X-Git-Url: https://git.distorted.org.uk/~mdw/catacomb/blobdiff_plain/a117c06f5ee62cbe7812769703eada01843f76ca..8c5956c14f5834a072e1a9345ae1f356b14164ca:/math/mpx-mul4-amd64-sse2.S

diff --git a/math/mpx-mul4-amd64-sse2.S b/math/mpx-mul4-amd64-sse2.S
index 64460ca9..d313765f 100644
--- a/math/mpx-mul4-amd64-sse2.S
+++ b/math/mpx-mul4-amd64-sse2.S
@@ -25,15 +25,13 @@
 /// MA 02111-1307, USA.
 
 ///--------------------------------------------------------------------------
-/// External definitions.
+/// Preliminaries.
 
 #include "config.h"
 #include "asm-common.h"
 
-///--------------------------------------------------------------------------
-/// Prologue.
-
 	.arch	pentium4
+
 	.text
 
 ///--------------------------------------------------------------------------
@@ -1329,7 +1327,7 @@ ENDFUNC
 #  define ARG6 STKARG(2)
 #  define ARG7 STKARG(3)
 #  define ARG8 STKARG(4)
-#  define STKARG_OFFSET 40
+#  define STKARG_OFFSET 224
 #endif
 #define STKARG(i) [rsp + STKARG_OFFSET + 8*(i)]
 
@@ -1386,7 +1384,7 @@ ENDFUNC
 	mov	rbx, r8
 	movdqu	xmm8, [r9]
 	movdqu	xmm10, [rax]
-	mov	r8, STKARG(1)
+	mov	r8d, STKARG(1)
 	mov	r9, STKARG(2)
 	mov	r10, rdx
 	mov	r11, rcx
@@ -1395,7 +1393,7 @@ ENDFUNC
   .ifeqs "\mode", "mont"
 	mov	rbx, rcx
 	movdqu	xmm8, [r8]
-	mov	r8, r9
+	mov	r8d, r9d
 	mov	r9, STKARG(0)
 	mov	r10, rdx
 	mov	rcx, rsi
@@ -1423,16 +1421,16 @@ ENDFUNC
 	mov	rbx, r9
 	movdqu	xmm8, [r10]
 	movdqu	xmm10, [r11]
-	mov	r8, STKARG(2)
-	mov	r9, STKARG(3)
 	mov	r11, r8
+	mov	r8d, STKARG(2)
+	mov	r9, STKARG(3)
   .endif
   .ifeqs "\mode", "smul"
 	mov	rdi, rcx
 	mov	rcx, rdx
 	mov	rbx, r8
 	movdqu	xmm10, [r9]
-	mov	r8, STKARG(0)
+	mov	r8d, STKARG(0)
 	mov	r9, STKARG(1)
   .endif
   .ifeqs "\mode", "mmul"
@@ -1443,10 +1441,10 @@ ENDFUNC
 	mov	rbx, STKARG(0)
 	movdqu	xmm8, [r10]
 	movdqu	xmm10, [r11]
-	mov	r8, STKARG(3)
-	mov	r9, STKARG(4)
 	mov	r10, r8
 	mov	r11, r9
+	mov	r8d, STKARG(3)
+	mov	r9, STKARG(4)
   .endif
   .ifeqs "\mode", "mont"
 	mov	r10, STKARG(0)
@@ -1454,9 +1452,9 @@ ENDFUNC
 	mov	rcx, rdx
 	mov	rbx, r9
 	movdqu	xmm8, [r10]
-	mov	r8, STKARG(1)
-	mov	r9, STKARG(2)
 	mov	r10, r8
+	mov	r8d, STKARG(1)
+	mov	r9, STKARG(2)
   .endif
 #endif
 
@@ -1550,6 +1548,16 @@ FUNC(test_mul4)
 	testepilogue
 ENDFUNC
 
+FUNC(test_mul4zc)
+	testprologue smul
+	testldcarry
+	testtop	nil
+	call	mul4zc
+	testtail
+	testcarryout
+	testepilogue
+ENDFUNC
+
 FUNC(test_mla4)
 	testprologue smul
 	testldcarry
@@ -1560,6 +1568,16 @@ FUNC(test_mla4)
 	testepilogue
 ENDFUNC
 
+FUNC(test_mla4zc)
+	testprologue smul
+	testldcarry
+	testtop	nil
+	call	mla4zc
+	testtail
+	testcarryout
+	testepilogue
+ENDFUNC
+
 FUNC(test_mmul4)
 	testprologue mmul
 	testtop	r11