https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69248
Bug ID: 69248
Summary: VFP register constraint 'w' is ignored when
optimizations are enabled
Product: gcc
Version: 5.3.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: inline-asm
Assignee: unassigned at gcc dot gnu.org
Reporter: vorfeed.canal at gmail dot com
Target Milestone: ---
Looks like optimizations confuse something in GCC and cause it to ignore proper
'w' constraint. Unoptimized version produces "vcvt.f64.s32 d16, s15" command
which is accepted by assembler, optimized version produces "vcvt.f64.s32 s16,
s15" which assembler does not like.
$ cat test2.cc
struct FILE;
int fprintf(FILE *stream, const char *format, ...);
extern FILE *gcc_tests_file;
int foo() {
do {
double result;
__asm__ __volatile__( "vcvt.f64.s32" " %[result], %[arg1]\n" :
[result]"=w"(result) : [arg1]"t"(0x01020304) );
fprintf(gcc_tests_file, "vcvt.f64.s32: ");
fprintf(gcc_tests_file, "%.8g", result);
} while (false);
}
$ arm-linux-gnueabihf-g++-5 -march=armv7-a -mfpu=vfpv3-d16 -mfloat-abi=softfp
-mfpu=neon -mtls-dialect=gnu -O0 test.cc -o-
.arch armv7-a
.fpu neon
.eabi_attribute 20, 1
.eabi_attribute 21, 1
.eabi_attribute 23, 3
.eabi_attribute 24, 1
.eabi_attribute 25, 1
.eabi_attribute 26, 2
.eabi_attribute 30, 6
.eabi_attribute 34, 1
.eabi_attribute 18, 4
.file "test2.cc"
.section .rodata
.align 2
.LC0:
.ascii "vcvt.f64.s32\000"
.align 2
.LC1:
.ascii "%.8g\000"
.text
.align 2
.global _Z3foov
.syntax unified
.arm
.type _Z3foov, %function
_Z3foov:
.fnstart
.LFB0:
@ args = 0, pretend = 0, frame = 8
@ frame_needed = 1, uses_anonymous_args = 0
push {fp, lr}
.save {fp, lr}
.setfp fp, sp, #4
add fp, sp, #4
.pad #8
sub sp, sp, #8
mov r3, #772
movt r3, 258
vmov s15, r3 @ int
.syntax divided
@ 8 "test2.cc" 1
vcvt.f64.s32 d16, s15
@ 0 "" 2
.syntax unified
vstr.64 d16, [fp, #-12]
movw r3, #:lower16:gcc_tests_file
movt r3, #:upper16:gcc_tests_file
ldr r3, [r3]
movw r1, #:lower16:.LC0
movt r1, #:upper16:.LC0
mov r0, r3
bl _Z7fprintfP4FILEPKcz
movw r3, #:lower16:gcc_tests_file
movt r3, #:upper16:gcc_tests_file
ldr r0, [r3]
ldrd r2, [fp, #-12]
movw r1, #:lower16:.LC1
movt r1, #:upper16:.LC1
bl _Z7fprintfP4FILEPKcz
nop
mov r0, r3
sub sp, fp, #4
@ sp needed
pop {fp, pc}
.fnend
.size _Z3foov, .-_Z3foov
.ident "GCC: (Ubuntu/Linaro 5.3.1-5ubuntu1) 5.3.1 20160101"
.section .note.GNU-stack,"",%progbits
$ arm-linux-gnueabihf-g++-5 -march=armv7-a -mfpu=vfpv3-d16 -mfloat-abi=softfp
-mfpu=neon -mtls-dialect=gnu -O3 test2.cc -o-
.arch armv7-a
.fpu neon
.eabi_attribute 20, 1
.eabi_attribute 21, 1
.eabi_attribute 23, 3
.eabi_attribute 24, 1
.eabi_attribute 25, 1
.eabi_attribute 26, 2
.eabi_attribute 30, 2
.eabi_attribute 34, 1
.eabi_attribute 18, 4
.file "test2.cc"
.text
.align 2
.global _Z3foov
.syntax unified
.arm
.type _Z3foov, %function
_Z3foov:
.fnstart
.LFB0:
@ args = 0, pretend = 0, frame = 0
@ frame_needed = 0, uses_anonymous_args = 0
mov r3, #772
movt r3, 258
push {r4, lr}
.save {r4, lr}
vmov s15, r3 @ int
vpush.64 {d8}
.vsave {d8}
.syntax divided
@ 8 "test2.cc" 1
vcvt.f64.s32 s16, s15
@ 0 "" 2
.syntax unified
movw r4, #:lower16:gcc_tests_file
movt r4, #:upper16:gcc_tests_file
movw r1, #:lower16:.LC0
movt r1, #:upper16:.LC0
ldr r0, [r4]
bl _Z7fprintfP4FILEPKcz
vmov r2, r3, d8
vldm sp!, {d8}
movw r1, #:lower16:.LC1
ldr r0, [r4]
movt r1, #:upper16:.LC1
pop {r4, lr}
b _Z7fprintfP4FILEPKcz
.fnend
.size _Z3foov, .-_Z3foov
.section .rodata.str1.4,"aMS",%progbits,1
.align 2
.LC0:
.ascii "vcvt.f64.s32\000"
.space 3
.LC1:
.ascii "%.8g\000"
.ident "GCC: (Ubuntu/Linaro 5.3.1-5ubuntu1) 5.3.1 20160101"
.section .note.GNU-stack,"",%progbits
