ish: save/restore FPU context only for the task uses FPU

Currently we save/retore FPU H/W context for every task on
every contxt switch. This hurts overall performance of ISH.

This patch allows save and restore FPU H/W context only for
a task that declares it uses FPU.

BRANCH=none
BUG=none
TEST=verified in Atlas platform

Change-Id: Ic2f0bbf59f655661e2dd788c688edc4e83068c1c
Signed-off-by: Hyungwoo Yang <hyungwoo.yang@intel.com>
Reviewed-on: https://chromium-review.googlesource.com/1448818
Commit-Ready: ChromeOS CL Exonerator Bot <chromiumos-cl-exonerator@appspot.gserviceaccount.com>
Tested-by: Hyungwoo Yang <hyungwoo.yang@intel.corp-partner.google.com>
Reviewed-by: Jett Rink <jettrink@chromium.org>

5 files changed, 73 insertions, 14 deletions

diff --git a/core/minute-ia/config_core.h b/core/minute-ia/config_core.h
index 20f4404d38..47121642a4 100644
--- a/core/minute-ia/config_core.h
+++ b/core/minute-ia/config_core.h
@@ -26,4 +26,9 @@
 #define ASM_LOCK_PREFIX
 #endif
 
+/*
+ * Flag indicates the task uses FPU H/W
+ */
+#define MIA_TASK_FLAG_USE_FPU                   0x00000001
+
 #endif /* __CROS_EC_CONFIG_CORE_H */
diff --git a/core/minute-ia/irq_handler.h b/core/minute-ia/irq_handler.h
index c561ee472b..4df2b14da6 100644
--- a/core/minute-ia/irq_handler.h
+++ b/core/minute-ia/irq_handler.h
@@ -9,10 +9,20 @@
 #define __CROS_EC_IRQ_HANDLER_H
 
 #include "registers.h"
+#include "task_defs.h"
 
 #ifdef CONFIG_FPU
-#define save_fpu_ctx	"fnsave 20(%eax)\n"
-#define rstr_fpu_ctx	"frstor 20(%eax)\n"
+#define save_fpu_ctx	"movl "USE_FPU_OFFSET_STR"(%eax), %ebx\n"	\
+			"test %ebx, %ebx\n"				\
+			"jz 9f\n"					\
+			"fnsave "FPU_CTX_OFFSET_STR"(%eax)\n"		\
+			"9:\n"
+
+#define rstr_fpu_ctx	"movl "USE_FPU_OFFSET_STR"(%eax), %ebx\n"	\
+			"test %ebx, %ebx\n"				\
+			"jz 9f\n"					\
+			"frstor "FPU_CTX_OFFSET_STR"(%eax)\n"		\
+			"9:\n"
 #else
 #define save_fpu_ctx
 #define rstr_fpu_ctx
diff --git a/core/minute-ia/switch.S b/core/minute-ia/switch.S
index eed2f7929f..27aba1a7cc 100644
--- a/core/minute-ia/switch.S
+++ b/core/minute-ia/switch.S
@@ -33,7 +33,11 @@ __task_start:
 	movl current_task, %eax
 	movl (%eax), %esp
 #ifdef CONFIG_FPU
-	frstor 20(%eax)
+	movl USE_FPU_OFFSET(%eax), %ebx
+	test %ebx, %ebx
+	jz 1f
+	frstor FPU_CTX_OFFSET(%eax)
+	1:
 #endif
 	movl $0x1, (%ebx)       # first task is ready. set start_called = 1
 	popa
@@ -112,7 +116,11 @@ __switchto:
 	movl current_task, %eax
 
 #ifdef CONFIG_FPU
-	fnsave 20(%eax)		# Save current FPU context at current->fp_ctx
+	movl USE_FPU_OFFSET(%eax), %ebx
+	test %ebx, %ebx
+	jz 2f
+	fnsave FPU_CTX_OFFSET(%eax) # Save current FPU context(current->fp_ctx)
+	2:
 #endif
 
 	# Save SP of current task and switch to new task
@@ -122,7 +130,10 @@ __switchto:
 	movl (%eax), %esp
 
 #ifdef CONFIG_FPU
-	frstor 20(%eax)		# Restore next FPU context
+	movl USE_FPU_OFFSET(%eax), %ebx
+	test %ebx, %ebx
+	jz 1f
+	frstor FPU_CTX_OFFSET(%eax)		# Restore next FPU context
 #endif
 
 	1:
diff --git a/core/minute-ia/task.c b/core/minute-ia/task.c
index 2836d40975..9975cc62d7 100644
--- a/core/minute-ia/task.c
+++ b/core/minute-ia/task.c
@@ -26,7 +26,7 @@
 #define STACK_UNUSED_VALUE 0xdeadd00d
 
 /* declare task routine prototypes */
-#define TASK(n, r, d, s) void r(void *);
+#define TASK(n, r, d, s, f) void r(void *);
 void __idle(void);
 CONFIG_TASK_LIST
 CONFIG_TEST_TASK_LIST
@@ -36,7 +36,7 @@ CONFIG_TEST_TASK_LIST
 extern volatile uint32_t __in_isr;
 
 /* Task names for easier debugging */
-#define TASK(n, r, d, s)  #n,
+#define TASK(n, r, d, s, f)  #n,
 static const char * const task_names[] = {
 	"<< idle >>",
 	CONFIG_TASK_LIST
@@ -90,17 +90,19 @@ static void task_exit_trap(void)
 }
 
 /* Startup parameters for all tasks. */
-#define TASK(n, r, d, s)  {	\
+#define TASK(n, r, d, s, f)  {	\
 	.r0 = (uint32_t)d,	\
 	.pc = (uint32_t)r,	\
 	.stack_size = s,	\
+	.flags = f,		\
 },
 static const struct {
 	uint32_t r0;
 	uint32_t pc;
 	uint16_t stack_size;
+	uint32_t flags;
 } tasks_init[] = {
-	TASK(IDLE, __idle, 0, IDLE_TASK_STACK_SIZE)
+	TASK(IDLE, __idle, 0, IDLE_TASK_STACK_SIZE, 0)
 	CONFIG_TASK_LIST
 	CONFIG_TEST_TASK_LIST
 };
@@ -115,9 +117,9 @@ BUILD_ASSERT(TASK_ID_COUNT < (1 << (sizeof(task_id_t) * 8)));
 
 
 /* Stacks for all tasks */
-#define TASK(n, r, d, s)  + s
+#define TASK(n, r, d, s, f)  + s
 uint8_t task_stacks[0
-		    TASK(IDLE, __idle, 0, IDLE_TASK_STACK_SIZE)
+		    TASK(IDLE, __idle, 0, IDLE_TASK_STACK_SIZE, 0)
 		    CONFIG_TASK_LIST
 		    CONFIG_TEST_TASK_LIST
 ] __aligned(8);
@@ -474,10 +476,18 @@ void task_print_list(void)
 {
 	int i;
 
+#ifdef CONFIG_FPU
+	ccputs("Task Ready Name         Events      Time (s)  "
+	       "  StkUsed UseFPU\n");
+#else
 	ccputs("Task Ready Name         Events      Time (s)  StkUsed\n");
+#endif
 
 	for (i = 0; i < TASK_ID_COUNT; i++) {
 		char is_ready = (tasks_ready & (1<<i)) ? 'R' : ' ';
+#ifdef CONFIG_FPU
+		char use_fpu = tasks[i].use_fpu ? 'Y' : 'N';
+#endif
 		uint32_t *sp;
 
 		int stackused = tasks_init[i].stack_size;
@@ -487,9 +497,16 @@ void task_print_list(void)
 		     sp++)
 			stackused -= sizeof(uint32_t);
 
+#ifdef CONFIG_FPU
+		ccprintf("%4d %c %-16s %08x %11.6ld  %3d/%3d %c\n", i, is_ready,
+			 task_names[i], tasks[i].events, tasks[i].runtime,
+			 stackused, tasks_init[i].stack_size, use_fpu);
+#else
 		ccprintf("%4d %c %-16s %08x %11.6ld  %3d/%3d\n", i, is_ready,
 			 task_names[i], tasks[i].events, tasks[i].runtime,
 			 stackused, tasks_init[i].stack_size);
+#endif
+
 		cflush();
 	}
 }
@@ -606,6 +623,9 @@ void task_pre_init(void)
 		/* Copy default x86 FPU state for each task */
 		memcpy(tasks[i].fp_ctx, default_fp_ctx,
 			sizeof(default_fp_ctx));
+
+		if (tasks_init[i].flags & MIA_TASK_FLAG_USE_FPU)
+			tasks[i].use_fpu = 1;
 #endif
 		/* Fill unused stack; also used to detect stack overflow. */
 		for (sp = stack_next; sp < (uint32_t *)tasks[i].sp; sp++)
@@ -618,7 +638,6 @@ void task_pre_init(void)
 
 	/* Initialize IRQs */
 	init_interrupts();
-
 }
 
 void task_clear_fp_used(void)
diff --git a/core/minute-ia/task_defs.h b/core/minute-ia/task_defs.h
index 2ec33e5c98..15aca6b8b5 100644
--- a/core/minute-ia/task_defs.h
+++ b/core/minute-ia/task_defs.h
@@ -6,8 +6,21 @@
 #ifndef __CROS_EC_TASK_DEFS_H
 #define __CROS_EC_TASK_DEFS_H
 
-#define FPU_CTX_SZ		108 /* 28 bytes header + 80 bytes registers */
-#define FPU_CTX_OFFSET		20  /* offsetof(task_, fp_ctx) */
+#ifdef CONFIG_FPU
+#define FPU_CTX_SZ		108  /* 28 bytes header + 80 bytes registers */
+#define USE_FPU_OFFSET		20   /* offsetof(task_, use_fpu */
+#define FPU_CTX_OFFSET		24   /* offsetof(task_, fp_ctx) */
+
+/*
+ * defines for inline asm
+ */
+#ifndef __ASSEMBLER__
+#include "common.h"
+
+#define USE_FPU_OFFSET_STR	STRINGIFY(USE_FPU_OFFSET)	/* "20" */
+#define FPU_CTX_OFFSET_STR	STRINGIFY(FPU_CTX_OFFSET)	/* "24" */
+#endif
+#endif /* CONFIG_FPU */
 
 #ifndef __ASSEMBLER__
 typedef union {
@@ -21,6 +34,7 @@ typedef union {
 		uint64_t runtime;	/* Time spent in task */
 		uint32_t *stack;	/* Start of stack */
 #ifdef CONFIG_FPU
+		uint32_t use_fpu;	/* set if task uses FPU */
 		uint8_t fp_ctx[FPU_CTX_SZ]; /* x87 FPU context */
 #endif
 	};