freebsd-dev/contrib/bmake/unit-tests/cond-op.mk
Simon J. Gerraty 129043849f Merge bmake-20211212
commit '2935fe8237c83c1dcb113dd5335733263e68e6fd'
2021-12-18 10:09:14 -08:00

146 lines
3.9 KiB
Makefile

# $NetBSD: cond-op.mk,v 1.15 2021/12/10 23:12:44 rillig Exp $
#
# Tests for operators like &&, ||, ! in .if conditions.
#
# See also:
# cond-op-and.mk
# cond-op-not.mk
# cond-op-or.mk
# cond-op-parentheses.mk
# In make, && binds more tightly than ||, like in C.
# If make had the same precedence for both && and ||, like in the shell,
# the result would be different.
# If || were to bind more tightly than &&, the result would be different
# as well.
.if !(1 || 1 && 0)
. error
.endif
# If make were to interpret the && and || operators like the shell, the
# previous condition would be interpreted as:
.if (1 || 1) && 0
. error
.endif
# The precedence of the ! operator is different from C though. It has a
# lower precedence than the comparison operators. Negating a condition
# does not need parentheses.
#
# This kind of condition looks so unfamiliar that it doesn't occur in
# practice.
.if !"word" == "word"
. error
.endif
# This is how the above condition is actually interpreted.
.if !("word" == "word")
. error
.endif
# TODO: Demonstrate that the precedence of the ! and == operators actually
# makes a difference. There is a simple example for sure, I just cannot
# wrap my head around it right now. See the truth table generator below
# for an example that doesn't require much thought.
# This condition is malformed because the '!' on the right-hand side must not
# appear unquoted. If any, it must be enclosed in quotes.
# In any case, it is not interpreted as a negation of an unquoted string.
# See CondParser_String.
.if "!word" == !word
. error
.endif
# Surprisingly, the ampersand and pipe are allowed in bare strings.
# That's another opportunity for writing confusing code.
# See CondParser_String, which only has '!' in the list of stop characters.
.if "a&&b||c" != a&&b||c
. error
.endif
# In the following malformed conditions, as soon as the parser sees the '$'
# after the '0' or the '1', it knows that the condition will be malformed.
# Therefore there is no point in evaluating the misplaced expression.
#
# Before cond.c 1.286 from 2021-12-10, the extra expression was evaluated
# nevertheless, since CondParser_Or and CondParser_And asked for the expanded
# next token, even though in this position of the condition, only comparison
# operators, TOK_AND, TOK_OR or TOK_RPAREN are allowed.
#
#
#
#
#
#
.undef ERR
.if 0 ${ERR::=evaluated}
. error
.endif
.if ${ERR:Uundefined} == undefined
. info A misplaced expression after 0 is not evaluated.
.endif
.undef ERR
.if 1 ${ERR::=evaluated}
. error
.endif
.if ${ERR:Uundefined} == undefined
. info A misplaced expression after 1 is not evaluated.
.endif
# Just in case that parsing should ever stop on the first error.
.info Parsing continues until here.
# Demonstration that '&&' has higher precedence than '||'.
.info A B C => (A || B) && C A || B && C A || (B && C)
.for a in 0 1
. for b in 0 1
. for c in 0 1
. for r1 in ${ ($a || $b) && $c :?1:0}
. for r2 in ${ $a || $b && $c :?1:0}
. for r3 in ${ $a || ($b && $c) :?1:0}
. info $a $b $c => ${r1} ${r2} ${r3}
. endfor
. endfor
. endfor
. endfor
. endfor
.endfor
# This condition is obviously malformed. It is properly detected and also
# was properly detected before 2021-01-19, but only because the left hand
# side of the '&&' evaluated to true.
.if 1 &&
. error
.else
. error
.endif
# This obviously malformed condition was not detected as such before cond.c
# 1.238 from 2021-01-19.
.if 0 &&
. error
.else
. error
.endif
# This obviously malformed condition was not detected as such before cond.c
# 1.238 from 2021-01-19.
.if 1 ||
. error
.else
. error
.endif
# This condition is obviously malformed. It is properly detected and also
# was properly detected before 2021-01-19, but only because the left hand
# side of the '||' evaluated to false.
.if 0 ||
. error
.else
. error
.endif
all:
@:;