Merge pull request #2881 from mattip/fninit

add fninit to reset fpu registers before assembler routines
2020-10-12 23:50:41 +02:00 · 2020-10-12 23:50:41 +02:00 · 0c84ffe05f
parent cb4274e3ad 403eb513a0
commit 0c84ffe05f
27 changed files with 107 additions and 0 deletions
--- a/kernel/x86_64/amax.S
+++ b/kernel/x86_64/amax.S
@ -55,6 +55,10 @@
 	PROLOGUE
 	PROFCODE
 	
+#ifdef WINDOWS_ABI
+	emms
+#endif
+
 	salq	$BASE_SHIFT, INCX

 	fldz
--- a/kernel/x86_64/asum.S
+++ b/kernel/x86_64/asum.S
@ -50,6 +50,10 @@
 	PROLOGUE
 	PROFCODE
 	
+#ifdef WINDOWS_ABI
+	emms
+#endif
+
 	fldz
 	testq	M, M
 	jle	.L999
--- a/kernel/x86_64/dot.S
+++ b/kernel/x86_64/dot.S
@ -50,6 +50,10 @@
 	PROLOGUE
 	PROFCODE

+#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
@ -60,6 +60,10 @@
 	PROLOGUE
 	PROFCODE

+#ifdef WINDOWS_ABI
+	emms
+#endif
+
 	salq	$BASE_SHIFT, INCX

 	fldz
--- a/kernel/x86_64/izamax.S
+++ b/kernel/x86_64/izamax.S
@ -60,6 +60,10 @@
 	PROLOGUE
 	PROFCODE

+#ifdef WINDOWS_ABI
+	emms
+#endif
+
 	salq	$ZBASE_SHIFT, INCX

 	fldz
--- a/kernel/x86_64/nrm2.S
+++ b/kernel/x86_64/nrm2.S
@ -50,6 +50,10 @@
 	PROLOGUE
 	PROFCODE

+#ifdef WINDOWS_ABI
+	emms
+#endif
+
 	fldz
 	testq	M, M
 	jle	.L999
--- a/kernel/x86_64/qconjg.S
+++ b/kernel/x86_64/qconjg.S
@ -42,6 +42,10 @@
 	PROLOGUE
 	PROFCODE

+#ifdef WINDOWS_ABI
+	emms
+#endif
+
 	fldz
 	FLD	1 * SIZE(ARG1)
 	fsubrp	%st, %st(1)
--- a/kernel/x86_64/qdot.S
+++ b/kernel/x86_64/qdot.S
@ -58,6 +58,10 @@

 	PROLOGUE

+#ifdef WINDOWS_ABI
+	emms
+#endif
+
 	pushl	%edi
 	pushl	%esi
 	pushl	%ebx
--- a/kernel/x86_64/qgemm_kernel_2x2.S
+++ b/kernel/x86_64/qgemm_kernel_2x2.S
@ -74,6 +74,10 @@
 	PROLOGUE
 	PROFCODE

+#ifdef WINDOWS_ABI
+	emms
+#endif
+
 	subq	$STACKSIZE, %rsp
 	movq	%rbx,  0(%rsp)
 	movq	%rbp,  8(%rsp)
--- a/kernel/x86_64/qgemv_n.S
+++ b/kernel/x86_64/qgemv_n.S
@ -76,6 +76,10 @@
 	PROLOGUE
 	PROFCODE

+#ifdef WINDOWS_ABI
+	emms
+#endif
+
 	subq	$STACKSIZE, %rsp
 	movq	%rbx,  0(%rsp)
 	movq	%rbp,  8(%rsp)
--- a/kernel/x86_64/qgemv_t.S
+++ b/kernel/x86_64/qgemv_t.S
@ -75,6 +75,10 @@
 	PROLOGUE
 	PROFCODE

+#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
@ -74,6 +74,10 @@
 	PROLOGUE
 	PROFCODE

+#ifdef WINDOWS_ABI
+	emms
+#endif
+
 	subq	$STACKSIZE, %rsp
 	movq	%rbx,  0(%rsp)
 	movq	%rbp,  8(%rsp)
--- a/kernel/x86_64/qtrsm_kernel_LT_2x2.S
+++ b/kernel/x86_64/qtrsm_kernel_LT_2x2.S
@ -74,6 +74,10 @@
 	PROLOGUE
 	PROFCODE

+#ifdef WINDOWS_ABI
+	emms
+#endif
+
 	subq	$STACKSIZE, %rsp
 	movq	%rbx,  0(%rsp)
 	movq	%rbp,  8(%rsp)
--- a/kernel/x86_64/qtrsm_kernel_RT_2x2.S
+++ b/kernel/x86_64/qtrsm_kernel_RT_2x2.S
@ -74,6 +74,10 @@
 	PROLOGUE
 	PROFCODE

+#ifdef WINDOWS_ABI
+	emms
+#endif
+
 	subq	$STACKSIZE, %rsp
 	movq	%rbx,  0(%rsp)
 	movq	%rbp,  8(%rsp)
--- a/kernel/x86_64/sum.S
+++ b/kernel/x86_64/sum.S
@ -50,6 +50,10 @@
 	PROLOGUE
 	PROFCODE

+#ifdef WINDOWS_ABI
+	emms
+#endif
+
 	fldz
 	testq	M, M
 	jle	.L999
--- a/kernel/x86_64/xdot.S
+++ b/kernel/x86_64/xdot.S
@ -59,6 +59,11 @@

 	PROFCODE

+#ifdef WINDOWS_ABI
+	emms
+#endif
+
+
 #define N	%ebx
 #define X	%esi
 #define INCX	%ecx
--- a/kernel/x86_64/xgemm3m_kernel_2x2.S
+++ b/kernel/x86_64/xgemm3m_kernel_2x2.S
@ -78,6 +78,10 @@
 	PROLOGUE
 	PROFCODE

+#ifdef WINDOWS_ABI
+	emms
+#endif
+
 	subq	$STACKSIZE, %rsp
 	movq	%rbx,  0(%rsp)
 	movq	%rbp,  8(%rsp)
--- a/kernel/x86_64/xgemm_kernel_1x1.S
+++ b/kernel/x86_64/xgemm_kernel_1x1.S
@ -97,6 +97,10 @@
 	PROLOGUE
 	PROFCODE

+#ifdef WINDOWS_ABI
+	emms
+#endif
+
 	subq	$STACKSIZE, %rsp
 	movq	%rbx,  0(%rsp)
 	movq	%rbp,  8(%rsp)
--- a/kernel/x86_64/xgemv_n.S
+++ b/kernel/x86_64/xgemv_n.S
@ -76,6 +76,10 @@
 	PROLOGUE
 	PROFCODE

+#ifdef WINDOWS_ABI
+	emms
+#endif
+
 	subq	$STACKSIZE, %rsp
 	movq	%rbx,  0(%rsp)
 	movq	%rbp,  8(%rsp)
--- a/kernel/x86_64/xgemv_t.S
+++ b/kernel/x86_64/xgemv_t.S
@ -75,6 +75,10 @@
 	PROLOGUE
 	PROFCODE

+#ifdef WINDOWS_ABI
+	emms
+#endif
+
 	subq	$STACKSIZE, %rsp
 	movq	%rbx,  0(%rsp)
 	movq	%rbp,  8(%rsp)
--- a/kernel/x86_64/xtrsm_kernel_LT_1x1.S
+++ b/kernel/x86_64/xtrsm_kernel_LT_1x1.S
@ -90,6 +90,10 @@
 	PROLOGUE
 	PROFCODE

+#ifdef WINDOWS_ABI
+	emms
+#endif
+
 	subq	$STACKSIZE, %rsp
 	movq	%rbx,  0(%rsp)
 	movq	%rbp,  8(%rsp)
--- a/kernel/x86_64/zamax.S
+++ b/kernel/x86_64/zamax.S
@ -55,6 +55,10 @@
 	PROLOGUE
 	PROFCODE

+#ifdef WINDOWS_ABI
+	emms
+#endif
+
 	salq	$ZBASE_SHIFT, INCX

 	fldz
--- a/kernel/x86_64/zasum.S
+++ b/kernel/x86_64/zasum.S
@ -50,6 +50,10 @@
 	PROLOGUE
 	PROFCODE

+#ifdef WINDOWS_ABI
+	emms
+#endif
+
 	fldz
 	testq	M, M
 	jle	.L999
--- a/kernel/x86_64/zdot.S
+++ b/kernel/x86_64/zdot.S
@ -55,6 +55,8 @@
 	PROFCODE

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

--- a/kernel/x86_64/znrm2.S
+++ b/kernel/x86_64/znrm2.S
@ -50,6 +50,10 @@
 	PROLOGUE
 	PROFCODE

+#ifdef WINDOWS_ABI
+	emms
+#endif
+
 	fldz
 	testq	M, M
 	jle	.L999
--- a/kernel/x86_64/zscal.S
+++ b/kernel/x86_64/zscal.S
@ -50,6 +50,10 @@
 	PROLOGUE
 	PROFCODE

+#ifdef WINDOWS_ABI
+	emms
+#endif
+
 	salq	$ZBASE_SHIFT, INCX

 	FLD	 8(%rsp)
--- a/kernel/x86_64/zsum.S
+++ b/kernel/x86_64/zsum.S
@ -50,6 +50,10 @@
 	PROLOGUE
 	PROFCODE

+#ifdef WINDOWS_ABI
+	emms
+#endif
+
 	fldz
 	testq	M, M
 	jle	.L999