[reactos] 02/03: [RTL/x64] Improve RtlCaptureContext

[Timo Kreuzer]

https://git.reactos.org/?p=reactos.git;a=commitdiff;h=abb338b13d1fce57c045f…

commit abb338b13d1fce57c045f38ca759ede1da3fa2dc
Author:     Timo Kreuzer <timo.kreuzer(a)reactos.org>
AuthorDate: Fri Mar 2 08:02:13 2018 +0100
Commit:     Timo Kreuzer <timo.kreuzer(a)reactos.org>
CommitDate: Sat May 29 21:20:48 2021 +0200

    [RTL/x64] Improve RtlCaptureContext
    
    Use movaps instead of movdqa, it does the same thing, but is one byte shorter.
    Shuffle instructions around a bit to maximize parallel execution.
---
 sdk/lib/rtl/amd64/except_asm.S | 118 ++++++++++++++++++++++-------------------
 1 file changed, 63 insertions(+), 55 deletions(-)

diff --git a/sdk/lib/rtl/amd64/except_asm.S b/sdk/lib/rtl/amd64/except_asm.S
index b56cc48f51c..ea2eb7f88c8 100644
--- a/sdk/lib/rtl/amd64/except_asm.S
+++ b/sdk/lib/rtl/amd64/except_asm.S
@@ -16,9 +16,9 @@
 .code64
 
 /*
- * VOID NTAPI
+ * VOID
  * RtlCaptureContext(
- *     PCONTEXT ContextRecord); <rcx>
+ *     _Out_ PCONTEXT ContextRecord@<rcx>);
  */
 PUBLIC RtlCaptureContext
 .PROC RtlCaptureContext
@@ -28,70 +28,78 @@ PUBLIC RtlCaptureContext
     .ALLOCSTACK 8
     .ENDPROLOG
 
-    /* Save the basic register context */
-    mov [rcx + CONTEXT_Rax], rax
-    mov [rcx + CONTEXT_Rcx], rcx
-    mov [rcx + CONTEXT_Rdx], rdx
+    /* Save rax first, we use it later to copy some data */
+    mov [rcx + CxRax], rax
 
-    /* Load rflags into rax */
-    mov rax, [rsp]
+    /* Set ContextFlags */
+    mov dword ptr [rcx + CxContextFlags], (CONTEXT_FULL or CONTEXT_SEGMENTS)
 
-    mov [rcx + CONTEXT_Rbx], rbx
-    mov [rcx + CONTEXT_Rsi], rsi
-    mov [rcx + CONTEXT_Rdi], rdi
+    /* Store the basic register context */
+    mov [rcx + CxRcx], rcx
+    mov [rcx + CxRdx], rdx
+    mov [rcx + CxRbx], rbx
+    mov [rcx + CxRsi], rsi
 
-    /* Store rflags */
-    mov [rcx + CONTEXT_EFlags], rax
+    /* Load return address in rax */
+    mov rax, [rsp + 8]
+
+    mov [rcx + CxRdi], rdi
+    mov [rcx + CxRbp], rbp
+    mov [rcx + CxR8], r8
+    mov [rcx + CxR9], r9
+    mov [rcx + CxR10], r10
+
+    /* Store the return address */
+    mov [rcx + CxRip], rax
 
-    mov [rcx + CONTEXT_Rbp], rbp
-    mov [rcx + CONTEXT_R8], r8
-    mov [rcx + CONTEXT_R9], r9
+    mov [rcx + CxR11], r11
+    mov [rcx + CxR12], r12
+    mov [rcx + CxR13], r13
+    mov [rcx + CxR14], r14
+    mov [rcx + CxR15], r15
 
     /* Load former stack pointer in rax */
     lea rax, [rsp + 16]
 
-    mov [rcx + CONTEXT_R10], r10
-    mov [rcx + CONTEXT_R11], r11
-    mov [rcx + CONTEXT_R12], r12
+    /* Store segment selectors */
+    mov [rcx + CxSegCs], cs
+    mov [rcx + CxSegDs], ds
+    mov [rcx + CxSegEs], es
+    mov [rcx + CxSegFs], fs
+    mov [rcx + CxSegGs], gs
+    mov [rcx + CxSegSs], ss
 
     /* Store stack pointer */
-    mov [rcx + CONTEXT_Rsp], rax
+    mov [rcx + CxRsp], rax
+
+    /* Store xmm registers */
+    movaps [rcx + CxXmm0], xmm0
+    movaps [rcx + CxXmm1], xmm1
+    movaps [rcx + CxXmm2], xmm2
+    movaps [rcx + CxXmm3], xmm3
+    movaps [rcx + CxXmm4], xmm4
+    movaps [rcx + CxXmm5], xmm5
+    movaps [rcx + CxXmm6], xmm6
+    movaps [rcx + CxXmm7], xmm7
+
+    /* Load rflags into eax */
+    mov eax, [rsp]
+
+    movaps [rcx + CxXmm8], xmm8
+    movaps [rcx + CxXmm9], xmm9
+    movaps [rcx + CxXmm10], xmm10
+    movaps [rcx + CxXmm11], xmm11
+    movaps [rcx + CxXmm12], xmm12
+    movaps [rcx + CxXmm13], xmm13
+    movaps [rcx + CxXmm14], xmm14
+    movaps [rcx + CxXmm15], xmm15
+
+    /* Store legacy floating point registers */
+    fxsave [rcx + CxFltSave]
+    stmxcsr [rcx + CxMxCsr]
 
-    mov [rcx + CONTEXT_R13], r13
-    mov [rcx + CONTEXT_R14], r14
-    mov [rcx + CONTEXT_R15], r15
-
-    /* Load return address in rax */
-    mov rax, [rsp + 8]
-
-    /* Safe segment selectors */
-    mov [rcx + CONTEXT_SegCs], cs
-    mov [rcx + CONTEXT_SegDs], ds
-    mov [rcx + CONTEXT_SegEs], es
-    mov [rcx + CONTEXT_SegFs], fs
-    mov [rcx + CONTEXT_SegGs], gs
-    mov [rcx + CONTEXT_SegSs], ss
-
-    /* Store return address */
-    mov [rcx + CONTEXT_Rip], rax
-
-    /* Safe xmm registers */
-    movdqa [rcx + CONTEXT_Xmm0], xmm0
-    movdqa [rcx + CONTEXT_Xmm1], xmm1
-    movdqa [rcx + CONTEXT_Xmm2], xmm2
-    movdqa [rcx + CONTEXT_Xmm3], xmm3
-    movdqa [rcx + CONTEXT_Xmm4], xmm4
-    movdqa [rcx + CONTEXT_Xmm5], xmm5
-    movdqa [rcx + CONTEXT_Xmm6], xmm6
-    movdqa [rcx + CONTEXT_Xmm7], xmm7
-    movdqa [rcx + CONTEXT_Xmm8], xmm8
-    movdqa [rcx + CONTEXT_Xmm9], xmm9
-    movdqa [rcx + CONTEXT_Xmm10], xmm10
-    movdqa [rcx + CONTEXT_Xmm11], xmm11
-    movdqa [rcx + CONTEXT_Xmm12], xmm12
-    movdqa [rcx + CONTEXT_Xmm13], xmm13
-    movdqa [rcx + CONTEXT_Xmm14], xmm14
-    movdqa [rcx + CONTEXT_Xmm15], xmm15
+    /* Store rflags */
+    mov [rcx + CxEFlags], eax
 
     /* Cleanup stack and return */
     add rsp, 8