use emms instead, add WIN guards

2020-10-12 18:15:01 +03:00 · 2020-10-12 18:15:01 +03:00 · 403eb513a0
parent a5b164946c
commit 403eb513a0
27 changed files with 87 additions and 29 deletions
--- a/kernel/x86_64/amax.S
+++ b/kernel/x86_64/amax.S
@ -55,7 +55,9 @@
 	PROLOGUE
 	PROFCODE
 	
-	fninit
+#ifdef WINDOWS_ABI
+	emms
+#endif

 	salq	$BASE_SHIFT, INCX

--- a/kernel/x86_64/asum.S
+++ b/kernel/x86_64/asum.S
@ -50,7 +50,10 @@
 	PROLOGUE
 	PROFCODE
 	
-	fninit
+#ifdef WINDOWS_ABI
+	emms
+#endif
+
 	fldz
 	testq	M, M
 	jle	.L999
--- a/kernel/x86_64/dot.S
+++ b/kernel/x86_64/dot.S
@ -49,7 +49,10 @@

 	PROLOGUE
 	PROFCODE
-	fninit
+
+#ifdef WINDOWS_ABI
+	emms
+#endif

 	salq	$BASE_SHIFT, INCX
 	salq	$BASE_SHIFT, INCY
--- a/kernel/x86_64/iamax.S
+++ b/kernel/x86_64/iamax.S
@ -59,7 +59,10 @@

 	PROLOGUE
 	PROFCODE
-	fninit
+
+#ifdef WINDOWS_ABI
+	emms
+#endif

 	salq	$BASE_SHIFT, INCX

--- a/kernel/x86_64/izamax.S
+++ b/kernel/x86_64/izamax.S
@ -59,7 +59,10 @@

 	PROLOGUE
 	PROFCODE
-	fninit
+
+#ifdef WINDOWS_ABI
+	emms
+#endif

 	salq	$ZBASE_SHIFT, INCX

--- a/kernel/x86_64/nrm2.S
+++ b/kernel/x86_64/nrm2.S
@ -50,7 +50,10 @@
 	PROLOGUE
 	PROFCODE

-	fninit
+#ifdef WINDOWS_ABI
+	emms
+#endif
+
 	fldz
 	testq	M, M
 	jle	.L999
--- a/kernel/x86_64/qconjg.S
+++ b/kernel/x86_64/qconjg.S
@ -41,7 +41,10 @@

 	PROLOGUE
 	PROFCODE
-	fninit
+
+#ifdef WINDOWS_ABI
+	emms
+#endif

 	fldz
 	FLD	1 * SIZE(ARG1)
--- a/kernel/x86_64/qdot.S
+++ b/kernel/x86_64/qdot.S
@ -58,7 +58,9 @@

 	PROLOGUE

-	fninit
+#ifdef WINDOWS_ABI
+	emms
+#endif

 	pushl	%edi
 	pushl	%esi
--- a/kernel/x86_64/qgemm_kernel_2x2.S
+++ b/kernel/x86_64/qgemm_kernel_2x2.S
@ -74,7 +74,9 @@
 	PROLOGUE
 	PROFCODE

-	fninit
+#ifdef WINDOWS_ABI
+	emms
+#endif

 	subq	$STACKSIZE, %rsp
 	movq	%rbx,  0(%rsp)
--- a/kernel/x86_64/qgemv_n.S
+++ b/kernel/x86_64/qgemv_n.S
@ -76,7 +76,9 @@
 	PROLOGUE
 	PROFCODE

-	fninit
+#ifdef WINDOWS_ABI
+	emms
+#endif

 	subq	$STACKSIZE, %rsp
 	movq	%rbx,  0(%rsp)
--- a/kernel/x86_64/qgemv_t.S
+++ b/kernel/x86_64/qgemv_t.S
@ -75,7 +75,10 @@
 	PROLOGUE
 	PROFCODE

-	fninit
+#ifdef WINDOWS_ABI
+	emms
+#endif
+
 	subq	$STACKSIZE, %rsp
 	movq	%rbx,  0(%rsp)
 	movq	%rbp,  8(%rsp)
--- a/kernel/x86_64/qtrsm_kernel_LN_2x2.S
+++ b/kernel/x86_64/qtrsm_kernel_LN_2x2.S
@ -74,7 +74,9 @@
 	PROLOGUE
 	PROFCODE

-	fninit
+#ifdef WINDOWS_ABI
+	emms
+#endif

 	subq	$STACKSIZE, %rsp
 	movq	%rbx,  0(%rsp)
--- a/kernel/x86_64/qtrsm_kernel_LT_2x2.S
+++ b/kernel/x86_64/qtrsm_kernel_LT_2x2.S
@ -74,7 +74,9 @@
 	PROLOGUE
 	PROFCODE

-	fninit
+#ifdef WINDOWS_ABI
+	emms
+#endif

 	subq	$STACKSIZE, %rsp
 	movq	%rbx,  0(%rsp)
--- a/kernel/x86_64/qtrsm_kernel_RT_2x2.S
+++ b/kernel/x86_64/qtrsm_kernel_RT_2x2.S
@ -74,8 +74,9 @@
 	PROLOGUE
 	PROFCODE

-	fninit
-
+#ifdef WINDOWS_ABI
+	emms
+#endif

 	subq	$STACKSIZE, %rsp
 	movq	%rbx,  0(%rsp)
--- a/kernel/x86_64/sum.S
+++ b/kernel/x86_64/sum.S
@ -50,7 +50,9 @@
 	PROLOGUE
 	PROFCODE

-	fninit
+#ifdef WINDOWS_ABI
+	emms
+#endif

 	fldz
 	testq	M, M
--- a/kernel/x86_64/xdot.S
+++ b/kernel/x86_64/xdot.S
@ -59,7 +59,9 @@

 	PROFCODE

-	fninit
+#ifdef WINDOWS_ABI
+	emms
+#endif


 #define N	%ebx
--- a/kernel/x86_64/xgemm3m_kernel_2x2.S
+++ b/kernel/x86_64/xgemm3m_kernel_2x2.S
@ -78,7 +78,9 @@
 	PROLOGUE
 	PROFCODE

-	fninit
+#ifdef WINDOWS_ABI
+	emms
+#endif

 	subq	$STACKSIZE, %rsp
 	movq	%rbx,  0(%rsp)
--- a/kernel/x86_64/xgemm_kernel_1x1.S
+++ b/kernel/x86_64/xgemm_kernel_1x1.S
@ -97,7 +97,9 @@
 	PROLOGUE
 	PROFCODE

-	fninit
+#ifdef WINDOWS_ABI
+	emms
+#endif

 	subq	$STACKSIZE, %rsp
 	movq	%rbx,  0(%rsp)
--- a/kernel/x86_64/xgemv_n.S
+++ b/kernel/x86_64/xgemv_n.S
@ -76,7 +76,9 @@
 	PROLOGUE
 	PROFCODE

-	fninit
+#ifdef WINDOWS_ABI
+	emms
+#endif

 	subq	$STACKSIZE, %rsp
 	movq	%rbx,  0(%rsp)
--- a/kernel/x86_64/xgemv_t.S
+++ b/kernel/x86_64/xgemv_t.S
@ -75,7 +75,9 @@
 	PROLOGUE
 	PROFCODE

-	fninit
+#ifdef WINDOWS_ABI
+	emms
+#endif

 	subq	$STACKSIZE, %rsp
 	movq	%rbx,  0(%rsp)
--- a/kernel/x86_64/xtrsm_kernel_LT_1x1.S
+++ b/kernel/x86_64/xtrsm_kernel_LT_1x1.S
@ -90,7 +90,9 @@
 	PROLOGUE
 	PROFCODE

-	fninit
+#ifdef WINDOWS_ABI
+	emms
+#endif

 	subq	$STACKSIZE, %rsp
 	movq	%rbx,  0(%rsp)
--- a/kernel/x86_64/zamax.S
+++ b/kernel/x86_64/zamax.S
@ -55,7 +55,9 @@
 	PROLOGUE
 	PROFCODE

-	fninit
+#ifdef WINDOWS_ABI
+	emms
+#endif

 	salq	$ZBASE_SHIFT, INCX

--- a/kernel/x86_64/zasum.S
+++ b/kernel/x86_64/zasum.S
@ -50,7 +50,9 @@
 	PROLOGUE
 	PROFCODE

-	fninit
+#ifdef WINDOWS_ABI
+	emms
+#endif

 	fldz
 	testq	M, M
--- a/kernel/x86_64/zdot.S
+++ b/kernel/x86_64/zdot.S
@ -54,9 +54,9 @@
 	PROLOGUE
 	PROFCODE

-	fninit
-
 #ifdef WINDOWS_ABI
+	emms
+
 	movq	40(%rsp), INCY
 #endif

--- a/kernel/x86_64/znrm2.S
+++ b/kernel/x86_64/znrm2.S
@ -50,7 +50,9 @@
 	PROLOGUE
 	PROFCODE

-	fninit
+#ifdef WINDOWS_ABI
+	emms
+#endif

 	fldz
 	testq	M, M
--- a/kernel/x86_64/zscal.S
+++ b/kernel/x86_64/zscal.S
@ -50,7 +50,9 @@
 	PROLOGUE
 	PROFCODE

-	fninit
+#ifdef WINDOWS_ABI
+	emms
+#endif

 	salq	$ZBASE_SHIFT, INCX

--- a/kernel/x86_64/zsum.S
+++ b/kernel/x86_64/zsum.S
@ -50,7 +50,9 @@
 	PROLOGUE
 	PROFCODE

-	fninit
+#ifdef WINDOWS_ABI
+	emms
+#endif

 	fldz
 	testq	M, M