X-Git-Url: https://git.distorted.org.uk/~mdw/catacomb/blobdiff_plain/0923a413958b0e778a3f059c76355ab58e5be414..1f1fd8845afef39000b08a0f79ff14fae7690d41:/base/asm-common.h

diff --git a/base/asm-common.h b/base/asm-common.h
index 22bb44d6..642820af 100644
--- 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
 ///
@@ -24,6 +24,9 @@
 /// Software Foundation, Inc., 59 Temple Place - Suite 330, Boston,
 /// MA 02111-1307, USA.
 
+#ifndef CATACOMB_ASM_COMMON_H
+#define CATACOMB_ASM_COMMON_H
+
 ///--------------------------------------------------------------------------
 /// General definitions.
 
@@ -217,11 +220,11 @@ 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.
 
@@ -255,10 +258,10 @@ name:
 #  define _DECOR_abcd_q(reg) r##reg##x
 #endif
 
-#define _DECOR_xp_b(reg) reg##l
 #define _DECOR_xp_w(reg) reg
 #define _DECOR_xp_d(reg) e##reg
 #if CPUFAM_AMD64
+#  define _DECOR_xp_b(reg) reg##l
 #  define _DECOR_xp_q(reg) r##reg
 #endif
 
@@ -276,22 +279,33 @@ name:
 #  define _DECOR_rn_r(reg) reg
 #endif
 
+#define _DECOR_mem_b(addr) byte ptr addr
+#define _DECOR_mem_w(addr) word ptr addr
+#define _DECOR_mem_d(addr) dword ptr addr
+#if CPUFAM_AMD64
+#  define _DECOR_mem_q(addr) qword ptr addr
+#endif
+
+#define _DECOR_imm_b(imm) byte imm
+#define _DECOR_imm_w(imm) word imm
+#define _DECOR_imm_d(imm) dword imm
+#if CPUFAM_AMD64
+#  define _DECOR_imm_q(imm) qword imm
+#endif
+
 #if CPUFAM_X86
 #  define _DECOR_abcd_r(reg) e##reg##x
 #  define _DECOR_xp_r(reg) e##reg
 #  define _DECOR_ip_r(reg) e##reg
+#  define _DECOR_mem_r(addr) dword ptr addr
+#  define _DECOR_imm_r(imm) dword imm
 #endif
 #if CPUFAM_AMD64
 #  define _DECOR_abcd_r(reg) r##reg##x
 #  define _DECOR_xp_r(reg) r##reg
 #  define _DECOR_ip_r(reg) r##reg
-#endif
-
-#define _DECOR_mem_b(addr) byte ptr addr
-#define _DECOR_mem_w(addr) word ptr addr
-#define _DECOR_mem_d(addr) dword ptr addr
-#if CPUFAM_AMD64
-#  define _DECOR_mem_q(addr) qword ptr addr
+#  define _DECOR_mem_r(addr) qword ptr addr
+#  define _DECOR_imm_r(imm) qword imm
 #endif
 
 // R_r(decor) applies decoration decor to register r, which is an internal
@@ -321,6 +335,9 @@ name:
 // address addr (which should supply its own square-brackets).
 #define MEM(decor, addr) _DECOR(mem, decor, addr)
 
+// Refer to an immediate datum of the type implied by decor.
+#define IMM(decor, imm) _DECOR(mem, decor, imm)
+
 // Applies decoration decor to assembler-level register name reg.
 #define _REGFORM(reg, decor) _GLUE(_REGFORM_, reg)(decor)
 
@@ -436,7 +453,7 @@ name:
 #define WHOLE(reg) _REGFORM(reg, r)
 
 // Stack management and unwinding.
-.macro	setfp	fp, offset = 0
+.macro	setfp	fp=R_bp(r), offset=0
   .if \offset == 0
 	mov	\fp, R_sp(r)
 #if __ELF__
@@ -459,7 +476,7 @@ name:
 	.macro dropfp; _dropfp	\fp, \offset; .endm
 .endm
 
-.macro	_dropfp	fp, offset = 0
+.macro	_dropfp	fp, offset=0
   .if \offset == 0
 	mov	R_sp(r), \fp
 #if __ELF__
@@ -706,12 +723,12 @@ name:
 #if WANT_PIC
 	ldr\cond \reg, .L$_leaextq$\@
 .L$_leaextq_pc$\@:
-	.if	.L$_pcoff == 8
+  .if .L$_pcoff == 8
 	ldr\cond \reg, [pc, \reg]
-	.else
+  .else
 	add\cond \reg, pc
 	ldr\cond \reg, [\reg]
-	.endif
+  .endif
   _LIT
 	.balign	4
 .L$_leaextq$\@:
@@ -972,7 +989,7 @@ name:
 #define QQ(qlo, qhi) D0(qlo)-D1(qhi)
 
 // Stack management and unwinding.
-.macro	setfp	fp, offset = 0
+.macro	setfp	fp=r11, offset=0
   .if \offset == 0
 	mov	\fp, sp
 	  .setfp \fp, sp
@@ -984,7 +1001,7 @@ name:
 	.L$_frameptr_p = -1
 .endm
 
-.macro	_dropfp	fp, offset = 0
+.macro	_dropfp	fp, offset=0
   .if \offset == 0
 	mov	sp, \fp
   .else
@@ -1005,12 +1022,12 @@ name:
 .endm
 
 .macro	pushreg	rr:vararg
-	stmfd	sp!, {\rr}
+	push	{\rr}
 	  .save {\rr}
 .endm
 
 .macro	popreg	rr:vararg
-	ldmfd	sp!, {\rr}
+	pop	{\rr}
 .endm
 
 .macro	pushvfp rr:vararg
@@ -1031,6 +1048,125 @@ name:
 #endif
 
 ///--------------------------------------------------------------------------
+/// AArch64-specific hacking.
+
+#if CPUFAM_ARM64
+
+// Set the function hooks.
+#define FUNC_PREHOOK(_) .balign 4
+#define FUNC_POSTHOOK(_) .cfi_startproc; .L$_prologue_p = -1
+#define ENDFUNC_HOOK(_) .cfi_endproc
+
+// Call external subroutine at ADDR, possibly via PLT.
+.macro	callext addr
+	bl	\addr
+.endm
+
+// Load address of external symbol ADDR into REG.
+.macro	leaext	reg, addr
+#if WANT_PIC
+	adrp	\reg, :got:\addr
+	ldr	\reg, [\reg, #:got_lo12:\addr]
+#else
+	adrp	\reg, \addr
+	add	\reg, \reg, #:lo12:\addr
+#endif
+.endm
+
+// Stack management and unwinding.
+.macro	setfp	fp=x29, offset=0
+  // If you're just going through the motions with a fixed-size stack frame,
+  // then you want to say `add x29, sp, #OFFSET' directly, which will avoid
+  // pointlessly restoring sp later.
+  .if \offset == 0
+	mov	\fp, sp
+	  .cfi_def_cfa_register \fp
+  .else
+	add	\fp, sp, #\offset
+	  .cfi_def_cfa_register \fp
+	  .cfi_adjust_cfa_offset -\offset
+  .endif
+	.macro dropfp; _dropfp	\fp, \offset; .endm
+	.L$_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=nil
+  .ifeqs "\y", "nil"
+	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=nil
+  .ifeqs "\y", "nil"
+	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=nil
+  .ifeqs "\z", "nil"
+	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=nil
+  .ifeqs "\z", "nil"
+	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.
@@ -1056,9 +1192,11 @@ name:
 #  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
 
 ///----- That's all, folks --------------------------------------------------
+
+#endif