x86 validator: Rewrite non-temporal stores into cached memory accesses

src/trusted/validator/validation_rewrite_test_data.S

+/*
+ * Copyright 2016 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.
+ */
+
+#include "native_client/src/include/build_config.h"
+#include "native_client/src/include/nacl_asm.h"
+
+#define GLOBAL_SYM(name) .global IDENTIFIER(name); IDENTIFIER(name):
+
+.data
+
+#if NACL_BUILD_SUBARCH == 32
+
+GLOBAL_SYM(no_rewrite_code)
+    movntq %mm0, (%ebx)
+GLOBAL_SYM(no_rewrite_code_end)
+
+GLOBAL_SYM(no_rewrite_code_post_rewrite)
+    movntq %mm0, (%ebx)
+GLOBAL_SYM(no_rewrite_code_post_rewrite_end)
+
+GLOBAL_SYM(movntq_code)
+    movntq %mm0, (%ebx)
+GLOBAL_SYM(movntq_code_end)
+
+GLOBAL_SYM(movntq_code_post_rewrite)
+    movq %mm0, (%ebx)
+GLOBAL_SYM(movntq_code_post_rewrite_end)
+
+GLOBAL_SYM(movntdq_code)
+    movntdq %xmm0, (%edx)
+GLOBAL_SYM(movntdq_code_end)
+
+GLOBAL_SYM(movntdq_code_post_rewrite)
+    movdqa %xmm0, (%edx)
+GLOBAL_SYM(movntdq_code_post_rewrite_end)
+
+#else
+
+GLOBAL_SYM(no_rewrite_code)
+    mov %edi,%edi
+    movnti %rax,0x68(%r15,%rdi,1)
+GLOBAL_SYM(no_rewrite_code_end)
+
+GLOBAL_SYM(no_rewrite_code_post_rewrite)
+    mov %edi,%edi
+    movnti %rax,0x68(%r15,%rdi,1)
+GLOBAL_SYM(no_rewrite_code_post_rewrite_end)
+
+GLOBAL_SYM(off_webstore_movnt_code)
+    mov %ebx,%ebx
+    movntq %mm0,(%rbx)
+GLOBAL_SYM(off_webstore_movnt_code_end)
+
+GLOBAL_SYM(off_webstore_movnt_code_post_rewrite)
+    mov %ebx,%ebx
+    movntq %mm0,(%rbx)
+GLOBAL_SYM(off_webstore_movnt_code_post_rewrite_end)
+
+GLOBAL_SYM(prefetchnta_code)
+    mov %edi,%edi
+    prefetchnta (%r15,%rdi,1)
+GLOBAL_SYM(prefetchnta_code_end)
+
+GLOBAL_SYM(prefetchnta_code_post_rewrite)
+    mov %edi,%edi
+    nop
+    nop
+    nop
+    nop
+    nop
+GLOBAL_SYM(prefetchnta_code_post_rewrite_end)
+
+GLOBAL_SYM(movntps_code)
+    mov %ebx,%ebx
+    movntps %xmm0,(%r15,%rbx,1)
+GLOBAL_SYM(movntps_code_end)
+
+GLOBAL_SYM(movntps_code_post_rewrite)
+    mov %ebx,%ebx
+    movaps %xmm0,(%r15,%rbx,1)
+GLOBAL_SYM(movntps_code_post_rewrite_end)
+
+GLOBAL_SYM(movnti_code)
+    mov %edi,%edi
+    movnti %rax,0x68(%r15,%rdi,1)
+GLOBAL_SYM(movnti_code_end)
+
+GLOBAL_SYM(movnti_code_post_rewrite)
+    mov %edi,%edi
+    mov %rax,0x68(%r15,%rdi,1)
+    nop
+GLOBAL_SYM(movnti_code_post_rewrite_end)
+
+GLOBAL_SYM(movnti_code2)
+    mov %edx,%edx
+    movnti %r9d,(%r15,%rdx,1)
+GLOBAL_SYM(movnti_code2_end)
+
+GLOBAL_SYM(movnti_code2_post_rewrite)
+    mov %edx,%edx
+    mov %r9d,(%r15,%rdx,1)
+    nop
+GLOBAL_SYM(movnti_code2_post_rewrite_end)
+
+GLOBAL_SYM(movnti_rip_relative_code)
+    movnti %rax,0x20(%rip)
+GLOBAL_SYM(movnti_rip_relative_code_end)
+
+GLOBAL_SYM(movnti_rip_relative_code_post_rewrite)
+    nop
+    mov %rax,0x20(%rip)
+GLOBAL_SYM(movnti_rip_relative_code_post_rewrite_end)
+
+GLOBAL_SYM(movntdq_code)
+    mov %edx,%edx
+    movntdq %xmm0,0x10(%r15,%rdx,1)
+GLOBAL_SYM(movntdq_code_end)
+
+GLOBAL_SYM(movntdq_code_post_rewrite)
+    mov %edx,%edx
+    movdqa %xmm0,0x10(%r15,%rdx,1)
+GLOBAL_SYM(movntdq_code_post_rewrite_end)
+
+GLOBAL_SYM(movntdq_code2)
+    mov %ecx,%ecx
+    movntdq %xmm15,(%r15,%rcx,1)
+GLOBAL_SYM(movntdq_code2_end)
+
+GLOBAL_SYM(movntdq_code2_post_rewrite)
+    mov %ecx,%ecx
+    movdqa %xmm15,(%r15,%rcx,1)
+GLOBAL_SYM(movntdq_code2_post_rewrite_end)
+
+GLOBAL_SYM(multiple_movnt_code)
+    mov %edi,%edi
+    movnti %rax,0x68(%r15,%rdi,1)
+    mov %edx,%edx
+    movntdq %xmm0,0x10(%r15,%rdx,1)
+GLOBAL_SYM(multiple_movnt_code_end)
+
+GLOBAL_SYM(multiple_movnt_code_post_rewrite)
+    mov %edi,%edi
+    mov %rax,0x68(%r15,%rdi,1)
+    nop
+    mov %edx,%edx
+    movdqa %xmm0,0x10(%r15,%rdx,1)
+GLOBAL_SYM(multiple_movnt_code_post_rewrite_end)
+
+    /* size = 32 bytes */
+GLOBAL_SYM(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)
+GLOBAL_SYM(one_bundle_movnt_code_end)
+
+    /* size = 32 bytes */
+GLOBAL_SYM(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)
+GLOBAL_SYM(one_bundle_movnt_code_post_rewrite_end)
+
+GLOBAL_SYM(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 */
+GLOBAL_SYM(last_movnti_cross_bundle_by_one_end)
+
+GLOBAL_SYM(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 */
+GLOBAL_SYM(last_movnti_cross_bundle_by_one_post_rewrite_end)
+
+#endif