https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120278
Bug ID: 120278
Summary: [9/10/11/12/13/14/15 Regression] Switch expansion
generates extra compares with -fno-jump-tables
Product: gcc
Version: 12.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: selectstriker2 at protonmail dot com
Target Milestone: ---
Created attachment 61423
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=61423&action=edit
Simple source file
Starting with GCC9, switch expansion to decision tree changed and has extra
comparison instructions that will only ever evaluate to false because the true
case has already been eliminated above.
This was noted when switching to the arm-none-eabi C compiler based on GCC9,
and exists in the x86_64 C compiler as well.
Godbolt example : https://godbolt.org/z/f1fzMhbEx
gcc main.c -fno-jump-tables
Correct in GCC8
.L10:
cmpl $2, -4(%rbp)
je .L2
cmpl $2, -4(%rbp)
jg .L3
cmpl $0, -4(%rbp)
je .L4
cmpl $1, -4(%rbp)
je .L5
jmp .L6
.L3:
cmpl $3, -4(%rbp)
je .L7
cmpl $4, -4(%rbp)
je .L8
jmp .L6
Extra/incorrect comparisons started in GCC9
.L9:
cmpl $4, -4(%rbp)
je .L2
cmpl $4, -4(%rbp)
jg .L3
cmpl $3, -4(%rbp)
je .L4
cmpl $3, -4(%rbp)
jg .L3 - already checked for ==4, >4, ==3 so >3 will always be
false
cmpl $2, -4(%rbp)
je .L5
cmpl $2, -4(%rbp)
jg .L3 - already checked for ==4, >4, ==3, ==2 so >2 will always
be false
cmpl $0, -4(%rbp)
je .L6
cmpl $1, -4(%rbp)
je .L7
jmp .L3
