation.h b/include/simdjson/fallback/bitmanipulation.h
index 59117ef8..c4dae452 100644
--- a/include/simdjson/fallback/bitmanipulation.h
+++ b/include/simdjson/fallback/bitmanipulation.h
@@ -1,7 +1,7 @@
 #ifndef SIMDJSON_FALLBACK_BITMANIPULATION_H
 #define SIMDJSON_FALLBACK_BITMANIPULATION_H
 
-#include "simdjson.h"
+#include "simdjson/base.h"
 #include <limits>
 
 namespace simdjson {
diff --git a/include/simdjson/fallback/stringparsing.h b/include/simdjson/fallback/stringparsing.h
index 97f04fd3..8a0951ef 100644
--- a/include/simdjson/fallback/stringparsing.h
+++ b/include/simdjson/fallback/stringparsing.h
@@ -1,7 +1,7 @@
 #ifndef SIMDJSON_FALLBACK_STRINGPARSING_H
 #define SIMDJSON_FALLBACK_STRINGPARSING_H
 
-#include "simdjson.h"
+#include "simdjson/base.h"
 
 namespace simdjson {
 namespace SIMDJSON_IMPLEMENTATION {
diff --git a/include/simdjson/generic/dom_parser_implementation.h b/include/simdjson/generic/dom_parser_implementation.h
index 25869567..d5ce9a43 100644
--- a/include/simdjson/generic/dom_parser_implementation.h
+++ b/include/simdjson/generic/dom_parser_implementation.h
@@ -1,4 +1,4 @@
-#include "simdjson.h"
+#include "simdjson/base.h"
 #include "simdjson/internal/isadetection.h"
 
 namespace simdjson {
diff --git a/include/simdjson/haswell.h b/include/simdjson/haswell.h
index 0195d423..38095418 100644
--- a/include/simdjson/haswell.h
+++ b/include/simdjson/haswell.h
@@ -3,22 +3,6 @@
 
 #include "simdjson/implementation-base.h"
 
-#ifdef SIMDJSON_WESTMERE_H
-#error "haswell.h must be included before westmere.h"
-#endif
-#ifdef SIMDJSON_FALLBACK_H
-#error "haswell.h must be included before fallback.h"
-#endif
-
-// Default Haswell to on if this is x86-64. Even if we're not compiled for it, it could be selected
-// at runtime.
-#ifndef SIMDJSON_IMPLEMENTATION_HASWELL
-#define SIMDJSON_IMPLEMENTATION_HASWELL (SIMDJSON_IS_X86_64)
-#endif
-// To see why  (__BMI__) && (__PCLMUL__) && (__LZCNT__) are not part of this next line, see
-// https://github.com/simdjson/simdjson/issues/1247
-#define SIMDJSON_CAN_ALWAYS_RUN_HASWELL ((SIMDJSON_IMPLEMENTATION_HASWELL) && (SIMDJSON_IS_X86_64) && (__AVX2__))
-
 #if SIMDJSON_IMPLEMENTATION_HASWELL
 
 #define SIMDJSON_TARGET_HASWELL SIMDJSON_TARGET_REGION("avx2,bmi,pclmul,lzcnt")
diff --git a/include/simdjson/haswell/intrinsics.h b/include/simdjson/haswell/intrinsics.h
index e118c2c7..d1383648 100644
--- a/include/simdjson/haswell/intrinsics.h
+++ b/include/simdjson/haswell/intrinsics.h
@@ -1,7 +1,7 @@
 #ifndef SIMDJSON_HASWELL_INTRINSICS_H
 #define SIMDJSON_HASWELL_INTRINSICS_H
 
-#include "simdjson.h"
+#include "simdjson/base.h"
 
 #ifdef SIMDJSON_VISUAL_STUDIO
 // under clang within visual studio, this will include <x86intrin.h>
diff --git a/include/simdjson/haswell/stringparsing.h b/include/simdjson/haswell/stringparsing.h
index 0141c780..ec56bc59 100644
--- a/include/simdjson/haswell/stringparsing.h
+++ b/include/simdjson/haswell/stringparsing.h
@@ -1,7 +1,7 @@
 #ifndef SIMDJSON_HASWELL_STRINGPARSING_H
 #define SIMDJSON_HASWELL_STRINGPARSING_H
 
-#include "simdjson.h"
+#include "simdjson/base.h"
 #include "simdjson/haswell/simd.h"
 #include "simdjson/haswell/bitmanipulation.h"
 
diff --git a/include/simdjson/implementation-base.h b/include/simdjson/implementation-base.h
index 7b95fa3d..bcdf3ee3 100644
--- a/include/simdjson/implementation-base.h
+++ b/include/simdjson/implementation-base.h
@@ -20,4 +20,35 @@
 #include "simdjson/internal/numberparsing_tables.h"
 #include "simdjson/internal/simdprune_tables.h"
 
+#ifndef SIMDJSON_IMPLEMENTATION_ARM64
+#define SIMDJSON_IMPLEMENTATION_ARM64 (SIMDJSON_IS_ARM64)
+#endif
+#define SIMDJSON_CAN_ALWAYS_RUN_ARM64 SIMDJSON_IMPLEMENTATION_ARM64 && SIMDJSON_IS_ARM64
+
+// Default Haswell to on if this is x86-64. Even if we're not compiled for it, it could be selected
+// at runtime.
+#ifndef SIMDJSON_IMPLEMENTATION_HASWELL
+#define SIMDJSON_IMPLEMENTATION_HASWELL (SIMDJSON_IS_X86_64)
+#endif
+// To see why  (__BMI__) && (__PCLMUL__) && (__LZCNT__) are not part of this next line, see
+// https://github.com/simdjson/simdjson/issues/1247
+#define SIMDJSON_CAN_ALWAYS_RUN_HASWELL ((SIMDJSON_IMPLEMENTATION_HASWELL) && (SIMDJSON_IS_X86_64) && (__AVX2__))
+
+// Default Westmere to on if this is x86-64, unless we'll always select Haswell.
+#ifndef SIMDJSON_IMPLEMENTATION_WESTMERE
+#define SIMDJSON_IMPLEMENTATION_WESTMERE (SIMDJSON_IS_X86_64 && !SIMDJSON_REQUIRES_HASWELL)
+#endif
+#define SIMDJSON_CAN_ALWAYS_RUN_WESTMERE (SIMDJSON_IMPLEMENTATION_WESTMERE && SIMDJSON_IS_X86_64 && __SSE4_2__ && __PCLMUL__)
+
+#ifndef SIMDJSON_IMPLEMENTATION_PPC64
+#define SIMDJSON_IMPLEMENTATION_PPC64 (SIMDJSON_IS_PPC64)
+#endif
+#define SIMDJSON_CAN_ALWAYS_RUN_PPC64 SIMDJSON_IMPLEMENTATION_PPC64 && SIMDJSON_IS_PPC64
+
+// Default Fallback to on unless a builtin implementation has already been selected.
+#ifndef SIMDJSON_IMPLEMENTATION_FALLBACK
+#define SIMDJSON_IMPLEMENTATION_FALLBACK 1 // (!SIMDJSON_CAN_ALWAYS_RUN_ARM64 && !SIMDJSON_CAN_ALWAYS_RUN_HASWELL && !SIMDJSON_CAN_ALWAYS_RUN_WESTMERE && !SIMDJSON_CAN_ALWAYS_RUN_PPC64)
+#endif
+#define SIMDJSON_CAN_ALWAYS_RUN_FALLBACK SIMDJSON_IMPLEMENTATION_FALLBACK
+
 #endif // SIMDJSON_IMPLEMENTATION_BASE_H
\ No newline at end of file
diff --git a/include/simdjson/internal/jsoncharutils_tables.h b/include/simdjson/internal/jsoncharutils_tables.h
index a53551cc..bc412b86 100644
--- a/include/simdjson/internal/jsoncharutils_tables.h
+++ b/include/simdjson/internal/jsoncharutils_tables.h
@@ -1,7 +1,7 @@
 #ifndef SIMDJSON_INTERNAL_JSONCHARUTILS_TABLES_H
 #define SIMDJSON_INTERNAL_JSONCHARUTILS_TABLES_H
 
-#include "simdjson.h"
+#include "simdjson/common_defs.h"
 
 #ifdef JSON_TEST_STRINGS
 void found_string(const uint8_t *buf, const uint8_t *parsed_begin,
diff --git a/include/simdjson/internal/numberparsing_tables.h b/include/simdjson/internal/numberparsing_tables.h
index 9f66b685..1762056f 100644
--- a/include/simdjson/internal/numberparsing_tables.h
+++ b/include/simdjson/internal/numberparsing_tables.h
@@ -1,7 +1,7 @@
 #ifndef SIMDJSON_INTERNAL_NUMBERPARSING_TABLES_H
 #define SIMDJSON_INTERNAL_NUMBERPARSING_TABLES_H
 
-#include "simdjson.h"
+#include "simdjson/base.h"
 
 namespace simdjson {
 namespace internal {
diff --git a/include/simdjson/ppc64.h b/include/simdjson/ppc64.h
index 315869be..c8d76ca3 100644
--- a/include/simdjson/ppc64.h
+++ b/include/simdjson/ppc64.h
@@ -3,15 +3,6 @@
 
 #include "simdjson/implementation-base.h"
 
-#ifdef SIMDJSON_FALLBACK_H
-#error "ppc64.h must be included before fallback.h"
-#endif
-
-#ifndef SIMDJSON_IMPLEMENTATION_PPC64
-#define SIMDJSON_IMPLEMENTATION_PPC64 (SIMDJSON_IS_PPC64)
-#endif
-#define SIMDJSON_CAN_ALWAYS_RUN_PPC64 SIMDJSON_IMPLEMENTATION_PPC64 && SIMDJSON_IS_PPC64
-
 #if SIMDJSON_IMPLEMENTATION_PPC64
 
 namespace simdjson {
diff --git a/include/simdjson/ppc64/implementation.h b/include/simdjson/ppc64/implementation.h
index 0620f3cc..77b0e830 100644
--- a/include/simdjson/ppc64/implementation.h
+++ b/include/simdjson/ppc64/implementation.h
@@ -1,7 +1,7 @@
 #ifndef SIMDJSON_PPC64_IMPLEMENTATION_H
 #define SIMDJSON_PPC64_IMPLEMENTATION_H
 
-#include "simdjson.h"
+#include "simdjson/base.h"
 #include "simdjson/internal/isadetection.h"
 
 namespace simdjson {
diff --git a/include/simdjson/ppc64/intrinsics.h b/include/simdjson/ppc64/intrinsics.h
index ab86ed84..00dc79c1 100644
--- a/include/simdjson/ppc64/intrinsics.h
+++ b/include/simdjson/ppc64/intrinsics.h
@@ -1,7 +1,7 @@
 #ifndef SIMDJSON_PPC64_INTRINSICS_H
 #define SIMDJSON_PPC64_INTRINSICS_H
 
-#include "simdjson.h"
+#include "simdjson/base.h"
 
 // This should be the correct header whether
 // you use visual studio or other compilers.
diff --git a/include/simdjson/ppc64/simd.h b/include/simdjson/ppc64/simd.h
index 6ebecf1e..47a7b9a3 100644
--- a/include/simdjson/ppc64/simd.h
+++ b/include/simdjson/ppc64/simd.h
@@ -1,7 +1,7 @@
 #ifndef SIMDJSON_PPC64_SIMD_H
 #define SIMDJSON_PPC64_SIMD_H
 
-#include "simdjson.h"
+#include "simdjson/base.h"
 #include "simdjson/internal/simdprune_tables.h"
 #include "simdjson/ppc64/bitmanipulation.h"
 #include <type_traits>
diff --git a/include/simdjson/ppc64/stringparsing.h b/include/simdjson/ppc64/stringparsing.h
index ca78df50..6654d3ff 100644
--- a/include/simdjson/ppc64/stringparsing.h
+++ b/include/simdjson/ppc64/stringparsing.h
@@ -1,7 +1,7 @@
 #ifndef SIMDJSON_PPC64_STRINGPARSING_H
 #define SIMDJSON_PPC64_STRINGPARSING_H
 
-#include "simdjson.h"
+#include "simdjson/base.h"
 #include "simdjson/ppc64/bitmanipulation.h"
 #include "simdjson/ppc64/simd.h"
 
diff --git a/include/simdjson/westmere.h b/include/simdjson/westmere.h
index 917031e2..d0a1f25e 100644
--- a/include/simdjson/westmere.h
+++ b/include/simdjson/westmere.h
@@ -3,16 +3,6 @@
 
 #include "simdjson/implementation-base.h"
 
-#ifdef SIMDJSON_FALLBACK_H
-#error "westmere.h must be included before fallback.h"
-#endif
-
-// Default Westmere to on if this is x86-64, unless we'll always select Haswell.
-#ifndef SIMDJSON_IMPLEMENTATION_WESTMERE
-#define SIMDJSON_IMPLEMENTATION_WESTMERE (SIMDJSON_IS_X86_64 && !SIMDJSON_REQUIRES_HASWELL)
-#endif
-#define SIMDJSON_CAN_ALWAYS_RUN_WESTMERE (SIMDJSON_IMPLEMENTATION_WESTMERE && SIMDJSON_IS_X86_64 && __SSE4_2__ && __PCLMUL__)
-
 #if SIMDJSON_IMPLEMENTATION_WESTMERE
 
 #define SIMDJSON_TARGET_WESTMERE SIMDJSON_TARGET_REGION("sse4.2,pclmul")
diff --git a/src/implementation.cpp b/src/implementation.cpp
index 4301b324..7bd68b78 100644
--- a/src/implementation.cpp
+++ b/src/implementation.cpp
@@ -1,4 +1,4 @@
-#include "simdjson.h"
+#include "simdjson/base.h"
 #include <initializer_list>
 
 namespace simdjson {
diff --git a/src/internal/error_tables.cpp b/src/internal/error_tables.cpp
index ce7da7bc..e8792603 100644
--- a/src/internal/error_tables.cpp
+++ b/src/internal/error_tables.cpp
@@ -1,4 +1,4 @@
-#include "simdjson.h"
+#include "simdjson/base.h"
 
 namespace simdjson {
 namespace internal {
diff --git a/src/internal/jsoncharutils_tables.cpp b/src/internal/jsoncharutils_tables.cpp
index 2cbe1199..deae5489 100644
--- a/src/internal/jsoncharutils_tables.cpp
+++ b/src/internal/jsoncharutils_tables.cpp
@@ -1,4 +1,4 @@
-#include "simdjson.h"
+#include "simdjson/base.h"
 
 namespace simdjson {
 namespace internal {
diff --git a/src/internal/numberparsing_tables.cpp b/src/internal/numberparsing_tables.cpp
index 4c98607c..4db818b4 100644
--- a/src/internal/numberparsing_tables.cpp
+++ b/src/internal/numberparsing_tables.cpp
@@ -1,4 +1,4 @@
-#include "simdjson.h"
+#include "simdjson/base.h"
 
 namespace simdjson {
 namespace internal {
diff --git a/src/simdjson.cpp b/src/simdjson.cpp
index 8d17791c..75799633 100644
--- a/src/simdjson.cpp
+++ b/src/simdjson.cpp
@@ -1,3 +1,7 @@
+// Instead of including simdjson.h, we include the stuff inside so we can get all the
+// implementations and not just builtin
+#define SIMDJSON_CPP
+
 #include "simdjson.h"
 
 SIMDJSON_PUSH_DISABLE_WARNINGS