From 6d2bd7f11dcd292461bbd66e487e4367f05f9fe8 Mon Sep 17 00:00:00 2001
From: Mark Wooding <mdw@distorted.org.uk>
Date: Wed, 30 Oct 2019 23:10:51 +0000
Subject: [PATCH] x86ish *.S: Use `stalloc' consistently to allocate space on
 the stack.

Also, prefer aligning afterwards, except where that won't work.
---
 math/mpx-mul4-x86-sse2.S   | 8 ++++----
 symm/chacha-x86ish-sse2.S  | 2 +-
 symm/gcm-x86ish-pclmul.S   | 6 +++---
 symm/salsa20-x86ish-sse2.S | 2 +-
 4 files changed, 9 insertions(+), 9 deletions(-)

diff --git a/math/mpx-mul4-x86-sse2.S b/math/mpx-mul4-x86-sse2.S
index ba7ae6a3..0e90541e 100644
--- a/math/mpx-mul4-x86-sse2.S
+++ b/math/mpx-mul4-x86-sse2.S
@@ -706,8 +706,8 @@ FUNC(mpx_umul4_x86_sse2)
 	pushreg	esi
 	pushreg	edi
 	setfp
+	stalloc	32
 	and	SP, ~15
-	sub	SP, 32
   endprologue
 
 	// Prepare for the first iteration.
@@ -821,8 +821,8 @@ FUNC(mpxmont_mul4_x86_sse2)
 	pushreg	esi
 	pushreg	edi
 	setfp
+	stalloc	112
 	and	SP, ~15
-	sub	SP, 112
   endprologue
 
 	// Establish the expanded operands.
@@ -968,7 +968,7 @@ FUNC(mpxmont_redc4_x86_sse2)
 	pushreg	edi
 	setfp
 	and	SP, ~15
-	sub	SP, 76
+	stalloc	76
   endprologue
 
 	// Establish the expanded operands and the blocks-of-4 dv limit.
@@ -1096,8 +1096,8 @@ ENDFUNC
 	pushreg	esi
 	pushreg	edi
 	setfp
+	stalloc	3*32 + 4*4
 	and	SP, ~15
-	sub	SP, 3*32 + 4*4
   endprologue
 	mov	eax, \n
 	mov	[SP + 104], eax
diff --git a/symm/chacha-x86ish-sse2.S b/symm/chacha-x86ish-sse2.S
index 33af65f0..974ec5b5 100644
--- a/symm/chacha-x86ish-sse2.S
+++ b/symm/chacha-x86ish-sse2.S
@@ -70,7 +70,7 @@ FUNC(chacha_core_x86ish_sse2)
 
 	pushreg	BP
 	setfp
-	sub	SP, 16
+	stalloc	16
 	mov	IN, [BP + 12]
 	mov	OUT, [BP + 16]
 	and	SP, ~15
diff --git a/symm/gcm-x86ish-pclmul.S b/symm/gcm-x86ish-pclmul.S
index 092242bc..5edf56e1 100644
--- a/symm/gcm-x86ish-pclmul.S
+++ b/symm/gcm-x86ish-pclmul.S
@@ -974,8 +974,8 @@ SSEFUNC(gcm_mulk_256b_x86ish_pclmul)
 	setfp
 	mov	A, [SP + 8]
 	mov	K, [SP + 12]
+	stalloc	16
 	and	SP, ~15
-	sub	SP, 16
 #endif
 #if CPUFAM_AMD64 && ABI_WIN
 	stalloc	3*16 + 8
@@ -1018,9 +1018,9 @@ SSEFUNC(gcm_mulk_256l_x86ish_pclmul)
 	setfp
 	mov	A, [SP + 8]
 	mov	K, [SP + 12]
-	and	SP, ~15
+	stalloc	16
 	ldgot	ecx
-	sub	SP, 16
+	and	SP, ~15
 #endif
 #if CPUFAM_AMD64 && ABI_WIN
 	stalloc	3*16 + 8
diff --git a/symm/salsa20-x86ish-sse2.S b/symm/salsa20-x86ish-sse2.S
index eb346afe..26bab892 100644
--- a/symm/salsa20-x86ish-sse2.S
+++ b/symm/salsa20-x86ish-sse2.S
@@ -70,7 +70,7 @@ FUNC(salsa20_core_x86ish_sse2)
 
 	pushreg	BP
 	setfp
-	sub	SP, 32
+	stalloc	32
 	mov	IN, [BP + 12]
 	mov	OUT, [BP + 16]
 	and	SP, ~15
-- 
2.11.0