[llvm-commits] [llvm] r65178 - in /llvm/trunk: lib/Transforms/Scalar/LoopStrengthReduce.cpp test/CodeGen/X86/loop-strength-reduce7.ll
Evan Cheng
evan.cheng at apple.com
Fri Feb 20 19:06:06 PST 2009
On Feb 20, 2009, at 2:31 PM, Dale Johannesen wrote:
>
> On Feb 20, 2009, at 2:16 PMPST, Evan Cheng wrote:
>
>> Author: evancheng
>> Date: Fri Feb 20 16:16:49 2009
>> New Revision: 65178
>>
>> URL: http://llvm.org/viewvc/llvm-project?rev=65178&view=rev
>> Log:
>> Fix strange logic in CollectIVUsers used to determine whether all
>> uses are
>> addresses, part 1. This fixes an obvious logic bug. Previously if
>> the only
>> in-loop use is a PHI, it would return AllUsesAreAddresses as true.
>
> Yes. The logic here goes back before I started working on it, but
> that's clearly intentional:
>
> // If one of the use if a PHI node and all other uses are
> addresses, still
> // allow iv reuse. Essentially we are trading one constant
> multiplication
> // for one fewer iv.
> if (NumPHI > 1)
> AllUsesAreAddresses = false;
>
> although whoever put that in probably should have changed the
> name...are you sure you want to do this?
Yes. It's saying if there are at least one address uses, then it's
willing to accept the tradeoff. However it was setting it to true even
when there is a single in-loop PHI use.
BTW, I did this way back. :-) I'm considering removing this hack. But
one ARM test case suffers a lot if this is removed.
Evan
>
>
>> Added:
>> llvm/trunk/test/CodeGen/X86/loop-strength-reduce7.ll
>> Modified:
>> llvm/trunk/lib/Transforms/Scalar/LoopStrengthReduce.cpp
>>
>> Modified: llvm/trunk/lib/Transforms/Scalar/LoopStrengthReduce.cpp
>> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/Scalar/LoopStrengthReduce.cpp?rev=65178&r1=65177&r2=65178&view=diff
>>
>> =
>> =
>> =
>> =
>> =
>> =
>> =
>> =
>> =
>> =====================================================================
>> --- llvm/trunk/lib/Transforms/Scalar/LoopStrengthReduce.cpp
>> (original)
>> +++ llvm/trunk/lib/Transforms/Scalar/LoopStrengthReduce.cpp Fri Feb
>> 20 16:16:49 2009
>> @@ -1439,6 +1439,7 @@
>> // fields of the BasedUsers. We do this so that it increases the
>> commonality
>> // of the remaining uses.
>> unsigned NumPHI = 0;
>> + bool HasAddress = false;
>> for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
>> // If the user is not in the current loop, this means it is
>> using the exit
>> // value of the IV. Do not put anything in the base, make sure
>> it's all in
>> @@ -1449,7 +1450,6 @@
>> UsersToProcess[i].Base =
>> SE->getIntegerSCEV(0, UsersToProcess[i].Base->getType());
>> } else {
>> -
>> // Addressing modes can be folded into loads and stores. Be
>> careful that
>> // the store is through the expression, not of the expression
>> though.
>> bool isPHI = false;
>> @@ -1462,6 +1462,9 @@
>>
>> // Not all uses are outside the loop.
>> AllUsesAreOutsideLoop = false;
>> +
>> + if (isAddress)
>> + HasAddress = true;
>>
>> // If this use isn't an address, then not all uses are
>> addresses.
>> if (!isAddress && !isPHI)
>> @@ -1478,6 +1481,10 @@
>> if (NumPHI > 1)
>> AllUsesAreAddresses = false;
>>
>> + // There are no in-loop address uses.
>> + if (AllUsesAreAddresses && (!HasAddress && !
>> AllUsesAreOutsideLoop))
>> + AllUsesAreAddresses = false;
>> +
>> return CommonExprs;
>> }
>>
>>
>> Added: llvm/trunk/test/CodeGen/X86/loop-strength-reduce7.ll
>> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/loop-strength-reduce7.ll?rev=65178&view=auto
>>
>> =
>> =
>> =
>> =
>> =
>> =
>> =
>> =
>> =
>> =====================================================================
>> --- llvm/trunk/test/CodeGen/X86/loop-strength-reduce7.ll (added)
>> +++ llvm/trunk/test/CodeGen/X86/loop-strength-reduce7.ll Fri Feb 20
>> 16:16:49 2009
>> @@ -0,0 +1,44 @@
>> +; RUN: llvm-as < %s | llc -march=x86 | not grep imul
>> +
>> +target triple = "i386-apple-darwin9.6"
>> + %struct.III_psy_xmin = type { [22 x double], [13 x [3 x double]] }
>> + %struct.III_scalefac_t = type { [22 x i32], [13 x [3 x i32]] }
>> + %struct.gr_info = type { i32, i32, i32, i32, i32, i32, i32, i32,
>> [3 x i32], [3 x i32], i32, i32, i32, i32, i32, i32, i32, i32, i32,
>> i32*, [4 x i32] }
>> + %struct.lame_global_flags = type { i32, i32, i32, i32, i32, i32,
>> i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32,
>> i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i8*,
>> i8*, i32, i32, float, i32, i32, i32, i32, i32, i32, i32, i32, i32,
>> i32, i32, i32, i32, float, i32, i32, i32, float, float, float,
>> float, i32, i32, i32, i32, i32, i32, i32, i32 }
>> +
>> +define fastcc void @outer_loop(%struct.lame_global_flags* nocapture
>> %gfp, double* nocapture %xr, i32 %targ_bits, double* nocapture
>> %best_noise, %struct.III_psy_xmin* nocapture %l3_xmin, i32*
>> nocapture %l3_enc, %struct.III_scalefac_t* nocapture %scalefac,
>> %struct.gr_info* nocapture %cod_info, i32 %ch) nounwind {
>> +entry:
>> + br label %bb4
>> +
>> +bb4: ; preds = %bb4, %entry
>> + br i1 true, label %bb5, label %bb4
>> +
>> +bb5: ; preds = %bb4
>> + br i1 true, label %bb28.i37, label %bb.i4
>> +
>> +bb.i4: ; preds = %bb.i4, %bb5
>> + br label %bb.i4
>> +
>> +bb28.i37: ; preds = %bb33.i47, %bb5
>> + %i.1.reg2mem.0.i = phi i32 [ %0, %bb33.i47 ], [ 0, %bb5 ] ; <i32>
>> [#uses=2]
>> + %0 = add i32 %i.1.reg2mem.0.i, 1 ; <i32> [#uses=2]
>> + br label %bb29.i38
>> +
>> +bb29.i38: ; preds = %bb33.i47, %bb28.i37
>> + %indvar32.i = phi i32 [ %indvar.next33.i, %bb33.i47 ], [ 0,
>> %bb28.i37 ] ; <i32> [#uses=2]
>> + %sfb.314.i = add i32 %indvar32.i, 0 ; <i32> [#uses=3]
>> + %1 = getelementptr [4 x [21 x double]]* null, i32 0, i32 %0, i32
>> %sfb.314.i ; <double*> [#uses=1]
>> + %2 = load double* %1, align 8 ; <double> [#uses=0]
>> + br i1 false, label %bb30.i41, label %bb33.i47
>> +
>> +bb30.i41: ; preds = %bb29.i38
>> + %3 = getelementptr %struct.III_scalefac_t* null, i32 0, i32 1, i32
>> %sfb.314.i, i32 %i.1.reg2mem.0.i ; <i32*> [#uses=1]
>> + store i32 0, i32* %3, align 4
>> + br label %bb33.i47
>> +
>> +bb33.i47: ; preds = %bb30.i41, %bb29.i38
>> + %4 = add i32 %sfb.314.i, 1 ; <i32> [#uses=1]
>> + %phitmp.i46 = icmp ugt i32 %4, 11 ; <i1> [#uses=1]
>> + %indvar.next33.i = add i32 %indvar32.i, 1 ; <i32> [#uses=1]
>> + br i1 %phitmp.i46, label %bb28.i37, label %bb29.i38
>> +}
>>
>>
>> _______________________________________________
>> llvm-commits mailing list
>> llvm-commits at cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
More information about the llvm-commits
mailing list