From a53643889dcb5be24b965af8fb0461fac3045344 Mon Sep 17 00:00:00 2001
From: YAMAMOTO Takashi <imuwoto@gmail.com>
Date: Mon, 5 Jun 2023 22:14:44 +0900
Subject: [PATCH] workaround x87 issues

---
 cmake/ToywasmConfig.cmake |  6 ++++++
 lib/endian.c              |  2 ++
 lib/endian.h              | 41 +++++++++++++++++++++++++++++++++++++++
 3 files changed, 49 insertions(+)

diff --git a/cmake/ToywasmConfig.cmake b/cmake/ToywasmConfig.cmake
index e7ed3514..a618f59f 100644
--- a/cmake/ToywasmConfig.cmake
+++ b/cmake/ToywasmConfig.cmake
@@ -61,6 +61,12 @@ endif()
 endif()
 endif()
 
+if(TRIPLET MATCHES "i386")
+# x87 doesn't preserve sNaN as IEEE 754 and wasm expect.
+# unfortunately, clang doesn't have -mno-fp-ret-in-387.
+set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -msse2 -mfpmath=sse")
+endif()
+
 if(CMAKE_C_COMPILER_TARGET MATCHES "wasm")
 if(TOYWASM_USE_TAILCALL)
 set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mtail-call")
diff --git a/lib/endian.c b/lib/endian.c
index 65405fbe..88d5c1ea 100644
--- a/lib/endian.c
+++ b/lib/endian.c
@@ -179,6 +179,7 @@ lef64_encode(void *p, double v)
         le64_encode(p, u.i);
 }
 
+#if !defined(__i386__)
 float
 lef32_decode(const void *p)
 {
@@ -200,3 +201,4 @@ lef64_decode(const void *p)
         u.i = le64_decode(p);
         return u.f;
 }
+#endif
diff --git a/lib/endian.h b/lib/endian.h
index 30f0f362..946290de 100644
--- a/lib/endian.h
+++ b/lib/endian.h
@@ -1,3 +1,5 @@
+#if !defined(_ENDIAN_H)
+#define _ENDIAN_H
 #include <stdint.h>
 
 uint8_t le8_to_host(uint8_t v);
@@ -23,5 +25,44 @@ uint64_t le64_decode(const void *p);
 void lef32_encode(void *p, float v);
 void lef64_encode(void *p, double v);
 
+/*
+ * x87 fld/fstp does not preserve sNaN. it breaks wasm semantics.
+ *
+ * while in later processors we can use XMM registers (eg. -msse2) which
+ * don't have the problem, x87 ST0 register is still used to return
+ * float/double function results as it's specified by the i386 ABI.
+ *
+ * while GCC has -mno-fp-ret-in-387 to alter the abi, Clang unfortunately
+ * doesn't seem to have an equivalent.
+ */
+#if defined(__i386__)
+__attribute__((always_inline, used)) static float
+lef32_decode(const void *p)
+{
+        union {
+                uint32_t i;
+                float f;
+        } u;
+        u.i = le32_decode(p);
+        return u.f;
+}
+#else
 float lef32_decode(const void *p);
+#endif
+
+#if defined(__i386__)
+__attribute__((always_inline, used)) static double
+lef64_decode(const void *p)
+{
+        union {
+                uint64_t i;
+                double f;
+        } u;
+        u.i = le64_decode(p);
+        return u.f;
+}
+#else
 double lef64_decode(const void *p);
+#endif
+
+#endif /* !defined(_ENDIAN_H) */