Add window alignment selection for s390x DFLTCC support.

deflate.c

@@ -71,7 +71,14 @@ const char PREFIX(deflate_copyright)[] = " deflate 1.3.1 Copyright 1995-2024 Jea
  */
 #ifdef S390_DFLTCC_DEFLATE
 #  include "arch/s390/dfltcc_deflate.h"
+/* DFLTCC instructions require window to be page-aligned */
+#  define PAD_WINDOW            PAD_4096
+#  define WINDOW_PAD_SIZE       4096
+#  define HINT_ALIGNED_WINDOW   HINT_ALIGNED_4096
 #else
+#  define PAD_WINDOW            PAD_64
+#  define WINDOW_PAD_SIZE       64
+#  define HINT_ALIGNED_WINDOW   HINT_ALIGNED_64
 /* Invoked at the beginning of deflateSetDictionary(). Useful for checking arch-specific window data. */
 #  define DEFLATE_SET_DICTIONARY_HOOK(strm, dict, dict_len) do {} while (0)
 /* Invoked at the beginning of deflateGetDictionary(). Useful for adjusting arch-specific window data. */
@@ -202,8 +209,8 @@ Z_INTERNAL deflate_allocs* alloc_deflate(PREFIX3(stream) *strm, int windowBits,
     int alloc_size = sizeof(deflate_allocs);
 
     /* Calculate relative buffer positions and paddings */
-    LOGSZP("window", window_size, PAD_64(currsize), PADSZ(currsize,64));
-    int window_pos = PAD_64(currsize);
+    LOGSZP("window", window_size, PAD_WINDOW(currsize), PADSZ(currsize,WINDOW_PAD_SIZE));
+    int window_pos = PAD_WINDOW(currsize);
     currsize = window_pos + window_size;
 
     LOGSZP("prev", prev_size, PAD_64(currsize), PADSZ(currsize,64));
@@ -226,24 +233,24 @@ Z_INTERNAL deflate_allocs* alloc_deflate(PREFIX3(stream) *strm, int windowBits,
     int alloc_pos = PAD_16(currsize);
     currsize = alloc_pos + alloc_size;
 
-    /* Add 64 to allow our alignment, and round size of buffer up to multiple of 64 */
-    int totalsize = PAD_64(63 + currsize);
+    /* Add 64-1 or 4096-1 to allow window alignment, and round size of buffer up to multiple of 64 */
+    int totalsize = PAD_64(currsize + (WINDOW_PAD_SIZE - 1));
 
     /* Allocate buffer, align to 64-byte cacheline, and zerofill the resulting buffer */
     char *origbuff = strm->zalloc(strm->opaque, 1, totalsize);
     if (origbuff == NULL)
         return NULL;
 
-    char *buff = (char *)HINT_ALIGNED_64((char *)PAD_64(origbuff));
-    LOGSZPL("Buffer alloc", totalsize, PADSZ((uintptr_t)origbuff,64), PADSZ(currsize,64));
+    char *buff = (char *)HINT_ALIGNED_WINDOW((char *)PAD_WINDOW(origbuff));
+    LOGSZPL("Buffer alloc", totalsize, PADSZ((uintptr_t)origbuff,WINDOW_PAD_SIZE), PADSZ(currsize,WINDOW_PAD_SIZE));
 
     /* Initialize alloc_bufs */
     deflate_allocs *alloc_bufs  = (struct deflate_allocs_s *)(buff + alloc_pos);
     alloc_bufs->buf_start = (char *)origbuff;
     alloc_bufs->zfree = strm->zfree;
 
     /* Assign buffers */
-    alloc_bufs->window = (unsigned char *)HINT_ALIGNED_64(buff + window_pos);
+    alloc_bufs->window = (unsigned char *)HINT_ALIGNED_WINDOW(buff + window_pos);
     alloc_bufs->prev = (Pos *)HINT_ALIGNED_64(buff + prev_pos);
     alloc_bufs->head = (Pos *)HINT_ALIGNED_64(buff + head_pos);
     alloc_bufs->pending_buf = (unsigned char *)HINT_ALIGNED_64(buff + pending_pos);

inflate.c

@@ -154,8 +154,8 @@ Z_INTERNAL inflate_allocs* alloc_inflate(PREFIX3(stream) *strm) {
     int alloc_size = sizeof(inflate_allocs);
 
     /* Calculate relative buffer positions and paddings */
-    LOGSZP("window", window_size, PAD_64(currsize), PADSZ(currsize,64));
-    int window_pos = PAD_64(currsize);
+    LOGSZP("window", window_size, PAD_WINDOW(currsize), PADSZ(currsize,WINDOW_PAD_SIZE));
+    int window_pos = PAD_WINDOW(currsize);
     currsize = window_pos + window_size;
 
     LOGSZP("state", state_size, PAD_64(currsize), PADSZ(currsize,64));
@@ -166,23 +166,23 @@ Z_INTERNAL inflate_allocs* alloc_inflate(PREFIX3(stream) *strm) {
     int alloc_pos = PAD_16(currsize);
     currsize = alloc_pos + alloc_size;
 
-    /* Add 64 to allow our alignment, and round size of buffer up to multiple of 64 */
-    int totalsize = PAD_64(currsize + 63);
+    /* Add 64-1 or 4096-1 to allow window alignment, and round size of buffer up to multiple of 64 */
+    int totalsize = PAD_64(currsize + (WINDOW_PAD_SIZE - 1));
 
     /* Allocate buffer, align to 64-byte cacheline, and zerofill the resulting buffer */
     char *origbuff = strm->zalloc(strm->opaque, 1, totalsize);
     if (origbuff == NULL)
         return NULL;
 
-    char *buff = (char *)HINT_ALIGNED_64((char *)PAD_64(origbuff));
-    LOGSZPL("Buffer alloc", totalsize, PADSZ((uintptr_t)origbuff,64), PADSZ(currsize,64));
+    char *buff = (char *)HINT_ALIGNED_WINDOW((char *)PAD_WINDOW(origbuff));
+    LOGSZPL("Buffer alloc", totalsize, PADSZ((uintptr_t)origbuff,WINDOW_PAD_SIZE), PADSZ(currsize,WINDOW_PAD_SIZE));
 
     /* Initialize alloc_bufs */
     inflate_allocs *alloc_bufs  = (struct inflate_allocs_s *)(buff + alloc_pos);
     alloc_bufs->buf_start = (char *)origbuff;
     alloc_bufs->zfree = strm->zfree;
 
-    alloc_bufs->window =  (unsigned char *)HINT_ALIGNED_64((buff + window_pos));
+    alloc_bufs->window =  (unsigned char *)HINT_ALIGNED_WINDOW((buff + window_pos));
     alloc_bufs->state = (inflate_state *)HINT_ALIGNED_64((buff + state_pos));
 
 #ifdef Z_MEMORY_SANITIZER

inflate_p.h

@@ -10,7 +10,14 @@
 /* Architecture-specific hooks. */
 #ifdef S390_DFLTCC_INFLATE
 #  include "arch/s390/dfltcc_inflate.h"
+/* DFLTCC instructions require window to be page-aligned */
+#  define PAD_WINDOW            PAD_4096
+#  define WINDOW_PAD_SIZE       4096
+#  define HINT_ALIGNED_WINDOW   HINT_ALIGNED_4096
 #else
+#  define PAD_WINDOW            PAD_64
+#  define WINDOW_PAD_SIZE       64
+#  define HINT_ALIGNED_WINDOW   HINT_ALIGNED_64
 /* Memory management for the window. Useful for allocation the aligned window. */
 #  define ZCOPY_WINDOW(dest, src, n) memcpy(dest, src, n)
 /* Invoked at the end of inflateResetKeep(). Useful for initializing arch-specific extension blocks. */

zbuild.h

@@ -207,16 +207,18 @@
 #else
 #  define HINT_ALIGNED(p,n) (p)
 #endif
-#define HINT_ALIGNED_64(p) HINT_ALIGNED((p),64)
 #define HINT_ALIGNED_16(p) HINT_ALIGNED((p),16)
+#define HINT_ALIGNED_64(p) HINT_ALIGNED((p),64)
+#define HINT_ALIGNED_4096(p) HINT_ALIGNED((p),4096)
 
 /* PADSZ returns needed bytes to pad bpos to pad size
  * PAD_NN calculates pad size and adds it to bpos, returning the result.
  * All take an integer or a pointer as bpos input.
  */
 #define PADSZ(bpos, pad) (((pad) - ((uintptr_t)(bpos) % (pad))) % (pad))
-#define PAD_64(bpos) ((bpos) + PADSZ((bpos),64))
 #define PAD_16(bpos) ((bpos) + PADSZ((bpos),16))
+#define PAD_64(bpos) ((bpos) + PADSZ((bpos),64))
+#define PAD_4096(bpos) ((bpos) + PADSZ((bpos),4096))
 
 /* Diagnostic functions */
 #ifdef ZLIB_DEBUG