Re: [llvm-commits] [llvm] r46397 - /llvm/trunk/test/CodeGen/PowerPC/2008-01-25-EmptyFunction.ll
On Jan 25, 2008, at 10:56 PM, Bill Wendling wrote:
> Author: void
> Date: Sat Jan 26 00:56:08 2008
> New Revision: 46397
>
> URL: http://llvm.org/viewvc/llvm-project?rev=46397&view=rev
> Log:
> Need to convert to LLVM code and not C.
This should also pass a darwin target triple to llc (or embed it in
the .ll file), because ppc linux doesn't need the nop.
It would also be nice to put a short comment in the test explaining
why we expect a nop in this crazy case :)
Thanks Bill,
-Chris
>
>
> Modified:
>llvm/trunk/test/CodeGen/PowerPC/2008-01-25-EmptyFunction.ll
>
> Modified: llvm/trunk/test/CodeGen/PowerPC/2008-01-25-EmptyFunction.ll
> URL:
> http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/PowerPC/2008-01-25-EmptyFunction.ll?rev=46397&r1=46396&r2=46397&view=diff
>
> =
> =
> =
> =
> =
> =
> =
> =
> ==
> --- llvm/trunk/test/CodeGen/PowerPC/2008-01-25-EmptyFunction.ll
> (original)
> +++ llvm/trunk/test/CodeGen/PowerPC/2008-01-25-EmptyFunction.ll Sat
> Jan 26 00:56:08 2008
> @@ -1,7 +1,6 @@
> -// RUN: %llvmgcc -O2 -S -o - -emit-llvm %s | grep llvm.trap
> -// RUN: %llvmgcc -O2 -S -o - -emit-llvm %s | grep unreachable
> +; RUN: llvm-as < %s | llc -march=ppc32 | grep nop
>
> -void bork() {
> - int *address = 0;
> - *address = 0;
> +define void @bork() noreturn nounwind {
> +entry:
> +unreachable
> }
>
>
> ___
> llvm-commits mailing list
> llvm-commits@cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
___
llvm-commits mailing list
llvm-commits@cs.uiuc.edu
http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
Re: [llvm-commits] [llvm-gcc-4.2] r46373 - /llvm-gcc-4.2/trunk/gcc/llvm-types.cpp
On Jan 25, 2008, at 9:58 PM, Duncan Sands wrote: >> If a function takes a byval parameter, it can't be readnone, we >> have to mark it readonly instead. This fixes >> test/CFrontend/2008-01-25-ByValReadNone.c > > Thanks for doing this! No problem, thanks for pointing out that it needed to be done. It would be nice if globalsmodref inferred that functions which only store to their byval args are really readnone. -Chris ___ llvm-commits mailing list llvm-commits@cs.uiuc.edu http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
Re: [llvm-commits] [llvm] r46394 - in /llvm/trunk/lib: Target/PowerPC/PPCAsmPrinter.cpp Transforms/Scalar/SimplifyCFG.cpp
On Jan 26, 2008, at 12:21 AM, Chris Lattner wrote:
>
> On Jan 25, 2008, at 10:51 PM, Bill Wendling wrote:
>
>> Author: void
>> Date: Sat Jan 26 00:51:24 2008
>> New Revision: 46394
>>
>> URL: http://llvm.org/viewvc/llvm-project?rev=46394&view=rev
>> Log:
>> If there are no machine instructions emitted for a function, then
>> insert
>> a "nop" instruction so that we don't have the function's label
>> associated
>> with something that it's not supposed to be associated with.
>
> Thanks Bill, this is looking better. Some thoughts:
>
>> + // If the function is empty, then we need to emit *something*.
>> Otherwise, the
>> + // function's label might be associated with something that it
>> wasn't meant to
>> + // be associated with. We emit a noop in this situation.
>> + MachineFunction::iterator I = MF.begin();
>> +
>> + if (++I == MF.end()) {
>> +MachineBasicBlock &MBB = MF.front();
>> +
>> +if (MBB.begin() == MBB.end())
>
> You can use MBB.empty()
>
Good point!
>>
>> + BuildMI(MBB, MBB.end(), TM.getInstrInfo()->get(PPC::NOP));
>> + }
>
> Is there any reason to actually make a machineinstr here? It seems
> like it would be possible to just emit "nop" to the .s file. The JIT
> doesn't need this.
>
I initially had that, but I thought that it would be better to do
things though the "print machine instruction" function just so I got
the syntax correct. But nop is probably safe to emit like that.
> Also, I think this is required on x86 and arm as well, as they have
> the same problem with -fomit-frame-pointer.
>
There's a nop for X86, but I don't see one for ARM in the .td file.
> Thanks for working on this!
>
No problem. :-)
-bw
___
llvm-commits mailing list
llvm-commits@cs.uiuc.edu
http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
[llvm-commits] [llvm] r46398 - in /llvm/trunk: lib/Target/PowerPC/PPCAsmPrinter.cpp lib/Target/X86/X86ATTAsmPrinter.cpp test/CodeGen/X86/2008-01-25-EmptyFunction.ll
Author: void
Date: Sat Jan 26 03:03:52 2008
New Revision: 46398
URL: http://llvm.org/viewvc/llvm-project?rev=46398&view=rev
Log:
If there's no instructions being emitted on X86 for a function, emit a
nop. Emit the nop directly for PPC.
Added:
llvm/trunk/test/CodeGen/X86/2008-01-25-EmptyFunction.ll
Modified:
llvm/trunk/lib/Target/PowerPC/PPCAsmPrinter.cpp
llvm/trunk/lib/Target/X86/X86ATTAsmPrinter.cpp
Modified: llvm/trunk/lib/Target/PowerPC/PPCAsmPrinter.cpp
URL:
http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/PowerPC/PPCAsmPrinter.cpp?rev=46398&r1=46397&r2=46398&view=diff
==
--- llvm/trunk/lib/Target/PowerPC/PPCAsmPrinter.cpp (original)
+++ llvm/trunk/lib/Target/PowerPC/PPCAsmPrinter.cpp Sat Jan 26 03:03:52 2008
@@ -819,12 +819,8 @@
// be associated with. We emit a noop in this situation.
MachineFunction::iterator I = MF.begin();
- if (++I == MF.end()) {
-MachineBasicBlock &MBB = MF.front();
-
-if (MBB.begin() == MBB.end())
- BuildMI(MBB, MBB.end(), TM.getInstrInfo()->get(PPC::NOP));
- }
+ if (++I == MF.end() && MF.front().empty())
+O << "\tnop\n";
// Print out code for the function.
for (MachineFunction::const_iterator I = MF.begin(), E = MF.end();
Modified: llvm/trunk/lib/Target/X86/X86ATTAsmPrinter.cpp
URL:
http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86ATTAsmPrinter.cpp?rev=46398&r1=46397&r2=46398&view=diff
==
--- llvm/trunk/lib/Target/X86/X86ATTAsmPrinter.cpp (original)
+++ llvm/trunk/lib/Target/X86/X86ATTAsmPrinter.cpp Sat Jan 26 03:03:52 2008
@@ -165,6 +165,16 @@
DW.BeginFunction(&MF);
}
+ if (Subtarget->isTargetDarwin()) {
+// If the function is empty, then we need to emit *something*. Otherwise,
+// the function's label might be associated with something that it wasn't
+// meant to be associated with. We emit a noop in this situation.
+MachineFunction::iterator I = MF.begin();
+
+if (++I == MF.end() && MF.front().empty())
+ O << "\tnop\n";
+ }
+
// Print out code for the function.
for (MachineFunction::const_iterator I = MF.begin(), E = MF.end();
I != E; ++I) {
@@ -173,8 +183,8 @@
printBasicBlockLabel(I, true);
O << '\n';
}
-for (MachineBasicBlock::const_iterator II = I->begin(), E = I->end();
- II != E; ++II) {
+for (MachineBasicBlock::const_iterator II = I->begin(), IE = I->end();
+ II != IE; ++II) {
// Print the assembly for the instruction.
O << "\t";
printMachineInstruction(II);
Added: llvm/trunk/test/CodeGen/X86/2008-01-25-EmptyFunction.ll
URL:
http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/2008-01-25-EmptyFunction.ll?rev=46398&view=auto
==
--- llvm/trunk/test/CodeGen/X86/2008-01-25-EmptyFunction.ll (added)
+++ llvm/trunk/test/CodeGen/X86/2008-01-25-EmptyFunction.ll Sat Jan 26 03:03:52
2008
@@ -0,0 +1,6 @@
+; RUN: llvm-as < %s | llc -march=x86 | grep nop
+
+define void @bork() noreturn nounwind {
+entry:
+unreachable
+}
___
llvm-commits mailing list
llvm-commits@cs.uiuc.edu
http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
Re: [llvm-commits] [llvm] r46394 - in /llvm/trunk/lib: Target/PowerPC/PPCAsmPrinter.cpp Transforms/Scalar/SimplifyCFG.cpp
On Jan 25, 2008, at 10:51 PM, Bill Wendling wrote:
> Author: void
> Date: Sat Jan 26 00:51:24 2008
> New Revision: 46394
>
> URL: http://llvm.org/viewvc/llvm-project?rev=46394&view=rev
> Log:
> If there are no machine instructions emitted for a function, then
> insert
> a "nop" instruction so that we don't have the function's label
> associated
> with something that it's not supposed to be associated with.
Thanks Bill, this is looking better. Some thoughts:
> + // If the function is empty, then we need to emit *something*.
> Otherwise, the
> + // function's label might be associated with something that it
> wasn't meant to
> + // be associated with. We emit a noop in this situation.
> + MachineFunction::iterator I = MF.begin();
> +
> + if (++I == MF.end()) {
> +MachineBasicBlock &MBB = MF.front();
> +
> +if (MBB.begin() == MBB.end())
You can use MBB.empty()
>
> + BuildMI(MBB, MBB.end(), TM.getInstrInfo()->get(PPC::NOP));
> + }
Is there any reason to actually make a machineinstr here? It seems
like it would be possible to just emit "nop" to the .s file. The JIT
doesn't need this.
Also, I think this is required on x86 and arm as well, as they have
the same problem with -fomit-frame-pointer.
Thanks for working on this!
-Chris
___
llvm-commits mailing list
llvm-commits@cs.uiuc.edu
http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
Re: [llvm-commits] [llvm] r46393 - in /llvm/trunk: lib/Transforms/Utils/InlineFunction.cpp test/CFrontend/2008-01-26-ReadOnlyByVal.c
On Jan 25, 2008, at 10:41 PM, Duncan Sands wrote:
> Author: baldrick
> Date: Sat Jan 26 00:41:49 2008
> New Revision: 46393
>
> URL: http://llvm.org/viewvc/llvm-project?rev=46393&view=rev
> Log:
> Create an explicit copy for byval parameters even
> when inlining a readonly function.
Are you sure this is correct? There is a difference between 'const',
'pure' and readnone/readonly. Further, isn't it always safe to elide
the copy for readnone functions?
-Chris
>
>
> Added:
>llvm/trunk/test/CFrontend/2008-01-26-ReadOnlyByVal.c
> Modified:
>llvm/trunk/lib/Transforms/Utils/InlineFunction.cpp
>
> Modified: llvm/trunk/lib/Transforms/Utils/InlineFunction.cpp
> URL:
> http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/Utils/InlineFunction.cpp?rev=46393&r1=46392&r2=46393&view=diff
>
> =
> =
> =
> =
> =
> =
> =
> =
> ==
> --- llvm/trunk/lib/Transforms/Utils/InlineFunction.cpp (original)
> +++ llvm/trunk/lib/Transforms/Utils/InlineFunction.cpp Sat Jan 26
> 00:41:49 2008
> @@ -240,12 +240,15 @@
> E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) {
> Value *ActualArg = *AI;
>
> - // When byval arguments actually inlined, we need to make the
> copy implied
> - // by them explicit. However, we don't do this if the callee
> is readonly
> - // or readnone, because the copy would be unneeded: the
> callee doesn't
> - // modify the struct.
> - if (CalledFunc->paramHasAttr(ArgNo+1, ParamAttr::ByVal) &&
> - !CalledFunc->onlyReadsMemory()) {
> + // When byval arguments are inlined, we need to make the copy
> implied
> + // by them explicit. It is tempting to think that this is
> not needed if
> + // the callee is readonly, because the callee doesn't modify
> the struct.
> + // However this would be wrong: readonly means that any
> writes the callee
> + // performs are not visible to the caller. But writes by the
> callee to
> + // an argument passed byval are by definition not visible to
> the caller!
> + // Since we allow this kind of readonly function, there needs
> to be an
> + // explicit copy in order to keep the writes invisible after
> inlining.
> + if (CalledFunc->paramHasAttr(ArgNo+1, ParamAttr::ByVal)) {
> const Type *AggTy = cast(I->getType())-
> >getElementType();
> const Type *VoidPtrTy = PointerType::getUnqual(Type::Int8Ty);
>
>
> Added: llvm/trunk/test/CFrontend/2008-01-26-ReadOnlyByVal.c
> URL:
> http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CFrontend/2008-01-26-ReadOnlyByVal.c?rev=46393&view=auto
>
> =
> =
> =
> =
> =
> =
> =
> =
> ==
> --- llvm/trunk/test/CFrontend/2008-01-26-ReadOnlyByVal.c (added)
> +++ llvm/trunk/test/CFrontend/2008-01-26-ReadOnlyByVal.c Sat Jan 26
> 00:41:49 2008
> @@ -0,0 +1,19 @@
> +// RUN: %llvmgcc %s -S -O1 -o - | llvm-as | opt -std-compile-opts |
> llvm-dis | not grep add
> +
> +struct S { int A; int B; int C; int D; };
> +
> +int f(struct S x) __attribute__ ((const));
> +
> +static int __attribute__ ((const)) g(struct S x) {
> + x.A = x.B;
> + return f(x);
> +}
> +
> +int h(void) {
> + struct S x;
> + int r;
> + x.A = 0;
> + x.B = 9;
> + r = g(x);
> + return r + x.A;
> +}
>
>
> ___
> llvm-commits mailing list
> llvm-commits@cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
___
llvm-commits mailing list
llvm-commits@cs.uiuc.edu
http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
Re: [llvm-commits] [llvm-gcc-4.2] r46373 - /llvm-gcc-4.2/trunk/gcc/llvm-types.cpp
> No problem, thanks for pointing out that it needed to be done. It > would be nice if globalsmodref inferred that functions which only > store to their byval args are really readnone. I guess you mean readonly, since readnone+byval doesn't really work. D. ___ llvm-commits mailing list llvm-commits@cs.uiuc.edu http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
Re: [llvm-commits] [llvm] r46393 - in /llvm/trunk: lib/Transforms/Utils/InlineFunction.cpp test/CFrontend/2008-01-26-ReadOnlyByVal.c
> > Create an explicit copy for byval parameters even > > when inlining a readonly function. > > Are you sure this is correct? There is a difference between 'const', > 'pure' and readnone/readonly. Further, isn't it always safe to elide > the copy for readnone functions? A readnone function could write to a byval parameter, using it as local storage (which it is). If you don't make an explicit copy when inlining, these become writes to the storage that the caller passed in. I'm pretty sure this is correct. D. ___ llvm-commits mailing list llvm-commits@cs.uiuc.edu http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
[llvm-commits] [llvm] r46399 - /llvm/trunk/test/Transforms/Inline/byval2.ll
Author: baldrick
Date: Sat Jan 26 06:33:01 2008
New Revision: 46399
URL: http://llvm.org/viewvc/llvm-project?rev=46399&view=rev
Log:
Invert this test, because it is wrong if we allow
readonly functions to use byval parameters as local
storage (how much do we want this?).
Modified:
llvm/trunk/test/Transforms/Inline/byval2.ll
Modified: llvm/trunk/test/Transforms/Inline/byval2.ll
URL:
http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/Inline/byval2.ll?rev=46399&r1=46398&r2=46399&view=diff
==
--- llvm/trunk/test/Transforms/Inline/byval2.ll (original)
+++ llvm/trunk/test/Transforms/Inline/byval2.ll Sat Jan 26 06:33:01 2008
@@ -1,7 +1,7 @@
-; RUN: llvm-as < %s | opt -inline | llvm-dis | not grep {llvm.memcpy}
+; RUN: llvm-as < %s | opt -inline | llvm-dis | grep {llvm.memcpy}
-; Inlining a byval struct should NOT cause an explicit copy
-; into an alloca if the function is readonly
+; Inlining a byval struct should cause an explicit copy
+; into an alloca even if the function is readonly
%struct.ss = type { i32, i64 }
@.str = internal constant [10 x i8] c"%d, %lld\0A\00" ; <[10 x i8]*>
[#uses=1]
___
llvm-commits mailing list
llvm-commits@cs.uiuc.edu
http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
Re: [llvm-commits] [llvm-testresults] Grawp-PIC i386 nightly tester results
On Jan 26, 2008, at 6:16 AM, Apache wrote: > > New Test Failures: > test/Transforms/Inline/byval2.ll [DEJAGNU] Hi Duncan, This is failing because inlining of readonly functions is making the memcpy after your change. Why is the memcpy needed here? Is this because stores to the byval argument are not counted as a write? We really need to pin down the semantics of byval and readonly/none. Getting a copy for this is a big bad change. What do you think? -Chris ___ llvm-commits mailing list llvm-commits@cs.uiuc.edu http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
[llvm-commits] [llvm] r46401 - in /llvm/trunk: lib/CodeGen/SelectionDAG/DAGCombiner.cpp test/CodeGen/X86/stack-align.ll
Author: lattner
Date: Sat Jan 26 13:45:50 2008
New Revision: 46401
URL: http://llvm.org/viewvc/llvm-project?rev=46401&view=rev
Log:
Infer alignment of loads and increase their alignment when we can tell they are
from the stack. This allows us to compile stack-align.ll to:
_test:
movsd LCPI1_0, %xmm0
movapd %xmm0, %xmm1
*** andpd 4(%esp), %xmm1
andpd _G, %xmm0
addsd %xmm1, %xmm0
movl20(%esp), %eax
movsd %xmm0, (%eax)
ret
instead of:
_test:
movsd LCPI1_0, %xmm0
** movsd 4(%esp), %xmm1
** andpd %xmm0, %xmm1
andpd _G, %xmm0
addsd %xmm1, %xmm0
movl20(%esp), %eax
movsd %xmm0, (%eax)
ret
Added:
llvm/trunk/test/CodeGen/X86/stack-align.ll
Modified:
llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
Modified: llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
URL:
http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp?rev=46401&r1=46400&r2=46401&view=diff
==
--- llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp (original)
+++ llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp Sat Jan 26 13:45:50 2008
@@ -18,6 +18,7 @@
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Target/TargetData.h"
+#include "llvm/Target/TargetFrameInfo.h"
#include "llvm/Target/TargetLowering.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Target/TargetOptions.h"
@@ -4073,13 +4074,41 @@
static unsigned InferAlignment(SDOperand Ptr, SelectionDAG &DAG) {
// If this is a direct reference to a stack slot, use information about the
// stack slot's alignment.
+ int FrameIdx = 1 << 31;
+ int64_t FrameOffset = 0;
if (FrameIndexSDNode *FI = dyn_cast(Ptr)) {
-return DAG.getMachineFunction().getFrameInfo()->
- getObjectAlignment(FI->getIndex());
+FrameIdx = FI->getIndex();
+ } else if (Ptr.getOpcode() == ISD::ADD &&
+ isa(Ptr.getOperand(1)) &&
+ isa(Ptr.getOperand(0))) {
+FrameIdx = cast(Ptr.getOperand(0))->getIndex();
+FrameOffset = Ptr.getConstantOperandVal(1);
+ }
+
+ if (FrameIdx != (1 << 31)) {
+// FIXME: Handle FI+CST.
+const MachineFrameInfo &MFI = *DAG.getMachineFunction().getFrameInfo();
+if (MFI.isFixedObjectIndex(FrameIdx)) {
+ int64_t ObjectOffset = MFI.getObjectOffset(FrameIdx);
+
+ // The alignment of the frame index can be determined from its offset
from
+ // the incoming frame position. If the frame object is at offset 32 and
+ // the stack is guaranteed to be 16-byte aligned, then we know that the
+ // object is 16-byte aligned.
+ unsigned StackAlign =
DAG.getTarget().getFrameInfo()->getStackAlignment();
+ unsigned Align = MinAlign(ObjectOffset, StackAlign);
+
+ // Finally, the frame object itself may have a known alignment. Factor
+ // the alignment + offset into a new alignment. For example, if we know
+ // the FI is 8 byte aligned, but the pointer is 4 off, we really have a
+ // 4-byte alignment of the resultant pointer. Likewise align 4 + 4-byte
+ // offset = 4-byte alignment, align 4 + 1-byte offset = align 1, etc.
+ unsigned FIInfoAlign = MinAlign(MFI.getObjectAlignment(FrameIdx),
+ FrameOffset);
+ return std::max(Align, FIInfoAlign);
+}
}
- // FIXME: Handle FI+CST.
-
return 0;
}
Added: llvm/trunk/test/CodeGen/X86/stack-align.ll
URL:
http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/stack-align.ll?rev=46401&view=auto
==
--- llvm/trunk/test/CodeGen/X86/stack-align.ll (added)
+++ llvm/trunk/test/CodeGen/X86/stack-align.ll Sat Jan 26 13:45:50 2008
@@ -0,0 +1,22 @@
+; RUN: llvm-as < %s | llc -relocation-model=static -mcpu=yonah | grep
{andpd.*4(%esp), %xmm}
+
+; The double argument is at 4(esp) which is 16-byte aligned, allowing us to
+; fold the load into the andpd.
+
+target datalayout =
"e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f80:128:128"
+target triple = "i686-apple-darwin8"
[EMAIL PROTECTED] = external global double
+
+define void @test({ double, double }* byval %z, double* %P) {
+entry:
+ %tmp = getelementptr { double, double }* %z, i32 0, i32 0
; [#uses=1]
+ %tmp1 = load double* %tmp, align 8 ; [#uses=1]
+ %tmp2 = tail call double @fabs( double %tmp1 ) ;
[#uses=1]
+ %tmp3 = load double* @G, align 16 ; [#uses=1]
+ %tmp4 = tail call double @fabs( double %tmp3 ) ;
[#uses=1]
+ %tmp6 = add double %tmp4, %tmp2 ; [#uses=1]
+ store double %tmp6, double* %P, align 8
+ ret void
+}
+
+declar
[llvm-commits] [llvm] r46400 - /llvm/trunk/test/CodeGen/X86/2007-08-10-LEA16Use32.ll
Author: lattner
Date: Sat Jan 26 13:35:46 2008
New Revision: 46400
URL: http://llvm.org/viewvc/llvm-project?rev=46400&view=rev
Log:
remove a useless xfailed test.
Removed:
llvm/trunk/test/CodeGen/X86/2007-08-10-LEA16Use32.ll
Removed: llvm/trunk/test/CodeGen/X86/2007-08-10-LEA16Use32.ll
URL:
http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/2007-08-10-LEA16Use32.ll?rev=46399&view=auto
==
--- llvm/trunk/test/CodeGen/X86/2007-08-10-LEA16Use32.ll (original)
+++ llvm/trunk/test/CodeGen/X86/2007-08-10-LEA16Use32.ll (removed)
@@ -1,27 +0,0 @@
-; RUN: llvm-as < %s | llc -march=x86 | grep {leal}
-; XFAIL: *
-; This test is XFAIL'd because strength-reduction was improved to
-; avoid emitting the lea, so it longer tests whether the 16-bit
-; lea is avoided.
-
[EMAIL PROTECTED] = global i16 0 ; [#uses=1]
[EMAIL PROTECTED] = global i16 0 ; [#uses=1]
-
-define void @_Z3fooi(i32 %N) {
-entry:
-%tmp1019 = icmp sgt i32 %N, 0 ; [#uses=1]
-br i1 %tmp1019, label %bb, label %return
-
-bb: ; preds = %bb, %entry
-%i.014.0 = phi i32 [ 0, %entry ], [ %indvar.next, %bb ] ;
[#uses=2]
-%tmp1 = trunc i32 %i.014.0 to i16 ; [#uses=2]
-volatile store i16 %tmp1, i16* @X, align 2
-%tmp34 = shl i16 %tmp1, 2 ; [#uses=1]
-volatile store i16 %tmp34, i16* @Y, align 2
-%indvar.next = add i32 %i.014.0, 1 ; [#uses=2]
-%exitcond = icmp eq i32 %indvar.next, %N;
[#uses=1]
-br i1 %exitcond, label %return, label %bb
-
-return: ; preds = %bb, %entry
-ret void
-}
___
llvm-commits mailing list
llvm-commits@cs.uiuc.edu
http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
[llvm-commits] [llvm] r46402 - /llvm/trunk/include/llvm/Support/MathExtras.h
Author: lattner
Date: Sat Jan 26 13:50:30 2008
New Revision: 46402
URL: http://llvm.org/viewvc/llvm-project?rev=46402&view=rev
Log:
make this 64-bit safe
Modified:
llvm/trunk/include/llvm/Support/MathExtras.h
Modified: llvm/trunk/include/llvm/Support/MathExtras.h
URL:
http://llvm.org/viewvc/llvm-project/llvm/trunk/include/llvm/Support/MathExtras.h?rev=46402&r1=46401&r2=46402&view=diff
==
--- llvm/trunk/include/llvm/Support/MathExtras.h (original)
+++ llvm/trunk/include/llvm/Support/MathExtras.h Sat Jan 26 13:50:30 2008
@@ -360,11 +360,11 @@
/// MinAlign - A and B are either alignments or offsets. Return the minimum
/// alignment that may be assumed after adding the two together.
-static inline unsigned MinAlign(unsigned A, unsigned B) {
+static inline uint64_t MinAlign(uint64_t A, uint64_t B) {
// The largest power of 2 that divides both A and B.
return (A | B) & -(A | B);
}
-
+
} // End llvm namespace
#endif
___
llvm-commits mailing list
llvm-commits@cs.uiuc.edu
http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
[llvm-commits] [llvm] r46403 - /llvm/trunk/test/CodeGen/X86/vec_align.ll
Author: lattner
Date: Sat Jan 26 14:06:45 2008
New Revision: 46403
URL: http://llvm.org/viewvc/llvm-project?rev=46403&view=rev
Log:
New test to verify that "merging 4 loads into a vec load" continues to work and
continues to infer alignment info.
Added:
llvm/trunk/test/CodeGen/X86/vec_align.ll
Added: llvm/trunk/test/CodeGen/X86/vec_align.ll
URL:
http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/vec_align.ll?rev=46403&view=auto
==
--- llvm/trunk/test/CodeGen/X86/vec_align.ll (added)
+++ llvm/trunk/test/CodeGen/X86/vec_align.ll Sat Jan 26 14:06:45 2008
@@ -0,0 +1,35 @@
+; RUN: llvm-as < %s | llc -mcpu=yonah -relocation-model=static | grep movaps |
count 2
+
+target datalayout =
"e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f80:128:128"
+target triple = "i686-apple-darwin8"
+
+%f4 = type <4 x float>
+
[EMAIL PROTECTED] = external global { float,float,float,float}, align 16
+
+define %f4 @test1(float %W, float %X, float %Y, float %Z) {
+%tmp = insertelement %f4 undef, float %W, i32 0
+%tmp2 = insertelement %f4 %tmp, float %X, i32 1
+%tmp4 = insertelement %f4 %tmp2, float %Y, i32 2
+%tmp6 = insertelement %f4 %tmp4, float %Z, i32 3
+ ret %f4 %tmp6
+}
+
+define %f4 @test2() {
+ %Wp = getelementptr { float,float,float,float}* @G, i32 0, i32 0
+ %Xp = getelementptr { float,float,float,float}* @G, i32 0, i32 1
+ %Yp = getelementptr { float,float,float,float}* @G, i32 0, i32 2
+ %Zp = getelementptr { float,float,float,float}* @G, i32 0, i32 3
+
+ %W = load float* %Wp
+ %X = load float* %Xp
+ %Y = load float* %Yp
+ %Z = load float* %Zp
+
+%tmp = insertelement %f4 undef, float %W, i32 0
+%tmp2 = insertelement %f4 %tmp, float %X, i32 1
+%tmp4 = insertelement %f4 %tmp2, float %Y, i32 2
+%tmp6 = insertelement %f4 %tmp4, float %Z, i32 3
+ ret %f4 %tmp6
+}
+
___
llvm-commits mailing list
llvm-commits@cs.uiuc.edu
http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
[llvm-commits] [llvm] r46404 - /llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
Author: lattner
Date: Sat Jan 26 14:07:42 2008
New Revision: 46404
URL: http://llvm.org/viewvc/llvm-project?rev=46404&view=rev
Log:
Remove some code for inferring alignment info from the x86 backend
now that the dag combiner does it.
Modified:
llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
Modified: llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
URL:
http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86ISelLowering.cpp?rev=46404&r1=46403&r2=46404&view=diff
==
--- llvm/trunk/lib/Target/X86/X86ISelLowering.cpp (original)
+++ llvm/trunk/lib/Target/X86/X86ISelLowering.cpp Sat Jan 26 14:07:42 2008
@@ -5499,16 +5499,7 @@
int64_t Offset;
if (isGAPlusOffset(Base, GV, Offset))
return (GV->getAlignment() >= 16 && (Offset % 16) == 0);
- else {
-assert(Base->getOpcode() == ISD::FrameIndex && "Unexpected base node!");
-int BFI = cast(Base)->getIndex();
-if (BFI < 0)
- // Fixed objects do not specify alignment, however the offsets are known.
- return ((Subtarget->getStackAlignment() % 16) == 0 &&
- (MFI->getObjectOffset(BFI) % 16) == 0);
-else
- return MFI->getObjectAlignment(BFI) >= 16;
- }
+ // DAG combine handles the stack object case.
return false;
}
___
llvm-commits mailing list
llvm-commits@cs.uiuc.edu
http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
[llvm-commits] [llvm] r46405 - /llvm/trunk/lib/Target/X86/README-SSE.txt
Author: lattner
Date: Sat Jan 26 14:12:07 2008
New Revision: 46405
URL: http://llvm.org/viewvc/llvm-project?rev=46405&view=rev
Log:
Add some notes.
Modified:
llvm/trunk/lib/Target/X86/README-SSE.txt
Modified: llvm/trunk/lib/Target/X86/README-SSE.txt
URL:
http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/README-SSE.txt?rev=46405&r1=46404&r2=46405&view=diff
==
--- llvm/trunk/lib/Target/X86/README-SSE.txt (original)
+++ llvm/trunk/lib/Target/X86/README-SSE.txt Sat Jan 26 14:12:07 2008
@@ -704,3 +704,21 @@
We should use movmskp{s|d} instead.
+//===-===//
+
+CodeGen/X86/vec_align.ll tests whether we can turn 4 scalar loads into a single
+(aligned) vector load. This functionality has a couple of problems.
+
+1. The code to infer alignment from loads of globals is in the X86 backend,
+ not the dag combiner. This is because dagcombine2 needs to be able to see
+ through the X86ISD::Wrapper node, which DAGCombine can't really do.
+2. The code for turning 4 x load into a single vector load is target
+ independent and should be moved to the dag combiner.
+3. The code for turning 4 x load into a vector load can only handle a direct
+ load from a global or a direct load from the stack. It should be
generalized
+ to handle any load from P, P+4, P+8, P+12, where P can be anything.
+4. The alignment inference code cannot handle loads from globals in non-static
+ mode because it doesn't look through the extra dyld stub load. If you try
+ vec_align.ll without -relocation-model=static, you'll see what I mean.
+
+//===-===//
___
llvm-commits mailing list
llvm-commits@cs.uiuc.edu
http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
Re: [llvm-commits] [llvm-gcc-4.2] r46381 - in /llvm-gcc-4.2/trunk/gcc: llvm-abi.h llvm-internal.h llvm-types.cpp
Can you get me a test case? I need to see what gcc is doing. Thanks, Evan On Jan 25, 2008, at 10:25 PM, Duncan Sands <[EMAIL PROTECTED]> wrote: > Hi Evan, dropping zero sized aggregates from types > http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20080121/057570.html > broke the Ada build. A zero-sized aggregate is used in a > COMPONENT_REF, so > llvm-convert barfs because it can't find the corresponding LLVM > field. All gcc > fields should have a corresponding LLVM field - the only exception > we make is > for fields at variable offsets. I would like to keep it that way > unless you > really can't solve this differently. > > Best wishes, > > Duncan. ___ llvm-commits mailing list llvm-commits@cs.uiuc.edu http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
[llvm-commits] [llvm] r46406 - in /llvm/trunk: lib/Transforms/Scalar/InstructionCombining.cpp test/Transforms/InstCombine/fpextend.ll
Author: lattner
Date: Sat Jan 26 23:29:54 2008
New Revision: 46406
URL: http://llvm.org/viewvc/llvm-project?rev=46406&view=rev
Log:
Fold fptrunc(add (fpextend x), (fpextend y)) -> add(x,y), as GCC does.
Added:
llvm/trunk/test/Transforms/InstCombine/fpextend.ll
Modified:
llvm/trunk/lib/Transforms/Scalar/InstructionCombining.cpp
Modified: llvm/trunk/lib/Transforms/Scalar/InstructionCombining.cpp
URL:
http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/Scalar/InstructionCombining.cpp?rev=46406&r1=46405&r2=46406&view=diff
==
--- llvm/trunk/lib/Transforms/Scalar/InstructionCombining.cpp (original)
+++ llvm/trunk/lib/Transforms/Scalar/InstructionCombining.cpp Sat Jan 26
23:29:54 2008
@@ -203,7 +203,7 @@
Instruction *visitTrunc(TruncInst &CI);
Instruction *visitZExt(ZExtInst &CI);
Instruction *visitSExt(SExtInst &CI);
-Instruction *visitFPTrunc(CastInst &CI);
+Instruction *visitFPTrunc(FPTruncInst &CI);
Instruction *visitFPExt(CastInst &CI);
Instruction *visitFPToUI(CastInst &CI);
Instruction *visitFPToSI(CastInst &CI);
@@ -7141,8 +7141,80 @@
return 0;
}
-Instruction *InstCombiner::visitFPTrunc(CastInst &CI) {
- return commonCastTransforms(CI);
+/// FitsInFPType - Return a Constant* for the specified FP constant if it fits
+/// in the specified FP type without changing its value.
+static Constant *FitsInFPType(ConstantFP *CFP, const Type *FPTy,
+ const fltSemantics &Sem) {
+ APFloat F = CFP->getValueAPF();
+ if (F.convert(Sem, APFloat::rmNearestTiesToEven) == APFloat::opOK)
+return ConstantFP::get(FPTy, F);
+ return 0;
+}
+
+/// LookThroughFPExtensions - If this is an fp extension instruction, look
+/// through it until we get the source value.
+static Value *LookThroughFPExtensions(Value *V) {
+ if (Instruction *I = dyn_cast(V))
+if (I->getOpcode() == Instruction::FPExt)
+ return LookThroughFPExtensions(I->getOperand(0));
+
+ // If this value is a constant, return the constant in the smallest FP type
+ // that can accurately represent it. This allows us to turn
+ // (float)((double)X+2.0) into x+2.0f.
+ if (ConstantFP *CFP = dyn_cast(V)) {
+if (CFP->getType() == Type::PPC_FP128Ty)
+ return V; // No constant folding of this.
+// See if the value can be truncated to float and then reextended.
+if (Value *V = FitsInFPType(CFP, Type::FloatTy, APFloat::IEEEsingle))
+ return V;
+if (CFP->getType() == Type::DoubleTy)
+ return V; // Won't shrink.
+if (Value *V = FitsInFPType(CFP, Type::DoubleTy, APFloat::IEEEdouble))
+ return V;
+// Don't try to shrink to various long double types.
+ }
+
+ return V;
+}
+
+Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
+ if (Instruction *I = commonCastTransforms(CI))
+return I;
+
+ // If we have fptrunc(add (fpextend x), (fpextend y)), where x and y are
+ // smaller than the destination type, we can eliminate the truncate by doing
+ // the add as the smaller type. This applies to add/sub/mul/div as well as
+ // many builtins (sqrt, etc).
+ BinaryOperator *OpI = dyn_cast(CI.getOperand(0));
+ if (OpI && OpI->hasOneUse()) {
+switch (OpI->getOpcode()) {
+default: break;
+case Instruction::Add:
+case Instruction::Sub:
+case Instruction::Mul:
+case Instruction::FDiv:
+case Instruction::FRem:
+ const Type *SrcTy = OpI->getType();
+ Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
+ Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
+ if (LHSTrunc->getType() != SrcTy &&
+ RHSTrunc->getType() != SrcTy) {
+unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
+// If the source types were both smaller than the destination type of
+// the cast, do this xform.
+if (LHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize &&
+RHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize) {
+ LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
+ CI.getType(), CI);
+ RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
+ CI.getType(), CI);
+ return BinaryOperator::create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
+}
+ }
+ break;
+}
+ }
+ return 0;
}
Instruction *InstCombiner::visitFPExt(CastInst &CI) {
Added: llvm/trunk/test/Transforms/InstCombine/fpextend.ll
URL:
http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/InstCombine/fpextend.ll?rev=46406&view=auto
==
--- llvm/trunk/test/Transforms/InstCombine/fpextend.ll (added)
+++ llvm/trunk/test/Transforms/InstCombine/fpextend.ll Sat Jan 26 23:29:54 2008
@@ -0,0 +1,36 @@
+; RUN: llvm-as < %s | opt -instcombine | llvm-dis | not grep fpext
[llvm-commits] [llvm] r46407 - /llvm/trunk/lib/CodeGen/AsmPrinter.cpp
Author: lattner
Date: Sun Jan 27 00:09:28 2008
New Revision: 46407
URL: http://llvm.org/viewvc/llvm-project?rev=46407&view=rev
Log:
For long double constants, print an approximation of their value to the .s file
to make it easier to read.
Modified:
llvm/trunk/lib/CodeGen/AsmPrinter.cpp
Modified: llvm/trunk/lib/CodeGen/AsmPrinter.cpp
URL:
http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/CodeGen/AsmPrinter.cpp?rev=46407&r1=46406&r2=46407&view=diff
==
--- llvm/trunk/lib/CodeGen/AsmPrinter.cpp (original)
+++ llvm/trunk/lib/CodeGen/AsmPrinter.cpp Sun Jan 27 00:09:28 2008
@@ -939,10 +939,13 @@
// api needed to prevent premature destruction
APInt api = CFP->getValueAPF().convertToAPInt();
const uint64_t *p = api.getRawData();
+ APFloat DoubleVal = CFP->getValueAPF();
+ DoubleVal.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven);
if (TD->isBigEndian()) {
O << TAI->getData16bitsDirective() << uint16_t(p[0] >> 48)
<< "\t" << TAI->getCommentString()
- << " long double most significant halfword\n";
+ << " long double most significant halfword of ~"
+ << DoubleVal.convertToDouble() << "\n";
O << TAI->getData16bitsDirective() << uint16_t(p[0] >> 32)
<< "\t" << TAI->getCommentString()
<< " long double next halfword\n";
@@ -958,7 +961,8 @@
} else {
O << TAI->getData16bitsDirective() << uint16_t(p[1])
<< "\t" << TAI->getCommentString()
- << " long double least significant halfword\n";
+ << " long double least significant halfword of ~"
+ << DoubleVal.convertToDouble() << "\n";
O << TAI->getData16bitsDirective() << uint16_t(p[0])
<< "\t" << TAI->getCommentString()
<< " long double next halfword\n";
___
llvm-commits mailing list
llvm-commits@cs.uiuc.edu
http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
[llvm-commits] [llvm] r46409 - /llvm/trunk/test/Transforms/CorrelatedExprs/
Author: void Date: Sun Jan 27 00:13:32 2008 New Revision: 46409 URL: http://llvm.org/viewvc/llvm-project?rev=46409&view=rev Log: The CorrelatedExpressions pass is now no more. Removed: llvm/trunk/test/Transforms/CorrelatedExprs/ ___ llvm-commits mailing list llvm-commits@cs.uiuc.edu http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
[llvm-commits] [llvm] r46410 - /llvm/trunk/include/llvm/CodeGen/SelectionDAGNodes.h
Author: lattner
Date: Sun Jan 27 00:19:08 2008
New Revision: 46410
URL: http://llvm.org/viewvc/llvm-project?rev=46410&view=rev
Log:
make isExactlyValue work for long double.
Modified:
llvm/trunk/include/llvm/CodeGen/SelectionDAGNodes.h
Modified: llvm/trunk/include/llvm/CodeGen/SelectionDAGNodes.h
URL:
http://llvm.org/viewvc/llvm-project/llvm/trunk/include/llvm/CodeGen/SelectionDAGNodes.h?rev=46410&r1=46409&r2=46410&view=diff
==
--- llvm/trunk/include/llvm/CodeGen/SelectionDAGNodes.h (original)
+++ llvm/trunk/include/llvm/CodeGen/SelectionDAGNodes.h Sun Jan 27 00:19:08 2008
@@ -1208,10 +1208,9 @@
/// convenient to write "2.0" and the like. Without this function we'd
/// have to duplicate its logic everywhere it's called.
bool isExactlyValue(double V) const {
-if (getValueType(0)==MVT::f64)
- return isExactlyValue(APFloat(V));
-else
- return isExactlyValue(APFloat((float)V));
+APFloat Tmp(V);
+Tmp.convert(Value.getSemantics(), APFloat::rmNearestTiesToEven);
+return isExactlyValue(Tmp);
}
bool isExactlyValue(const APFloat& V) const;
___
llvm-commits mailing list
llvm-commits@cs.uiuc.edu
http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
[llvm-commits] [llvm] r46411 - /llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
Author: lattner
Date: Sun Jan 27 00:19:31 2008
New Revision: 46411
URL: http://llvm.org/viewvc/llvm-project?rev=46411&view=rev
Log:
Use fldz and fld1 for long double constants instead of a constant pool load.
Modified:
llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
Modified: llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
URL:
http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86ISelLowering.cpp?rev=46411&r1=46410&r2=46411&view=diff
==
--- llvm/trunk/lib/Target/X86/X86ISelLowering.cpp (original)
+++ llvm/trunk/lib/Target/X86/X86ISelLowering.cpp Sun Jan 27 00:19:31 2008
@@ -455,7 +455,20 @@
addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
setOperationAction(ISD::UNDEF, MVT::f80, Expand);
setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
- setOperationAction(ISD::ConstantFP, MVT::f80, Expand);
+ {
+setOperationAction(ISD::ConstantFP, MVT::f80, Expand);
+APFloat TmpFlt(+0.0);
+TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven);
+addLegalFPImmediate(TmpFlt); // FLD0
+TmpFlt.changeSign();
+addLegalFPImmediate(TmpFlt); // FLD0/FCHS
+APFloat TmpFlt2(+1.0);
+TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven);
+addLegalFPImmediate(TmpFlt2); // FLD1
+TmpFlt2.changeSign();
+addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
+ }
+
if (!UnsafeFPMath) {
setOperationAction(ISD::FSIN , MVT::f80 , Expand);
setOperationAction(ISD::FCOS , MVT::f80 , Expand);
___
llvm-commits mailing list
llvm-commits@cs.uiuc.edu
http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
[llvm-commits] [llvm-gcc-4.2] r46412 - in /llvm-gcc-4.2/trunk/gcc: config/i386/llvm-i386-target.h config/i386/llvm-i386.cpp llvm-abi.h llvm-convert.cpp
Author: lattner
Date: Sun Jan 27 01:18:20 2008
New Revision: 46412
URL: http://llvm.org/viewvc/llvm-project?rev=46412&view=rev
Log:
Change x86-32 aggregate passing code to pass many aggregates "by pieces"
instead of passing them with byval. This allows us to get significantly
better code without affecting the ABI (because x86-32 passes stuff on the
stack). Most significantly, we pass _Complex as a two scalar fp values
now. For example, before we compiled ccosf to:
_ccosf:
pushl %ebp
movl%esp, %ebp
pushl %edi
pushl %esi
subl$16, %esp
movss 12(%ebp), %xmm0
xorps LCPI1_0, %xmm0
movss 8(%ebp), %xmm1
movss %xmm0, -16(%ebp)
movss %xmm1, -12(%ebp)
leal-16(%ebp), %esi
movl$2, %ecx
movl%esp, %edi
rep;movsl
call_ccoshf
addl$16, %esp
popl%esi
popl%edi
popl%ebp
ret
now we get:
_ccosf:
pushl %ebp
movl%esp, %ebp
subl$8, %esp
movss 8(%ebp), %xmm0
movss %xmm0, 4(%esp)
movss 12(%ebp), %xmm0
xorps LCPI1_0, %xmm0
movss %xmm0, (%esp)
call_ccoshf
addl$8, %esp
popl%ebp
ret
Evan, please review this. Thanks!
Modified:
llvm-gcc-4.2/trunk/gcc/config/i386/llvm-i386-target.h
llvm-gcc-4.2/trunk/gcc/config/i386/llvm-i386.cpp
llvm-gcc-4.2/trunk/gcc/llvm-abi.h
llvm-gcc-4.2/trunk/gcc/llvm-convert.cpp
Modified: llvm-gcc-4.2/trunk/gcc/config/i386/llvm-i386-target.h
URL:
http://llvm.org/viewvc/llvm-project/llvm-gcc-4.2/trunk/gcc/config/i386/llvm-i386-target.h?rev=46412&r1=46411&r2=46412&view=diff
==
--- llvm-gcc-4.2/trunk/gcc/config/i386/llvm-i386-target.h (original)
+++ llvm-gcc-4.2/trunk/gcc/config/i386/llvm-i386-target.h Sun Jan 27 01:18:20
2008
@@ -62,20 +62,24 @@
} \
}
-extern bool llvm_x86_should_pass_aggregate_in_memory(tree);
+#ifdef LLVM_ABI_H
+extern bool llvm_x86_should_pass_aggregate_in_memory(tree, const Type *);
-#define LLVM_SHOULD_PASS_AGGREGATE_USING_BYVAL_ATTR(X) \
- llvm_x86_should_pass_aggregate_in_memory(X)
+#define LLVM_SHOULD_PASS_AGGREGATE_USING_BYVAL_ATTR(X, TY) \
+ llvm_x86_should_pass_aggregate_in_memory(X, TY)
-#ifdef LLVM_ABI_H
extern bool
-llvm_x86_64_should_pass_aggregate_in_mixed_regs(tree,
+llvm_x86_64_should_pass_aggregate_in_mixed_regs(tree, const Type *Ty,
+std::vector&);
+extern bool
+llvm_x86_32_should_pass_aggregate_in_mixed_regs(tree, const Type *Ty,
std::vector&);
-#define LLVM_SHOULD_PASS_AGGREGATE_IN_MIXED_REGS(T, E) \
- (TARGET_64BIT && \
- llvm_x86_64_should_pass_aggregate_in_mixed_regs((T), (E)))
+#define LLVM_SHOULD_PASS_AGGREGATE_IN_MIXED_REGS(T, TY, E)\
+ (TARGET_64BIT ? \
+ llvm_x86_64_should_pass_aggregate_in_mixed_regs((T), (TY), (E)) : \
+ llvm_x86_32_should_pass_aggregate_in_mixed_regs((T), (TY), (E)))
#endif /* LLVM_ABI_H */
/* LLVM LOCAL end (ENTIRE FILE!) */
Modified: llvm-gcc-4.2/trunk/gcc/config/i386/llvm-i386.cpp
URL:
http://llvm.org/viewvc/llvm-project/llvm-gcc-4.2/trunk/gcc/config/i386/llvm-i386.cpp?rev=46412&r1=46411&r2=46412&view=diff
==
--- llvm-gcc-4.2/trunk/gcc/config/i386/llvm-i386.cpp (original)
+++ llvm-gcc-4.2/trunk/gcc/config/i386/llvm-i386.cpp Sun Jan 27 01:18:20 2008
@@ -691,7 +691,7 @@
/* Target hook for llvm-abi.h. It returns true if an aggregate of the
specified type should be passed in memory. */
-bool llvm_x86_should_pass_aggregate_in_memory(tree TreeType) {
+bool llvm_x86_should_pass_aggregate_in_memory(tree TreeType, const Type *Ty) {
enum machine_mode Mode = ix86_getNaturalModeForType(TreeType);
HOST_WIDE_INT Bytes =
(Mode == BLKmode) ? int_size_in_bytes(TreeType) : (int)
GET_MODE_SIZE(Mode);
@@ -706,8 +706,10 @@
if (llvm_x86_is_all_integer_types(Ty))
return false;
}
- if (!TARGET_64BIT)
-return true;
+ if (!TARGET_64BIT) {
+std::vector Elts;
+return !llvm_x86_32_should_pass_aggregate_in_mixed_regs(TreeType, Ty,
Elts);
+ }
return llvm_x86_64_should_pass_aggregate_in_memory(TreeType, Mode);
}
@@ -716,7 +718,7 @@
It also returns a vector of types that correspond to the registers used
for parameter passing. This is only called for x86-64. */
bool
-llvm_x86_64_should_pass_aggregate_in_mixed_regs(tree TreeType,
+llvm_x86_64_should_pass_aggregate_in_mixed_regs(tree TreeType, const Type *Ty,
std::vector
&Elts){
enum
[llvm-commits] [llvm] r46413 - /llvm/trunk/lib/Target/X86/README-SSE.txt
Author: lattner
Date: Sun Jan 27 01:31:41 2008
New Revision: 46413
URL: http://llvm.org/viewvc/llvm-project?rev=46413&view=rev
Log:
add a note
Modified:
llvm/trunk/lib/Target/X86/README-SSE.txt
Modified: llvm/trunk/lib/Target/X86/README-SSE.txt
URL:
http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/README-SSE.txt?rev=46413&r1=46412&r2=46413&view=diff
==
--- llvm/trunk/lib/Target/X86/README-SSE.txt (original)
+++ llvm/trunk/lib/Target/X86/README-SSE.txt Sun Jan 27 01:31:41 2008
@@ -722,3 +722,42 @@
vec_align.ll without -relocation-model=static, you'll see what I mean.
//===-===//
+
+We should lower store(fneg(load p), q) into an integer load+xor+store, which
+eliminates a constant pool load. For example, consider:
+
+define i64 @ccosf(float %z.0, float %z.1) nounwind readonly {
+entry:
+ %tmp6 = sub float -0.00e+00, %z.1 ; [#uses=1]
+ %tmp20 = tail call i64 @ccoshf( float %tmp6, float %z.0 ) nounwind
readonly ; [#uses=1]
+ ret i64 %tmp20
+}
+
+This currently compiles to:
+
+LCPI1_0: # <4 x float>
+ .long 2147483648 # float -0
+ .long 2147483648 # float -0
+ .long 2147483648 # float -0
+ .long 2147483648 # float -0
+_ccosf:
+ subl$12, %esp
+ movss 16(%esp), %xmm0
+ movss %xmm0, 4(%esp)
+ movss 20(%esp), %xmm0
+ xorps LCPI1_0, %xmm0
+ movss %xmm0, (%esp)
+ callL_ccoshf$stub
+ addl$12, %esp
+ ret
+
+Note the load into xmm0, then xor (to negate), then store. In PIC mode,
+this code computes the pic base and does two loads to do the constant pool
+load, so the improvement is much bigger.
+
+The tricky part about this xform is that the argument load/store isn't exposed
+until post-legalize, and at that point, the fneg has been custom expanded into
+an X86 fxor. This means that we need to handle this case in the x86 backend
+instead of in target independent code.
+
+//===-===//
___
llvm-commits mailing list
llvm-commits@cs.uiuc.edu
http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
