http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43716
--- Comment #39 from Dominique d'Humieres <dominiq at lps dot ens.fr>
2012-07-22 10:47:42 UTC ---
> If I compile with -fno-tree-loop-if-convert the exception occurs
> somewhere else:
It occurs in s00018 in the loop between lines 3782 to 3796
DO j = 1 , n1
dtpar = V0012
IF ( i.EQ.NIV .AND. j.EQ.ISLni ) THEN
k1 = 2
DO k = 1 , k1
dtpa = V0012
IF ( DABS(DTMi(k)).GT.eps ) &
& dtpa = 2.*sens*DABS(TMI(k)/DTMi(k))
dtpar = DMIN1(dtpar,dtpa)
ENDDO
ELSEIF ( DABS(DTGai(i,j)).GT.eps ) THEN
dtpar = sens*DABS(TGAi(i,j)/DTGai(i,j))
ENDIF
DTTemp = DMIN1(DTTemp,dtpar)
ENDDO
Before the loop i=2, n1=2, NIV=0, ISLni=0, eps=1.0e-8, sens=0.01,
DTTemp=0.05, DTGai(2,1)=3.4750524086054519, DTGai(2,2)=2.8245839253949421,
TGAi(2,1)=905.00000000000000, and TGAi(2,1)=1040.0000000000000. So
(i.EQ.NIV .AND. j.EQ.ISLni) is false, DTGai(2,:)>eps and dtpar>20.0>DTTemp,
hence the loop does nothing and should not generate any "Floating-point
exception" as it does.
If I read correctly the assembly bellow, the loop
k1 = 2
DO k = 1 , k1
dtpa = V0012
IF ( DABS(DTMi(k)).GT.eps ) &
& dtpa = 2.*sens*DABS(TMI(k)/DTMi(k))
dtpar = DMIN1(dtpar,dtpa)
ENDDO
is unrolled and moved outside the 'DO j = 1 , n1' loop when s00018.f90 is
compiled with -O2 -ffpe-trap=invalid -fno-trapping-math. The worst part of
the optimization and the cause of the floating-point exception seems that
the test 'IF ( DABS(DTMi(k)).GT.eps )' is removed (or delayed, even if I
replace eps
with 1.0) while both TMI and DTMI are filled with zeroes.
Note that if I replace
IF ( DABS(DTMi(k)).GT.eps ) &
& dtpa = 2.*sens*DABS(TMI(k)/DTMi(k))
with
dtpa = 2.*sens*DABS(TMI(k))/max(DABS(DTMi(k)),eps)
there is no exception.
IMO computing TMI(k)/DTMi(k) without the IF is over-agressive: if
DABS(DTMi(k)).LT.eps, dtpa will have to be recomputed. In addition, if the
k loop is moved outside the j loop, it should at least protected by
IF ( i.EQ.NIV), or better by
IF ( i.EQ.NIV .AND. ISLni.GE.1 .AND. ISLni.LE.n1).
Also since the ELSEIF depends on j and the IF is taken at most once,
moving the k loop outside the j loop seems dubious.
Note that the exception is (should be) harmless as it occurs in a dead
branch.
L29:
movl 36404(%r9,%rdi,4), %ecx # aaa77.nasl, n1
movsd %xmm9, 936(%r11) # x, aaa11.dtphy
movsd 856(%r11), %xmm8 # aaa11.v22202, sens
movl $1, 396(%rsp) #, j
testl %ecx, %ecx # n1
jle L41 #,
movsd 19416(%r9), %xmm7 # aaa77.dtmi, D.2813
movapd %xmm8, %xmm11 # sens, D.2817
movl $0, %edx #, tmp908
movsd 19400(%r9), %xmm0 # aaa77.tmi, x
addsd %xmm8, %xmm11 # sens, D.2817
movapd %xmm7, %xmm10 # D.2813, D.2814
movq _aaa13_@GOTPCREL(%rip), %rax #,
divsd %xmm7, %xmm0 # D.2813, x
andpd %xmm1, %xmm10 # tmp1323, D.2814
movsd 19424(%r9), %xmm12 # aaa77.dtmi, D.2813
movsd 19408(%r9), %xmm7 # aaa77.tmi, x
movsd 944(%r11), %xmm6 # aaa11.dttemp, x
movl 240(%rax), %esi # aaa13.niv, pretmp.160
movl 244(%rax), %r14d # aaa13.islni, D.2806
xorl %eax, %eax # pretmp.165
divsd %xmm12, %xmm7 # D.2813, x
andpd %xmm1, %xmm0 # tmp1323, x
mulsd %xmm11, %xmm0 # D.2817, x
ucomisd %xmm0, %xmm2 # x, x
seta %al #, pretmp.165
ucomisd %xmm3, %xmm10 # tmp1324, D.2814
cmovbe %edx, %eax # pretmp.165,, tmp908, pretmp.165
xorl %edx, %edx #
ucomisd %xmm2, %xmm2 # x, x
setp %dl #,
addl $1, %ecx #,
movl %ecx, 12(%rsp) #, %sfp
xorl %ecx, %ecx #
movl %edx, 336(%rsp) #, %sfp
leaq 30032(%r9,%rdi,8), %rdx #, ivtmp.264
andpd %xmm1, %xmm7 # tmp1323, x
mulsd %xmm11, %xmm7 # D.2817, x
movapd %xmm3, %xmm11 # tmp1324,
cmpltsd %xmm10, %xmm11 #, D.2814,
movapd %xmm11, %xmm10 #, tmp923
andpd %xmm11, %xmm0 # tmp923, tmp924
movapd %xmm2, %xmm11 # x, x
andnpd %xmm2, %xmm10 # x, tmp921
orpd %xmm0, %xmm10 # tmp924, tmp921
ucomisd %xmm10, %xmm10 # tmp921, tmp921
setp %cl #,
orl 336(%rsp), %eax # %sfp, pretmp.165
movl %ecx, 364(%rsp) #, %sfp
movl %eax, 336(%rsp) # pretmp.165, %sfp
je L34 #,
movapd %xmm10, %xmm11 # tmp921, x
L34:
movapd %xmm3, %xmm0 # tmp1324,
andpd %xmm1, %xmm12 # tmp1323, D.2814
xorl %eax, %eax #
cmpltsd %xmm12, %xmm0 #, D.2814,
movapd %xmm1, %xmm10 # tmp1323, tmp1385
movl 12(%rsp), %ecx # %sfp, D.3602
movapd %xmm0, %xmm12 #, tmp933
andpd %xmm0, %xmm7 # tmp933, tmp934
andnpd %xmm2, %xmm12 # x, x
orpd %xmm7, %xmm12 # tmp934, x
ucomisd %xmm12, %xmm11 # x, x
seta %al #,
movl %eax, 360(%rsp) #, %sfp
movl $1, %eax #, prephitmp.161
jmp L42 #
.align 4,0x90
L255:
movsd -10560(%rdx), %xmm0 # MEM[base: D.3603_170, offset: -10560B], x
divsd %xmm7, %xmm0 # D.2831, x
andpd %xmm10, %xmm0 # tmp1385, x
mulsd %xmm8, %xmm0 # sens, x
L37:
ucomisd %xmm0, %xmm6 # x, x
ja L160 #,
ucomisd %xmm6, %xmm6 # x, x
jnp L39 #,
L160:
movapd %xmm0, %xmm6 # x, x
L39:
addl $1, %eax #, prephitmp.161
addq $176, %rdx #, ivtmp.264
cmpl %ecx, %eax # D.3602, prephitmp.161
je L253 #,
L42:
cmpl %esi, %r10d # pretmp.160, prephitmp.132
je L254 #,
L35:
movsd (%rdx), %xmm7 # MEM[base: D.3603_170, offset: 0B], D.2831
movapd %xmm7, %xmm0 # D.2831, D.2832
andpd %xmm1, %xmm0 # tmp1323, D.2832
ucomisd %xmm3, %xmm0 # tmp1324, D.2832
ja L255 #,
movapd %xmm2, %xmm0 # x, x
jmp L37 #
.align 4,0x90
L254:
cmpl %eax, %r14d # prephitmp.161, D.2806
jne L35 #,
cmpl $0, 336(%rsp) #, %sfp
je L238 #,
cmpl $0, 364(%rsp) #, %sfp
jne L158 #,
L238:
cmpl $0, 360(%rsp) #, %sfp
movapd %xmm11, %xmm0 # x, x
je L37 #,
L158:
movapd %xmm12, %xmm0 # x, x
jmp L37 #
.align 4,0x90
L253:
movl %eax, 396(%rsp) # prephitmp.161, j
movsd %xmm6, 944(%r11) # x, aaa11.dttemp
