diff --git a/linux-user/i386/signal.c b/linux-user/i386/signal.c
index 857a126699..1917fc0a20 100644
--- a/linux-user/i386/signal.c
+++ b/linux-user/i386/signal.c
@@ -837,6 +837,9 @@ restore_sigcontext(CPUX86State *env, struct target_sigcontext *sc)
         unlock_user_struct(fpstate, fpstate_addr, 0);
 #ifdef CONFIG_LATX
         load_xmm_from_env(env);
+        if (option_softfpu) {
+            cpu_x86_set_host_rounding_from_mxcsr(env);
+        }
 #endif
     } else {
         err = 0;
diff --git a/target/i386/cpu.h b/target/i386/cpu.h
index 7a3f8aeb89..444cdd80de 100644
--- a/target/i386/cpu.h
+++ b/target/i386/cpu.h
@@ -1934,6 +1934,9 @@ void cpu_x86_fxsave(CPUX86State *s, target_ulong ptr);
 void cpu_x86_fxrstor(CPUX86State *s, target_ulong ptr);
 void cpu_x86_xsave(CPUX86State *s, target_ulong ptr);
 void cpu_x86_xrstor(CPUX86State *s, target_ulong ptr);
+#ifdef CONFIG_LATX
+void cpu_x86_set_host_rounding_from_mxcsr(CPUX86State *s);
+#endif
 
 /* you can call this signal handler from your SIGBUS and SIGSEGV
    signal handlers to inform the virtual CPU of exceptions. non zero
diff --git a/target/i386/latx/include/translate.h b/target/i386/latx/include/translate.h
index 38200bc429..ff709c4e5b 100644
--- a/target/i386/latx/include/translate.h
+++ b/target/i386/latx/include/translate.h
@@ -1728,6 +1728,8 @@ void tr_lat_spin_unlock(IR2_OPND lat_lock_addr);
 
 void gen_softfpu_helper_prologue(IR1_INST *pir1);
 void gen_softfpu_helper_epilogue(IR1_INST *pir1);
+void gen_softfpu_x87_fcsr_enter(void);
+void gen_softfpu_x87_fcsr_exit(void);
 void update_fcsr_rm(IR2_OPND control_word, IR2_OPND fcsr);
 
 bool ir1_need_reserve_h128(IR1_INST *ir1);
diff --git a/target/i386/latx/translator/tr-fctrl.c b/target/i386/latx/translator/tr-fctrl.c
index 36773c5f27..d017a2c5ba 100644
--- a/target/i386/latx/translator/tr-fctrl.c
+++ b/target/i386/latx/translator/tr-fctrl.c
@@ -111,6 +111,24 @@ void update_fcsr_rm(IR2_OPND control_word, IR2_OPND fcsr)
     ra_free_temp(temp_cw);
 }
 
+void gen_softfpu_x87_fcsr_exit(void)
+{
+    IR2_OPND mxcsr_opnd = ra_alloc_itemp();
+    IR2_OPND cw_opnd = ra_alloc_itemp();
+    IR2_OPND fcsr_opnd = ra_alloc_itemp();
+
+    la_ld_wu(mxcsr_opnd, env_ir2_opnd, lsenv_offset_of_mxcsr(lsenv));
+    la_bstrpick_d(cw_opnd, mxcsr_opnd, 14, 13);
+    la_slli_d(cw_opnd, cw_opnd, 10);
+    la_movfcsr2gr(fcsr_opnd, fcsr_ir2_opnd);
+    update_fcsr_rm(cw_opnd, fcsr_opnd);
+    la_movgr2fcsr(fcsr_ir2_opnd, fcsr_opnd);
+
+    ra_free_temp(mxcsr_opnd);
+    ra_free_temp(cw_opnd);
+    ra_free_temp(fcsr_opnd);
+}
+
 void update_fcsr_by_sw(IR2_OPND sw)
 {
     IR2_OPND old_fcsr = ra_alloc_itemp();
@@ -234,6 +252,9 @@ bool translate_ldmxcsr(IR1_INST *pir1)
 
     tr_gen_call_to_helper1((ADDR)update_mxcsr_status, 1,
                            LOAD_HELPER_UPDATE_MXCSR_STATUS);
+    if (option_softfpu) {
+        gen_softfpu_x87_fcsr_exit();
+    }
 
     return true;
 }
diff --git a/target/i386/latx/translator/tr-softfpu.c b/target/i386/latx/translator/tr-softfpu.c
index 849cb968d0..f63ffbd95c 100644
--- a/target/i386/latx/translator/tr-softfpu.c
+++ b/target/i386/latx/translator/tr-softfpu.c
@@ -338,7 +338,9 @@ static void la_update_fp_status(IR2_OPND cw_opnd)
     IR2_OPND itemp = ra_alloc_itemp();
     IR2_OPND tmp_fcsr = ra_alloc_itemp();
 
-    IR2_OPND label_float64_float80 = ra_alloc_label();
+    IR2_OPND label_float32 = ra_alloc_label();
+    IR2_OPND label_float64 = ra_alloc_label();
+    IR2_OPND label_precision_done = ra_alloc_label();
 
     int fp_status_offset = lsenv_offset_of_fp_status(lsenv);
     int round_mode_offset = fp_status_offset +
@@ -359,24 +361,59 @@ static void la_update_fp_status(IR2_OPND cw_opnd)
      *          x86      env
      *
      * 32       00       10
+     * reserved 01       00
      * 64       10       01
-     * 80       11       10
+     * 80       11       00
      *
      */
 
     /* Precision Control (9, 8)*/
     la_bstrpick_d(itemp, cw_opnd, 9, 8);
-    la_bne(itemp, zero_ir2_opnd, label_float64_float80);
-    li_wu(itemp, 1);
+    la_beq(itemp, zero_ir2_opnd, label_float32);
+    li_wu(tmp_fcsr, 2);
+    la_beq(itemp, tmp_fcsr, label_float64);
+    li_wu(itemp, floatx80_precision_x);
+    la_b(label_precision_done);
 
-    la_label(label_float64_float80);
-    la_xori(itemp, itemp, 3);
+    la_label(label_float32);
+    li_wu(itemp, floatx80_precision_s);
+    la_b(label_precision_done);
+
+    la_label(label_float64);
+    li_wu(itemp, floatx80_precision_d);
+
+    la_label(label_precision_done);
     la_st_b(itemp, env_ir2_opnd, round_precision_offset);
 
     ra_free_temp(itemp);
     ra_free_temp(tmp_fcsr);
 }
 
+static void la_update_fp_status_from_env(void)
+{
+    IR2_OPND cw_opnd = ra_alloc_itemp();
+
+    la_ld_hu(cw_opnd, env_ir2_opnd,
+             lsenv_offset_of_control_word(lsenv));
+    la_update_fp_status(cw_opnd);
+    ra_free_temp(cw_opnd);
+}
+
+void gen_softfpu_x87_fcsr_enter(void)
+{
+    IR2_OPND cw_opnd = ra_alloc_itemp();
+    IR2_OPND fcsr_opnd = ra_alloc_itemp();
+
+    la_ld_hu(cw_opnd, env_ir2_opnd,
+             lsenv_offset_of_control_word(lsenv));
+    la_movfcsr2gr(fcsr_opnd, fcsr_ir2_opnd);
+    update_fcsr_rm(cw_opnd, fcsr_opnd);
+    la_movgr2fcsr(fcsr_ir2_opnd, fcsr_opnd);
+
+    ra_free_temp(cw_opnd);
+    ra_free_temp(fcsr_opnd);
+}
+
 __attribute__((unused))
 void gen_softfpu_helper_prologue(IR1_INST *pir1)
 {
@@ -1924,17 +1961,12 @@ static bool translate_fldcw_softfpu(IR1_INST *pir1)
 {
     IR1_OPND *opnd0 = ir1_get_opnd(pir1, 0);
     IR2_OPND mem_opnd = convert_mem_no_offset(opnd0);
+    IR2_OPND new_cw = ra_alloc_itemp();
 
-    if (option_softfpu == 2) {
-        IR2_OPND new_cw = ra_alloc_itemp();
-        la_ld_hu(new_cw, mem_opnd, 0);
-        la_st_h(new_cw, env_ir2_opnd, lsenv_offset_of_control_word(lsenv));
-        la_update_fp_status(new_cw);
-        ra_free_temp(new_cw);
-
-    } else {
-        gen_softfpu_helper2m_16u((ADDR)helper_fldcw, mem_opnd);
-    }
+    la_ld_hu(new_cw, mem_opnd, 0);
+    la_st_h(new_cw, env_ir2_opnd, lsenv_offset_of_control_word(lsenv));
+    la_update_fp_status(new_cw);
+    ra_free_temp(new_cw);
     return true;
 }
 
@@ -2020,6 +2052,7 @@ static bool translate_fldenv_softfpu(IR1_INST *pir1)
 
     } else {
         gen_softfpu_helper3i((ADDR)helper_fldenv, mem_opnd, data32);
+        la_update_fp_status_from_env();
     }
     return true;
 }
@@ -2325,6 +2358,7 @@ static bool translate_fninit_softfpu(IR1_INST *pir1)
         ra_free_temp(temp);
     } else {
         gen_softfpu_helper1((ADDR)helper_fninit);
+        la_update_fp_status_from_env();
     }
     return true;
 }
@@ -2474,6 +2508,7 @@ static bool translate_fnsave_softfpu(IR1_INST *pir1)
     } else {
         IR2_OPND mem_opnd = convert_mem_no_offset(opnd0);
         gen_softfpu_helper3i((ADDR)helper_fsave, mem_opnd, data32);
+        la_update_fp_status_from_env();
     }
     return true;
 }
@@ -2618,6 +2653,7 @@ static bool translate_frstor_softfpu(IR1_INST *pir1)
     } else {
         IR2_OPND mem_opnd = convert_mem_no_offset(opnd0);
         gen_softfpu_helper3i((ADDR)helper_frstor, mem_opnd, data32);
+        la_update_fp_status_from_env();
 
     }
     return true;
@@ -3131,6 +3167,7 @@ static bool translate_fxrstor_softfpu(IR1_INST *pir1)
     IR2_OPND mem_opnd = convert_mem_no_offset(opnd0);
 
     gen_softfpu_helper2m_ptr((ADDR)helper_fxrstor, mem_opnd);
+    la_update_fp_status_from_env();
     return true;
 }
 
@@ -3227,6 +3264,7 @@ static bool translate_xrstor_softfpu(IR1_INST *pir1)
     la_bstrins_d(temp_rfbm, eax_opnd, 31, 0);
     la_bstrins_d(temp_rfbm, edx_opnd, 63, 32);
     gen_softfpu_helper3_ll((ADDR)helper_xrstor, mem_opnd, temp_rfbm);
+    gen_softfpu_x87_fcsr_exit();
     return true;
 }
 #endif
@@ -3268,7 +3306,7 @@ TRANS_FPU_WRAP_GEN(fisub);
 TRANS_FPU_WRAP_GEN(fisubr);
 TRANS_FPU_WRAP_GEN(fld1);
 TRANS_FPU_WRAP_GEN(fld);
-TRANS_FPU_WRAP_GEN(fldcw);
+TRANS_FPU_WRAP_GEN_NO_PROLOGUE(fldcw);
 TRANS_FPU_WRAP_GEN(fldenv);
 TRANS_FPU_WRAP_GEN(fldl2e);
 TRANS_FPU_WRAP_GEN(fldl2t);
diff --git a/target/i386/latx/translator/translate.c b/target/i386/latx/translator/translate.c
index 38266d7bbc..75a1aa739b 100644
--- a/target/i386/latx/translator/translate.c
+++ b/target/i386/latx/translator/translate.c
@@ -2101,6 +2101,108 @@ typedef enum SoftFPURegionClass {
     SOFTFPU_REGION_TRANSPARENT,
 } SoftFPURegionClass;
 
+static bool is_softfpu_x87_insn(IR1_OPCODE opcode)
+{
+    switch (opcode) {
+    case dt_X86_INS_WAIT:
+    case dt_X86_INS_F2XM1:
+    case dt_X86_INS_FABS:
+    case dt_X86_INS_FADD:
+    case dt_X86_INS_FADDP:
+    case dt_X86_INS_FBLD:
+    case dt_X86_INS_FBSTP:
+    case dt_X86_INS_FCHS:
+    case dt_X86_INS_FCMOVB:
+    case dt_X86_INS_FCMOVBE:
+    case dt_X86_INS_FCMOVE:
+    case dt_X86_INS_FCMOVNB:
+    case dt_X86_INS_FCMOVNBE:
+    case dt_X86_INS_FCMOVNE:
+    case dt_X86_INS_FCMOVNU:
+    case dt_X86_INS_FCMOVU:
+    case dt_X86_INS_FCOM:
+    case dt_X86_INS_FCOMI:
+    case dt_X86_INS_FCOMIP:
+    case dt_X86_INS_FCOMP:
+    case dt_X86_INS_FCOMPP:
+    case dt_X86_INS_FCOS:
+    case dt_X86_INS_FDECSTP:
+    case dt_X86_INS_FDIV:
+    case dt_X86_INS_FDIVP:
+    case dt_X86_INS_FDIVR:
+    case dt_X86_INS_FDIVRP:
+    case dt_X86_INS_FFREE:
+    case dt_X86_INS_FFREEP:
+    case dt_X86_INS_FIADD:
+    case dt_X86_INS_FICOM:
+    case dt_X86_INS_FICOMP:
+    case dt_X86_INS_FIDIV:
+    case dt_X86_INS_FIDIVR:
+    case dt_X86_INS_FILD:
+    case dt_X86_INS_FIMUL:
+    case dt_X86_INS_FINCSTP:
+    case dt_X86_INS_FIST:
+    case dt_X86_INS_FISTP:
+    case dt_X86_INS_FISTTP:
+    case dt_X86_INS_FISUB:
+    case dt_X86_INS_FISUBR:
+    case dt_X86_INS_FLD1:
+    case dt_X86_INS_FLD:
+    case dt_X86_INS_FLDCW:
+    case dt_X86_INS_FLDENV:
+    case dt_X86_INS_FLDL2E:
+    case dt_X86_INS_FLDL2T:
+    case dt_X86_INS_FLDLG2:
+    case dt_X86_INS_FLDLN2:
+    case dt_X86_INS_FLDPI:
+    case dt_X86_INS_FLDZ:
+    case dt_X86_INS_FMUL:
+    case dt_X86_INS_FMULP:
+    case dt_X86_INS_FNCLEX:
+    case dt_X86_INS_FNINIT:
+    case dt_X86_INS_FNOP:
+    case dt_X86_INS_FNSAVE:
+    case dt_X86_INS_FNSTCW:
+    case dt_X86_INS_FNSTENV:
+    case dt_X86_INS_FNSTSW:
+    case dt_X86_INS_FPATAN:
+    case dt_X86_INS_FPREM1:
+    case dt_X86_INS_FPREM:
+    case dt_X86_INS_FPTAN:
+    case dt_X86_INS_FRNDINT:
+    case dt_X86_INS_FRSTOR:
+    case dt_X86_INS_FSCALE:
+    case dt_X86_INS_FSETPM:
+    case dt_X86_INS_FSIN:
+    case dt_X86_INS_FSINCOS:
+    case dt_X86_INS_FSQRT:
+    case dt_X86_INS_FST:
+    case dt_X86_INS_FSTP:
+    case dt_X86_INS_FSUB:
+    case dt_X86_INS_FSUBP:
+    case dt_X86_INS_FSUBR:
+    case dt_X86_INS_FSUBRP:
+    case dt_X86_INS_FTST:
+    case dt_X86_INS_FUCOM:
+    case dt_X86_INS_FUCOMI:
+    case dt_X86_INS_FUCOMIP:
+    case dt_X86_INS_FUCOMP:
+    case dt_X86_INS_FUCOMPP:
+    case dt_X86_INS_FXAM:
+    case dt_X86_INS_FXCH:
+    case dt_X86_INS_FXRSTOR:
+    case dt_X86_INS_FXRSTOR64:
+    case dt_X86_INS_FXSAVE:
+    case dt_X86_INS_FXSAVE64:
+    case dt_X86_INS_FXTRACT:
+    case dt_X86_INS_FYL2X:
+    case dt_X86_INS_FYL2XP1:
+        return true;
+    default:
+        return false;
+    }
+}
+
 static uint32_t softfpu_fast_mask(IR1_OPCODE opcode)
 {
     switch (opcode) {
@@ -2310,6 +2412,7 @@ int tr_ir2_generate(struct TranslationBlock *tb)
     IR1_INST *pir1 = tb_ir1_inst(tb, 0);
 
     bool reduce_proepo = false;
+    bool x87_fcsr_active = false;
     int softfpu_region_end = -1;
 
     lsenv->tr_data->softfpu_region_active = false;
@@ -2419,6 +2522,16 @@ int tr_ir2_generate(struct TranslationBlock *tb)
         }
 #endif
 
+        bool x87_insn = option_softfpu &&
+                         is_softfpu_x87_insn(ir1_opcode(pir1));
+        if (x87_insn && !x87_fcsr_active) {
+            gen_softfpu_x87_fcsr_enter();
+            x87_fcsr_active = true;
+        } else if (!x87_insn && x87_fcsr_active) {
+            gen_softfpu_x87_fcsr_exit();
+            x87_fcsr_active = false;
+        }
+
         if (option_softfpu == 2 && !reduce_proepo &&
             softfpu_region_class(ir1_opcode(pir1)) ==
                 SOFTFPU_REGION_REQUIRED) {
@@ -2455,6 +2568,9 @@ int tr_ir2_generate(struct TranslationBlock *tb)
 
         pir1++;
     }
+    if (x87_fcsr_active) {
+        gen_softfpu_x87_fcsr_exit();
+    }
 #ifdef CONFIG_LATX_DEBUG
     if (option_dump_ir1) {
         pir1 = tb_ir1_inst(tb, 0);
diff --git a/target/i386/tcg/fpu_helper.c b/target/i386/tcg/fpu_helper.c
index 1efc9d70b7..1af410363b 100644
--- a/target/i386/tcg/fpu_helper.c
+++ b/target/i386/tcg/fpu_helper.c
@@ -763,6 +763,34 @@ uint32_t helper_fnstcw(CPUX86State *env)
     return env->fpuc;
 }
 
+static int x86_rounding_mode_to_host(unsigned mode)
+{
+    static int host_round_mode[4] = {
+        FE_TONEAREST,
+        FE_DOWNWARD,
+        FE_UPWARD,
+        FE_TOWARDZERO
+    };
+
+    assert(mode < ARRAY_SIZE(host_round_mode));
+    return host_round_mode[mode];
+}
+
+#ifdef CONFIG_LATX
+static unsigned x86_rounding_mode_to_fcsr(unsigned mode)
+{
+    static const unsigned fcsr_round_mode[4] = {
+        0, /* round to nearest */
+        3, /* round down */
+        2, /* round up */
+        1, /* round toward zero */
+    };
+
+    assert(mode < ARRAY_SIZE(fcsr_round_mode));
+    return fcsr_round_mode[mode];
+}
+#endif
+
 static void set_x86_rounding_mode(unsigned mode, float_status *status)
 {
     static FloatRoundMode x86_round_mode[4] = {
@@ -774,13 +802,7 @@ static void set_x86_rounding_mode(unsigned mode, float_status *status)
     assert(mode < ARRAY_SIZE(x86_round_mode));
     set_float_rounding_mode(x86_round_mode[mode], status);
     if (option_set_rounding_opt) {
-        static int round_mode_enum[4] = {
-            FE_TONEAREST,
-            FE_DOWNWARD,
-            FE_UPWARD,
-            FE_TOWARDZERO
-        };
-        fesetround(round_mode_enum[mode]);
+        fesetround(x86_rounding_mode_to_host(mode));
     }
 }
 
@@ -2856,7 +2878,7 @@ void helper_fxsave(CPUX86State *env, target_ulong ptr)
     do_xsave_fpu(env, ptr, ra);
 
     if (env->cr[4] & CR4_OSFXSR_MASK) {
-        cpu_stl_data_ra(env, ptr + XO(legacy.mxcsr_mask), 0x0000ffff, ra);
+        do_xsave_mxcsr(env, ptr, ra);
         /* Fast FXSAVE leaves out the XMM registers */
         if (!(env->efer & MSR_EFER_FFXSR)
             || (env->hflags & HF_CPL_MASK)
@@ -3312,6 +3334,17 @@ void update_mxcsr_status(CPUX86State *env)
     set_flush_to_zero((mxcsr & SSE_FZ) ? 1 : 0, &env->sse_status);
 }
 
+#ifdef CONFIG_LATX
+void cpu_x86_set_host_rounding_from_mxcsr(CPUX86State *env)
+{
+    unsigned mode = (env->mxcsr & SSE_RC_MASK) >> SSE_RC_SHIFT;
+
+    env->fcsr = (env->fcsr & ~(3U << 8)) |
+                (x86_rounding_mode_to_fcsr(mode) << 8);
+    fesetround(x86_rounding_mode_to_host(mode));
+}
+#endif
+
 void update_mxcsr_from_sse_status(CPUX86State *env)
 {
     uint8_t flags = get_float_exception_flags(&env->sse_status);
diff --git a/tests/integration/registrations/process/meson.build b/tests/integration/registrations/process/meson.build
index bf2f8406bc..5fee4cf1c9 100644
--- a/tests/integration/registrations/process/meson.build
+++ b/tests/integration/registrations/process/meson.build
@@ -1,4 +1,20 @@
 if 'x86_64-linux-user' in target_dirs
+  latx_integration_tests += [{
+    'name': 'test-softfpu-control-rounding-x86_64',
+    'runner': find_program('../../test-softfpu-control-rounding-x86_64.sh'),
+    'args': [
+      emulators['latx-x86_64'],
+      files('../../softfpu-control-rounding-x86_64.S'),
+    ],
+  }]
+  latx_integration_tests += [{
+    'name': 'test-softfpu-signal-rounding-x86_64',
+    'runner': find_program('../../test-softfpu-signal-rounding-x86_64.sh'),
+    'args': [
+      emulators['latx-x86_64'],
+      files('../../softfpu-signal-rounding-x86_64.S'),
+    ],
+  }]
   latx_integration_tests += [{
     'name': 'test-proc-readdir',
     'runner': find_program('../../test-proc-readdir.sh'),
@@ -86,6 +102,48 @@ if 'x86_64-linux-user' in target_dirs
 endif
 
 if 'i386-linux-user' in target_dirs
+  latx_integration_tests += [{
+    'name': 'test-softfpu-ldmxcsr-rounding-i386',
+    'runner': find_program('../../test-softfpu-ldmxcsr-rounding-i386.sh'),
+    'args': [
+      emulators['latx-i386'],
+      files('../../softfpu-ldmxcsr-rounding-i386.S'),
+    ],
+  }]
+  latx_integration_tests += [{
+    'name': 'test-softfpu-fxsave-mxcsr-i386',
+    'runner': find_program('../../test-softfpu-fxsave-mxcsr-i386.sh'),
+    'args': [
+      emulators['latx-i386'],
+      files('../../softfpu-fxsave-mxcsr-i386.S'),
+    ],
+  }]
+  latx_integration_tests += [{
+    'name': 'test-softfpu-control-rounding-i386',
+    'runner': find_program('../../test-softfpu-control-rounding-i386.sh'),
+    'args': [
+      emulators['latx-i386'],
+      files('../../softfpu-control-rounding-i386.S'),
+    ],
+  }]
+  if 'CONFIG_LATX_AVX_OPT' in config_host
+    latx_integration_tests += [{
+      'name': 'test-softfpu-xrstor-rounding-i386',
+      'runner': find_program('../../test-softfpu-xrstor-rounding-i386.sh'),
+      'args': [
+        emulators['latx-i386'],
+        files('../../softfpu-xrstor-rounding-i386.S'),
+      ],
+    }]
+  endif
+  latx_integration_tests += [{
+    'name': 'test-softfpu-signal-rounding-i386',
+    'runner': find_program('../../test-softfpu-signal-rounding-i386.sh'),
+    'args': [
+      emulators['latx-i386'],
+      files('../../softfpu-signal-rounding-i386.S'),
+    ],
+  }]
   latx_integration_tests += [{
     'name': 'test-softfpu-eflags-region-i386',
     'runner': find_program('../../test-softfpu-eflags-region-i386.sh'),
diff --git a/tests/integration/softfpu-control-rounding-i386.S b/tests/integration/softfpu-control-rounding-i386.S
new file mode 100644
index 0000000000..8b6dc948f3
--- /dev/null
+++ b/tests/integration/softfpu-control-rounding-i386.S
@@ -0,0 +1,145 @@
+.equ __NR_exit, 1
+
+.macro check_down fail
+    fldl 8(%esp)
+    fcos
+    fstps 4(%esp)
+    cmpl $0x3f7fffff, 4(%esp)
+    jne \fail
+.endm
+
+.macro check_sse_up fail
+    movss 16(%esp), %xmm0
+    addss 20(%esp), %xmm0
+    movss %xmm0, 4(%esp)
+    cmpl $0x3f800001, 4(%esp)
+    jne \fail
+.endm
+
+.macro check_nearest fail
+    fldl 8(%esp)
+    fcos
+    fstps 4(%esp)
+    cmpl $0x3f800000, 4(%esp)
+    jne \fail
+.endm
+
+.section .text
+.global _start
+.type _start, @function
+_start:
+    and $-16, %esp
+    sub $2048, %esp
+
+    movw $0x077f, 0(%esp)
+    movl $0, 8(%esp)
+    movl $0x3d700000, 12(%esp)
+    movl $0x3f800000, 16(%esp)
+    movl $0x33800000, 20(%esp)
+
+    /* Reserved PC=01 follows QEMU's default extended-precision behavior. */
+    movw $0x017f, 128(%esp)
+    fldcw 128(%esp)
+    fld1
+    fldl 8(%esp)
+    faddp
+    fstpl 136(%esp)
+    cmpl $0x00001000, 136(%esp)
+    jne .Lfail_reserved_pc
+    cmpl $0x3ff00000, 140(%esp)
+    jne .Lfail_reserved_pc
+
+    /* FLDCW must update both softfloat state and the host FCSR. */
+    fninit
+    fldcw 0(%esp)
+    check_down .Lfail_fldcw
+
+    /* FNINIT restores round-to-nearest. */
+    fldcw 0(%esp)
+    fninit
+    check_nearest .Lfail_fninit
+
+    /* FLDENV restores the x87 control word. */
+    fldcw 0(%esp)
+    fnstenv 32(%esp)
+    fninit
+    fldenv 32(%esp)
+    check_down .Lfail_fldenv
+
+    /* FNSAVE resets the x87 control word; FRSTOR restores it. */
+    fldcw 0(%esp)
+    fnsave 64(%esp)
+    check_nearest .Lfail_fnsave
+    frstor 64(%esp)
+    check_down .Lfail_frstor
+
+    /* FXRSTOR restores independent x87 and MXCSR rounding controls. */
+    fldcw 0(%esp)
+    movl $0x5f80, 24(%esp)
+    ldmxcsr 24(%esp)
+    stmxcsr 28(%esp)
+    movl 28(%esp), %eax
+    and $0x6000, %eax
+    cmp $0x4000, %eax
+    jne .Lfail_ldmxcsr
+    fxsave 512(%esp)
+    movl 536(%esp), %eax
+    and $0x6000, %eax
+    cmp $0x4000, %eax
+    jne .Lfail_fxsave_mxcsr
+    fninit
+    movl $0x3f80, 24(%esp)
+    ldmxcsr 24(%esp)
+    fxrstor 512(%esp)
+    check_down .Lfail_fxrstor
+    stmxcsr 28(%esp)
+    movl 28(%esp), %eax
+    and $0x6000, %eax
+    cmp $0x4000, %eax
+    jne .Lfail_fxrstor_saved_mxcsr
+    check_sse_up .Lfail_fxrstor_mxcsr
+    check_down .Lfail_fxrstor_x87_again
+
+    xor %ebx, %ebx
+    jmp .Lexit
+.Lfail_fldcw:
+    mov $1, %ebx
+    jmp .Lexit
+.Lfail_fninit:
+    mov $2, %ebx
+    jmp .Lexit
+.Lfail_fldenv:
+    mov $3, %ebx
+    jmp .Lexit
+.Lfail_fnsave:
+    mov $4, %ebx
+    jmp .Lexit
+.Lfail_frstor:
+    mov $5, %ebx
+    jmp .Lexit
+.Lfail_fxrstor:
+    mov $6, %ebx
+    jmp .Lexit
+.Lfail_ldmxcsr:
+    mov $7, %ebx
+    jmp .Lexit
+.Lfail_fxsave_mxcsr:
+    mov $8, %ebx
+    jmp .Lexit
+.Lfail_fxrstor_saved_mxcsr:
+    mov $9, %ebx
+    jmp .Lexit
+.Lfail_fxrstor_mxcsr:
+    mov $10, %ebx
+    jmp .Lexit
+.Lfail_fxrstor_x87_again:
+    mov $11, %ebx
+    jmp .Lexit
+.Lfail_reserved_pc:
+    mov $12, %ebx
+.Lexit:
+    mov $__NR_exit, %eax
+    int $0x80
+.size _start, .-_start
+
+.section .note.GNU-stack,"",@progbits
diff --git a/tests/integration/softfpu-control-rounding-x86_64.S b/tests/integration/softfpu-control-rounding-x86_64.S
new file mode 100644
index 0000000000..6907b6ca49
--- /dev/null
+++ b/tests/integration/softfpu-control-rounding-x86_64.S
@@ -0,0 +1,78 @@
+.equ __NR_exit, 60
+
+.macro check_down fail
+    fldl 8(%rsp)
+    fcos
+    fstps 4(%rsp)
+    cmpl $0x3f7fffff, 4(%rsp)
+    jne \fail
+.endm
+
+.macro check_sse_up fail
+    movss 16(%rsp), %xmm0
+    addss 20(%rsp), %xmm0
+    movss %xmm0, 4(%rsp)
+    cmpl $0x3f800001, 4(%rsp)
+    jne \fail
+.endm
+
+.section .text
+.global _start
+.type _start, @function
+_start:
+    and $-64, %rsp
+    sub $2048, %rsp
+
+    movw $0x077f, 0(%rsp)
+    movl $0, 8(%rsp)
+    movl $0x3d700000, 12(%rsp)
+    movl $0x3f800000, 16(%rsp)
+    movl $0x33800000, 20(%rsp)
+
+    /* Reserved PC=01 follows QEMU's default extended-precision behavior. */
+    movw $0x017f, 128(%rsp)
+    fldcw 128(%rsp)
+    fld1
+    fldl 8(%rsp)
+    faddp
+    fstpl 136(%rsp)
+    cmpl $0x00001000, 136(%rsp)
+    jne .Lfail_reserved_pc
+    cmpl $0x3ff00000, 140(%rsp)
+    jne .Lfail_reserved_pc
+
+    fldcw 0(%rsp)
+    movl $0x5f80, 24(%rsp)
+    ldmxcsr 24(%rsp)
+    fxsave64 512(%rsp)
+    movl 536(%rsp), %eax
+    and $0x6000, %eax
+    cmp $0x4000, %eax
+    jne .Lfail_fxsave_mxcsr
+
+    fninit
+    movl $0x3f80, 24(%rsp)
+    ldmxcsr 24(%rsp)
+    fxrstor64 512(%rsp)
+    check_down .Lfail_fxrstor
+    check_sse_up .Lfail_fxrstor_mxcsr
+
+    xor %edi, %edi
+    jmp .Lexit
+.Lfail_fxsave_mxcsr:
+    mov $1, %edi
+    jmp .Lexit
+.Lfail_fxrstor:
+    mov $2, %edi
+    jmp .Lexit
+.Lfail_fxrstor_mxcsr:
+    mov $3, %edi
+    jmp .Lexit
+.Lfail_reserved_pc:
+    mov $4, %edi
+.Lexit:
+    mov $__NR_exit, %eax
+    syscall
+.size _start, .-_start
+
+.section .note.GNU-stack,"",@progbits
diff --git a/tests/integration/softfpu-fxsave-mxcsr-i386.S b/tests/integration/softfpu-fxsave-mxcsr-i386.S
new file mode 100644
index 0000000000..1680355268
--- /dev/null
+++ b/tests/integration/softfpu-fxsave-mxcsr-i386.S
@@ -0,0 +1,29 @@
+.equ __NR_exit, 1
+
+.section .text
+.global _start
+.type _start, @function
+_start:
+    and $-16, %esp
+    sub $1024, %esp
+
+    movl $0x5f80, 0(%esp)
+    ldmxcsr 0(%esp)
+    movl $0, 536(%esp)
+    fxsave 512(%esp)
+
+    movl 536(%esp), %eax
+    and $0x6000, %eax
+    cmp $0x4000, %eax
+    jne .Lfail
+
+    xor %ebx, %ebx
+    jmp .Lexit
+.Lfail:
+    mov $1, %ebx
+.Lexit:
+    mov $__NR_exit, %eax
+    int $0x80
+.size _start, .-_start
+
+.section .note.GNU-stack,"",@progbits
diff --git a/tests/integration/softfpu-ldmxcsr-rounding-i386.S b/tests/integration/softfpu-ldmxcsr-rounding-i386.S
new file mode 100644
index 0000000000..4db5726309
--- /dev/null
+++ b/tests/integration/softfpu-ldmxcsr-rounding-i386.S
@@ -0,0 +1,30 @@
+.equ __NR_exit, 1
+
+.section .text
+.global _start
+.type _start, @function
+_start:
+    and $-16, %esp
+    sub $32, %esp
+
+    movl $0x3f800000, 0(%esp)
+    movl $0x33800000, 4(%esp)
+    movl $0x5f80, 8(%esp)
+    ldmxcsr 8(%esp)
+
+    movss 0(%esp), %xmm0
+    addss 4(%esp), %xmm0
+    movss %xmm0, 12(%esp)
+    cmpl $0x3f800001, 12(%esp)
+    jne .Lfail
+
+    xor %ebx, %ebx
+    jmp .Lexit
+.Lfail:
+    mov $1, %ebx
+.Lexit:
+    mov $__NR_exit, %eax
+    int $0x80
+.size _start, .-_start
+
+.section .note.GNU-stack,"",@progbits
diff --git a/tests/integration/softfpu-signal-rounding-i386.S b/tests/integration/softfpu-signal-rounding-i386.S
new file mode 100644
index 0000000000..a994652651
--- /dev/null
+++ b/tests/integration/softfpu-signal-rounding-i386.S
@@ -0,0 +1,108 @@
+.equ __NR_exit, 1
+.equ __NR_getpid, 20
+.equ __NR_kill, 37
+.equ __NR_rt_sigreturn, 173
+.equ __NR_rt_sigaction, 174
+
+.equ SIGUSR1, 10
+.equ SA_SIGINFO, 4
+.equ SA_RESTORER, 0x04000000
+
+.macro check_down fail
+    fldl 8(%esp)
+    fcos
+    fstps 4(%esp)
+    cmpl $0x3f7fffff, 4(%esp)
+    jne \fail
+.endm
+
+.macro check_sse_up fail
+    movss 16(%esp), %xmm0
+    addss 20(%esp), %xmm0
+    movss %xmm0, 4(%esp)
+    cmpl $0x3f800001, 4(%esp)
+    jne \fail
+.endm
+
+.section .text
+.global _start
+.type _start, @function
+_start:
+    and $-16, %esp
+    sub $64, %esp
+
+    movw $0x077f, 0(%esp)
+    movl $0, 8(%esp)
+    movl $0x3d700000, 12(%esp)
+    movl $0x3f800000, 16(%esp)
+    movl $0x33800000, 20(%esp)
+    movl $0x5f80, 24(%esp)
+
+    mov $__NR_rt_sigaction, %eax
+    mov $SIGUSR1, %ebx
+    mov $signal_action, %ecx
+    xor %edx, %edx
+    mov $8, %esi
+    int $0x80
+    test %eax, %eax
+    js .Lfail_sigaction
+
+    fldcw 0(%esp)
+    ldmxcsr 24(%esp)
+    mov $__NR_getpid, %eax
+    int $0x80
+    mov %eax, %ebx
+    mov $SIGUSR1, %ecx
+    mov $__NR_kill, %eax
+    int $0x80
+    test %eax, %eax
+    js .Lfail_signal
+
+    check_sse_up .Lfail_mxcsr
+    check_down .Lfail_x87
+    xor %ebx, %ebx
+    jmp .Lexit
+.Lfail_sigaction:
+    mov $1, %ebx
+    jmp .Lexit
+.Lfail_signal:
+    mov $2, %ebx
+    jmp .Lexit
+.Lfail_x87:
+    mov $3, %ebx
+    jmp .Lexit
+.Lfail_mxcsr:
+    mov $4, %ebx
+.Lexit:
+    mov $__NR_exit, %eax
+    int $0x80
+.size _start, .-_start
+
+.type signal_handler, @function
+signal_handler:
+    fldcw signal_nearest_cw
+    ldmxcsr signal_down_mxcsr
+    ret
+.size signal_handler, .-signal_handler
+
+.type signal_restorer, @function
+signal_restorer:
+    mov $__NR_rt_sigreturn, %eax
+    int $0x80
+    ud2
+.size signal_restorer, .-signal_restorer
+
+.section .data
+.align 8
+signal_action:
+    .long signal_handler
+    .long SA_SIGINFO | SA_RESTORER
+    .long signal_restorer
+    .quad 0
+signal_nearest_cw:
+    .short 0x037f
+.align 4
+signal_down_mxcsr:
+    .long 0x3f80
+
+.section .note.GNU-stack,"",@progbits
diff --git a/tests/integration/softfpu-signal-rounding-x86_64.S b/tests/integration/softfpu-signal-rounding-x86_64.S
new file mode 100644
index 0000000000..cd6d9db990
--- /dev/null
+++ b/tests/integration/softfpu-signal-rounding-x86_64.S
@@ -0,0 +1,108 @@
+.equ __NR_rt_sigaction, 13
+.equ __NR_rt_sigreturn, 15
+.equ __NR_getpid, 39
+.equ __NR_kill, 62
+.equ __NR_exit, 60
+
+.equ SIGUSR1, 10
+.equ SA_SIGINFO, 4
+.equ SA_RESTORER, 0x04000000
+
+.macro check_down fail
+    fldl 8(%rsp)
+    fcos
+    fstps 4(%rsp)
+    cmpl $0x3f7fffff, 4(%rsp)
+    jne \fail
+.endm
+
+.macro check_sse_up fail
+    movss 16(%rsp), %xmm0
+    addss 20(%rsp), %xmm0
+    movss %xmm0, 4(%rsp)
+    cmpl $0x3f800001, 4(%rsp)
+    jne \fail
+.endm
+
+.section .text
+.global _start
+.type _start, @function
+_start:
+    and $-16, %rsp
+    sub $64, %rsp
+
+    movw $0x077f, 0(%rsp)
+    movl $0, 8(%rsp)
+    movl $0x3d700000, 12(%rsp)
+    movl $0x3f800000, 16(%rsp)
+    movl $0x33800000, 20(%rsp)
+    movl $0x5f80, 24(%rsp)
+
+    mov $__NR_rt_sigaction, %eax
+    mov $SIGUSR1, %edi
+    lea signal_action(%rip), %rsi
+    xor %edx, %edx
+    mov $8, %r10d
+    syscall
+    test %rax, %rax
+    js .Lfail_sigaction
+
+    fldcw 0(%rsp)
+    ldmxcsr 24(%rsp)
+    mov $__NR_getpid, %eax
+    syscall
+    mov %eax, %edi
+    mov $SIGUSR1, %esi
+    mov $__NR_kill, %eax
+    syscall
+    test %rax, %rax
+    js .Lfail_signal
+
+    check_sse_up .Lfail_mxcsr
+    check_down .Lfail_x87
+    xor %edi, %edi
+    jmp .Lexit
+.Lfail_sigaction:
+    mov $1, %edi
+    jmp .Lexit
+.Lfail_signal:
+    mov $2, %edi
+    jmp .Lexit
+.Lfail_x87:
+    mov $3, %edi
+    jmp .Lexit
+.Lfail_mxcsr:
+    mov $4, %edi
+.Lexit:
+    mov $__NR_exit, %eax
+    syscall
+.size _start, .-_start
+
+.type signal_handler, @function
+signal_handler:
+    fldcw signal_nearest_cw(%rip)
+    ldmxcsr signal_down_mxcsr(%rip)
+    ret
+.size signal_handler, .-signal_handler
+
+.type signal_restorer, @function
+signal_restorer:
+    mov $__NR_rt_sigreturn, %eax
+    syscall
+    ud2
+.size signal_restorer, .-signal_restorer
+
+.section .data
+.align 8
+signal_action:
+    .quad signal_handler
+    .quad SA_SIGINFO | SA_RESTORER
+    .quad signal_restorer
+    .quad 0
+signal_nearest_cw:
+    .short 0x037f
+.align 4
+signal_down_mxcsr:
+    .long 0x3f80
+
+.section .note.GNU-stack,"",@progbits
diff --git a/tests/integration/softfpu-xrstor-rounding-i386.S b/tests/integration/softfpu-xrstor-rounding-i386.S
new file mode 100644
index 0000000000..bca0883f4d
--- /dev/null
+++ b/tests/integration/softfpu-xrstor-rounding-i386.S
@@ -0,0 +1,98 @@
+.equ __NR_exit, 1
+
+.macro check_down fail
+    fldl 8(%esp)
+    fcos
+    fstps 4(%esp)
+    cmpl $0x3f7fffff, 4(%esp)
+    jne \fail
+.endm
+
+.macro check_sse_nearest fail
+    movss 16(%esp), %xmm0
+    addss 20(%esp), %xmm0
+    movss %xmm0, 4(%esp)
+    cmpl $0x3f800000, 4(%esp)
+    jne \fail
+.endm
+
+.macro check_sse_up fail
+    movss 16(%esp), %xmm0
+    addss 20(%esp), %xmm0
+    movss %xmm0, 4(%esp)
+    cmpl $0x3f800001, 4(%esp)
+    jne \fail
+.endm
+
+.section .text
+.global _start
+.type _start, @function
+_start:
+    and $-64, %esp
+    sub $2048, %esp
+
+    movw $0x077f, 0(%esp)
+    movl $0, 8(%esp)
+    movl $0x3d700000, 12(%esp)
+    movl $0x3f800000, 16(%esp)
+    movl $0x33800000, 20(%esp)
+
+    xor %ecx, %ecx
+    xgetbv
+    and $3, %eax
+    cmp $3, %eax
+    jne .Lfail_xcr0
+
+    fldcw 0(%esp)
+    movl $0x5f80, 24(%esp)
+    ldmxcsr 24(%esp)
+    xor %ecx, %ecx
+    xgetbv
+    xsave 512(%esp)
+
+    movl 536(%esp), %ecx
+    and $0x6000, %ecx
+    cmp $0x4000, %ecx
+    jne .Lfail_xsave_mxcsr
+
+    fninit
+    movl $0x1f80, 24(%esp)
+    ldmxcsr 24(%esp)
+    mov $1, %eax
+    xor %edx, %edx
+    xrstor 512(%esp)
+    /* Check SSE before an x87 exit can repair a stale host FCSR. */
+    check_sse_nearest .Lfail_xrstor_mask
+    check_down .Lfail_xrstor_fp
+
+    mov $3, %eax
+    xor %edx, %edx
+    xrstor 512(%esp)
+    check_sse_up .Lfail_xrstor_both_mxcsr
+    check_down .Lfail_xrstor_both_x87
+
+    xor %ebx, %ebx
+    jmp .Lexit
+.Lfail_xcr0:
+    mov $1, %ebx
+    jmp .Lexit
+.Lfail_xsave_mxcsr:
+    mov $2, %ebx
+    jmp .Lexit
+.Lfail_xrstor_fp:
+    mov $3, %ebx
+    jmp .Lexit
+.Lfail_xrstor_mask:
+    mov $4, %ebx
+    jmp .Lexit
+.Lfail_xrstor_both_x87:
+    mov $5, %ebx
+    jmp .Lexit
+.Lfail_xrstor_both_mxcsr:
+    mov $6, %ebx
+.Lexit:
+    mov $__NR_exit, %eax
+    int $0x80
+.size _start, .-_start
+
+.section .note.GNU-stack,"",@progbits
diff --git a/tests/integration/test-softfpu-control-rounding-i386.sh b/tests/integration/test-softfpu-control-rounding-i386.sh
new file mode 100755
index 0000000000..75dec8bb6f
--- /dev/null
+++ b/tests/integration/test-softfpu-control-rounding-i386.sh
@@ -0,0 +1,35 @@
+#!/bin/sh
+set -eu
+
+emulator=$(readlink -f "$1")
+source_file=$(readlink -f "$2")
+workdir=$(mktemp -d)
+trap 'rm -rf "$workdir"' EXIT HUP INT TERM
+
+if command -v clang-19 >/dev/null 2>&1; then
+    clang=clang-19
+elif command -v clang >/dev/null 2>&1; then
+    clang=clang
+else
+    echo "SKIP: clang is required to build the i386 guest"
+    exit 77
+fi
+if ! "$clang" --target=i386-linux-gnu -fuse-ld=lld -nostdlib -static \
+        -Wl,--build-id=none -x assembler /dev/null \
+        -o "$workdir/linker-check" >/dev/null 2>&1; then
+    echo "SKIP: clang with ld.lld is required to build the i386 guest"
+    exit 77
+fi
+
+"$clang" --target=i386-linux-gnu -fuse-ld=lld -nostdlib -static \
+    -Wl,--build-id=none "$source_file" \
+    -o "$workdir/softfpu-control-rounding-i386"
+
+for mode in 1 2; do
+    LATX_AOT=0 LATX_MT=0 LATX_SOFTFPU=$mode LATX_SOFTFPU_FAST=0 \
+        "$emulator" "$workdir/softfpu-control-rounding-i386"
+done
+LATX_AOT=0 LATX_MT=0 LATX_SOFTFPU=2 LATX_SOFTFPU_FAST=-1 \
+    "$emulator" "$workdir/softfpu-control-rounding-i386"
+
+echo "PASS: softfpu restores independent x87 and SSE rounding state"
diff --git a/tests/integration/test-softfpu-control-rounding-x86_64.sh b/tests/integration/test-softfpu-control-rounding-x86_64.sh
new file mode 100755
index 0000000000..173b779306
--- /dev/null
+++ b/tests/integration/test-softfpu-control-rounding-x86_64.sh
@@ -0,0 +1,35 @@
+#!/bin/sh
+set -eu
+
+emulator=$(readlink -f "$1")
+source_file=$(readlink -f "$2")
+workdir=$(mktemp -d)
+trap 'rm -rf "$workdir"' EXIT HUP INT TERM
+
+if command -v clang-19 >/dev/null 2>&1; then
+    clang=clang-19
+elif command -v clang >/dev/null 2>&1; then
+    clang=clang
+else
+    echo "SKIP: clang is required to build the x86_64 guest"
+    exit 77
+fi
+if ! "$clang" --target=x86_64-linux-gnu -fuse-ld=lld -nostdlib -static \
+        -Wl,--build-id=none -x assembler /dev/null \
+        -o "$workdir/linker-check" >/dev/null 2>&1; then
+    echo "SKIP: clang with ld.lld is required to build the x86_64 guest"
+    exit 77
+fi
+
+"$clang" --target=x86_64-linux-gnu -fuse-ld=lld -nostdlib -static \
+    -Wl,--build-id=none "$source_file" \
+    -o "$workdir/softfpu-control-rounding-x86_64"
+
+for mode in 1 2; do
+    LATX_AOT=0 LATX_MT=0 LATX_SOFTFPU=$mode LATX_SOFTFPU_FAST=0 \
+        "$emulator" "$workdir/softfpu-control-rounding-x86_64"
+done
+LATX_AOT=0 LATX_MT=0 LATX_SOFTFPU=2 LATX_SOFTFPU_FAST=-1 \
+    "$emulator" "$workdir/softfpu-control-rounding-x86_64"
+
+echo "PASS: x86_64 softfpu restores independent rounding state"
diff --git a/tests/integration/test-softfpu-eflags-region-i386.sh b/tests/integration/test-softfpu-eflags-region-i386.sh
index 534f279d1d..25de724a75 100755
--- a/tests/integration/test-softfpu-eflags-region-i386.sh
+++ b/tests/integration/test-softfpu-eflags-region-i386.sh
@@ -14,6 +14,12 @@ else
     echo "SKIP: clang is required to build the i386 guest"
     exit 77
 fi
+if ! "$clang" --target=i386-linux-gnu -fuse-ld=lld -nostdlib -static \
+        -Wl,--build-id=none -x assembler /dev/null \
+        -o "$workdir/linker-check" >/dev/null 2>&1; then
+    echo "SKIP: clang with ld.lld is required to build the i386 guest"
+    exit 77
+fi
 
 "$clang" --target=i386-linux-gnu -fuse-ld=lld -nostdlib -static \
     -Wl,--build-id=none "$source_file" \
diff --git a/tests/integration/test-softfpu-fxsave-mxcsr-i386.sh b/tests/integration/test-softfpu-fxsave-mxcsr-i386.sh
new file mode 100755
index 0000000000..44f4c8f0bf
--- /dev/null
+++ b/tests/integration/test-softfpu-fxsave-mxcsr-i386.sh
@@ -0,0 +1,33 @@
+#!/bin/sh
+set -eu
+
+emulator=$(readlink -f "$1")
+source_file=$(readlink -f "$2")
+workdir=$(mktemp -d)
+trap 'rm -rf "$workdir"' EXIT HUP INT TERM
+
+if command -v clang-19 >/dev/null 2>&1; then
+    clang=clang-19
+elif command -v clang >/dev/null 2>&1; then
+    clang=clang
+else
+    echo "SKIP: clang is required to build the i386 guest"
+    exit 77
+fi
+if ! "$clang" --target=i386-linux-gnu -fuse-ld=lld -nostdlib -static \
+        -Wl,--build-id=none -x assembler /dev/null \
+        -o "$workdir/linker-check" >/dev/null 2>&1; then
+    echo "SKIP: clang with ld.lld is required to build the i386 guest"
+    exit 77
+fi
+
+"$clang" --target=i386-linux-gnu -fuse-ld=lld -nostdlib -static \
+    -Wl,--build-id=none "$source_file" \
+    -o "$workdir/softfpu-fxsave-mxcsr-i386"
+
+for mode in 1 2; do
+    LATX_AOT=0 LATX_MT=0 LATX_SOFTFPU=$mode LATX_SOFTFPU_FAST=0 \
+        "$emulator" "$workdir/softfpu-fxsave-mxcsr-i386"
+done
+
+echo "PASS: softfpu FXSAVE records MXCSR"
diff --git a/tests/integration/test-softfpu-helper-store-gpr-x86_64.sh b/tests/integration/test-softfpu-helper-store-gpr-x86_64.sh
index 1a5b0872d0..6052b89401 100755
--- a/tests/integration/test-softfpu-helper-store-gpr-x86_64.sh
+++ b/tests/integration/test-softfpu-helper-store-gpr-x86_64.sh
@@ -14,6 +14,12 @@ else
     echo "SKIP: clang is required to build the x86_64 guest"
     exit 77
 fi
+if ! "$clang" --target=x86_64-linux-gnu -fuse-ld=lld -nostdlib -static \
+        -Wl,--build-id=none -x assembler /dev/null \
+        -o "$workdir/linker-check" >/dev/null 2>&1; then
+    echo "SKIP: clang with ld.lld is required to build the x86_64 guest"
+    exit 77
+fi
 
 "$clang" --target=x86_64-linux-gnu -fuse-ld=lld -nostdlib -static \
     -Wl,--build-id=none "$source_file" \
diff --git a/tests/integration/test-softfpu-ldmxcsr-rounding-i386.sh b/tests/integration/test-softfpu-ldmxcsr-rounding-i386.sh
new file mode 100755
index 0000000000..fc731811af
--- /dev/null
+++ b/tests/integration/test-softfpu-ldmxcsr-rounding-i386.sh
@@ -0,0 +1,33 @@
+#!/bin/sh
+set -eu
+
+emulator=$(readlink -f "$1")
+source_file=$(readlink -f "$2")
+workdir=$(mktemp -d)
+trap 'rm -rf "$workdir"' EXIT HUP INT TERM
+
+if command -v clang-19 >/dev/null 2>&1; then
+    clang=clang-19
+elif command -v clang >/dev/null 2>&1; then
+    clang=clang
+else
+    echo "SKIP: clang is required to build the i386 guest"
+    exit 77
+fi
+if ! "$clang" --target=i386-linux-gnu -fuse-ld=lld -nostdlib -static \
+        -Wl,--build-id=none -x assembler /dev/null \
+        -o "$workdir/linker-check" >/dev/null 2>&1; then
+    echo "SKIP: clang with ld.lld is required to build the i386 guest"
+    exit 77
+fi
+
+"$clang" --target=i386-linux-gnu -fuse-ld=lld -nostdlib -static \
+    -Wl,--build-id=none "$source_file" \
+    -o "$workdir/softfpu-ldmxcsr-rounding-i386"
+
+for mode in 1 2; do
+    LATX_AOT=0 LATX_MT=0 LATX_SOFTFPU=$mode LATX_SOFTFPU_FAST=0 \
+        "$emulator" "$workdir/softfpu-ldmxcsr-rounding-i386"
+done
+
+echo "PASS: softfpu LDMXCSR updates host rounding"
diff --git a/tests/integration/test-softfpu-region-planner-x86_64.sh b/tests/integration/test-softfpu-region-planner-x86_64.sh
index a6d1b3ac6c..eaa716caf6 100755
--- a/tests/integration/test-softfpu-region-planner-x86_64.sh
+++ b/tests/integration/test-softfpu-region-planner-x86_64.sh
@@ -14,6 +14,12 @@ else
     echo "SKIP: clang is required to build the x86_64 guest"
     exit 77
 fi
+if ! "$clang" --target=x86_64-linux-gnu -fuse-ld=lld -nostdlib -static \
+        -Wl,--build-id=none -x assembler /dev/null \
+        -o "$workdir/linker-check" >/dev/null 2>&1; then
+    echo "SKIP: clang with ld.lld is required to build the x86_64 guest"
+    exit 77
+fi
 
 "$clang" --target=x86_64-linux-gnu -fuse-ld=lld -nostdlib -static \
     -Wl,--build-id=none "$source_file" \
diff --git a/tests/integration/test-softfpu-shared-region-gpr-x86_64.sh b/tests/integration/test-softfpu-shared-region-gpr-x86_64.sh
index 636c43169a..b9d98fecc2 100755
--- a/tests/integration/test-softfpu-shared-region-gpr-x86_64.sh
+++ b/tests/integration/test-softfpu-shared-region-gpr-x86_64.sh
@@ -14,6 +14,12 @@ else
     echo "SKIP: clang is required to build the x86_64 guest"
     exit 77
 fi
+if ! "$clang" --target=x86_64-linux-gnu -fuse-ld=lld -nostdlib -static \
+        -Wl,--build-id=none -x assembler /dev/null \
+        -o "$workdir/linker-check" >/dev/null 2>&1; then
+    echo "SKIP: clang with ld.lld is required to build the x86_64 guest"
+    exit 77
+fi
 
 "$clang" --target=x86_64-linux-gnu -fuse-ld=lld -nostdlib -static \
     -Wl,--build-id=none "$source_file" \
diff --git a/tests/integration/test-softfpu-signal-rounding-i386.sh b/tests/integration/test-softfpu-signal-rounding-i386.sh
new file mode 100755
index 0000000000..fc7cb9f138
--- /dev/null
+++ b/tests/integration/test-softfpu-signal-rounding-i386.sh
@@ -0,0 +1,35 @@
+#!/bin/sh
+set -eu
+
+emulator=$(readlink -f "$1")
+source_file=$(readlink -f "$2")
+workdir=$(mktemp -d)
+trap 'rm -rf "$workdir"' EXIT HUP INT TERM
+
+if command -v clang-19 >/dev/null 2>&1; then
+    clang=clang-19
+elif command -v clang >/dev/null 2>&1; then
+    clang=clang
+else
+    echo "SKIP: clang is required to build the i386 guest"
+    exit 77
+fi
+if ! "$clang" --target=i386-linux-gnu -fuse-ld=lld -nostdlib -static \
+        -Wl,--build-id=none -x assembler /dev/null \
+        -o "$workdir/linker-check" >/dev/null 2>&1; then
+    echo "SKIP: clang with ld.lld is required to build the i386 guest"
+    exit 77
+fi
+
+"$clang" --target=i386-linux-gnu -fuse-ld=lld -nostdlib -static \
+    -Wl,--build-id=none "$source_file" \
+    -o "$workdir/softfpu-signal-rounding-i386"
+
+for mode in 1 2; do
+    LATX_AOT=0 LATX_MT=0 LATX_SOFTFPU=$mode LATX_SOFTFPU_FAST=0 \
+        "$emulator" "$workdir/softfpu-signal-rounding-i386"
+done
+LATX_AOT=0 LATX_MT=0 LATX_SOFTFPU=2 LATX_SOFTFPU_FAST=-1 \
+    "$emulator" "$workdir/softfpu-signal-rounding-i386"
+
+echo "PASS: i386 sigreturn restores softfpu rounding state"
diff --git a/tests/integration/test-softfpu-signal-rounding-x86_64.sh b/tests/integration/test-softfpu-signal-rounding-x86_64.sh
new file mode 100755
index 0000000000..a3fe0310a5
--- /dev/null
+++ b/tests/integration/test-softfpu-signal-rounding-x86_64.sh
@@ -0,0 +1,35 @@
+#!/bin/sh
+set -eu
+
+emulator=$(readlink -f "$1")
+source_file=$(readlink -f "$2")
+workdir=$(mktemp -d)
+trap 'rm -rf "$workdir"' EXIT HUP INT TERM
+
+if command -v clang-19 >/dev/null 2>&1; then
+    clang=clang-19
+elif command -v clang >/dev/null 2>&1; then
+    clang=clang
+else
+    echo "SKIP: clang is required to build the x86_64 guest"
+    exit 77
+fi
+if ! "$clang" --target=x86_64-linux-gnu -fuse-ld=lld -nostdlib -static \
+        -Wl,--build-id=none -x assembler /dev/null \
+        -o "$workdir/linker-check" >/dev/null 2>&1; then
+    echo "SKIP: clang with ld.lld is required to build the x86_64 guest"
+    exit 77
+fi
+
+"$clang" --target=x86_64-linux-gnu -fuse-ld=lld -nostdlib -static \
+    -Wl,--build-id=none "$source_file" \
+    -o "$workdir/softfpu-signal-rounding-x86_64"
+
+for mode in 1 2; do
+    LATX_AOT=0 LATX_MT=0 LATX_SOFTFPU=$mode LATX_SOFTFPU_FAST=0 \
+        "$emulator" "$workdir/softfpu-signal-rounding-x86_64"
+done
+LATX_AOT=0 LATX_MT=0 LATX_SOFTFPU=2 LATX_SOFTFPU_FAST=-1 \
+    "$emulator" "$workdir/softfpu-signal-rounding-x86_64"
+
+echo "PASS: x86_64 sigreturn restores softfpu rounding state"
diff --git a/tests/integration/test-softfpu-xgetbv-gpr-x86_64.sh b/tests/integration/test-softfpu-xgetbv-gpr-x86_64.sh
index a763c05d2d..e8d3da9ce5 100755
--- a/tests/integration/test-softfpu-xgetbv-gpr-x86_64.sh
+++ b/tests/integration/test-softfpu-xgetbv-gpr-x86_64.sh
@@ -14,6 +14,12 @@ else
     echo "SKIP: clang is required to build the x86_64 guest"
     exit 77
 fi
+if ! "$clang" --target=x86_64-linux-gnu -fuse-ld=lld -nostdlib -static \
+        -Wl,--build-id=none -x assembler /dev/null \
+        -o "$workdir/linker-check" >/dev/null 2>&1; then
+    echo "SKIP: clang with ld.lld is required to build the x86_64 guest"
+    exit 77
+fi
 
 "$clang" --target=x86_64-linux-gnu -fuse-ld=lld -nostdlib -static \
     -Wl,--build-id=none "$source_file" \
diff --git a/tests/integration/test-softfpu-xrstor-rounding-i386.sh b/tests/integration/test-softfpu-xrstor-rounding-i386.sh
new file mode 100755
index 0000000000..06d0be445b
--- /dev/null
+++ b/tests/integration/test-softfpu-xrstor-rounding-i386.sh
@@ -0,0 +1,37 @@
+#!/bin/sh
+set -eu
+
+emulator=$(readlink -f "$1")
+source_file=$(readlink -f "$2")
+workdir=$(mktemp -d)
+trap 'rm -rf "$workdir"' EXIT HUP INT TERM
+
+if command -v clang-19 >/dev/null 2>&1; then
+    clang=clang-19
+elif command -v clang >/dev/null 2>&1; then
+    clang=clang
+else
+    echo "SKIP: clang is required to build the i386 guest"
+    exit 77
+fi
+if ! "$clang" --target=i386-linux-gnu -fuse-ld=lld -nostdlib -static \
+        -Wl,--build-id=none -x assembler /dev/null \
+        -o "$workdir/linker-check" >/dev/null 2>&1; then
+    echo "SKIP: clang with ld.lld is required to build the i386 guest"
+    exit 77
+fi
+
+"$clang" --target=i386-linux-gnu -fuse-ld=lld -nostdlib -static \
+    -Wl,--build-id=none "$source_file" \
+    -o "$workdir/softfpu-xrstor-rounding-i386"
+
+for mode in 1 2; do
+    LATX_AOT=0 LATX_AVX_CPUID=1 LATX_MT=0 LATX_SOFTFPU=$mode \
+        LATX_SOFTFPU_FAST=0 \
+        "$emulator" "$workdir/softfpu-xrstor-rounding-i386"
+done
+LATX_AOT=0 LATX_AVX_CPUID=1 LATX_MT=0 LATX_SOFTFPU=2 \
+    LATX_SOFTFPU_FAST=-1 \
+    "$emulator" "$workdir/softfpu-xrstor-rounding-i386"
+
+echo "PASS: softfpu XRSTOR preserves independent rounding controls"