Rewrite non-temporal instructions

src/trusted/validator/validation_rewrite_64_test_data.S

Index: src/trusted/validator/validation_rewrite_64_test_data.S
diff --git a/src/trusted/validator/validation_rewrite_64_test_data.S b/src/trusted/validator/validation_rewrite_64_test_data.S
new file mode 100644
index 0000000000000000000000000000000000000000..96852d98a2e6de79804e6183a7bfccc44aeb795c
--- /dev/null
+++ b/src/trusted/validator/validation_rewrite_64_test_data.S
@@ -0,0 +1,167 @@
+/*
+ * Copyright 2015 The Native Client Authors. All rights reserved.
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+

[Mark Seaborn, 2015/08/14 21:12:59]

Does the test have an example of a non-temporal write that the rewriter doesn't
handle?

[ruiq, 2015/08/15 04:46:40]

Done.

+.data
+
+/* This macro is used to define two global symbols for the same byte sequence:

[Mark Seaborn, 2015/08/14 21:12:59]

Nit: Use the NaCl style for multiline comments, with "/*" and "*/" on separate
lines, like this:
/*
 * Blah...
 * blah.
 */

[ruiq, 2015/08/15 04:46:40]

Done.

+ * 'name' and '_name'. This is to work around the linkage issue across OS and

[Mark Seaborn, 2015/08/14 21:12:59]

There's already a macro for doing this, "IDENTIFIER", in src/include/nacl_asm.h.

[ruiq, 2015/08/15 04:46:40]

Done.

+ * compilers. */
+.macro global_bytes name
+    .global \name
+\name:
+    .global _\name
+_\name:
+.endm
+
+    global_bytes no_rewrite_code
+    mov %edi,%edi
+    movnti %rax,0x68(%r15,%rdi,1)
+    .byte 0

[Mark Seaborn, 2015/08/14 21:12:59]

What is this for?

Oh, you're using strlen() on x86 code?  That's not a good idea, because there
might be zero bytes inside the code.

[ruiq, 2015/08/15 04:46:40]

Now using two labels: xx_code, and xx_code_end.

+
+    global_bytes no_rewrite_code_post_rewrite
+    mov %edi,%edi
+    movnti %rax,0x68(%r15,%rdi,1)
+    .byte 0
+
+    global_bytes prefetchnta_code
+    mov %edi,%edi
+    prefetchnta (%r15,%rdi,1)
+    .byte 0
+
+    global_bytes prefetchnta_code_post_rewrite
+    mov %edi,%edi
+    nop
+    nop
+    nop
+    nop
+    nop
+    .byte 0
+
+    global_bytes movntps_code
+    mov %ebx,%ebx
+    movntps %xmm0,(%r15,%rbx,1)
+    .byte 0
+
+    global_bytes movntps_code_post_rewrite
+    mov %ebx,%ebx
+    movaps %xmm0,(%r15,%rbx,1)
+    .byte 0
+
+    global_bytes movnti_code
+    mov %edi,%edi
+    movnti %rax,0x68(%r15,%rdi,1)
+    .byte 0
+
+    global_bytes movnti_code_post_rewrite
+    mov %edi,%edi
+    mov %rax,0x68(%r15,%rdi,1)
+    nop
+    .byte 0
+
+    global_bytes movnti_code2
+    mov %edx,%edx
+    movnti %r9d,(%r15,%rdx,1)
+    .byte 0
+
+    global_bytes movnti_code2_post_rewrite
+    mov %edx,%edx
+    mov %r9d,(%r15,%rdx,1)
+    nop
+    .byte 0
+
+    global_bytes movnti_rip_relative_code
+    movnti %rax,0x20(%rip)
+    .byte 0
+
+    global_bytes movnti_rip_relative_code_post_rewrite
+    nop
+    mov %rax,0x20(%rip)
+    .byte 0
+
+    global_bytes movntdq_code
+    mov %edx,%edx
+    movntdq %xmm0,0x10(%r15,%rdx,1)
+    .byte 0
+
+    global_bytes movntdq_code_post_rewrite
+    mov %edx,%edx
+    movdqa %xmm0,0x10(%r15,%rdx,1)
+    .byte 0
+
+    global_bytes movntdq_code2
+    mov %ecx,%ecx
+    movntdq %xmm15,(%r15,%rcx,1)
+    .byte 0
+
+    global_bytes movntdq_code2_post_rewrite
+    mov %ecx,%ecx
+    movdqa %xmm15,(%r15,%rcx,1)
+    .byte 0
+
+    global_bytes multiple_movnt_code
+    mov %edi,%edi
+    movnti %rax,0x68(%r15,%rdi,1)
+    mov %edx,%edx
+    movntdq %xmm0,0x10(%r15,%rdx,1)
+    .byte 0
+
+    global_bytes multiple_movnt_code_post_rewrite
+    mov %edi,%edi
+    mov %rax,0x68(%r15,%rdi,1)
+    nop
+    mov %edx,%edx
+    movdqa %xmm0,0x10(%r15,%rdx,1)
+    .byte 0
+
+    /* size = 32 bytes */
+    global_bytes one_bundle_movnt_code
+    mov %ecx,%ecx
+    movntdq %xmm15,(%r15,%rcx,1)
+    mov %ecx,%ecx
+    movntdq %xmm15,(%r15,%rcx,1)
+    mov %ecx,%ecx
+    movntdq %xmm15,(%r15,%rcx,1)
+    mov %ecx,%ecx
+    movntdq %xmm15,(%r15,%rcx,1)
+    .byte 0
+
+    /* size = 32 bytes */
+    global_bytes one_bundle_movnt_code_post_rewrite
+    mov %ecx,%ecx
+    movdqa %xmm15,(%r15,%rcx,1)
+    mov %ecx,%ecx
+    movdqa %xmm15,(%r15,%rcx,1)
+    mov %ecx,%ecx
+    movdqa %xmm15,(%r15,%rcx,1)
+    mov %ecx,%ecx
+    movdqa %xmm15,(%r15,%rcx,1)
+    .byte 0
+
+    global_bytes last_movnti_cross_bundle_by_one
+    mov %ecx,%ecx                 /* offset 0, length 2 */
+    movntdq %xmm15,(%r15,%rcx,1)  /* offset 2, length 6 */
+    mov %ecx,%ecx                 /* offset 8, length 2 */
+    movntdq %xmm15,(%r15,%rcx,1)  /* offset 10, length 6 */
+    mov %ecx,%ecx                 /* offset 16, length 2 */
+    movntdq %xmm15,(%r15,%rcx,1)  /* offset 18, length 6 */
+    nop                           /* offset 24, length 1 */
+    mov %edx,%edx                 /* offset 25, length 2 */
+    /* the last byte of the following instruction is at offset 32 */
+    movnti %r9d,(%r15,%rdx,1)     /* offset 27, length 5 */
+    .byte 0
+
+    global_bytes last_movnti_cross_bundle_by_one_post_rewrite
+    mov %ecx,%ecx                 /* offset 0, length 2 */
+    movdqa %xmm15,(%r15,%rcx,1)   /* offset 2, length 6 */
+    mov %ecx,%ecx                 /* offset 8, length 2 */
+    movdqa %xmm15,(%r15,%rcx,1)   /* offset 10, length 6 */
+    mov %ecx,%ecx                 /* offset 16, length 2 */
+    movdqa %xmm15,(%r15,%rcx,1)   /* offset 18, length 6 */
+    nop                           /* offset 24, length 1 */
+    mov %edx,%edx                 /* offset 25, length 2 */
+    mov %r9d,(%r15,%rdx,1)        /* offset 27, length 4 */
+    nop                           /* offset 32, length 1 */
+    .byte 0