Update agbabi's memcpy

agb/build.rs

@@ -12,6 +12,7 @@ fn main() {
     println!("cargo:rerun-if-changed=gba.ld");
     println!("cargo:rerun-if-changed=gba_mb.ld");
     println!("cargo:rerun-if-changed=src/asm_include.s");
+    println!("cargo:rerun-if-changed=src/agbabi/macros.inc");
     println!("cargo:rerun-if-changed=gfx/test_logo.png");
 
     println!("cargo:rerun-if-changed=build.rs");

agb/src/agbabi/macros.inc

@@ -0,0 +1,43 @@
+/*
+===============================================================================
+
+ ARM assembly support macros
+
+ Copyright (C) 2021-2022 agbabi contributors
+ For conditions of distribution and use, see copyright notice in LICENSE.md
+
+===============================================================================
+*/
+
+// Shift and test upper two bits, clobbering \reg
+// Use mi for first bit, cs for second bit
+.macro joaobapt_test_lsl reg shift = #0
+    movs    \reg, \reg, lsl \shift
+.endm
+
+// Test lowest two bits, clobbering \reg
+// Use mi for low bit, cs for high bit
+.macro joaobapt_test reg
+    joaobapt_test_lsl \reg, #31
+.endm
+
+// Test lowest two bits of \src, result stored in \dst
+// Use mi for low bit, cs for high bit
+.macro joaobapt_test_into dst, src
+    movs    \dst, \src, lsl #31
+.endm
+
+// Branches depending on lowest two bits, clobbering \reg
+// b_mi = low bit case, b_cs = high bit case
+.macro joaobapt_switch reg, b_mi, b_cs
+    joaobapt_test \reg
+    bmi     \b_mi
+    bcs     \b_cs
+.endm
+
+// Branches depending on alignment of \a and \b, clobbering \scratch
+// b_byte = off-by-byte case, b_half = off-by-half case
+.macro align_switch a, b, scratch, b_byte, b_half
+    eor     \scratch, \a, \b
+    joaobapt_switch \scratch, \b_byte, \b_half
+.endm

agb/src/agbabi/memcpy.s

@@ -1,19 +1,18 @@
 /*
 ===============================================================================
-
  ABI:
     __aeabi_memcpy, __aeabi_memcpy4, __aeabi_memcpy8
  Standard:
     memcpy
  Support:
-    __agbabi_memcpy2
+    __agbabi_memcpy2, __agbabi_memcpy1
-
  Copyright (C) 2021-2022 agbabi contributors
  For conditions of distribution and use, see copyright notice in LICENSE.md
-
 ===============================================================================
 */
 
+.include "src/agbabi/macros.inc"
+
     .arm
     .align 2
 
@@ -22,95 +21,97 @@
 __agbabi_memcpy:
     .global __aeabi_memcpy
 __aeabi_memcpy:
-    // Check pointer alignment
+    // >6-bytes is roughly the threshold when byte-by-byte copy is slower
-    eor     r3, r1, r0
+    cmp     r2, #6
-    // JoaoBapt carry & sign bit test
+    ble     __agbabi_memcpy1
-    movs    r3, r3, lsl #31
-    bmi     .Lcopy1
-    bcs     .Lcopy2
 
-.Lcopy4:
+    align_switch r0, r1, r3, __agbabi_memcpy1, .Lcopy_halves
-    // Handle <= 2 byte copies byte-by-byte
-    cmp     r2, #2
-    ble     .Lcopy1
 
-    // Copy half and byte head
+    // Check if r0 (or r1) needs word aligning
-    rsb     r3, r0, #4
+    rsbs    r3, r0, #4
-    // JoaoBapt carry & sign bit test
+    joaobapt_test r3
-    movs    r3, r3, lsl #31
+
+    // Copy byte head to align
     ldrmib  r3, [r1], #1
     strmib  r3, [r0], #1
     submi   r2, r2, #1
+    // r0, r1 are now half aligned
+
+    // Copy half head to align
     ldrcsh  r3, [r1], #2
     strcsh  r3, [r0], #2
     subcs   r2, r2, #2
-    // Fallthrough
+    // r0, r1 are now word aligned
 
     .global __aeabi_memcpy8
 __aeabi_memcpy8:
     .global __aeabi_memcpy4
 __aeabi_memcpy4:
-    // Copy 8 words
+    cmp     r2, #32
-    movs    r12, r2, lsr #5
+    blt     .Lcopy_words
-    beq     .Lskip32
+
-    lsl     r3, r12, #5
+    // Word aligned, 32-byte copy
-    sub     r2, r2, r3
     push    {r4-r10}
-.LcopyWords8:
+.Lloop_32:
-    ldmia   r1!, {r3-r10}
+    subs    r2, r2, #32
-    stmia   r0!, {r3-r10}
+    ldmgeia r1!, {r3-r10}
-    subs    r12, r12, #1
+    stmgeia r0!, {r3-r10}
-    bne     .LcopyWords8
+    bgt     .Lloop_32
     pop     {r4-r10}
-.Lskip32:
+    bxeq    lr
 
-    // Copy words
+    // < 32 bytes remaining to be copied
-    movs    r12, r2, lsr #2
+    add     r2, r2, #32
-.LcopyWords:
-    subs    r12, r12, #1
-    ldrhs   r3, [r1], #4
-    strhs   r3, [r0], #4
-    bhs     .LcopyWords
 
-    // Copy half and byte tail
+.Lcopy_words:
-    // JoaoBapt carry & sign bit test
+    cmp     r2, #4
-    movs    r3, r2, lsl #31
+    blt     .Lcopy_halves
+.Lloop_4:
+    subs    r2, r2, #4
+    ldrge   r3, [r1], #4
+    strge   r3, [r0], #4
+    bgt     .Lloop_4
+    bxeq    lr
+
+    // Copy byte & half tail
+    // This test still works when r2 is negative
+    joaobapt_test r2
+    // Copy half
     ldrcsh  r3, [r1], #2
     strcsh  r3, [r0], #2
+    // Copy byte
     ldrmib  r3, [r1]
     strmib  r3, [r0]
     bx      lr
 
-.Lcopy2:
+.Lcopy_halves:
-    // Copy byte head
+    // Copy byte head to align
     tst     r0, #1
-    cmpne   r2, #0
     ldrneb  r3, [r1], #1
     strneb  r3, [r0], #1
     subne   r2, r2, #1
-    // Fallthrough
+    // r0, r1 are now half aligned
 
     .global __agbabi_memcpy2
 __agbabi_memcpy2:
-    // Copy halves
+    subs    r2, r2, #2
-    movs    r12, r2, lsr #1
+    ldrgeh  r3, [r1], #2
-.LcopyHalves:
+    strgeh  r3, [r0], #2
-    subs    r12, r12, #1
+    bgt     __agbabi_memcpy2
-    ldrhsh  r3, [r1], #2
+    bxeq    lr
-    strhsh  r3, [r0], #2
-    bhs     .LcopyHalves
 
     // Copy byte tail
-    tst     r2, #1
+    adds    r2, r2, #1
-    ldrneb  r3, [r1]
+    ldreqb  r3, [r1]
-    strneb  r3, [r0]
+    streqb  r3, [r0]
     bx      lr
 
-.Lcopy1:
+    .global __agbabi_memcpy1
+__agbabi_memcpy1:
     subs    r2, r2, #1
-    ldrhsb  r3, [r1], #1
+    ldrgeb  r3, [r1], #1
-    strhsb  r3, [r0], #1
+    strgeb  r3, [r0], #1
-    bhs     .Lcopy1
+    bgt     __agbabi_memcpy1
     bx      lr
 
     .section .iwram.memcpy, "ax", %progbits