Hi Jingyue,
When I look at the cost model of GEP for the following IR on AArch64, I have a
question about getGEPCost() that you implemented in 2015.
%"class.boost::array.16" = type { [24 x i32] }
%arrayidx.i17 = getelementptr inbounds %"class.boost::array.16",
%"class.boost::array.16"* %moves, i64 0, i32 0, i64 %conv7
store i32 %add, i32* %arrayidx.i17, align 4, !tbaa !18
GEP can be folded into the Store and the assembly is simple as
str w24, [x20, x8, lsl #2]
But the cost model says the GEP cannot be folded.
getGEPCost() calls isLegalAddressingMode() in the following way to check the GEP
can be folded or not
static_cast<T *>(this)->isLegalAddressingMode(
PointerType::get(*GTI, AS), const_cast<GlobalValue *>(BaseGV),
BaseOffset, HasBaseReg, Scale, AS)
It seems to me that the first argument should be *GTI, rather than
PointerType::get(*GTI, AS). In my example, *GTI is i32 and
PointerType::get(*GTI, AS) is i32*. The current implementation fails the last
check of AArch64TargetLowering::isLegalAddressingMode() which is
AM.Scale > 0 && (uint64_t)AM.Scale == NumBytes
Because AM.Scale = 4 and NumBytes = 8
I checked the implementation of isLegalAddressingMode() in several different
backends and several other passes (LoopStrengthReduce and CodeGenPrepare) that
call isLegalAddressingMode(). It seems my understanding is correct. Would you
please take a look at this?
A simplified test case can be
gep.ll
%"class.boost::array.16" = type { [24 x i32] }
define void @test_geps(%"class.boost::array.16"* %moves, i64 %conv7) {
%arrayidx.i17 = getelementptr inbounds %"class.boost::array.16",
%"class.boost::array.16"* %moves, i64 0, i32 0, i64 %conv7
ret void
}
Just run it with following command and the cost of GEP should be 0.
opt -cost-model -analyze -mtriple=aarch64-gnu-linux gep.ll
Thank you in advance,
Haicheng Wu
Employee of Qualcomm Datacenter Technologies, Inc.
Qualcomm Datacenter Technologies, Inc. as an affiliate of Qualcomm Technologies,
Inc. Qualcomm Technologies, Inc. is a member of the Code Aurora Forum, a Linux
Foundation Collaborative Project.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20161118/7439fd25/attachment.html>