base/rsvr.[ch]: New hack for buffering input to block-oriented functions.

[catacomb] / base / asm-common.h

diff --git a/base/asm-common.h b/base/asm-common.h
index f4c4f6e..d6a8b01 100644 (file)
--- a/base/asm-common.h
+++ b/base/asm-common.h
@@ -1,6 +1,6 @@
 /// -*- mode: asm; asm-comment-char: ?/ -*-
 ///
-/// Fancy SIMD implementation of Salsa20
+/// Common definitions for asesembler source files
 ///
 /// (c) 2015 Straylight/Edgeware
 ///
@@ -217,11 +217,11 @@ name:
 #  define INTADDR__1(addr, got) addr
 #endif
 
-// Permutations for SIMD instructions.  SHUF(D, C, B, A) is an immediate,
-// suitable for use in `pshufd' or `shufpd', which copies element D
-// (0 <= D < 4) of the source to element 3 of the destination, element C to
-// element 2, element B to element 1, and element A to element 0.
-#define SHUF(d, c, b, a) (64*(d) + 16*(c) + 4*(b) + (a))
+// Permutations for SIMD instructions.  SHUF(A, B, C, D) is an immediate,
+// suitable for use in `pshufd' or `shufpd', which copies element A
+// (0 <= A < 4) of the source to element 0 of the destination, element B to
+// element 1, element C to element 2, and element D to element 3.
+#define SHUF(a, b, c, d) ((a) + 4*(b) + 16*(c) + 64*(d))
 
 // Map register names to their individual pieces.
 
@@ -1031,6 +1031,125 @@ name:
 #endif
 
 ///--------------------------------------------------------------------------
+/// AArch64-specific hacking.
+
+#if CPUFAM_ARM64
+
+// Set the function hooks.
+#define FUNC_PREHOOK(_) .balign 4
+#define FUNC_POSTHOOK(_) .cfi_startproc; .L$_prologue_p = -1
+#define ENDFUNC_HOOK(_) .cfi_endproc
+
+// Call external subroutine at ADDR, possibly via PLT.
+.macro callext addr
+       bl      \addr
+.endm
+
+// Load address of external symbol ADDR into REG.
+.macro leaext  reg, addr
+#if WANT_PIC
+       adrp    \reg, :got:\addr
+       ldr     \reg, [\reg, #:got_lo12:\addr]
+#else
+       adrp    \reg, \addr
+       add     \reg, \reg, #:lo12:\addr
+#endif
+.endm
+
+// Stack management and unwinding.
+.macro setfp   fp, offset = 0
+  // If you're just going through the motions with a fixed-size stack frame,
+  // then you want to say `add x29, sp, #OFFSET' directly, which will avoid
+  // pointlessly restoring sp later.
+  .if \offset == 0
+       mov     \fp, sp
+         .cfi_def_cfa_register \fp
+  .else
+       add     \fp, sp, #\offset
+         .cfi_def_cfa_register \fp
+         .cfi_adjust_cfa_offset -\offset
+  .endif
+       .macro dropfp; _dropfp  \fp, \offset; .endm
+       .L$_frameptr_p = -1
+.endm
+
+.macro _dropfp fp, offset = 0
+  .if \offset == 0
+       mov     sp, \fp
+         .cfi_def_cfa_register sp
+  .else
+       sub     sp, \fp, #\offset
+         .cfi_def_cfa_register sp
+         .cfi_adjust_cfa_offset +\offset
+  .endif
+       .purgem dropfp
+       .L$_frameptr_p = 0
+.endm
+
+.macro stalloc n
+       sub     sp, sp, #\n
+         .cfi_adjust_cfa_offset +\n
+.endm
+
+.macro stfree  n
+       add     sp, sp, #\n
+         .cfi_adjust_cfa_offset -\n
+.endm
+
+.macro pushreg x, y=
+  .ifeqs "\y", ""
+       str     \x, [sp, #-16]!
+         .cfi_adjust_cfa_offset +16
+         .cfi_rel_offset \x, 0
+  .else
+       stp     \x, \y, [sp, #-16]!
+         .cfi_adjust_cfa_offset +16
+         .cfi_rel_offset \x, 0
+         .cfi_rel_offset \y, 8
+  .endif
+.endm
+
+.macro popreg  x, y=
+  .ifeqs "\y", ""
+       ldr     \x, [sp], #16
+         .cfi_restore \x
+         .cfi_adjust_cfa_offset -16
+  .else
+       ldp     \x, \y, [sp], #16
+         .cfi_restore \x
+         .cfi_restore \y
+         .cfi_adjust_cfa_offset -16
+  .endif
+.endm
+
+.macro savereg x, y, z=
+  .ifeqs "\z", ""
+       str     \x, [sp, #\y]
+         .cfi_rel_offset \x, \y
+  .else
+       stp     \x, \y, [sp, #\z]
+         .cfi_rel_offset \x, \z
+         .cfi_rel_offset \y, \z + 8
+  .endif
+.endm
+
+.macro rstrreg x, y, z=
+  .ifeqs "\z", ""
+       ldr     \x, [sp, #\y]
+         .cfi_restore \x
+  .else
+       ldp     \x, \y, [sp, #\z]
+         .cfi_restore \x
+         .cfi_restore \y
+  .endif
+.endm
+
+.macro endprologue
+.endm
+
+#endif
+
+///--------------------------------------------------------------------------
 /// Final stuff.
 
 // Default values for the various hooks.