make HiPE work on x86_64 when PIE is enabled

Currently HiPE amd64 assumes the runtime system code is loaded into
the low 2G of the address space. However, this is not the case when
PIE is enabled, it is loaded into a random location. So trampolines
are required to call BIFs, and also we have first to load the address
of sse2_fnegate_mask to a regisiter before xorpd in fchs.

1 files changed, 85 insertions, 21 deletions

diff --git a/erts/emulator/hipe/hipe_amd64.c b/erts/emulator/hipe/hipe_amd64.c
index e3cff4a4ba..bdeae23c7c 100644
--- a/erts/emulator/hipe/hipe_amd64.c
+++ b/erts/emulator/hipe/hipe_amd64.c
@@ -28,6 +28,7 @@
 #include "error.h"
 #include "bif.h"
 #include "big.h"	/* term_to_Sint() */
+#include "erl_binary.h"
 
 #include "hipe_arch.h"
 #include "hipe_bif0.h"
@@ -52,9 +53,9 @@ int hipe_patch_insn(void *address, Uint64 value, Eterm type)
     switch (type) {
       case am_closure:
       case am_constant:
+      case am_c_const:
 	*(Uint64*)address = value;
 	break;
-      case am_c_const:
       case am_atom:
 	/* check that value fits in an unsigned imm32 */
 	/* XXX: are we sure it's not really a signed imm32? */
@@ -71,14 +72,19 @@ int hipe_patch_insn(void *address, Uint64 value, Eterm type)
 
 int hipe_patch_call(void *callAddress, void *destAddress, void *trampoline)
 {
-    Sint rel32;
+    Sint64 destOffset = (Sint64)destAddress - (Sint64)callAddress - 4;
 
-    ASSERT(trampoline == NULL);
+    if ((destOffset < -0x80000000L) || (destOffset >= 0x80000000L)) {
+        destOffset = ((Sint64)trampoline - 2) - (Sint64)callAddress - 4;
 
-    rel32 = (Sint)destAddress - (Sint)callAddress - 4;
-    if ((Sint)(Sint32)rel32 != rel32)
-	return -1;
-    *(Uint32*)callAddress = (Uint32)rel32;
+        if ((destOffset < -0x80000000L) || (destOffset >= 0x80000000L))
+            return -1;
+
+        *(void**)trampoline = destAddress;
+        hipe_flush_icache_word(trampoline);
+    }
+
+    *(Uint32*)callAddress = (Uint32)destOffset;
     hipe_flush_icache_word(callAddress);
     return 0;
 }
@@ -96,12 +102,70 @@ static void *alloc_code(unsigned int alloc_bytes)
     return erts_alloc(ERTS_ALC_T_HIPE_EXEC, alloc_bytes);
 }
 
+static int check_callees(Eterm callees)
+{
+    Eterm *tuple;
+    Uint arity;
+    Uint i;
+
+    if (is_not_tuple(callees))
+	return -1;
+    tuple = tuple_val(callees);
+    arity = arityval(tuple[0]);
+    for (i = 1; i <= arity; ++i) {
+	Eterm mfa = tuple[i];
+	if (is_atom(mfa))
+	    continue;
+	if (is_not_tuple(mfa) ||
+	    tuple_val(mfa)[0] != make_arityval(3) ||
+	    is_not_atom(tuple_val(mfa)[1]) ||
+	    is_not_atom(tuple_val(mfa)[2]) ||
+	    is_not_small(tuple_val(mfa)[3]) ||
+	    unsigned_val(tuple_val(mfa)[3]) > 255)
+	    return -1;
+    }
+    return arity;
+}
+
+#define TRAMPOLINE_BYTES 12
+
+static void generate_trampolines(unsigned char *address,
+                                 int nrcallees, Eterm callees,
+                                 unsigned char **trampvec)
+{
+    unsigned char *trampoline = address;
+    int i;
+
+    for(i = 0; i < nrcallees; ++i) {
+        trampoline[0] = 0x48;           /* movabsq $..., %rax; */
+        trampoline[1] = 0xb8;
+        *((Uint64*)(trampoline+2)) = 0; /* callee's address */
+        trampoline[10] = 0xff;          /* jmpq *%rax */
+        trampoline[11] = 0xe0;
+        trampvec[i] = trampoline+2;
+        trampoline += TRAMPOLINE_BYTES;
+    }
+    hipe_flush_icache_range(address, nrcallees*TRAMPOLINE_BYTES);
+}
+
 void *hipe_alloc_code(Uint nrbytes, Eterm callees, Eterm *trampolines, Process *p)
 {
-    if (is_not_nil(callees))
+    int nrcallees;
+    Eterm trampvecbin;
+    unsigned char **trampvec;
+    unsigned char *address;
+
+    nrcallees = check_callees(callees);
+    if (nrcallees < 0)
 	return NULL;
-    *trampolines = NIL;
-    return alloc_code(nrbytes);
+
+    trampvecbin = new_binary(p, NULL, nrcallees*sizeof(unsigned char*));
+    trampvec = (unsigned char **)binary_bytes(trampvecbin);
+
+    address = alloc_code(nrbytes + nrcallees*TRAMPOLINE_BYTES);
+    generate_trampolines(address + nrbytes, nrcallees, callees, trampvec);
+    *trampolines = trampvecbin;
+    return address;
 }
 
 void hipe_free_code(void* code, unsigned int bytes)
@@ -129,10 +193,9 @@ void *hipe_make_native_stub(void *callee_exp, unsigned int beamArity)
      */
     unsigned int codeSize;
     unsigned char *code, *codep;
-    unsigned int callEmuOffset;
 
-    codeSize =	/* 23, 26, 29, or 32 bytes */
-      23 +	/* 23 when all offsets are 8-bit */
+    codeSize =	/* 30, 33, 36, or 39 bytes */
+      30 +	/* 30 when all offsets are 8-bit */
       (P_CALLEE_EXP >= 128 ? 3 : 0) +
       ((P_CALLEE_EXP + 4) >= 128 ? 3 : 0) +
       (P_ARITY >= 128 ? 3 : 0);
@@ -197,14 +260,15 @@ void *hipe_make_native_stub(void *callee_exp, unsigned int beamArity)
     codep[0] = beamArity;
     codep += 1;
 
-    /* jmp callemu; 5 bytes */
-    callEmuOffset = (unsigned char*)nbif_callemu - (code + codeSize);
-    codep[0] = 0xe9;
-    codep[1] =  callEmuOffset        & 0xFF;
-    codep[2] = (callEmuOffset >>  8) & 0xFF;
-    codep[3] = (callEmuOffset >> 16) & 0xFF;
-    codep[4] = (callEmuOffset >> 24) & 0xFF;
-    codep += 5;
+    /* jmp callemu; 12 bytes */
+    codep[0] = 0x48;
+    codep[1] = 0xb8;
+    codep += 2;
+    *(Uint64*)codep = (Uint64)nbif_callemu;
+    codep += 8;
+    codep[0] = 0xff;
+    codep[1] = 0xe0;
+    codep += 2;
 
     ASSERT(codep == code + codeSize);