vbox的更動 96242 路徑 trunk/src/VBox/Runtime/common/math/cosf.asm

時間撮記:

2022-8-17 上午01:59:06 (3 年 以前)

作者:

vboxsync

svn:sync-xref-src-repo-rev:

153054

訊息:

IPRT/nocrt: Adapted the reworked sin and cos code for sinf and cosf; fixed a few cos bugs and added tests for cos. bugref:10261

檔案:

• trunk/src/VBox/Runtime/common/math/cosf.asm (複製) (複製 從 trunk/src/VBox/Runtime/common/math/cos.asm ) (10 筆差異)

trunk/src/VBox/Runtime/common/math/cosf.asm

@@ -1 +1 @@
 ; $Id$
 ;; @file
-; IPRT - No-CRT cos - AMD64 & X86.
+; IPRT - No-CRT cosf - AMD64 & X86.
 ;
 
@@ -34 +34 @@
 
 ;;
-; Compute the cosine of rd, measured in radians.
+; Compute the cosine of rf, measured in radians.
 ;
 ; @returns  st(0) / xmm0
-; @param    rd      [rbp + xCB*2] / xmm0
-;
-RT_NOCRT_BEGINPROC cos
+; @param    rf      [rbp + xCB*2] / xmm0
+;
+RT_NOCRT_BEGINPROC cosf
         push    xBP
         SEH64_PUSH_xBP
@@ -64 +64 @@
         ;
 %ifdef RT_ARCH_AMD64
-        movsd   [xBP - 10h], xmm0
-        fld     qword [xBP - 10h]
+        movss   [xBP - 10h], xmm0
+        fld     dword [xBP - 10h]
 %else
-        fld     qword [xBP + xCB*2]
+        fld     dword [xBP + xCB*2]
 %endif
 
@@ -74 +74 @@
         ; works reliably, so outside that we'll use the FSIN instruction instead
         ; as it has a larger good range (-5pi/4 to 1pi/4 for cosine).
-        ; Input conversion follows: cos(x) = sin(x + pi/2)
+        ; Input conversion follows: cosf(x) = sinf(x + pi/2)
         ;
         ; We examin the input and weed out non-finit numbers first.
@@ -110 +110 @@
         fld     qword [.s_r64TinyCosTo1 xWrtRIP]
         fcomip  st1
-        jbe      .zero_extra_pop
+        ja      .zero_extra_pop
 
 .not_that_tiny_input:
@@ -147 +147 @@
         ;
 .do_sine:
-        mov     ecx, 1                      ; double
+        mov     ecx, 0                      ; double
         extern  NAME(rtNoCrtMathSinCore)
         call    NAME(rtNoCrtMathSinCore)
@@ -156 +156 @@
 .return_val:
 %ifdef RT_ARCH_AMD64
-        fstp    qword [xBP - 10h]
-        movsd   xmm0, [xBP - 10h]
+        fstp    dword [xBP - 10h]
+        movss   xmm0, [xBP - 10h]
 %endif
 %ifdef RT_OS_WINDOWS
@@ -167 +167 @@
 
         ;
-        ; cos(+/-0) = +1.0
+        ; cosf(+/-0) = +1.0
         ;
 .zero_extra_pop:
@@ -190 +190 @@
         ;
 ALIGNCODE(8)
-        ; About 2**-27. When fabs(input) is below this limit we can consider cos(input) ~= 1.0.
+        ; About 2**-18. When fabs(input) is below this limit we can consider cosf(input) ~= 1.0.
 .s_r64TinyCosTo1:
-        dq  7.4505806e-9
+        dq  0.000244140625
 
         ; The absolute limit for the range which FCOS is expected to produce reasonable results.
@@ -200 +200 @@
 .s_r64Two:
         dq  2.0
-ENDPROC   RT_NOCRT(cos)
-
+ENDPROC   RT_NOCRT(cosf)
+