Refs #154. Fixed a SEGFAULT bug of dgemv_t when m is very large.

It overflowed the internal buffer. Thus, we split vector x into blocks when m is very large.

Thank @wangqian for this patch.

diff --git a/kernel/x86_64/dgemv_t.S b/kernel/x86_64/dgemv_t.S
index 0719207..02601be 100644 (file)
--- a/kernel/x86_64/dgemv_t.S
+++ b/kernel/x86_64/dgemv_t.S
@@ -47,7 +47,7 @@
 
 #ifndef WINDOWS_ABI
 
-#define STACKSIZE      64
+#define STACKSIZE      128
        
 #define OLD_M    %rdi
 #define OLD_N    %rsi
@@ -57,7 +57,10 @@
 #define STACK_Y                16 + STACKSIZE(%rsp)
 #define STACK_INCY     24 + STACKSIZE(%rsp)
 #define STACK_BUFFER   32 + STACKSIZE(%rsp)
-
+#define MMM    56(%rsp)
+#define NN     64(%rsp)
+#define AA     72(%rsp)
+#define LDAX   80(%rsp)
 #else
 
 #define STACKSIZE      256
@@ -132,12 +135,44 @@
        movq    OLD_LDA,      LDA
        movq    OLD_X,        X
 #else
-       movq    OLD_M,        M
-       movq    OLD_N,        N
-       movq    OLD_A,        A
-       movq    OLD_LDA,      LDA
+       movq    OLD_M,        MMM
+       movq    OLD_N,        NN
+       movq    OLD_A,        AA
+       movq    OLD_LDA,      LDAX
+#endif
+#ifdef HAVE_SSE3
+#ifndef WINDOWS_ABI
+       movddup %xmm0, ALPHA
+#else
+       movddup %xmm3, ALPHA
 #endif
+#else
+#ifndef WINDOWS_ABI
+       movapd  %xmm0, ALPHA
+#else
+       movapd  %xmm3, ALPHA
+#endif
+       unpcklpd ALPHA, ALPHA
+#endif
+
+
 
+.L0x:
+       xorq    M,M
+       addq    $1,M
+       salq    $22,M
+       subq    M,MMM
+       jge .L00
+
+       movq    MMM,%rax
+       addq    M,%rax
+       jle     .L999x
+       movq    %rax,M
+
+.L00:  
+       movq    LDAX,LDA
+       movq    NN,N
+       movq    AA,A
        movq    STACK_INCX,   INCX
        movq    STACK_Y,      Y
        movq    STACK_INCY,   INCY
@@ -153,21 +188,6 @@
 
        subq    $-16 * SIZE, A
 
-#ifdef HAVE_SSE3
-#ifndef WINDOWS_ABI
-       movddup %xmm0, ALPHA
-#else
-       movddup %xmm3, ALPHA
-#endif
-#else
-#ifndef WINDOWS_ABI
-       movapd  %xmm0, ALPHA
-#else
-       movapd  %xmm3, ALPHA
-#endif
-       unpcklpd ALPHA, ALPHA
-#endif
-
        testq   M, M
        jle     .L999
        testq   N, N
@@ -854,7 +874,6 @@
 
 .L21:
 #endif
-
        subq    $4, N
 
        leaq    16 * SIZE(BUFFER), X1
@@ -2461,6 +2480,12 @@
        ALIGN_4
 
 .L999:
+       leaq    (, M, SIZE), %rax
+       addq %rax,AA
+       jmp .L0x;
+       ALIGN_4
+
+.L999x:
        movq      0(%rsp), %rbx
        movq      8(%rsp), %rbp
        movq     16(%rsp), %r12