Merge branch 'main' into fast-math

Reviewed-on: TCShenanigans/symphony_stdlib#13

README.md

@@ -3,7 +3,7 @@
 This is a standard library for symphony.
 It is both intended as a practical toolkit to develop more complex software as well as a teaching resource.
 
-If you just want to use the standard library [[stdlib.asm]] is your main header, include it after your code.
+If you just want to use the standard library [[src/stdlib.asm]] is your main header, include it after your code.
 
 If you are using it as a learning resource have a look at the [teaching folder](teaching).
 

examples/README.md

@@ -0,0 +1,3 @@
+# Examples
+
+Examples of how to use the standard library to accomplish a task.

src/array.asm

@@ -0,0 +1,77 @@
+; Returns the index of the first element matching the provided predicate function (or -1 if not found)
+; Arguments:
+; r1 - The array pointer
+; r2 - The array length (number of items)
+; r3 - The stride (size of one item) - either 1, 2 or 4 (bytes)
+; r4 - The predicate
+; r5 - Predicate context
+; Result:
+; r1 - The index of the first element matching the provided predicate function (or -1 if not found)
+; Clobbers: r2, r3, r4, r5, r6, + what the predicate clobbers
+; Info:
+; The predicate function should follow the stdlib calling convention
+; The predicate receives two arguments (the value and the predicate context) and should return either a zero when the value is not the one we search for
+; , or any other result if it is the searched-for item.
+pub find_index:
+    push r12 ; We will store the predicate pointer here
+    push r11 ; We will store the current pointer here
+    push r10 ; We will store the stride here
+    push r9 ; We will store the final address here
+    push r8 ; We will store the mask here
+
+    mov r12, r4
+    mov r11, r1
+    mov r10, r3
+    mov r9, r2
+    lsr r6, r3, 1 ; We turn the stride into a byte shift
+    lsl r9, r9, r6 ; We calculate bytes left
+    add r9, r9, r1 ; We add the start address to get the final address
+
+    push r13 ; We save up the return address because we will provide our own to the predicate
+    push r1 ; We need the array pointer to calculate the item index
+    counter r13
+    add r13, r13, 52 ; Point to just after the predicate call - we can set this up now so we don't waste loop cycles
+
+    nand r8, zr, zr ; We create a mask of 0xFFFFFFFF
+    mov r6, 4
+    sub r6, r6, r3 ; We create a "negative stride", e.g. 4 -> 0, 2 -> 2, 1 -> 3
+    lsl r6, r6, 3
+    lsr r8, r8, r6 ; We shift the mask by the negative stride to obtain the proper mask for a value
+    ; e.g. stride 4 -> mask is 0xFFFFFFFF
+    ; stride 2 -> mask is 0x0000FFFF
+    ; stride 1 -> mask is 0x000000FF
+
+    push r5 ; We save the predicate context on the stack
+    find_index_loop:
+        load_32 r1, [r11] ; We load the element
+        and r1, r1, r8 ; We mask it to handle stride 2 and 1 cases
+        load_32 r2, [sp] ; We load the predicate context into r2
+        jmp r12 ; We call the predicate
+        cmp r1, zr
+        jne find_index_found_item ; If we found the item, we jump out
+        ; If we didn't, move to next item
+        add r11, r11, r10 ; We add the stride to the pointer
+        cmp r11, r9 ; We compare with the final address
+        jne find_index_loop ; If we did not reach the end we jump back into the loop
+    
+    find_index_not_found:
+        add sp, sp, 8 ; The predicate context and old array pointer are not useful
+        pop r13 ; We get our return address
+        nand r1, zr, zr ; We put -1 in r1
+        jmp find_index_postamble
+    
+    find_index_found_item:
+        add sp, sp, 4 ; The predicate context is not useful
+        pop r1 ; We get the array pointer
+        pop r13 ; We get our return address
+        sub r1, r11, r1 ; We calculate the bytes from the start
+        lsr r10, r10, 1 ; We shift the stride to get the amount to shift the bytes for
+        lsr r1, r1, r10 ; We shift to get the index of the item
+
+    find_index_postamble:
+        pop r8
+        pop r9
+        pop r10
+        pop r11
+        pop r12
+        jmp r13 ; Return

src/imath.asm

@@ -4,22 +4,53 @@
 ; r2 - The second value
 ; Result:
 ; r1 - The lower 32 bits of the result
-; Clobbers: r2, r3, r4, r5
+; Clobbers: r2, r3, r4, r5, r6
 pub mul_low:
-	mov r3, 0 ; result
-	mov r4, 31 ; loop counter
+	cmp r1, r2
+	jbe mul_low_noswap
+	xor r1, r2, r1
+	xor r2, r1, r2
+	xor r1, r2, r1
+	; Passthrough
 
-	mul_low_loop:
-		asr r5, r2, 31
-		and r5, r5, r1
-		lsl r5, r5, r4
-		add r3, r3, r5
-		lsl r2, r2, 1
-		sub r4, r4, 1
-		cmp r4, 0
-		jge mul_low_loop
-	mov r1, r3
+; Multiplies r1 and r2, returning the lower part of the result
+; This method assumes r1 is smaller than r2, which results in faster execution
+; Arguments:
+; r1 - The first value
+; r2 - The second value
+; Result:
+; r1 - The lower 32 bits of the result
+; Clobbers: r2, r3, r4, r5, r6
+pub mul_low_noswap:
+	mov r4, 0 ; r4 has the result
+	mov r6, mul_loop_end
+	add r3, r2, r2
 	
+	mul_loop:
+		and r5, r1, 14 ; Taking the first four bits, discarding the odd
+		lsl r5, r5, 1 ; 0000 - 0, 0001 - 2, 0010 - 4, 0011 - 8, etc.
+		sub r5, r6, r5
+		jmp r5
+		
+		add r4, r4, r3
+		add r4, r4, r3
+		add r4, r4, r3
+		add r4, r4, r3
+		add r4, r4, r3
+		add r4, r4, r3
+		add r4, r4, r3
+		mul_loop_end:
+			mov flags, r1
+			jne mul_skip_one
+			add r4, r4, r2
+			mul_skip_one:
+			lsl r2, r2, 4
+			lsl r3, r3, 4
+			lsr r1, r1, 4
+			cmp r1, zr
+			jne mul_loop
+
+	mov r1, r4
 	jmp r13
 
 ; Calculates the absolute value of the value provided in the r1 register

src/mem.asm

@@ -0,0 +1,174 @@
+; int compare(uint8_t* a, uint8_t* b, size_t count);
+;   Compares two memory segment of equal length lexicographically.
+;
+; Arguments:
+; - `r1`: A pointer to the first memory segment.
+; - `r2`: A pointer to the second memory segment.
+; - `r3`: The size of both memory segments.
+; Results:
+; - `r1`:
+;   -  `0` if both segments are equal.
+;   - `<0` if the first segment is less than the second segment.
+;   - `>0` if the first segment is greater than the second segment.
+;
+pub compare:
+    ; Exclusive end point of the first segment.
+    add r3, r3, r1
+    sub r3, r3, 4
+_compare__loop:
+    load_32 r4, [r1]
+    add r1, r1, 4
+    load_32 r5, [r2]
+    add r2, r2, 4
+    ; Comparing two sequences of 4 bytes lexicographically is equivalent to
+    ; comparing the corresponding big endian 32 bit words.
+    cmp r4, r5
+    jne _compare__break
+    ; Check if there are enough bytes left to continue with the vectorized loop.
+    cmp r1, r3
+    jbe _compare__loop
+    ; `r3 + 4 - r1 = <remaining byte count> = r3 - r1 mod 4`
+    sub flags, r3, r1
+    ; Check if one of the lowest 2 bits is non-zero
+    jbe _compare__rem
+    ; If not, we are done. Both segments are equal.
+    mov r1, 0
+    jmp r13
+_compare__break:
+    ; `flags` is the comparison result in the format of `cmp`. Convert it to the desired format.
+    ; 00 => 0x40000000 > 0
+    ; 01 => 0x00000000 = 0
+    ; 10 => 0xC0000000 < 0
+    xor r1, flags, 1
+    lsl r1, r1, 30
+    jmp r13
+_compare__rem:
+    ; Compute `S = 8*(4 - <remaining byte count>)` and
+    ; [r1] >> S, [r2] >> S
+    mov r3, 8
+    load_32 r4, [r1]
+    sub r3, r3, flags
+    load_32 r5, [r2]
+    lsl r3, r3, 3
+    lsr r4, r4, r3
+    lsr r5, r5, r3
+    ; Compare both values, now with garbage bytes removed.
+    cmp r4, r5
+    jmp _compare__break
+
+
+; void copy(void* src, void* dest, size_t count);
+;   Copies `count` bytes from `src` to `dest`. The two memory segments must not overlap.
+;
+; Arguments:
+; - `r1`: Pointer to the memory segment to be copied.
+; - `r2`: Pointer to the memory segment to be copied into.
+; - `r3`: Byte size of both the `src` and `dest` segments.
+;
+pub copy:
+    ; Exclusive end point of the source segment.
+    add r3, r1, r3
+    ; Last index from where we can safely copy 8 bytes per loop iteration.
+    sub r3, r3, 8
+    jmp _copy__loop_entry
+_copy__loop:
+    ; Copy 8 bytes from `src` to `dest`.
+    load_32 flags, [r1]
+    add r1, r1, 4
+    store_32 [r2], flags
+    add r2, r2, 4
+    load_32 flags, [r1]
+    add r1, r1, 4
+    store_32 [r2], flags
+    add r2, r2, 4
+_copy__loop_entry:
+    ; Check if we can process more data in the vectorized loop.
+    cmp r1, r3
+    jbe _copy__loop
+    ; The remaining amount of bytes `R` is `R = r3 + 8 - r1 = r3 - r1 mod 8`.
+    sub flags, r3, r1
+    ; Test if `R` is not a multiple of `4`, i.e. the lowest 2 bits are non-zero.
+    jbe _copy__rem
+    ; `R` is a multiple of `4`. Special case this.
+    ; Check if `R` is `0`, i.e. the third bit is also 0. In that case, we are already done.
+    ; There are no conditional indirect jumps, so we can't return immediately.
+    jge _copy__ret
+    ; `R = 4`. No need to update `r1` or `r2`, we don't need them anymore.
+    load_32 flags, [r1]
+    store_32 [r2], flags
+_copy__ret:
+    ; Return
+    jmp r13
+_copy__rem:
+    ; Optimize the remaining cases for code size.
+    ; End point of the source segment.
+    add r3, r3, 8
+    ; We already handled the case `R = 0` earlier,
+    ; so no bounds check needed for the first iteration.
+_copy__rem_loop:
+    ; Copy 1 byte.
+    load_8 flags, [r1]
+    add r1, r1, 1
+    store_8 [r2], flags
+    add r2, r2, 1
+    ; Check if we are still within the bounds.
+    cmp r1, r3
+    jb _copy__rem_loop
+    jmp r13
+    
+; void fill32(uint8_t* dest, size_t count, uint32_t value);
+;   Fills `count` bytes in `dest` with `value`. If `count` is not a multiple of 4, 
+;   the least significant bytes of `value` are cut off for the last entry.
+;
+; Arguments:
+; - `r1`: A pointer to the destination segment.
+; - `r2`: The size of the destination segment.
+; - `r3`: The 32 bit value that the segment is filled with.
+;
+pub fill32:
+    ; Exclusive end point of the destination segment.
+    add r2, r1, r2
+    ; Last index from where we can safely write 8 bytes per loop iteration.
+    sub r2, r2, 8
+    jmp _fill32__entry
+_fill32__loop:
+	; Set 8 bytes per loop iteraion.
+    store_32 [r1], r3
+    add r1, r1, 4
+    store_32 [r1], r3
+    add r1, r1, 4
+_fill32__entry:
+    ; Check if we can process more data in the vectorized loop.
+    cmp r1, r2
+    jbe _fill32__loop
+    ; The remaining amount of bytes `R` is `R = r2 + 8 - r1 = r2 - r1 mod 8`.
+    sub flags, r2, r1
+    ; Check if the third bit of the remainder is cleared.
+    jge _fill32__r4
+    ; Otherwise set 4 bytes.
+    store_32 [r1], r3
+    add r1, r1, 4
+_fill32__r4:
+    ; Check if the two least significant bits of the remainder are zero.
+    ja _fill32__ret
+    ; Handle the remaining bytes `R` individually, in reverse order.
+    add r2, r2, 4
+    ; `r1 + 4 - r2 = 4 - R`.
+    sub flags, r1, r2
+    ; Exclusive end point of the destination segment.
+    add r2, r2, 4
+    ; Shift out the least significant `8*(4 - R)` bits of the value.
+    lsl flags, flags, 3
+    lsr r3, r3, flags
+    jmp _fill32__loop2_entry
+_fill32__loop2:
+    sub r2, r2, 1
+    ; Write the least significant byte of the value...
+    store_8 [r2], r3
+    ; and then shift it out.
+    lsr r3, r3, 8
+_fill32__loop2_entry:
+    cmp r1, r2
+    jb _fill32__loop2
+_fill32__ret:
+    jmp r13

src/stdlib.asm

@@ -1,6 +1,8 @@
 pub include bit
 pub include imath
+pub include array
 pub include console
+pub include mem
 
 ; Needs to be last!
 pub include LUTs

src/unsafe_mem.asm

@@ -0,0 +1,77 @@
+; int compare(uint8_t* a, uint8_t* b, size_t count);
+;   Compares two memory segment of equal length lexicographically.
+;   Temporarily modifies the byte at address `a + count`.
+;
+; Arguments:
+; - `r1`: A pointer to the first memory segment.
+; - `r2`: A pointer to the second memory segment.
+; - `r3`: The size of both memory segments.
+; Results:
+; - `r1`:
+;   -  `0` if both segments are equal.
+;   - `<0` if the first segment is less than the second segment.
+;   - `>0` if the first segment is greater than the second segment.
+pub compare:
+    cmp r1, r2
+    je _compare__is_eq
+    ; Exclusive end point of the second segment.
+    add r4, r3, r2
+    ; Exclusive end point of the first segment.
+    add r3, r3, r1
+    load_8 r6, [r3]
+    load_8 flags, [r4]
+    ; Check if the first bytes behind the sequences are equal.
+    cmp flags, r6
+    jne _compare__loop
+    ; Change the byte directly behind the first segment.
+    xor r4, r6, 1
+    ; This would be problematic if someone calls compare with a first segment
+    ; whose end point overlaps the program memory of this function.
+    store_8 [r3], r4
+_compare__loop:
+    load_32 r4, [r1]
+    add r1, r1, 4
+    load_32 r5, [r2]
+    add r2, r2, 4
+    ; Comparing two sequences of 4 bytes lexicographically is equivalent to
+    ; comparing the corresponding big endian 32 bit words.
+    cmp r4, r5
+    je _compare__loop
+    ; We overshot in the loop; decrement r1 again. (Only by 2, we backtrack the rest if necessary later)
+    sub r1, r1, 2
+    ; Restore the byte we changed.
+    store_8 [r3], r6
+    ; We encountered two different words. Figure out what byte they differ on.
+    xor r4, r4, r5
+    ; Store the flags for later, to figure out the return value.
+    mov r5, flags
+    ; Check if at least one of the two most significant bytes is not 0.
+    cmp r4, 0xffff
+    jbe _compare__low2
+    ; If it is, backtrack the remaining 2 indices.
+    ; Shift the most significant bytes to the least significant ones.
+    sub r1, r1, 2
+    lsr r4, r4, 16
+_compare__low2:
+    ; r1 now points to a non-zero 16 bit value.
+    ; If the 16 bit value at r1-2 is in-bounds, then it is 0.
+    ; Check if the most significant byte of the 16 bit value is 0.
+    cmp r4, 0xff
+    ja _compare__low1
+    ; If it is, our target is the least significant byte.
+    add r1, r1, 1
+_compare__low1:
+    ; Otherwise, the target is that non-zero byte.
+    ; Check if the target is out of bounds, i.e. the loop terminated through the "bounds check".
+    cmp r1, r3
+    jae _compare__is_eq
+    ; r5 is the comparison result in the format of `cmp`. Convert it to the desired format.
+    ; 00 => 0x40000000 > 0
+    ; 01 => 0x00000000 = 0
+    ; 10 => 0xC0000000 < 0
+    xor r1, r5, 1
+    lsl r1, r1, 30
+    jmp r13
+_compare__is_eq:
+    mov r1, 0
+    jmp r13

tests/README.md

@@ -0,0 +1,3 @@
+# Tests
+
+Tests for the standard library go here, tests are allowed to depend on the recommended spec.isa changes.