Roll src/third_party/boringssl/src dc8c1d962..0f28691d3

third_party/boringssl/mac-x86_64/crypto/bn/x86_64-mont5.S

Index: third_party/boringssl/mac-x86_64/crypto/bn/x86_64-mont5.S
diff --git a/third_party/boringssl/mac-x86_64/crypto/bn/x86_64-mont5.S b/third_party/boringssl/mac-x86_64/crypto/bn/x86_64-mont5.S
index a154cc8dfd550c58da1614d78e171bd08ca8800a..4646c3d9e2b24cfe08a4c28585610a26fbc46174 100644
--- a/third_party/boringssl/mac-x86_64/crypto/bn/x86_64-mont5.S
+++ b/third_party/boringssl/mac-x86_64/crypto/bn/x86_64-mont5.S
@@ -8,16 +8,15 @@
 
 .p2align	6
 _bn_mul_mont_gather5:
+	movl	%r9d,%r9d
+	movq	%rsp,%rax
 	testl	$7,%r9d
 	jnz	L$mul_enter
 	jmp	L$mul4x_enter
 
 .p2align	4
 L$mul_enter:
-	movl	%r9d,%r9d
-	movq	%rsp,%rax
 	movd	8(%rsp),%xmm5
-	leaq	L$inc(%rip),%r10
 	pushq	%rbx
 	pushq	%rbp
 	pushq	%r12
@@ -25,13 +24,39 @@ L$mul_enter:
 	pushq	%r14
 	pushq	%r15
 
-	leaq	2(%r9),%r11
-	negq	%r11
-	leaq	-264(%rsp,%r11,8),%rsp
-	andq	$-1024,%rsp
+	negq	%r9
+	movq	%rsp,%r11
+	leaq	-280(%rsp,%r9,8),%r10
+	negq	%r9
+	andq	$-1024,%r10
+
+
 
+
+
+
+
+
+
+	subq	%r10,%r11
+	andq	$-4096,%r11
+	leaq	(%r10,%r11,1),%rsp
+	movq	(%rsp),%r11
+	cmpq	%r10,%rsp
+	ja	L$mul_page_walk
+	jmp	L$mul_page_walk_done
+
+L$mul_page_walk:
+	leaq	-4096(%rsp),%rsp
+	movq	(%rsp),%r11
+	cmpq	%r10,%rsp
+	ja	L$mul_page_walk
+L$mul_page_walk_done:
+
+	leaq	L$inc(%rip),%r10
 	movq	%rax,8(%rsp,%r9,8)
 L$mul_body:
+
 	leaq	128(%rdx),%r12
 	movdqa	0(%r10),%xmm0
 	movdqa	16(%r10),%xmm1
@@ -370,16 +395,17 @@ L$sub:	sbbq	(%rcx,%r14,8),%rax
 
 	sbbq	$0,%rax
 	xorq	%r14,%r14
+	andq	%rax,%rsi
+	notq	%rax
+	movq	%rdi,%rcx
+	andq	%rax,%rcx
 	movq	%r9,%r15
+	orq	%rcx,%rsi
 .p2align	4
 L$copy:
-	movq	(%rsp,%r14,8),%rsi
-	movq	(%rdi,%r14,8),%rcx
-	xorq	%rcx,%rsi
-	andq	%rax,%rsi
-	xorq	%rcx,%rsi
+	movq	(%rsi,%r14,8),%rax
 	movq	%r14,(%rsp,%r14,8)
-	movq	%rsi,(%rdi,%r14,8)
+	movq	%rax,(%rdi,%r14,8)
 	leaq	1(%r14),%r14
 	subq	$1,%r15
 	jnz	L$copy
@@ -400,15 +426,16 @@ L$mul_epilogue:
 
 .p2align	5
 bn_mul4x_mont_gather5:
-L$mul4x_enter:
 .byte	0x67
 	movq	%rsp,%rax
+L$mul4x_enter:
 	pushq	%rbx
 	pushq	%rbp
 	pushq	%r12
 	pushq	%r13
 	pushq	%r14
 	pushq	%r15
+L$mul4x_prologue:
 
 .byte	0x67
 	shll	$3,%r9d
@@ -425,24 +452,41 @@ L$mul4x_enter:
 
 
 	leaq	-320(%rsp,%r9,2),%r11
+	movq	%rsp,%rbp
 	subq	%rdi,%r11
 	andq	$4095,%r11
 	cmpq	%r11,%r10
 	jb	L$mul4xsp_alt
-	subq	%r11,%rsp
-	leaq	-320(%rsp,%r9,2),%rsp
+	subq	%r11,%rbp
+	leaq	-320(%rbp,%r9,2),%rbp
 	jmp	L$mul4xsp_done
 
 .p2align	5
 L$mul4xsp_alt:
 	leaq	4096-320(,%r9,2),%r10
-	leaq	-320(%rsp,%r9,2),%rsp
+	leaq	-320(%rbp,%r9,2),%rbp
 	subq	%r10,%r11
 	movq	$0,%r10
 	cmovcq	%r10,%r11
-	subq	%r11,%rsp
+	subq	%r11,%rbp
 L$mul4xsp_done:
-	andq	$-64,%rsp
+	andq	$-64,%rbp
+	movq	%rsp,%r11
+	subq	%rbp,%r11
+	andq	$-4096,%r11
+	leaq	(%r11,%rbp,1),%rsp
+	movq	(%rsp),%r10
+	cmpq	%rbp,%rsp
+	ja	L$mul4x_page_walk
+	jmp	L$mul4x_page_walk_done
+
+L$mul4x_page_walk:
+	leaq	-4096(%rsp),%rsp
+	movq	(%rsp),%r10
+	cmpq	%rbp,%rsp
+	ja	L$mul4x_page_walk
+L$mul4x_page_walk_done:
+
 	negq	%r9
 
 	movq	%rax,40(%rsp)
@@ -1001,6 +1045,7 @@ _bn_power5:
 	pushq	%r13
 	pushq	%r14
 	pushq	%r15
+L$power5_prologue:
 
 	shll	$3,%r9d
 	leal	(%r9,%r9,2),%r10d
@@ -1015,24 +1060,41 @@ _bn_power5:
 
 
 	leaq	-320(%rsp,%r9,2),%r11
+	movq	%rsp,%rbp
 	subq	%rdi,%r11
 	andq	$4095,%r11
 	cmpq	%r11,%r10
 	jb	L$pwr_sp_alt
-	subq	%r11,%rsp
-	leaq	-320(%rsp,%r9,2),%rsp
+	subq	%r11,%rbp
+	leaq	-320(%rbp,%r9,2),%rbp
 	jmp	L$pwr_sp_done
 
 .p2align	5
 L$pwr_sp_alt:
 	leaq	4096-320(,%r9,2),%r10
-	leaq	-320(%rsp,%r9,2),%rsp
+	leaq	-320(%rbp,%r9,2),%rbp
 	subq	%r10,%r11
 	movq	$0,%r10
 	cmovcq	%r10,%r11
-	subq	%r11,%rsp
+	subq	%r11,%rbp
 L$pwr_sp_done:
-	andq	$-64,%rsp
+	andq	$-64,%rbp
+	movq	%rsp,%r11
+	subq	%rbp,%r11
+	andq	$-4096,%r11
+	leaq	(%r11,%rbp,1),%rsp
+	movq	(%rsp),%r10
+	cmpq	%rbp,%rsp
+	ja	L$pwr_page_walk
+	jmp	L$pwr_page_walk_done
+
+L$pwr_page_walk:
+	leaq	-4096(%rsp),%rsp
+	movq	(%rsp),%r10
+	cmpq	%rbp,%rsp
+	ja	L$pwr_page_walk
+L$pwr_page_walk_done:
+
 	movq	%r9,%r10
 	negq	%r9
 
@@ -1943,6 +2005,7 @@ bn_from_mont8x:
 	pushq	%r13
 	pushq	%r14
 	pushq	%r15
+L$from_prologue:
 
 	shll	$3,%r9d
 	leaq	(%r9,%r9,2),%r10
@@ -1957,24 +2020,41 @@ bn_from_mont8x:
 
 
 	leaq	-320(%rsp,%r9,2),%r11
+	movq	%rsp,%rbp
 	subq	%rdi,%r11
 	andq	$4095,%r11
 	cmpq	%r11,%r10
 	jb	L$from_sp_alt
-	subq	%r11,%rsp
-	leaq	-320(%rsp,%r9,2),%rsp
+	subq	%r11,%rbp
+	leaq	-320(%rbp,%r9,2),%rbp
 	jmp	L$from_sp_done
 
 .p2align	5
 L$from_sp_alt:
 	leaq	4096-320(,%r9,2),%r10
-	leaq	-320(%rsp,%r9,2),%rsp
+	leaq	-320(%rbp,%r9,2),%rbp
 	subq	%r10,%r11
 	movq	$0,%r10
 	cmovcq	%r10,%r11
-	subq	%r11,%rsp
+	subq	%r11,%rbp
 L$from_sp_done:
-	andq	$-64,%rsp
+	andq	$-64,%rbp
+	movq	%rsp,%r11
+	subq	%rbp,%r11
+	andq	$-4096,%r11
+	leaq	(%r11,%rbp,1),%rsp
+	movq	(%rsp),%r10
+	cmpq	%rbp,%rsp
+	ja	L$from_page_walk
+	jmp	L$from_page_walk_done
+
+L$from_page_walk:
+	leaq	-4096(%rsp),%rsp
+	movq	(%rsp),%r10
+	cmpq	%rbp,%rsp
+	ja	L$from_page_walk
+L$from_page_walk_done:
+
 	movq	%r9,%r10
 	negq	%r9