Fix fast ARM mult/square asm for Xcode

This commit is contained in:
Ken MacKay
2015-10-12 21:29:57 -07:00
parent a2c5132576
commit 1341d3dce5
4 changed files with 1032 additions and 1026 deletions
+38 -38
View File
@@ -40,12 +40,12 @@ else:
print ""
emit("mov r10, #0")
emit("umull r11, r9, r3, r7")
emit("adds r12, r11")
emit("adc r9, #0")
emit("adds r12, r12, r11")
emit("adc r9, r9, #0")
emit("umull r11, r14, r4, r6")
emit("adds r12, r11")
emit("adcs r9, r14")
emit("adc r10, #0")
emit("adds r12, r12, r11")
emit("adcs r9, r9, r14")
emit("adc r10, r10, #0")
emit("stmia r0!, {r12}")
print ""
@@ -55,9 +55,9 @@ else:
emit("mov r%s, #0", acc[2])
for i in xrange(0, 3):
emit("umull r%s, r%s, r%s, r%s", acc[3], acc[4], rx[i], ry[2 - i])
emit("adds r%s, r%s", acc[0], acc[3])
emit("adcs r%s, r%s", acc[1], acc[4])
emit("adc r%s, #0", acc[2])
emit("adds r%s, r%s, r%s", acc[0], acc[0], acc[3])
emit("adcs r%s, r%s, r%s", acc[1], acc[1], acc[4])
emit("adc r%s, r%s, #0", acc[2], acc[2])
emit("stmia r0!, {r%s}", acc[0])
print ""
acc = acc[1:] + acc[:1]
@@ -65,16 +65,16 @@ else:
emit("mov r%s, #0", acc[2])
for i in xrange(0, 2):
emit("umull r%s, r%s, r%s, r%s", acc[3], acc[4], rx[i + 1], ry[2 - i])
emit("adds r%s, r%s", acc[0], acc[3])
emit("adcs r%s, r%s", acc[1], acc[4])
emit("adc r%s, #0", acc[2])
emit("adds r%s, r%s, r%s", acc[0], acc[0], acc[3])
emit("adcs r%s, r%s, r%s", acc[1], acc[1], acc[4])
emit("adc r%s, r%s, #0", acc[2], acc[2])
emit("stmia r0!, {r%s}", acc[0])
print ""
acc = acc[1:] + acc[:1]
emit("umull r%s, r%s, r%s, r%s", acc[3], acc[4], rx[init_size-1], ry[init_size-1])
emit("adds r%s, r%s", acc[0], acc[3])
emit("adc r%s, r%s", acc[1], acc[4])
emit("adds r%s, r%s, r%s", acc[0], acc[0], acc[3])
emit("adc r%s, r%s, r%s", acc[1], acc[1], acc[4])
emit("stmia r0!, {r%s}", acc[0])
emit("stmia r0!, {r%s}", acc[1])
print ""
@@ -98,12 +98,12 @@ for row in xrange(full_rows):
print ""
emit("mov r10, #0")
emit("umull r11, r9, r3, r7")
emit("adds r12, r11")
emit("adc r9, #0")
emit("adds r12, r12, r11")
emit("adc r9, r9, #0")
emit("umull r11, r14, r4, r6")
emit("adds r12, r11")
emit("adcs r9, r14")
emit("adc r10, #0")
emit("adds r12, r12, r11")
emit("adcs r9, r9, r14")
emit("adc r10, r10, #0")
emit("stmia r0!, {r12}")
print ""
@@ -111,9 +111,9 @@ for row in xrange(full_rows):
emit("mov r%s, #0", acc[2])
for i in xrange(0, 3):
emit("umull r%s, r%s, r%s, r%s", acc[3], acc[4], rx[i], ry[2 - i])
emit("adds r%s, r%s", acc[0], acc[3])
emit("adcs r%s, r%s", acc[1], acc[4])
emit("adc r%s, #0", acc[2])
emit("adds r%s, r%s, r%s", acc[0], acc[0], acc[3])
emit("adcs r%s, r%s, r%s", acc[1], acc[1], acc[4])
emit("adc r%s, r%s, #0", acc[2], acc[2])
emit("stmia r0!, {r%s}", acc[0])
print ""
acc = acc[1:] + acc[:1]
@@ -126,13 +126,13 @@ for row in xrange(full_rows):
emit("mov r%s, #0", acc[2])
for i in xrange(0, 3):
emit("umull r%s, r%s, r%s, r%s", acc[3], acc[4], x_regs[i], ry[2 - i])
emit("adds r%s, r%s", acc[0], acc[3])
emit("adcs r%s, r%s", acc[1], acc[4])
emit("adc r%s, #0", acc[2])
emit("adds r%s, r%s, r%s", acc[0], acc[0], acc[3])
emit("adcs r%s, r%s, r%s", acc[1], acc[1], acc[4])
emit("adc r%s, r%s, #0", acc[2], acc[2])
emit("ldr r%s, [r0]", acc[3]) # load stored value from initial block, and add to accumulator
emit("adds r%s, r%s", acc[0], acc[3])
emit("adcs r%s, #0", acc[1])
emit("adc r%s, #0", acc[2])
emit("adds r%s, r%s, r%s", acc[0], acc[0], acc[3])
emit("adcs r%s, r%s, #0", acc[1], acc[1])
emit("adc r%s, r%s, #0", acc[2], acc[2])
emit("stmia r0!, {r%s}", acc[0])
print ""
acc = acc[1:] + acc[:1]
@@ -145,13 +145,13 @@ for row in xrange(full_rows):
emit("mov r%s, #0", acc[2])
for i in xrange(0, 3):
emit("umull r%s, r%s, r%s, r%s", acc[3], acc[4], x_regs[i], y_regs[2 - i])
emit("adds r%s, r%s", acc[0], acc[3])
emit("adcs r%s, r%s", acc[1], acc[4])
emit("adc r%s, #0", acc[2])
emit("adds r%s, r%s, r%s", acc[0], acc[0], acc[3])
emit("adcs r%s, r%s, r%s", acc[1], acc[1], acc[4])
emit("adc r%s, r%s, #0", acc[2], acc[2])
emit("ldr r%s, [r0]", acc[3]) # load stored value from initial block, and add to accumulator
emit("adds r%s, r%s", acc[0], acc[3])
emit("adcs r%s, #0", acc[1])
emit("adc r%s, #0", acc[2])
emit("adds r%s, r%s, r%s", acc[0], acc[0], acc[3])
emit("adcs r%s, r%s, #0", acc[1], acc[1])
emit("adc r%s, r%s, #0", acc[2], acc[2])
emit("stmia r0!, {r%s}", acc[0])
print ""
acc = acc[1:] + acc[:1]
@@ -160,16 +160,16 @@ for row in xrange(full_rows):
emit("mov r%s, #0", acc[2])
for i in xrange(0, 2):
emit("umull r%s, r%s, r%s, r%s", acc[3], acc[4], x_regs[i + 1], y_regs[2 - i])
emit("adds r%s, r%s", acc[0], acc[3])
emit("adcs r%s, r%s", acc[1], acc[4])
emit("adc r%s, #0", acc[2])
emit("adds r%s, r%s, r%s", acc[0], acc[0], acc[3])
emit("adcs r%s, r%s, r%s", acc[1], acc[1], acc[4])
emit("adc r%s, r%s, #0", acc[2], acc[2])
emit("stmia r0!, {r%s}", acc[0])
print ""
acc = acc[1:] + acc[:1]
emit("umull r%s, r%s, r%s, r%s", acc[3], acc[4], x_regs[2], y_regs[2])
emit("adds r%s, r%s", acc[0], acc[3])
emit("adc r%s, r%s", acc[1], acc[4])
emit("adds r%s, r%s, r%s", acc[0], acc[0], acc[3])
emit("adc r%s, r%s, r%s", acc[1], acc[1], acc[4])
emit("stmia r0!, {r%s}", acc[0])
emit("stmia r0!, {r%s}", acc[1])
print ""
+59 -59
View File
@@ -23,15 +23,15 @@ def emit(line, *args):
def mulacc(acc, r1, r2):
if size <= 6:
emit("umull r1, r14, r%s, r%s", r1, r2)
emit("adds r%s, r1", acc[0])
emit("adcs r%s, r14", acc[1])
emit("adc r%s, #0", acc[2])
emit("adds r%s, r%s, r1", acc[0], acc[0])
emit("adcs r%s, r%s, r14", acc[1], acc[1])
emit("adc r%s, r%s, #0", acc[2], acc[2])
else:
emit("mov r14, r%s", acc[1])
emit("umlal r%s, r%s, r%s, r%s", acc[0], acc[1], r1, r2)
emit("cmp r14, r%s", acc[1])
emit("it hi")
emit("adchi r%s, #0", acc[2])
emit("adchi r%s, r%s, #0", acc[2], acc[2])
r = [2, 3, 4, 5, 6, 7]
@@ -62,13 +62,13 @@ elif init_size == 2:
print ""
emit("umull r12, r10, r2, r6")
emit("adds r9, r12")
emit("adc r10, #0")
emit("adds r9, r9, r12")
emit("adc r10, r10, #0")
emit("stmia r0!, {r9}")
print ""
emit("umull r8, r9, r3, r6")
emit("adds r10, r8")
emit("adds r10, r10, r8")
emit("adc r11, r9, #0")
emit("stmia r0!, {r10, r11}")
print ""
@@ -85,26 +85,26 @@ emit("stmia r0!, {r11}")
print ""
emit("mov r9, #0")
emit("umull r10, r11, r2, r3")
emit("adds r12, r10")
emit("adds r12, r12, r10")
emit("adcs r8, r11, #0")
emit("adc r9, #0")
emit("adds r12, r10")
emit("adcs r8, r11")
emit("adc r9, #0")
emit("adc r9, r9, #0")
emit("adds r12, r12, r10")
emit("adcs r8, r8, r11")
emit("adc r9, r9, #0")
emit("stmia r0!, {r12}")
print ""
emit("mov r10, #0")
emit("umull r11, r12, r2, r4")
emit("adds r11, r11")
emit("adcs r12, r12")
emit("adc r10, #0")
emit("adds r8, r11")
emit("adcs r9, r12")
emit("adc r10, #0")
emit("adds r11, r11, r11")
emit("adcs r12, r12, r12")
emit("adc r10, r10, #0")
emit("adds r8, r8, r11")
emit("adcs r9, r9, r12")
emit("adc r10, r10, #0")
emit("umull r11, r12, r3, r3")
emit("adds r8, r11")
emit("adcs r9, r12")
emit("adc r10, #0")
emit("adds r8, r8, r11")
emit("adcs r9, r9, r12")
emit("adc r10, r10, #0")
emit("stmia r0!, {r8}")
print ""
@@ -121,18 +121,18 @@ for i in xrange(3, s):
for j in xrange(1, (i+1)//2):
mulacc(acc, r[j], r[i-j])
# multiply by 2
emit("adds r%s, r%s", acc[0], acc[0])
emit("adcs r%s, r%s", acc[1], acc[1])
emit("adc r%s, r%s", acc[2], acc[2])
emit("adds r%s, r%s, r%s", acc[0], acc[0], acc[0])
emit("adcs r%s, r%s, r%s", acc[1], acc[1], acc[1])
emit("adc r%s, r%s, r%s", acc[2], acc[2], acc[2])
# add equal word (if any)
if ((i+1) % 2) != 0:
mulacc(acc, r[i//2], r[i//2])
# add old accumulator
emit("adds r%s, r%s", acc[0], old_acc[0])
emit("adcs r%s, r%s", acc[1], old_acc[1])
emit("adc r%s, #0", acc[2])
emit("adds r%s, r%s, r%s", acc[0], acc[0], old_acc[0])
emit("adcs r%s, r%s, r%s", acc[1], acc[1], old_acc[1])
emit("adc r%s, r%s, #0", acc[2], acc[2])
# store
emit("stmia r0!, {r%s}", acc[0])
@@ -155,23 +155,23 @@ for i in xrange(init_size):
mulacc(acc, regs[j], regs[limit-j])
emit("ldr r14, [r0]") # load stored value from initial block, and add to accumulator
emit("adds r%s, r14", acc[0])
emit("adcs r%s, #0", acc[1])
emit("adc r%s, #0", acc[2])
emit("adds r%s, r%s, r14", acc[0], acc[0])
emit("adcs r%s, r%s, #0", acc[1], acc[1])
emit("adc r%s, r%s, #0", acc[2], acc[2])
# multiply by 2
emit("adds r%s, r%s", acc[0], acc[0])
emit("adcs r%s, r%s", acc[1], acc[1])
emit("adc r%s, r%s", acc[2], acc[2])
emit("adds r%s, r%s, r%s", acc[0], acc[0], acc[0])
emit("adcs r%s, r%s, r%s", acc[1], acc[1], acc[1])
emit("adc r%s, r%s, r%s", acc[2], acc[2], acc[2])
# add equal word
if limit == 4:
mulacc(acc, regs[2], regs[2])
# add old accumulator
emit("adds r%s, r%s", acc[0], old_acc[0])
emit("adcs r%s, r%s", acc[1], old_acc[1])
emit("adc r%s, #0", acc[2])
emit("adds r%s, r%s, r%s", acc[0], acc[0], old_acc[0])
emit("adcs r%s, r%s, r%s", acc[1], acc[1], old_acc[1])
emit("adc r%s, r%s, #0", acc[2], acc[2])
# store
emit("stmia r0!, {r%s}", acc[0])
@@ -189,18 +189,18 @@ for i in xrange(1, s-3):
mulacc(acc, regs[i+j], regs[s - 1 - j])
# multiply by 2
emit("adds r%s, r%s", acc[0], acc[0])
emit("adcs r%s, r%s", acc[1], acc[1])
emit("adc r%s, r%s", acc[2], acc[2])
emit("adds r%s, r%s, r%s", acc[0], acc[0], acc[0])
emit("adcs r%s, r%s, r%s", acc[1], acc[1], acc[1])
emit("adc r%s, r%s, r%s", acc[2], acc[2], acc[2])
# add equal word (if any)
if ((s-i) % 2) != 0:
mulacc(acc, regs[i + (s-i)//2], regs[i + (s-i)//2])
# add old accumulator
emit("adds r%s, r%s", acc[0], old_acc[0])
emit("adcs r%s, r%s", acc[1], old_acc[1])
emit("adc r%s, #0", acc[2])
emit("adds r%s, r%s, r%s", acc[0], acc[0], old_acc[0])
emit("adcs r%s, r%s, r%s", acc[1], acc[1], old_acc[1])
emit("adc r%s, r%s, #0", acc[2], acc[2])
# store
emit("stmia r0!, {r%s}", acc[0])
@@ -209,34 +209,34 @@ for i in xrange(1, s-3):
acc = acc[1:] + acc[:1]
emit("mov r%s, #0", acc[2])
emit("umull r1, r%s, r%s, r%s", old_acc[1], regs[s - 3], regs[s - 1])
emit("adds r1, r1")
emit("adcs r%s, r%s", old_acc[1], old_acc[1])
emit("adc r%s, #0", acc[2])
emit("adds r%s, r1", acc[0])
emit("adcs r%s, r%s", acc[1], old_acc[1])
emit("adc r%s, #0", acc[2])
emit("adds r1, r1, r1")
emit("adcs r%s, r%s, r%s", old_acc[1], old_acc[1], old_acc[1])
emit("adc r%s, r%s, #0", acc[2], acc[2])
emit("adds r%s, r%s, r1", acc[0], acc[0])
emit("adcs r%s, r%s, r%s", acc[1], acc[1], old_acc[1])
emit("adc r%s, r%s, #0", acc[2], acc[2])
emit("umull r1, r%s, r%s, r%s", old_acc[1], regs[s - 2], regs[s - 2])
emit("adds r%s, r1", acc[0])
emit("adcs r%s, r%s", acc[1], old_acc[1])
emit("adc r%s, #0", acc[2])
emit("adds r%s, r%s, r1", acc[0], acc[0])
emit("adcs r%s, r%s, r%s", acc[1], acc[1], old_acc[1])
emit("adc r%s, r%s, #0", acc[2], acc[2])
emit("stmia r0!, {r%s}", acc[0])
print ""
acc = acc[1:] + acc[:1]
emit("mov r%s, #0", acc[2])
emit("umull r1, r%s, r%s, r%s", old_acc[1], regs[s - 2], regs[s - 1])
emit("adds r1, r1")
emit("adcs r%s, r%s", old_acc[1], old_acc[1])
emit("adc r%s, #0", acc[2])
emit("adds r%s, r1", acc[0])
emit("adcs r%s, r%s", acc[1], old_acc[1])
emit("adc r%s, #0", acc[2])
emit("adds r1, r1, r1")
emit("adcs r%s, r%s, r%s", old_acc[1], old_acc[1], old_acc[1])
emit("adc r%s, r%s, #0", acc[2], acc[2])
emit("adds r%s, r%s, r1", acc[0], acc[0])
emit("adcs r%s, r%s, r%s", acc[1], acc[1], old_acc[1])
emit("adc r%s, r%s, #0", acc[2], acc[2])
emit("stmia r0!, {r%s}", acc[0])
print ""
acc = acc[1:] + acc[:1]
emit("umull r1, r%s, r%s, r%s", old_acc[1], regs[s - 1], regs[s - 1])
emit("adds r%s, r1", acc[0])
emit("adcs r%s, r%s", acc[1], old_acc[1])
emit("adds r%s, r%s, r1", acc[0], acc[0])
emit("adcs r%s, r%s, r%s", acc[1], acc[1], old_acc[1])
emit("stmia r0!, {r%s}", acc[0])
emit("stmia r0!, {r%s}", acc[1])