https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70581
Bug ID: 70581
Summary: [6 regression] gcc.dg/lto/simd-function FAILs
Product: gcc
Version: 6.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: ro at gcc dot gnu.org
Target Milestone: ---
Host: i386-pc-solaris2.1[012]
Target: i386-pc-solaris2.1[012]
Build: i386-pc-solaris2.1[012]
After 20160401, the gcc.dg/lto/simd-function testcase began to FAIL on
Solaris/x86
with /bin/as:
FAIL: gcc.dg/lto/simd-function c_lto_simd-function_0.o assemble, -fopenmp-simd
-O3 -ffast-math -mavx2 -flto -flto-partition=max
UNRESOLVED: gcc.dg/lto/simd-function
c_lto_simd-function_0.o-c_lto_simd-function_0.o execute -fopenmp-simd -O3
-ffast-math -mavx2 -flto -flto-partition=max
UNRESOLVED: gcc.dg/lto/simd-function
c_lto_simd-function_0.o-c_lto_simd-function_0.o link -fopenmp-simd -O3
-ffast-math -mavx2 -flto -flto-partition=max
Assembler: simd-function_0.c
"/var/tmp//ccw.aGWM.s", line 850 : Invalid instruction argument
Near line: " vmulps %zmm1, %zmm0, %zmm0"
"/var/tmp//ccw.aGWM.s", line 858 : Illegal mnemonic
Near line: " vpxord %zmm3, %zmm3, %zmm3"
"/var/tmp//ccw.aGWM.s", line 858 : Syntax error
Near line: " vpxord %zmm3, %zmm3, %zmm3"
"/var/tmp//ccw.aGWM.s", line 860 : Invalid instruction argument
Near line: " vmulps %zmm1, %zmm0, %zmm0"
"/var/tmp//ccw.aGWM.s", line 863 : Illegal mnemonic
Near line: " vpcmpd $4, %zmm3, %zmm2, %k1"
"/var/tmp//ccw.aGWM.s", line 863 : Syntax error
Near line: " vpcmpd $4, %zmm3, %zmm2, %k1"
"/var/tmp//ccw.aGWM.s", line 866 : Illegal mnemonic
Near line: " kmovw %k1, %eax"
"/var/tmp//ccw.aGWM.s", line 866 : Syntax error
Near line: " kmovw %k1, %eax"
"/var/tmp//ccw.aGWM.s", line 870 : Invalid instruction argument
Near line: " vmovaps %zmm0, -120(%ebp){%k1}"
"/var/tmp//ccw.aGWM.s", line 872 : Invalid instruction argument
Near line: " vmovaps -120(%ebp), %zmm0"
for both 32 and 64-bit. It seems the compiler started to emit avx512f insns
despite the -mavx2. Either the testcase needs to be amended requiring the
corresponding effective-target, or the compiler fixed not to emit those unless
told to...
Here's a diff between the assembler outputs from 20160401 and 20160407:
--- /homes/ro/simd-function_0.s 2016-04-07 18:05:16.449629000 +0200
+++ simd-function_0.s 2016-04-07 18:06:05.170292167 +0200
[...]
@@ -846,6 +846,39 @@
leal -4(%ecx), %esp
ret
.size _ZGVdM8vv_my_mul, .-_ZGVdM8vv_my_mul
+ .p2align 4,,15
+ .globl _ZGVeN16vv_my_mul
+ .type _ZGVeN16vv_my_mul, @function
+_ZGVeN16vv_my_mul:
+ vmulps %zmm1, %zmm0, %zmm0
+ ret
+ .size _ZGVeN16vv_my_mul, .-_ZGVeN16vv_my_mul
+ .p2align 4,,15
+ .globl _ZGVeM16vv_my_mul
+ .type _ZGVeM16vv_my_mul, @function
+_ZGVeM16vv_my_mul:
+ leal 4(%esp), %ecx
+ vpxord %zmm3, %zmm3, %zmm3
+ andl $-64, %esp
+ vmulps %zmm1, %zmm0, %zmm0
+ pushl -4(%ecx)
+ pushl %ebp
+ vpcmpd $4, %zmm3, %zmm2, %k1
+ movl %esp, %ebp
+ pushl %ecx
+ kmovw %k1, %eax
+ subl $308, %esp
+ testw %ax, %ax
+ je .L89
+ vmovaps %zmm0, -120(%ebp){%k1}
+.L89:
+ vmovaps -120(%ebp), %zmm0
+ addl $308, %esp
+ popl %ecx
+ popl %ebp
+ leal -4(%ecx), %esp
+ ret
+ .size _ZGVeM16vv_my_mul, .-_ZGVeM16vv_my_mul
.comm x,16384,32
.section .rodata
.align 4
Rainer
