utils,proxy: Add basic SIMD register fetch and mutation support

Signed-off-by: Hector Martin <marcan@marcan.st>

proxyclient/proxy.py

@@ -486,6 +486,8 @@ class M1N1Proxy:
     P_VECTOR = 0x00b
     P_GL1_CALL = 0x00c
     P_GL2_CALL = 0x00d
+    P_GET_SIMD_STATE = 0x00e
+    P_PUT_SIMD_STATE = 0x00f
 
     P_WRITE64 = 0x100
     P_WRITE32 = 0x101
@@ -698,6 +700,10 @@ class M1N1Proxy:
         if len(args) > 4:
             raise ValueError("Too many arguments")
         return self.request(self.P_GL2_CALL, addr, *args)
+    def get_simd_state(self, buf):
+        self.request(self.P_GET_SIMD_STATE, buf)
+    def put_simd_state(self, buf):
+        self.request(self.P_PUT_SIMD_STATE, buf)
 
     def write64(self, addr, data):
         if addr & 7:

proxyclient/proxyutils.py

@@ -7,6 +7,12 @@ from malloc import Heap
 import adt
 from contextlib import contextmanager
 
+SIMD_B = Array(32, Array(16, Int8ul))
+SIMD_H = Array(32, Array(8, Int16ul))
+SIMD_S = Array(32, Array(4, Int32ul))
+SIMD_D = Array(32, Array(2, Int64ul))
+SIMD_Q = Array(32, BytesInteger(16, swapped=True))
+
 class ProxyUtils(object):
     CODE_BUFFER_SIZE = 0x10000
     def __init__(self, p, heap_size=1024 * 1024 * 1024):
@@ -44,6 +50,10 @@ class ProxyUtils(object):
         self.adt_data = None
         self.adt = LazyADT(self)
 
+        self.simd_buf = self.malloc(32 * 16)
+        self.simd_type = None
+        self.simd = None
+
         self.exec_modes = {
             None: (self.proxy.call, REGION_RX_EL1),
             "el2": (self.proxy.call, REGION_RX_EL1),
@@ -208,6 +218,42 @@ class ProxyUtils(object):
         finally:
             self.proxy.mmu_restore(flags)
 
+    def push_simd(self):
+        if self.simd is not None:
+            data = self.simd_type.build(self.simd)
+            self.iface.writemem(self.simd_buf, data)
+            self.proxy.put_simd_state(self.simd_buf)
+            self.simd = self.simd_type = None
+
+    def get_simd(self, simd_type):
+        if self.simd is not None and self.simd_type is not simd_type:
+            data = self.simd_type.build(self.simd)
+            self.simd = simd_type.parse(data)
+            self.simd_type = simd_type
+        elif self.simd is None:
+            self.proxy.get_simd_state(self.simd_buf)
+            data = self.iface.readmem(self.simd_buf, 32 * 16)
+            self.simd = simd_type.parse(data)
+            self.simd_type = simd_type
+
+        return self.simd
+
+    @property
+    def b(self):
+        return self.get_simd(SIMD_B)
+    @property
+    def h(self):
+        return self.get_simd(SIMD_H)
+    @property
+    def s(self):
+        return self.get_simd(SIMD_S)
+    @property
+    def d(self):
+        return self.get_simd(SIMD_D)
+    @property
+    def q(self):
+        return self.get_simd(SIMD_Q)
+
 class LazyADT:
     def __init__(self, utils):
         self.__dict__["_utils"] = utils

proxyclient/shell.py

@@ -31,6 +31,14 @@ class HistoryConsole(code.InteractiveConsole):
         type, value, tb = sys.exc_info()
         traceback.print_exception(type, value, tb)
 
+    def runcode(self, code):
+        super().runcode(code)
+        if "mon" in self.locals:
+            self.locals["mon"].poll()
+        if "u" in self.locals:
+            self.locals["u"].push_simd()
+
+
 class ExitConsole(SystemExit):
     pass
 
@@ -38,11 +46,6 @@ def run_shell(locals, msg=None, exitmsg=None):
     saved_display = sys.displayhook
     try:
         def display(val):
-            try:
-                global mon
-                mon.poll()
-            except NameError:
-                pass
             if isinstance(val, int):
                 builtins._ = val
                 print(hex(val))
@@ -89,7 +92,7 @@ def run_shell(locals, msg=None, exitmsg=None):
 
 if __name__ == "__main__":
     from setup import *
-    locals = __main__.__dict__
+    locals = dict(__main__.__dict__)
 
     from tgtypes import *
 

src/proxy.c

@@ -81,6 +81,10 @@ int proxy_process(ProxyRequest *request, ProxyReply *reply)
             reply->retval = el1_call((void *)request->args[0], request->args[1], request->args[2],
                                      request->args[3], request->args[4]);
             break;
+        case P_VECTOR:
+            next_stage.entry = (generic_func *)request->args[0];
+            memcpy(next_stage.args, &request->args[1], 4 * sizeof(u64));
+            return 1;
         case P_GL1_CALL:
             reply->retval = gl1_call((void *)request->args[0], request->args[1], request->args[2],
                                      request->args[3], request->args[4]);
@@ -89,10 +93,12 @@ int proxy_process(ProxyRequest *request, ProxyReply *reply)
             reply->retval = gl2_call((void *)request->args[0], request->args[1], request->args[2],
                                      request->args[3], request->args[4]);
             break;
-        case P_VECTOR:
+        case P_GET_SIMD_STATE:
-            next_stage.entry = (generic_func *)request->args[0];
+            get_simd_state((void *)request->args[0]);
-            memcpy(next_stage.args, &request->args[1], 4 * sizeof(u64));
+            break;
-            return 1;
+        case P_PUT_SIMD_STATE:
+            put_simd_state((void *)request->args[0]);
+            break;
 
         case P_WRITE64:
             exc_guard = GUARD_SKIP;

src/proxy.h

@@ -20,6 +20,8 @@ typedef enum {
     P_VECTOR,
     P_GL1_CALL,
     P_GL2_CALL,
+    P_GET_SIMD_STATE,
+    P_PUT_SIMD_STATE,
 
     P_WRITE64 = 0x100, // Generic register functions
     P_WRITE32,

src/utils.h

@@ -355,6 +355,9 @@ void memcpy16(void *dst, void *src, size_t size);
 void memset8(void *dst, u8 value, size_t size);
 void memcpy8(void *dst, void *src, size_t size);
 
+void get_simd_state(void *state);
+void put_simd_state(void *state);
+
 void hexdump(const void *d, size_t len);
 void regdump(u64 addr, size_t len);
 int sprintf(char *str, const char *fmt, ...);

src/utils_asm.S

@@ -93,3 +93,45 @@ memset8:
     bne     1b
 2:
     ret
+
+.globl get_simd_state
+.type get_simd_state, @function
+get_simd_state:
+    stp q0, q1, [x0], #32
+    stp q2, q3, [x0], #32
+    stp q4, q5, [x0], #32
+    stp q6, q7, [x0], #32
+    stp q8, q9, [x0], #32
+    stp q10, q11, [x0], #32
+    stp q12, q13, [x0], #32
+    stp q14, q15, [x0], #32
+    stp q16, q17, [x0], #32
+    stp q18, q19, [x0], #32
+    stp q20, q21, [x0], #32
+    stp q22, q23, [x0], #32
+    stp q24, q25, [x0], #32
+    stp q26, q27, [x0], #32
+    stp q28, q29, [x0], #32
+    stp q30, q31, [x0], #32
+    ret
+
+.globl put_simd_state
+.type put_simd_state, @function
+put_simd_state:
+    ldp q0, q1, [x0], #32
+    ldp q2, q3, [x0], #32
+    ldp q4, q5, [x0], #32
+    ldp q6, q7, [x0], #32
+    ldp q8, q9, [x0], #32
+    ldp q10, q11, [x0], #32
+    ldp q12, q13, [x0], #32
+    ldp q14, q15, [x0], #32
+    ldp q16, q17, [x0], #32
+    ldp q18, q19, [x0], #32
+    ldp q20, q21, [x0], #32
+    ldp q22, q23, [x0], #32
+    ldp q24, q25, [x0], #32
+    ldp q26, q27, [x0], #32
+    ldp q28, q29, [x0], #32
+    ldp q30, q31, [x0], #32
+    ret