Merge pull request #3817 from Sonicadvance1/fix_x87_integer_indefinite

Softfloat: Fixes Integer indefinite return for 16-bit signed values

FEXCore/Source/Common/SoftFloat.h

@@ -452,9 +452,8 @@ struct FEX_PACKED X80SoftFloat {
 
   operator int16_t() const {
     auto rv = extF80_to_i32(*this, softfloat_roundingMode, false);
-    if (rv > INT16_MAX) {
-      return INT16_MAX;
-    } else if (rv < INT16_MIN) {
+    if (rv > INT16_MAX || rv < INT16_MIN) {
+      ///< Indefinite value for 16-bit conversions.
       return INT16_MIN;
     } else {
       return rv;

FEXCore/Source/Interface/Core/Interpreter/Fallbacks/F80Fallbacks.h

@@ -94,9 +94,8 @@ struct OpHandlers<IR::OP_F80CVTINT> {
     LoadDeferredFCW(NewFCW);
     auto rv = extF80_to_i32(src, softfloat_round_minMag, false);
 
-    if (rv > INT16_MAX) {
-      return INT16_MAX;
-    } else if (rv < INT16_MIN) {
+    if (rv > INT16_MAX || rv < INT16_MIN) {
+      ///< Indefinite value for 16-bit conversions.
       return INT16_MIN;
     } else {
       return rv;

unittests/ASM/FEX_bugs/x87_integer_indefinite.asm

@@ -0,0 +1,87 @@
+%ifdef CONFIG
+{
+  "RegData": {
+    "RAX": "0x8000",
+    "RBX": "0x8000",
+    "RCX": "0x80000000",
+    "RDX": "0x80000000",
+    "RSI": "0x8000000000000000",
+    "RDI": "0x8000000000000000"
+  }
+}
+%endif
+
+; FEX-Emu had a bug where x87 float to integer conversions weren't converting to the correct "integer indefinite" value for 16-bit conversions.
+; Test 16-bit, 32-bit, and 64-bit to ensure correct "integer indefinite" results for all.
+; The definition for "integer indefinite" is the smallest negative integer that can be represented.
+; This is regardless of the input value being positive or negative.
+fninit
+
+; 16-bit
+fld qword [rel .double_larger_than_int16]
+fistp word [rel .data_res_pos_16]
+
+fld qword [rel .double_smaller_than_int16]
+fistp word [rel .data_res_neg_16]
+
+; 32-bit
+fld qword [rel .double_larger_than_int32]
+fistp dword [rel .data_res_pos_32]
+
+fld qword [rel .double_smaller_than_int32]
+fistp dword [rel .data_res_neg_32]
+
+; 64-bit
+fld qword [rel .double_larger_than_int64]
+fistp qword [rel .data_res_pos_64]
+
+fld qword [rel .double_smaller_than_int64]
+fistp qword [rel .data_res_neg_64]
+
+; Load the results
+movzx rax, word [rel .data_res_pos_16]
+movzx rbx, word [rel .data_res_neg_16]
+
+mov ecx, dword [rel .data_res_pos_32]
+mov edx, dword [rel .data_res_neg_32]
+
+mov rsi, qword [rel .data_res_pos_64]
+mov rdi, qword [rel .data_res_neg_64]
+
+hlt
+
+; One-integer larger than what int16_t can hold
+.double_larger_than_int16:
+dq 32768.0
+; One-integer smaller than what int16_t can hold
+.double_smaller_than_int16:
+dq -32769.0
+
+; One-integer larger than what int32_t can hold
+.double_larger_than_int32:
+dq 2147483648.0
+; One-integer smaller than what int32_t can hold
+.double_smaller_than_int32:
+dq -2147483649.0
+
+; One-integer larger than what int64_t can hold
+.double_larger_than_int64:
+dq 9223372036854775808.0
+; One-integer smaller than what int64_t can hold
+.double_smaller_than_int64:
+dq -9223372036854775809.0
+
+.data_res_pos_16:
+dw -1
+.data_res_neg_16:
+dw -1
+
+.data_res_pos_32:
+dd -1
+.data_res_neg_32:
+dd -1
+
+.data_res_pos_64:
+dq -1
+.data_res_neg_64:
+dq -1