https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94927
Bug ID: 94927
Summary: cortex-m7: Incorrect insn_reservation for
vmaxnm/vminnm instructions
Product: gcc
Version: 7.3.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: rtl-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: stas.ibragimov at gmail dot com
Target Milestone: ---
There is no any pipeline usage description for vmaxnm and vminnm instructions
in cortex-m7.md. This instruction is described as smax<mode>3 and smin<mode>3
in vfp.md and has type f_minmax<vfp_type>. But in cortex-m7.md there is not any
insn_reservations for this type.
So, pipeline usage for code with this instructions is calculated wrong.
For example:
//test.c
#include <math.h>
float test(int a, int b) {
float max = fmaxf(a,b);
float min = fminf(a,b);
float x=min+1;
return max/min-x;
}
Compile:
arm-none-eabi-gcc test.c -c -da -O3 -mcpu=cortex-m7 -mfpu=fpv5-d16
-mfloat-abi=hard -mthumb -ffinite-math-only -fno-signed-zeros -save-temps
test.c.293r.sched2 (wrong places is marked with (!!)):
;; 0--> b 0: i 24 s15=r0
:cm7_i0|cm7_i1,cm7_a0|cm7_a1
;; 0--> b 0: i 12 s12=1.0e+0
:cm7_i0|cm7_i1,cm7_fpu
;; 2--> b 0: i 7 s14=flt(s15)
:cm7_i0|cm7_i1,cm7_fpu
;; 2--> b 0: i 25 s15=r1
:cm7_i0|cm7_i1,cm7_a0|cm7_a1
;; 4--> b 0: i 8 s13=flt(s15)
:cm7_i0|cm7_i1,cm7_fpu
;; 7--> b 0: i 9 s15=smin(s14,s13) :nothing
(!!)
;; 7--> b 0: i 10 s14=smax(s14,s13) :nothing
(!!)
;; 8--> b 0: i 11 s0=s14/s15
:cm7_i0|cm7_i1,cm7_fpu*5
;; 13--> b 0: i 13 s15=s15+s12
:cm7_i0|cm7_i1,cm7_fpu
;; 24--> b 0: i 19 s0=s0-s15
:cm7_i0|cm7_i1,cm7_fpu
;; 24--> b 0: i 20 use s0 :nothing
;; 24--> b 0: i 28 simple_return
:cm7_i0|cm7_i1,cm7_branch
