| Summary: | Excessive loop versioning done by vectorization + predictive commoning | ||
|---|---|---|---|
| Product: | gcc | Reporter: | Richard Biener <rguenth> |
| Component: | tree-optimization | Assignee: | Not yet assigned to anyone <unassigned> |
| Status: | RESOLVED WORKSFORME | ||
| Severity: | normal | Keywords: | missed-optimization |
| Priority: | P3 | ||
| Version: | 4.7.0 | ||
| Target Milestone: | --- | ||
| Host: | Target: | ||
| Build: | Known to work: | ||
| Known to fail: | Last reconfirmed: | ||
| Bug Depends on: | |||
| Bug Blocks: | 53947 | ||
On the trunk I only see one copy of the loop:
.L11:
movups (%rbx,%rax), %xmm7
movups 0(%rbp,%rax), %xmm0
movups (%r9,%rax), %xmm1
subps %xmm7, %xmm0
movups (%r11,%rax), %xmm7
addq $16, %rax
addps %xmm7, %xmm1
mulps %xmm6, %xmm0
mulps %xmm1, %xmm0
movaps %xmm0, %xmm1
addss %xmm2, %xmm1
movaps %xmm0, %xmm2
shufps $85, %xmm0, %xmm2
addss %xmm2, %xmm1
movaps %xmm0, %xmm2
unpckhps %xmm0, %xmm2
shufps $255, %xmm0, %xmm0
addss %xmm2, %xmm1
movaps %xmm1, %xmm2
addss %xmm0, %xmm2
cmpq %rax, %r10
jne .L11
pcom and vect have been exchanged, now we no longer perform predictive commoning. We're also better in identifying single-iteration loops. Let's close the issue. |
SUBROUTINE ANYAVG(NLVLS,HTS,PARRAY,ZBOT,NDXABV,ZTOP,NDXBLW,VALAVG) IMPLICIT NONE INTEGER I , NLVLS , NDXABV , NDXBLW REAL HTS(NLVLS) , PARRAY(NLVLS) , ZBOT , ZTOP , SUM , VALAVG REAL VALBOT , VALTOP IF ( ZBOT.LT.0.5 ) THEN ZBOT = 0.5 NDXABV = 2 ENDIF IF ( ZTOP.LT.0.51 ) THEN ZTOP = 0.51 NDXBLW = 2 ENDIF IF ( NDXBLW.LE.NDXABV ) GOTO 200 DO I = NDXABV + 1 , NDXBLW SUM = SUM + (HTS(I)-HTS(I-1))*0.5*(PARRAY(I)+PARRAY(I-1)) ENDDO 200 CONTINUE VALAVG = SUM/(ZTOP-ZBOT) END ends up with 5 loop copies.