Hi Uros, on 2021/7/15 下午3:17, Uros Bizjak wrote: > On Thu, Jul 15, 2021 at 9:07 AM Kewen.Lin <li...@linux.ibm.com> wrote: >> >> on 2021/7/14 下午3:45, Kewen.Lin via Gcc-patches wrote: >>> on 2021/7/14 下午2:38, Richard Biener wrote: >>>> On Tue, Jul 13, 2021 at 4:59 PM Kewen.Lin <li...@linux.ibm.com> wrote: >>>>> >>>>> on 2021/7/13 下午8:42, Richard Biener wrote: >>>>>> On Tue, Jul 13, 2021 at 12:25 PM Kewen.Lin <li...@linux.ibm.com> wrote: >>>> >>>>> I guess the proposed IFN would be directly mapped for [us]mul_highpart? >>>> >>>> Yes. >>>> >>> >>> Thanks for confirming! The related patch v2 is attached and the testing >>> is ongoing. >>> >> >> It's bootstrapped & regtested on powerpc64le-linux-gnu P9 and >> aarch64-linux-gnu. But on x86_64-redhat-linux there are XPASSes as below: >> >> XFAIL->XPASS: gcc.target/i386/pr100637-3w.c scan-assembler pmulhuw >> XFAIL->XPASS: gcc.target/i386/pr100637-3w.c scan-assembler pmulhuw >> XFAIL->XPASS: gcc.target/i386/pr100637-3w.c scan-assembler pmulhw >> XFAIL->XPASS: gcc.target/i386/pr100637-3w.c scan-assembler pmulhw > > These XFAILs should be removed after your patch. > I'm curious whether it's intentional not to specify -fno-vect-cost-model for this test case. As noted above, this case is sensitive on how we cost mult_highpart. Without cost modeling, the XFAILs can be removed only with this mul_highpart pattern support, no matter how we model it (x86 part of this patch exists or not).
> This is PR100696 [1], we want PMULH.W here, so x86 part of the patch > is actually not needed. > Thanks for the information! The justification for the x86 part is that: the IFN_MULH essentially covers MULT_HIGHPART_EXPR with mul_highpart optab support, i386 port has already customized costing for MULT_HIGHPART_EXPR (should mean/involve the case with mul_highpart optab support), if we don't follow the same way for IFN_MULH, I'm worried that we may cost the IFN_MULH wrongly. If taking IFN_MULH as normal stmt is a right thing (we shouldn't cost it specially), it at least means we have to adjust ix86_multiplication_cost for MULT_HIGHPART_EXPR when it has direct mul_highpart optab support, I think they should be costed consistently. Does it sound reasonable? BR, Kewen > [1] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100696 > > Uros. > >> They weren't exposed in the testing run with the previous patch which >> doesn't use IFN way. By investigating it, the difference comes from >> the different costing on MULT_HIGHPART_EXPR and IFN_MULH. >> >> For MULT_HIGHPART_EXPR, it's costed by 16 from below call: >> >> case MULT_EXPR: >> case WIDEN_MULT_EXPR: >> case MULT_HIGHPART_EXPR: >> stmt_cost = ix86_multiplication_cost (ix86_cost, mode); >> >> While for IFN_MULH, it's costed by 4 as normal stmt so the total cost >> becomes profitable and the expected vectorization happens. >> >> One conservative fix seems to make IFN_MULH costing go through the >> unique cost interface for multiplication, that is: >> >> case CFN_MULH: >> stmt_cost = ix86_multiplication_cost (ix86_cost, mode); >> break; >> >> As the test case marks the checks as "xfail", probably it's good to >> revisit the costing on mul_highpart to ensure it's not priced more. >> >> The attached patch also addressed Richard S.'s review comments on >> two reformatting hunks. Is it ok for trunk? >> >> BR, >> Kewen >> ----- >> gcc/ChangeLog: >> >> * internal-fn.c (first_commutative_argument): Add info for IFN_MULH. >> * internal-fn.def (IFN_MULH): New internal function. >> * tree-vect-patterns.c (vect_recog_mulhs_pattern): Add support to >> recog normal multiply highpart as IFN_MULH. >> * config/i386/i386.c (ix86_add_stmt_cost): Adjust for combined >> function CFN_MULH.
