/// -*- mode: asm; asm-comment-char: ?/ -*-
///
-/// Fancy SIMD implementation of Salsa20
+/// Common definitions for asesembler source files
///
/// (c) 2015 Straylight/Edgeware
///
#endif
#define DATA .data
-// Announcing an external function.
-#define FUNC(name) \
- .globl F(name); \
+// Announcing an internal function.
+#define INTFUNC(name) \
TYPE_FUNC(name); \
.macro ENDFUNC; _ENDFUNC(name); .endm; \
+ .L$_prologue_p = 0; .L$_frameptr_p = 0; \
FUNC_PREHOOK(name); \
-F(name): \
+name: \
FUNC_POSTHOOK(name)
+// Announcing an external function.
+#define FUNC(name) \
+ .globl F(name); \
+INTFUNC(F(name))
+
// Marking the end of a function.
#define _ENDFUNC(name) \
+ .if ~ .L$_prologue_p; .error "Missing `endprologue'"; .endif; \
+ .if .L$_frameptr_p; .purgem dropfp; .endif; \
.purgem ENDFUNC; \
SIZE_OBJ(name); \
ENDFUNC_HOOK(name); \
#if CPUFAM_X86 || CPUFAM_AMD64
+// Word size.
+#if CPUFAM_X86
+# define WORDSZ 4
+#endif
+#if CPUFAM_AMD64
+# define WORDSZ 8
+#endif
+
// Set the function hooks.
#define FUNC_PREHOOK(_) .balign 16
// `.seh_pushreg' and friends, and `.seh_endprologue'.
#endif
+#if __ELF__
+# define FUNC_POSTHOOK(_) .cfi_startproc
+# define ENDFUNC_HOOK(_) .cfi_endproc
+#endif
+
// Don't use the wretched AT&T syntax. It's festooned with pointless
// punctuation, and all of the data movement is backwards. Ugh!
.intel_syntax noprefix
#endif
#define WHOLE(reg) _REGFORM(reg, r)
+// Stack management and unwinding.
+.macro setfp fp, offset = 0
+ .if \offset == 0
+ mov \fp, R_sp(r)
+#if __ELF__
+ .cfi_def_cfa_register \fp
+#endif
+#if ABI_WIN && CPUFAM_AMD64
+ .seh_setframe \fp, 0
+#endif
+ .else
+ lea \fp, [R_sp(r) + \offset]
+#if __ELF__
+ .cfi_def_cfa_register \fp
+ .cfi_adjust_cfa_offset -\offset
+#endif
+#if ABI_WIN && CPUFAM_AMD64
+ .seh_setframe \fp, \offset
+#endif
+ .endif
+ .L$_frameptr_p = -1
+ .macro dropfp; _dropfp \fp, \offset; .endm
+.endm
+
+.macro _dropfp fp, offset = 0
+ .if \offset == 0
+ mov R_sp(r), \fp
+#if __ELF__
+ .cfi_def_cfa_register R_sp(r)
+#endif
+ .else
+ lea R_sp(r), [\fp - \offset]
+#if __ELF__
+ .cfi_def_cfa_register R_sp(r)
+ .cfi_adjust_cfa_offset +\offset
+#endif
+ .endif
+ .L$_frameptr_p = 0
+ .purgem dropfp
+.endm
+
+.macro stalloc n
+ sub R_sp(r), \n
+#if __ELF__
+ .cfi_adjust_cfa_offset +\n
+#endif
+#if ABI_WIN && CPUFAM_AMD64
+ .seh_stackalloc \n
+#endif
+.endm
+
+.macro stfree n
+ add R_sp(r), \n
+#if __ELF__
+ .cfi_adjust_cfa_offset -\n
+#endif
+.endm
+
+.macro pushreg r
+ push \r
+#if __ELF__
+ .cfi_adjust_cfa_offset +WORDSZ
+ .cfi_rel_offset \r, 0
+#endif
+#if ABI_WIN && CPUFAM_AMD64
+ .seh_pushreg \r
+#endif
+.endm
+
+.macro popreg r
+ pop \r
+#if __ELF__
+ .cfi_adjust_cfa_offset -WORDSZ
+ .cfi_restore \r
+#endif
+.endm
+
+.macro savexmm r, offset
+ movdqa [R_sp(r) + \offset], \r
+#if ABI_WIN && CPUFAM_AMD64
+ .seh_savexmm \r, \offset
+#endif
+.endm
+
+.macro rstrxmm r, offset
+ movdqa \r, [R_sp(r) + \offset]
+.endm
+
+.macro endprologue
+#if ABI_WIN && CPUFAM_AMD64
+ .seh_endprologue
+#endif
+ .L$_prologue_p = -1
+.endm
+
#endif
#if CPUFAM_X86
ARM
// Set the function hooks.
-#define FUNC_PREHOOK(_) .balign 4
-#define ENDFUNC_HOOK(name) .ltorg
+#define FUNC_PREHOOK(_) .balign 4; .fnstart
+#define ENDFUNC_HOOK(_) .fnend; .ltorg
// Call external subroutine at ADDR, possibly via PLT.
.macro callext addr, cond=
// Macros for converting vldm/vstm ranges.
#define QQ(qlo, qhi) D0(qlo)-D1(qhi)
+// Stack management and unwinding.
+.macro setfp fp, offset = 0
+ .if \offset == 0
+ mov \fp, sp
+ .setfp \fp, sp
+ .else
+ add \fp, sp, #\offset
+ .setfp \fp, sp, #\offset
+ .endif
+ .macro dropfp; _dropfp \fp, \offset; .endm
+ .L$_frameptr_p = -1
+.endm
+
+.macro _dropfp fp, offset = 0
+ .if \offset == 0
+ mov sp, \fp
+ .else
+ sub sp, \fp, #\offset
+ .endif
+ .purgem dropfp
+ .L$_frameptr_p = 0
+.endm
+
+.macro stalloc n
+ sub sp, sp, #\n
+ .pad #\n
+.endm
+
+.macro stfree n
+ add sp, sp, #\n
+ .pad #-\n
+.endm
+
+.macro pushreg rr:vararg
+ stmfd sp!, {\rr}
+ .save {\rr}
+.endm
+
+.macro popreg rr:vararg
+ ldmfd sp!, {\rr}
+.endm
+
+.macro pushvfp rr:vararg
+ vstmdb sp!, {\rr}
+ .vsave {\rr}
+.endm
+
+.macro popvfp rr:vararg
+ vldmia sp!, {\rr}
+.endm
+
+.macro endprologue
+.endm
+
+// No need for prologue markers on ARM.
+#define FUNC_POSTHOOK(_) .L$_prologue_p = -1
+
+#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
///--------------------------------------------------------------------------
// Default values for the various hooks.
#ifndef FUNC_PREHOOK
-# define FUNC_PREHOOK(name)
+# define FUNC_PREHOOK(_)
#endif
#ifndef FUNC_POSTHOOK
-# define FUNC_POSTHOOK(name)
+# define FUNC_POSTHOOK(_)
#endif
#ifndef ENDFUNC_HOOK
-# define ENDFUNC_HOOK(name)
+# define ENDFUNC_HOOK(_)
#endif
#ifndef F
# define SIZE_OBJ(name)
#endif
-#if __ELF__ && defined(WANT_EXECUTABLE_STACK)
+#if __ELF__ && !defined(WANT_EXECUTABLE_STACK)
.pushsection .note.GNU-stack, "", _SECTTY(progbits)
.popsection
#endif
~mdw / catacomb / blobdiff