From 376ffbcf9aabae6f47d62ba4734ae1bb230ebce3 Mon Sep 17 00:00:00 2001
From: William Wilgus <wilgus.william@gmail.com>
Date: Thu, 10 Mar 2022 15:48:51 -0500
Subject: [PATCH] ARM support, optimize popcount fn

Change-Id: Iec02d0b5973721a3943b9c23ced3afc721cd3753
---
 lib/arm_support/support-arm.S | 43 +++++++++++++++++++----------------
 1 file changed, 24 insertions(+), 19 deletions(-)

diff --git a/lib/arm_support/support-arm.S b/lib/arm_support/support-arm.S
index 442a629fca..f99d086b0b 100644
--- a/lib/arm_support/support-arm.S
+++ b/lib/arm_support/support-arm.S
@@ -705,6 +705,9 @@ __aeabi_idivmod:
 /*
  * int __popcountsi2(unsigned int x)
  * int __popcountdi2(unsigned long x)
+ *  x = x - ((x >> 1) & 0x55555555);
+ *  x = (x & 0x33333333) + ((x >> 2) & 0x33333333);
+ *  c = ((x + (x >> 4) & 0xF0F0F0F) * 0x1010101) >> 24;
  */
     .section    .text.__popcountsi2, "ax", %progbits
     .global     __popcountsi2
@@ -712,23 +715,25 @@ __aeabi_idivmod:
     .global     __popcountdi2
     .type       __popcountdi2, %function
     .set        __popcountdi2, __popcountsi2
+
 __popcountsi2:
-    mov r1, #0x33           @ r1 = 0x33333333
-    orr r1, r1, r1, lsl #8  @ ...
-    orr r1, r1, r1, lsl #16 @ ...
-    eor r2, r1, r1, lsl #1  @ r2 = 0x55555555
-    and r2, r2, r0, lsr #1  @ r2 = (x >> 1) & 0x55555555
-    sub r0, r0, r2          @ x = x - ((x >> 1) & 0x55555555)
-    and r2, r1, r0          @ r2 = x & 0x33333333
-    and r1, r1, r0, lsr #2  @ r1 = (x >> 2) & 0x33333333
-    add r0, r2, r1          @ x = (x & 0x33333333) + ((x >> 2) & 0x33333333)
-    mov r1, #0x0f           @ r1 = 0x0f0f0f0f
-    orr r1, r1, r1, lsl #8  @ ...
-    orr r1, r1, r1, lsl #16 @ ...
-    add r0, r0, lsr #4      @ x = x + (x >> 4)
-    and r0, r0, r1          @ x = (x + (x >> 4)) & 0x0f0f0f0f
-    add r0, r0, lsr #16     @ x = x + (x >> 16)
-    add r0, r0, lsr #8      @ x = x + (x >> 8)
-    and r0, r0, #0x3f       @ x &= 0x3f
-    bx  lr                  @ return x
-   .size    __popcountsi2, .-__popcountsi2
+    ldr     r2, .L2            @ r2 = 0x55555555
+    ldr     r3, .L2+4          @ r3 = 0x33333333
+    and     r2, r2, r0, lsr #1 @ r2 = (x >> 1)
+    rsb     r2, r2, r0         @ x = x - ((x >> 1) & 0x55555555)
+    and     r0, r2, r3
+    and     r3, r3, r2, lsr #2 @ r3 = (x >> 2) 
+    add     r0, r0, r3
+    ldr     r3, .L2+8          @ r3 = 0xF0F0F0F
+    add     r0, r0, r0, lsr #4 @ x = x + (x >> 4)
+    and     r3, r0, r3
+    add     r3, r3, r3, asl #8
+    add     r3, r3, r3, asl #16
+    mov     r0, r3, lsr #24    @ (r3 >> 24)
+    bx      lr
+.L2:
+    .word   0x55555555
+    .word   0x33333333
+    .word   0xF0F0F0F
+    .size    __popcountsi2, .-__popcountsi2
+