diff --git a/exe/ops/ops.exe b/exe/ops/ops.exe
index 4743f225d..44632266f 100755
Binary files a/exe/ops/ops.exe and b/exe/ops/ops.exe differ
diff --git a/exe/ops/util.cc b/exe/ops/util.cc
index 770d8c0db..106f6d622 100644
--- a/exe/ops/util.cc
+++ b/exe/ops/util.cc
@@ -27,10 +27,9 @@ void print(uint32_t x) {
 void print(double f) {
   // To print a float, we multiply by 1000 and print as decimal.
 
-  // TODO: retrowin32 doesn't support the fpu op that "<" needs.
-  // bool neg = f < 0;
-  // if (f < 0)
-  //   f = -f;
+  bool neg = f < 0;
+  if (f < 0)
+    f = -f;
   uint32_t x = (uint32_t)(f * 1000.0);
   char buf[64];
   size_t i = sizeof(buf);
@@ -48,9 +47,9 @@ void print(double f) {
       }
     }
   }
-  // if (neg) {
-  //   buf[--i] = '-';
-  // }
+  if (neg) {
+    buf[--i] = '-';
+  }
   print(std::string_view(&buf[i], sizeof(buf) - i));
 }
 
diff --git a/x86/src/ops/fpu.rs b/x86/src/ops/fpu.rs
index 64e73e2bd..5a412b52d 100644
--- a/x86/src/ops/fpu.rs
+++ b/x86/src/ops/fpu.rs
@@ -1,5 +1,5 @@
 use super::helpers::*;
-use crate::{fpu, x86::CPU};
+use crate::{fpu, registers::Flags, x86::CPU};
 use iced_x86::{Instruction, Register};
 use memory::{Extensions, Mem};
 
@@ -382,6 +382,32 @@ pub fn fucomp_st0_sti(cpu: &mut CPU, mem: Mem, instr: &Instruction) {
     // TODO: raise the invalid-arithmetic-operand exception when appropriate.
 }
 
+pub fn fcomi_st0_sti(cpu: &mut CPU, _mem: Mem, instr: &Instruction) {
+    let x = *cpu.fpu.st0();
+    let y = *cpu.fpu.get(instr.op1_register());
+    //cpu.flags.set(Flags::PF, false);
+    if x > y {
+        cpu.flags.set(Flags::ZF, false);
+        cpu.flags.set(Flags::CF, false);
+    } else if x < y {
+        cpu.flags.set(Flags::ZF, false);
+        cpu.flags.set(Flags::CF, true);
+    } else {
+        cpu.flags.set(Flags::ZF, true);
+        cpu.flags.set(Flags::CF, false);
+    }
+}
+
+pub fn fucomi_st0_sti(cpu: &mut CPU, mem: Mem, instr: &Instruction) {
+    fcomi_st0_sti(cpu, mem, instr);
+    // TODO: raise the invalid-arithmetic-operand exception when appropriate.
+}
+
+pub fn fucomip_st0_sti(cpu: &mut CPU, mem: Mem, instr: &Instruction) {
+    fucomi_st0_sti(cpu, mem, instr);
+    cpu.fpu.pop();
+}
+
 pub fn frndint(cpu: &mut CPU, _mem: Mem, _instr: &Instruction) {
     let x = cpu.fpu.st0();
     *x = x.round();
@@ -403,3 +429,10 @@ pub fn fldcw_m2byte(cpu: &mut CPU, mem: Mem, instr: &Instruction) {
     // TODO: control word
     mem.get_pod::<u16>(x86_addr(cpu, instr));
 }
+
+pub fn fcmovnbe_st0_sti(cpu: &mut CPU, _mem: Mem, instr: &Instruction) {
+    if !cpu.flags.contains(Flags::CF) && !cpu.flags.contains(Flags::ZF) {
+        let y = *cpu.fpu.get(instr.op1_register());
+        *cpu.fpu.st0() = y;
+    }
+}
diff --git a/x86/src/ops/table.rs b/x86/src/ops/table.rs
index c51a5ee3a..ed4fd1edb 100644
--- a/x86/src/ops/table.rs
+++ b/x86/src/ops/table.rs
@@ -400,12 +400,19 @@ pub unsafe fn init_op_tab() {
     OP_TAB[iced_x86::Code::Fcomp_m64fp as usize] = Some(ops::fcomp_m64fp);
     OP_TAB[iced_x86::Code::Fcomp_st0_sti as usize] = Some(ops::fcomp_st0_sti);
     OP_TAB[iced_x86::Code::Fucomp_st0_sti as usize] = Some(ops::fucomp_st0_sti);
+    OP_TAB[iced_x86::Code::Fcomi_st0_sti as usize] = Some(ops::fcomi_st0_sti);
+    OP_TAB[iced_x86::Code::Fucomi_st0_sti as usize] = Some(ops::fucomi_st0_sti);
+    OP_TAB[iced_x86::Code::Fucomip_st0_sti as usize] = Some(ops::fucomip_st0_sti);
+
     OP_TAB[iced_x86::Code::Frndint as usize] = Some(ops::frndint);
     OP_TAB[iced_x86::Code::Fnstsw_AX as usize] = Some(ops::fnstsw_ax);
     OP_TAB[iced_x86::Code::Fnstcw_m2byte as usize] = Some(ops::fnstcw_m2byte);
     OP_TAB[iced_x86::Code::Fldcw_m2byte as usize] = Some(ops::fldcw_m2byte);
     OP_TAB[iced_x86::Code::Fclex as usize] = Some(ops::nop);
     OP_TAB[iced_x86::Code::Fnclex as usize] = Some(ops::nop);
+
+    OP_TAB[iced_x86::Code::Fcmovnbe_st0_sti as usize] = Some(ops::fcmovnbe_st0_sti);
+
     OP_TAB[iced_x86::Code::Wait as usize] = Some(ops::nop);
 
     OP_TAB[iced_x86::Code::Pushad as usize] = Some(ops::pushad);