Merge branch 'mcount-update'

A couple of fixes and update during investigation of #1949.

Signed-off-by: Namhyung Kim <[email protected]>

arch/x86_64/mcount.S

@@ -58,8 +58,15 @@ GLOBAL(mcount)
 	/* save rax (implicit argument for variadic functions) */
 	push %rax
 
+	/* save scratch registers due to -fipa-ra */
+	push %r10
+	push %r11
+
 	call mcount_entry
 
+	pop  %r11
+	pop  %r10
+
 	pop  %rax
 
 	/* restore original stack pointer */
@@ -88,10 +95,19 @@ END(mcount)
  */
 ENTRY(mcount_return)
 	.cfi_startproc
-	sub $48, %rsp
-	.cfi_def_cfa_offset 48
-
-	movq   %rdi,  32(%rsp)
+	sub $96, %rsp
+	.cfi_def_cfa_offset 96
+
+	/* save all caller-saved registers due to -fipa-ra */
+	movq   %r11,  80(%rsp)
+	movq   %r10,  72(%rsp)
+	movq   %r9,   64(%rsp)
+	movq   %r8,   56(%rsp)
+	movq   %rdi,  48(%rsp)
+	movq   %rsi,  40(%rsp)
+	movq   %rcx,  32(%rsp)
+
+	/* below are used to carry return value */
 	movdqu %xmm0, 16(%rsp)
 	movq   %rdx,   8(%rsp)
 	movq   %rax,   0(%rsp)
@@ -113,14 +129,21 @@ ENTRY(mcount_return)
 	movq    0(%rsp), %rsp
 
 	/* restore original return address in parent */
-	movq    %rax, 40(%rsp)
+	movq    %rax, 88(%rsp)
 
 	movq    0(%rsp), %rax
 	movq    8(%rsp), %rdx
 	movdqu 16(%rsp), %xmm0
-	movq   32(%rsp), %rdi
 
-	add $40, %rsp
+	movq   32(%rsp), %rcx
+	movq   40(%rsp), %rsi
+	movq   48(%rsp), %rdi
+	movq   56(%rsp), %r8
+	movq   64(%rsp), %r9
+	movq   72(%rsp), %r10
+	movq   80(%rsp), %r11
+
+	add    $88, %rsp
 	.cfi_def_cfa_offset 8
 	retq
 	.cfi_endproc

libmcount/mcount.c

@@ -1201,7 +1201,7 @@ void mcount_entry_filter_record(struct mcount_thread_data *mtdp, struct mcount_r
 #define FLAGS_TO_CHECK (TRIGGER_FL_RECOVER | TRIGGER_FL_TRACE_ON | TRIGGER_FL_TRACE_OFF)
 
 		if (tr->flags & FLAGS_TO_CHECK) {
-			if (tr->flags & TRIGGER_FL_RECOVER) {
+			if ((tr->flags & TRIGGER_FL_RECOVER) && !mcount_estimate_return) {
 				mcount_rstack_restore(mtdp);
 				*rstack->parent_loc = mcount_return_fn;
 				rstack->flags |= MCOUNT_FL_RECOVER;

libmcount/mcount.h

@@ -49,11 +49,10 @@ struct mcount_ret_stack {
 	unsigned long parent_ip;
 	unsigned long child_ip;
 	enum mcount_rstack_flag flags;
+	unsigned dyn_idx;
 	/* time in nsec (CLOCK_MONOTONIC) */
 	uint64_t start_time;
 	uint64_t end_time;
-	int tid;
-	unsigned dyn_idx;
 	uint64_t filter_time;
 	unsigned filter_size;
 	unsigned short depth;

libmcount/record.c

@@ -905,7 +905,7 @@ static int record_event(struct mcount_thread_data *mtdp, struct mcount_event *ev
 		rec->data += 4; /* set 'more' bit in uftrace_record */
 
 		*(uint16_t *)ptr = data_size;
-		memcpy(ptr + 2, event->data, data_size);
+		mcount_memcpy1(ptr + 2, event->data, data_size);
 	}
 
 	curr_buf->size += size;