PERF: read_csv macro updates

[WillAyd]

This gives us a few percentage points back in read_csv performance, especially with larger files. Credit to @Dr-Irv for the idea

[Dr-Irv]

Thanks for the shout-out @WillAyd . Back in the day, I was part of a team writing performance-critical code in C, and we used to do optimizations like this all the time.

[WillAyd]

If anyone cared for the details here is the verbose annotated assembly output (from -fverbose-asm gcc flag) for before:

.L246:
# pandas/_libs/src/parser/tokenizer.c:929:                 } else if (IS_DELIMITER(c)) {
	.loc 3 929 28 is_stmt 1
	movq	-136(%rbp), %rax	# self, tmp683
	movl	188(%rax), %eax	# self_452(D)->delim_whitespace, _180
# pandas/_libs/src/parser/tokenizer.c:929:                 } else if (IS_DELIMITER(c)) {
	.loc 3 929 27
	testl	%eax, %eax	# _180
	jne	.L247	#,
# pandas/_libs/src/parser/tokenizer.c:929:                 } else if (IS_DELIMITER(c)) {
	.loc 3 929 28 discriminator 1
	movq	-136(%rbp), %rax	# self, tmp684
	movzbl	184(%rax), %eax	# self_452(D)->delimiter, _181
	cmpb	%al, -49(%rbp)	# _181, c
	je	.L248	#,
.L247:
# pandas/_libs/src/parser/tokenizer.c:929:                 } else if (IS_DELIMITER(c)) {
	.loc 3 929 28 is_stmt 0 discriminator 3
	movq	-136(%rbp), %rax	# self, tmp685
	movl	188(%rax), %eax	# self_452(D)->delim_whitespace, _182
	testl	%eax, %eax	# _182
	je	.L249	#,
# pandas/_libs/src/parser/tokenizer.c:929:                 } else if (IS_DELIMITER(c)) {
	.loc 3 929 28 discriminator 4
	call	__ctype_b_loc@PLT	#
	movq	(%rax), %rdx	# *_183, _184
	movsbq	-49(%rbp), %rax	# c, _185
	addq	%rax, %rax	# _186
	addq	%rdx, %rax	# _184, _187
	movzwl	(%rax), %eax	# *_187, _188
	movzwl	%ax, %eax	# _188, _189
	andl	$1, %eax	#, _190
	testl	%eax, %eax	# _190
	je	.L249	#,
.L248:
# pandas/_libs/src/parser/tokenizer.c:930:                     if (self->delim_whitespace) {

and after

.L243:
# pandas/_libs/src/parser/tokenizer.c:930:                 } else if (IS_DELIMITER(c)) {
	.loc 3 930 27 is_stmt 1
	movzbl	-57(%rbp), %eax	# c, tmp657
	cmpb	-41(%rbp), %al	# delimiter, tmp657
	je	.L244	#,
# pandas/_libs/src/parser/tokenizer.c:930:                 } else if (IS_DELIMITER(c)) {
	.loc 3 930 28 discriminator 1
	cmpl	$0, -40(%rbp)	#, delim_whitespace
	je	.L245	#,
# pandas/_libs/src/parser/tokenizer.c:930:                 } else if (IS_DELIMITER(c)) {
	.loc 3 930 28 is_stmt 0 discriminator 2
	call	__ctype_b_loc@PLT	#
	movq	(%rax), %rdx	# *_171, _172
	movsbq	-57(%rbp), %rax	# c, _173
	addq	%rax, %rax	# _174
	addq	%rdx, %rax	# _172, _175
	movzwl	(%rax), %eax	# *_175, _176
	movzwl	%ax, %eax	# _176, _177
	andl	$1, %eax	#, _178
	testl	%eax, %eax	# _178
	je	.L245	#,
.L244:
# pandas/_libs/src/parser/tokenizer.c:931:                     if (self->delim_whitespace) {

Very low level...but this is executed for potentially every character in a file

[lithomas1]

Can you add a whatsnew?

[lithomas1]

It's probably worth clarifying that this is for the C engine only.

[mroeschke]

Thanks @WillAyd