This function replaces a particular instruction with another
> - instruction. The following example illustrates the replacement of
> one
> - AllocaInst with another.
> + instruction, inserting the new instruction into the basic block at
> + the location where the old instruction was, and replacing any uses
> + of the old instruction with the new instruction. The following
> + example illustrates the replacement of one
>
>
>
> --
> Stefanus Du Toit
> RapidMind Inc.
> phone: +1 519 885 5455 x116 -- fax: +1 519 885 1463
>
>
>
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>
From ggreif at gmail.com Mon Sep 15 09:00:19 2008
From: ggreif at gmail.com (Gabor Greif)
Date: Mon, 15 Sep 2008 14:00:19 -0000
Subject: [llvm-commits] [llvm] r56215 -
/llvm/branches/ggreif/use-diet/lib/VMCore/getValue.cpp
Message-ID: <200809151400.m8FE0JwV000396@zion.cs.uiuc.edu>
Author: ggreif
Date: Mon Sep 15 09:00:15 2008
New Revision: 56215
URL: http://llvm.org/viewvc/llvm-project?rev=56215&view=rev
Log:
add some more tests
Modified:
llvm/branches/ggreif/use-diet/lib/VMCore/getValue.cpp
Modified: llvm/branches/ggreif/use-diet/lib/VMCore/getValue.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/branches/ggreif/use-diet/lib/VMCore/getValue.cpp?rev=56215&r1=56214&r2=56215&view=diff
==============================================================================
--- llvm/branches/ggreif/use-diet/lib/VMCore/getValue.cpp (original)
+++ llvm/branches/ggreif/use-diet/lib/VMCore/getValue.cpp Mon Sep 15 09:00:15 2008
@@ -164,8 +164,8 @@
--digits;
Acc = (Acc << 1) | (Tag & 1);
if (Cushion <= 0) {
- PrevU = U;
- U = stripTag(U->Next);
+ PrevU = U;
+ U = stripTag(U->Next);
}
continue;
}
@@ -363,6 +363,7 @@
struct UseChain<0> {
Use first;
UseChain(Value *V) : first(V) {}
+ UseChain(Use &U) : first(U) {}
};
template
@@ -372,6 +373,9 @@
UseChain(Value *V)
: first(rest.first)
, rest(V) {}
+ UseChain(Use &U)
+ : first(rest.first)
+ , rest(U) {}
};
UseChain<30> uc31((Value*)0xCAFEBABCUL);
@@ -432,6 +436,17 @@
Use ms35(m34);
Use ms31(m30);
Use ms32(ms31);
+
+ UseChain<24, 0xCAFEBABCUL> uc25;
+ Use& u25(uc25.first);
+ T3::UseChain<10> m11s24dS(u25);
+ Use& m36(m11s24dS.first);
+
+
+ T3::UseChain<10> s10S((Value*)0xCAFEBABCUL);
+ UseChain<20, 0xCAFEBABCUL> d20ss10S(s10S.first);
+ T3::UseChain<20> s20sd20ss10S(d20ss10S.first);
+ Use& m53(s20sd20ss10S.first);
}
@@ -476,6 +491,22 @@
return 4;
if ((Value*)0xCAFEBABCUL != T4::ms32.getValue()) // check the mutated value
return 4;
+
+ T4::m36.showWaymarks();
+ if ((Value*)0xCAFEBABCUL != T4::m36.getValue())
+ return 4;
+ T4::m36.showWaymarks();
+ if ((Value*)0xCAFEBABCUL != T4::m36.getValue()) // check the mutated value
+ return 4;
+ T4::m36.showWaymarks();
+
+ T4::m53.showWaymarks();
+ if ((Value*)0xCAFEBABCUL != T4::m53.getValue())
+ return 4;
+ T4::m53.showWaymarks();
+ if ((Value*)0xCAFEBABCUL != T4::m53.getValue()) // check the mutated value
+ return 4;
+ T4::m53.showWaymarks();
}
#endif
From ggreif at gmail.com Mon Sep 15 10:16:58 2008
From: ggreif at gmail.com (Gabor Greif)
Date: Mon, 15 Sep 2008 15:16:58 -0000
Subject: [llvm-commits] [llvm] r56216 - in /llvm/branches/ggreif/use-diet:
include/llvm/Use.h lib/VMCore/getValue.cpp
Message-ID: <200809151516.m8FFGxML002692@zion.cs.uiuc.edu>
Author: ggreif
Date: Mon Sep 15 10:16:57 2008
New Revision: 56216
URL: http://llvm.org/viewvc/llvm-project?rev=56216&view=rev
Log:
make nilUse inline and prepare for getting rid of PrevU
Modified:
llvm/branches/ggreif/use-diet/include/llvm/Use.h
llvm/branches/ggreif/use-diet/lib/VMCore/getValue.cpp
Modified: llvm/branches/ggreif/use-diet/include/llvm/Use.h
URL: http://llvm.org/viewvc/llvm-project/llvm/branches/ggreif/use-diet/include/llvm/Use.h?rev=56216&r1=56215&r2=56216&view=diff
==============================================================================
--- llvm/branches/ggreif/use-diet/include/llvm/Use.h (original)
+++ llvm/branches/ggreif/use-diet/include/llvm/Use.h Mon Sep 15 10:16:57 2008
@@ -69,7 +69,10 @@
class UseWaymark;
friend class UseWaymark;
Value *getValue() const;
- static Use *nilUse(const Value*); // marks the end of the def/use chain
+ /// nilUse - returns a 'token' that marks the end of the def/use chain
+ static Use *nilUse(const Value *V) {
+ return addTag((Use*)V, fullStopTagN);
+ }
static bool isNil(Use *U) { return extractTag(U) == fullStopTagN; }
void showWaymarks() const;
static bool isStop(Use *U) {
Modified: llvm/branches/ggreif/use-diet/lib/VMCore/getValue.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/branches/ggreif/use-diet/lib/VMCore/getValue.cpp?rev=56216&r1=56215&r2=56216&view=diff
==============================================================================
--- llvm/branches/ggreif/use-diet/lib/VMCore/getValue.cpp (original)
+++ llvm/branches/ggreif/use-diet/lib/VMCore/getValue.cpp Mon Sep 15 10:16:57 2008
@@ -62,7 +62,10 @@
/// punchAwayDigits -- ensure that repainted area
/// begins with a stop
///
-static inline void punchAwayDigits(Use *PrevU) {
+static inline void punchAwayDigits(Use *PrevU, Use **Uprev) {
+ if (PrevU)
+ assert(&PrevU->Next == stripTag(Uprev) && "U->Prev differs from PrevU?");
+
if (PrevU)
PrevU->Next = stripTag(PrevU->Next);
}
@@ -149,7 +152,7 @@
switch (Tag) {
case Use::fullStopTagN:
if (Cushion <= 0) {
- punchAwayDigits(PrevU);
+ punchAwayDigits(PrevU, U->Prev);
repaintByCalculating(reinterpret_cast(Next), U);
}
return reinterpret_cast(Next);
@@ -191,7 +194,7 @@
while (1) {
switch (Tag) {
case Use::fullStopTagN: {
- punchAwayDigits(PrevU);
+ punchAwayDigits(PrevU, U->Prev);
repaintByCalculating(reinterpret_cast(Next), U);
return reinterpret_cast(Next);
}
@@ -203,7 +206,7 @@
while (1) {
if (!digits) {
- punchAwayDigits(PrevU);
+ punchAwayDigits(PrevU, U->Prev);
repaintByCopying(Tagspace, U);
return reinterpret_cast(Acc << 2);
}
@@ -216,7 +219,7 @@
Next = stripTag(Next);
switch (Tag) {
case Use::fullStopTagN: {
- punchAwayDigits(PrevU);
+ punchAwayDigits(PrevU, U->Prev);
repaintByCalculating(reinterpret_cast(Next), U);
return reinterpret_cast(Next);
}
@@ -289,11 +292,6 @@
}
}
-/*inline*/Use *Use::nilUse(const Value *V) {
- // return 0;
- return addTag((Use*)V, fullStopTagN);
-}
-
static bool again(false);
Value *Use::get() const {
From gohman at apple.com Mon Sep 15 11:10:51 2008
From: gohman at apple.com (Dan Gohman)
Date: Mon, 15 Sep 2008 16:10:51 -0000
Subject: [llvm-commits] [llvm] r56221 - /llvm/trunk/utils/llvm.grm
Message-ID: <200809151610.m8FGAprU004754@zion.cs.uiuc.edu>
Author: djg
Date: Mon Sep 15 11:10:51 2008
New Revision: 56221
URL: http://llvm.org/viewvc/llvm-project?rev=56221&view=rev
Log:
Update the LLVM polygen grammar for recent language changes:
x86_ssecallcc, function notes, and some whitespace adjustments.
Modified:
llvm/trunk/utils/llvm.grm
Modified: llvm/trunk/utils/llvm.grm
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/utils/llvm.grm?rev=56221&r1=56220&r2=56221&view=diff
==============================================================================
--- llvm/trunk/utils/llvm.grm (original)
+++ llvm/trunk/utils/llvm.grm Mon Sep 15 11:10:51 2008
@@ -123,6 +123,7 @@
coldcc |
"x86_stdcallcc" |
"x86_fastcallcc" |
+ "x86_ssecallcc" |
cc EUINT64VAL ;
ParamAttr ::= zeroext
@@ -149,6 +150,19 @@
OptFuncAttrs ::= + _ | OptFuncAttrs FuncAttr ;
+FuncNoteList ::= FuncNote
+ | FuncNoteList ^ "," FuncNote
+ ;
+
+FuncNote ::= inline ^ "=" ^ never
+ | inline ^ "=" ^ always
+ | "opt_size"
+ ;
+
+OptFuncNotes ::= _
+ | notes ^ "(" ^ FuncNoteList ^ ")"
+ ;
+
OptGC ::= + _ | gc STRINGCONSTANT ;
OptAlign ::= + _ | align EUINT64VAL ;
@@ -175,9 +189,9 @@
| "[" ^ EUINT64VAL "x" Types ^ "]"
| "<" ^ EUINT64VAL "x" Types ^ ">"
| "{" TypeListI "}"
- | "{" "}"
+ | "{" ^ "}"
| "<" ^ "{" TypeListI "}" ^ ">"
- | "<" ^ "{" "}" ^ ">"
+ | "<" ^ "{" ^ "}" ^ ">"
;
ArgType ::= Types OptParamAttrs ;
@@ -191,13 +205,13 @@
TypeListI ::= Types | TypeListI ^ "," Types ;
ConstVal::= Types "[" ^ ConstVector ^ "]"
- | Types "[" "]"
+ | Types "[" ^ "]"
| Types "c" ^ STRINGCONSTANT
| Types "<" ^ ConstVector ^ ">"
| Types "{" ConstVector "}"
- | Types "{" "}"
+ | Types "{" ^ "}"
| Types "<" ^ "{" ConstVector "}" ^ ">"
- | Types "<" ^ "{" "}" ^ ">"
+ | Types "<" ^ "{" ^ "}" ^ ">"
| Types null
| Types undef
| Types SymbolicValueRef
@@ -271,7 +285,7 @@
ArgList ::= ArgListH | ArgListH ^ "," "..." | "..." | _ ;
FunctionHeaderH ::= OptCallingConv ResultTypes GlobalName "(" ^ ArgList ^ ")"
- OptFuncAttrs OptSection OptAlign OptGC ;
+ OptFuncAttrs OptSection OptAlign OptGC OptFuncNotes ;
BEGIN ::= ( begin | "{" ) ^ "\n";
@@ -296,12 +310,12 @@
| zeroinitializer
| "<" ConstVector ">"
| "[" ConstVector "]"
- | "[" "]"
+ | "[" ^ "]"
| "c" ^ STRINGCONSTANT
| "{" ConstVector "}"
- | "{" "}"
+ | "{" ^ "}"
| "<" ^ "{" ConstVector "}" ^ ">"
- | "<" ^ "{" "}" ^ ">"
+ | "<" ^ "{" ^ "}" ^ ">"
| ConstExpr
| asm OptSideEffect STRINGCONSTANT ^ "," STRINGCONSTANT ;
@@ -330,8 +344,9 @@
| br label ValueRef
| br INTTYPE ValueRef ^ "," label ValueRef ^ "," label ValueRef
| switch IntType ValueRef ^ "," label ValueRef "[" JumpTable "]"
- | switch IntType ValueRef ^ "," label ValueRef "[" "]"
- | invoke OptCallingConv ResultTypes ValueRef "(" ^ ParamList ^ ")" OptFuncAttrs
+ | switch IntType ValueRef ^ "," label ValueRef "[" ^ "]"
+ | invoke OptCallingConv ResultTypes ValueRef ^ "(" ^ ParamList ^ ")"
+ OptFuncAttrs
to label ValueRef unwind label ValueRef
| unwind
| unreachable ) ^ "\n";
@@ -370,7 +385,7 @@
| insertelement ResolvedVal ^ "," ResolvedVal ^ "," ResolvedVal
| shufflevector ResolvedVal ^ "," ResolvedVal ^ "," ResolvedVal
| phi PHIList
- | OptTailCall OptCallingConv ResultTypes ValueRef "(" ^ ParamList ^ ")"
+ | OptTailCall OptCallingConv ResultTypes ValueRef ^ "(" ^ ParamList ^ ")"
OptFuncAttrs
| MemoryInst ;
From baldrick at free.fr Mon Sep 15 13:42:40 2008
From: baldrick at free.fr (Duncan Sands)
Date: Mon, 15 Sep 2008 18:42:40 -0000
Subject: [llvm-commits] [llvm] r56222 - /llvm/trunk/tools/opt/opt.cpp
Message-ID: <200809151842.m8FIgerD009679@zion.cs.uiuc.edu>
Author: baldrick
Date: Mon Sep 15 13:42:38 2008
New Revision: 56222
URL: http://llvm.org/viewvc/llvm-project?rev=56222&view=rev
Log:
End of the GlobalsModRef experiment.
Modified:
llvm/trunk/tools/opt/opt.cpp
Modified: llvm/trunk/tools/opt/opt.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/tools/opt/opt.cpp?rev=56222&r1=56221&r2=56222&view=diff
==============================================================================
--- llvm/trunk/tools/opt/opt.cpp (original)
+++ llvm/trunk/tools/opt/opt.cpp Mon Sep 15 13:42:38 2008
@@ -285,8 +285,6 @@
addPass(PM, createLoopDeletionPass()); // Delete dead loops
addPass(PM, createLoopUnrollPass()); // Unroll small loops
addPass(PM, createInstructionCombiningPass()); // Clean up after the unroller
- addPass(PM, createGlobalsModRefPass()); // Alias analysis of globals
- addPass(PM, createMarkModRefPass()); // Mark functions readonly/readnone
addPass(PM, createGVNPass()); // Remove redundancies
addPass(PM, createMemCpyOptPass()); // Remove memcpy / form memset
addPass(PM, createSCCPPass()); // Constant prop with SCCP
From clattner at apple.com Mon Sep 15 13:54:29 2008
From: clattner at apple.com (Chris Lattner)
Date: Mon, 15 Sep 2008 11:54:29 -0700
Subject: [llvm-commits] [llvm] r56189 - in /llvm/trunk:
include/llvm/CodeGen/ lib/CodeGen/ lib/Target/ARM/
lib/Target/PowerPC/ lib/Target/Sparc/ lib/Target/X86/
In-Reply-To: <200809131758.m8DHwOc8023051@zion.cs.uiuc.edu>
References: <200809131758.m8DHwOc8023051@zion.cs.uiuc.edu>
Message-ID:
On Sep 13, 2008, at 10:58 AM, Dan Gohman wrote:
> Author: djg
> Date: Sat Sep 13 12:58:21 2008
> New Revision: 56189
>
> URL: http://llvm.org/viewvc/llvm-project?rev=56189&view=rev
> Log:
> Remove isImm(), isReg(), and friends, in favor of
> isImmediate(), isRegister(), and friends, to avoid confusion
> about having two different names with the same meaning. I'm
> not attached to the longer names, and would be ok with
> changing to the shorter names if others prefer it.
Thanks for doing this Dan, but could you go the other way? The short
names are just as legible and are more concise. I started converting
all the old ones over to the short ones before but apparently got
distracted and never finished, :(
-Chris
From baldrick at free.fr Mon Sep 15 13:57:31 2008
From: baldrick at free.fr (Duncan Sands)
Date: Mon, 15 Sep 2008 18:57:31 -0000
Subject: [llvm-commits] [llvm-gcc-4.2] r56223 - in /llvm-gcc-4.2/trunk/gcc:
llvm-backend.cpp llvm-linker-hack.cpp
Message-ID: <200809151857.m8FIvVtK010206@zion.cs.uiuc.edu>
Author: baldrick
Date: Mon Sep 15 13:57:29 2008
New Revision: 56223
URL: http://llvm.org/viewvc/llvm-project?rev=56223&view=rev
Log:
End of the GlobalsModRef experiment.
Modified:
llvm-gcc-4.2/trunk/gcc/llvm-backend.cpp
llvm-gcc-4.2/trunk/gcc/llvm-linker-hack.cpp
Modified: llvm-gcc-4.2/trunk/gcc/llvm-backend.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm-gcc-4.2/trunk/gcc/llvm-backend.cpp?rev=56223&r1=56222&r2=56223&view=diff
==============================================================================
--- llvm-gcc-4.2/trunk/gcc/llvm-backend.cpp (original)
+++ llvm-gcc-4.2/trunk/gcc/llvm-backend.cpp Mon Sep 15 13:57:29 2008
@@ -30,7 +30,6 @@
#include "llvm/PassManager.h"
#include "llvm/ValueSymbolTable.h"
#include "llvm/Analysis/LoopPass.h"
-#include "llvm/Analysis/Passes.h"
#include "llvm/Analysis/Verifier.h"
#include "llvm/Assembly/Writer.h"
#include "llvm/Assembly/PrintModulePass.h"
@@ -430,10 +429,6 @@
if (flag_unroll_loops)
PM->add(createLoopUnrollPass()); // Unroll small loops
PM->add(createInstructionCombiningPass()); // Clean up after the unroller
- if (flag_unit_at_a_time) {
- PM->add(createGlobalsModRefPass()); // Alias analysis of globals
- PM->add(createMarkModRefPass()); // Mark functions readonly/readnone
- }
PM->add(createGVNPass()); // Remove redundancies
PM->add(createMemCpyOptPass()); // Remove memcpy / form memset
PM->add(createSCCPPass()); // Constant prop with SCCP
Modified: llvm-gcc-4.2/trunk/gcc/llvm-linker-hack.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm-gcc-4.2/trunk/gcc/llvm-linker-hack.cpp?rev=56223&r1=56222&r2=56223&view=diff
==============================================================================
--- llvm-gcc-4.2/trunk/gcc/llvm-linker-hack.cpp (original)
+++ llvm-gcc-4.2/trunk/gcc/llvm-linker-hack.cpp Mon Sep 15 13:57:29 2008
@@ -22,7 +22,6 @@
#include "llvm/Module.h"
#include "llvm/ModuleProvider.h"
-#include "llvm/Analysis/Passes.h"
#include "llvm/Analysis/Verifier.h"
#include "llvm/Target/TargetMachineRegistry.h"
#include "llvm/Transforms/Scalar.h"
@@ -86,8 +85,6 @@
llvm::createMemCpyOptPass();
llvm::createDeadTypeEliminationPass();
llvm::createLoopDeletionPass();
- llvm::createGlobalsModRefPass();
- llvm::createMarkModRefPass();
std::string Err;
llvm::TargetMachineRegistry::getClosestStaticTargetForModule(*MP->getModule(),
From ggreif at gmail.com Mon Sep 15 14:15:30 2008
From: ggreif at gmail.com (Gabor Greif)
Date: Mon, 15 Sep 2008 19:15:30 -0000
Subject: [llvm-commits] [llvm] r56224 - in /llvm/branches/ggreif/use-diet:
include/llvm/Use.h lib/VMCore/getValue.cpp
Message-ID: <200809151915.m8FJFVOo010714@zion.cs.uiuc.edu>
Author: ggreif
Date: Mon Sep 15 14:15:29 2008
New Revision: 56224
URL: http://llvm.org/viewvc/llvm-project?rev=56224&view=rev
Log:
simplify
Modified:
llvm/branches/ggreif/use-diet/include/llvm/Use.h
llvm/branches/ggreif/use-diet/lib/VMCore/getValue.cpp
Modified: llvm/branches/ggreif/use-diet/include/llvm/Use.h
URL: http://llvm.org/viewvc/llvm-project/llvm/branches/ggreif/use-diet/include/llvm/Use.h?rev=56224&r1=56223&r2=56224&view=diff
==============================================================================
--- llvm/branches/ggreif/use-diet/include/llvm/Use.h (original)
+++ llvm/branches/ggreif/use-diet/include/llvm/Use.h Mon Sep 15 14:15:29 2008
@@ -115,17 +115,9 @@
bool P[] = { true, false, false, true };
return P[T];
}
- inline Value *getFastValueMaybe() const;
public:
-
-
operator Value*() const { return get(); }
- Value *get() const; /*{
- if (Value *V = getFastValueMaybe())
- return V;
- else
- return Val1; // for now :-)
- }*/
+ inline Value *get() const;
User *getUser() const;
const Use* getImpliedUser() const;
static Use *initTags(Use *Start, Use *Stop, ptrdiff_t Done = 0);
@@ -247,14 +239,12 @@
unsigned getOperandNo() const;
};
-Value *Use::getFastValueMaybe() const {
- if (fullStopTagN == extractTag(Next)) {
- return reinterpret_cast(stripTag(Next));
- }
- return 0;
+Value *Use::get() const {
+ return fullStopTagN == extractTag(Next)
+ ? reinterpret_cast(stripTag(Next))
+ : (Val1 == getValue() ? Val1 : 0); // should crash if not equal!
}
-
template<> struct simplify_type > {
typedef User* SimpleType;
Modified: llvm/branches/ggreif/use-diet/lib/VMCore/getValue.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/branches/ggreif/use-diet/lib/VMCore/getValue.cpp?rev=56224&r1=56223&r2=56224&view=diff
==============================================================================
--- llvm/branches/ggreif/use-diet/lib/VMCore/getValue.cpp (original)
+++ llvm/branches/ggreif/use-diet/lib/VMCore/getValue.cpp Mon Sep 15 14:15:29 2008
@@ -63,11 +63,11 @@
/// begins with a stop
///
static inline void punchAwayDigits(Use *PrevU, Use **Uprev) {
+ Uprev = stripTag(Uprev);
if (PrevU)
- assert(&PrevU->Next == stripTag(Uprev) && "U->Prev differs from PrevU?");
+ assert(&PrevU->Next == Uprev && "U->Prev differs from PrevU?");
- if (PrevU)
- PrevU->Next = stripTag(PrevU->Next);
+ *Uprev = stripTag(*Uprev);
}
@@ -292,20 +292,6 @@
}
}
-static bool again(false);
-
-Value *Use::get() const {
- Value *V(Val1);
- // if ((size_t)V == 0x5b0d150)
- // getValue();
- Value *ValComp(getValue());
- if (V != ValComp)
- assert(V == ValComp && "Computed Value wrong?");
- if (again)
- getValue();
- return V;
-}
-
static char TagChar(int Tag) {
return "s10S"[Tag];
}
From bruno.cardoso at gmail.com Mon Sep 15 14:38:14 2008
From: bruno.cardoso at gmail.com (Bruno Cardoso Lopes)
Date: Mon, 15 Sep 2008 19:38:14 -0000
Subject: [llvm-commits] [llvm] r56225 -
/llvm/trunk/test/CodeGen/Mips/2008-08-08-bswap.ll
Message-ID: <200809151938.m8FJcEXA011389@zion.cs.uiuc.edu>
Author: bruno
Date: Mon Sep 15 14:38:11 2008
New Revision: 56225
URL: http://llvm.org/viewvc/llvm-project?rev=56225&view=rev
Log:
Added testcase for bswap allegrexel intrinsic
Added:
llvm/trunk/test/CodeGen/Mips/2008-08-08-bswap.ll
Added: llvm/trunk/test/CodeGen/Mips/2008-08-08-bswap.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/Mips/2008-08-08-bswap.ll?rev=56225&view=auto
==============================================================================
--- llvm/trunk/test/CodeGen/Mips/2008-08-08-bswap.ll (added)
+++ llvm/trunk/test/CodeGen/Mips/2008-08-08-bswap.ll Mon Sep 15 14:38:11 2008
@@ -0,0 +1,12 @@
+; RUN: llvm-as < %s | llc -march=mips | grep wsbw | count 1
+
+target datalayout = "e-p:32:32:32-i1:8:8-i8:8:32-i16:16:32-i32:32:32-i64:32:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64"
+target triple = "psp"
+
+define i32 @__bswapsi2(i32 %u) nounwind {
+entry:
+ tail call i32 @llvm.bswap.i32( i32 %u ) ; :0 [#uses=1]
+ ret i32 %0
+}
+
+declare i32 @llvm.bswap.i32(i32) nounwind readnone
From gohman at apple.com Mon Sep 15 14:46:04 2008
From: gohman at apple.com (Dan Gohman)
Date: Mon, 15 Sep 2008 19:46:04 -0000
Subject: [llvm-commits] [llvm] r56226 - in /llvm/trunk:
include/llvm/CodeGen/SelectionDAGNodes.h
lib/CodeGen/SelectionDAG/SelectionDAG.cpp test/CodeGen/ARM/cse-libcalls.ll
Message-ID: <200809151946.m8FJk4g1011624@zion.cs.uiuc.edu>
Author: djg
Date: Mon Sep 15 14:46:03 2008
New Revision: 56226
URL: http://llvm.org/viewvc/llvm-project?rev=56226&view=rev
Log:
Re-enable SelectionDAG CSE for calls. It matters in the case of
libcalls, as in this testcase on ARM.
Added:
llvm/trunk/test/CodeGen/ARM/cse-libcalls.ll
Modified:
llvm/trunk/include/llvm/CodeGen/SelectionDAGNodes.h
llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
Modified: llvm/trunk/include/llvm/CodeGen/SelectionDAGNodes.h
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/include/llvm/CodeGen/SelectionDAGNodes.h?rev=56226&r1=56225&r2=56226&view=diff
==============================================================================
--- llvm/trunk/include/llvm/CodeGen/SelectionDAGNodes.h (original)
+++ llvm/trunk/include/llvm/CodeGen/SelectionDAGNodes.h Mon Sep 15 14:46:03 2008
@@ -2205,7 +2205,8 @@
/// Set this call to not be marked as a tail call. Normally setter
/// methods in SDNodes are unsafe because it breaks the CSE map,
- /// but we don't CSE calls so it's ok in this case.
+ /// but we don't include the tail call flag for calls so it's ok
+ /// in this case.
void setNotTailCall() { IsTailCall = false; }
SDValue getChain() const { return getOperand(0); }
Modified: llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAG.cpp?rev=56226&r1=56225&r2=56226&view=diff
==============================================================================
--- llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAG.cpp (original)
+++ llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAG.cpp Mon Sep 15 14:46:03 2008
@@ -423,6 +423,12 @@
ID.AddPointer(CP->getConstVal());
break;
}
+ case ISD::CALL: {
+ const CallSDNode *Call = cast(N);
+ ID.AddInteger(Call->getCallingConv());
+ ID.AddInteger(Call->isVarArg());
+ break;
+ }
case ISD::LOAD: {
const LoadSDNode *LD = cast(N);
ID.AddInteger(LD->getAddressingMode());
@@ -636,7 +642,6 @@
// not subject to CSE.
if (!Erased && N->getValueType(N->getNumValues()-1) != MVT::Flag &&
!N->isMachineOpcode() &&
- N->getOpcode() != ISD::CALL &&
N->getOpcode() != ISD::DBG_LABEL &&
N->getOpcode() != ISD::DBG_STOPPOINT &&
N->getOpcode() != ISD::EH_LABEL &&
@@ -662,7 +667,6 @@
switch (N->getOpcode()) {
default: break;
- case ISD::CALL:
case ISD::HANDLENODE:
case ISD::DBG_LABEL:
case ISD::DBG_STOPPOINT:
@@ -3310,13 +3314,23 @@
SelectionDAG::getCall(unsigned CallingConv, bool IsVarArgs, bool IsTailCall,
SDVTList VTs,
const SDValue *Operands, unsigned NumOperands) {
- // Do not CSE calls. Note that in addition to being a compile-time
- // optimization (since attempting CSE of calls is unlikely to be
- // meaningful), we actually depend on this behavior. CallSDNode can
- // be mutated, which is only safe if calls are not CSE'd.
+ // Do not include isTailCall in the folding set profile.
+ FoldingSetNodeID ID;
+ AddNodeIDNode(ID, ISD::CALL, VTs, Operands, NumOperands);
+ ID.AddInteger(CallingConv);
+ ID.AddInteger(IsVarArgs);
+ void *IP = 0;
+ if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) {
+ // Instead of including isTailCall in the folding set, we just
+ // set the flag of the existing node.
+ if (!IsTailCall)
+ cast(E)->setNotTailCall();
+ return SDValue(E, 0);
+ }
SDNode *N = NodeAllocator.Allocate();
new (N) CallSDNode(CallingConv, IsVarArgs, IsTailCall,
VTs, Operands, NumOperands);
+ CSEMap.InsertNode(N, IP);
AllNodes.push_back(N);
return SDValue(N, 0);
}
Added: llvm/trunk/test/CodeGen/ARM/cse-libcalls.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/ARM/cse-libcalls.ll?rev=56226&view=auto
==============================================================================
--- llvm/trunk/test/CodeGen/ARM/cse-libcalls.ll (added)
+++ llvm/trunk/test/CodeGen/ARM/cse-libcalls.ll Mon Sep 15 14:46:03 2008
@@ -0,0 +1,30 @@
+; RUN: llvm-as < %s | llc -march=arm | grep {bl.\*__ltdf} | count 1
+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 = "i386-apple-darwin8"
+
+; Without CSE of libcalls, there are two calls in the output instead of one.
+
+define i32 @u_f_nonbon(double %lambda) nounwind {
+entry:
+ %tmp19.i.i = load double* null, align 4 ; [#uses=2]
+ %tmp6.i = fcmp olt double %tmp19.i.i, 1.000000e+00 ; [#uses=1]
+ %dielectric.0.i = select i1 %tmp6.i, double 1.000000e+00, double %tmp19.i.i ; [#uses=1]
+ %tmp10.i4 = fdiv double 0x4074C2D71F36262D, %dielectric.0.i ; [#uses=1]
+ br i1 false, label %bb28.i, label %bb508.i
+
+bb28.i: ; preds = %bb28.i, %entry
+ br i1 false, label %bb502.loopexit.i, label %bb28.i
+
+bb.nph53.i: ; preds = %bb502.loopexit.i
+ %tmp354.i = sub double -0.000000e+00, %tmp10.i4 ; [#uses=0]
+ br label %bb244.i
+
+bb244.i: ; preds = %bb244.i, %bb.nph53.i
+ br label %bb244.i
+
+bb502.loopexit.i: ; preds = %bb28.i
+ br i1 false, label %bb.nph53.i, label %bb508.i
+
+bb508.i: ; preds = %bb502.loopexit.i, %entry
+ ret i32 1
+}
From ggreif at gmail.com Mon Sep 15 15:01:55 2008
From: ggreif at gmail.com (Gabor Greif)
Date: Mon, 15 Sep 2008 20:01:55 -0000
Subject: [llvm-commits] [llvm] r56227 - in /llvm/branches/ggreif/use-diet:
include/llvm/Use.h lib/VMCore/getValue.cpp
Message-ID: <200809152001.m8FK1tj6012105@zion.cs.uiuc.edu>
Author: ggreif
Date: Mon Sep 15 15:01:53 2008
New Revision: 56227
URL: http://llvm.org/viewvc/llvm-project?rev=56227&view=rev
Log:
more cleanups, enable other efficiency hack
Modified:
llvm/branches/ggreif/use-diet/include/llvm/Use.h
llvm/branches/ggreif/use-diet/lib/VMCore/getValue.cpp
Modified: llvm/branches/ggreif/use-diet/include/llvm/Use.h
URL: http://llvm.org/viewvc/llvm-project/llvm/branches/ggreif/use-diet/include/llvm/Use.h?rev=56227&r1=56226&r2=56227&view=diff
==============================================================================
--- llvm/branches/ggreif/use-diet/include/llvm/Use.h (original)
+++ llvm/branches/ggreif/use-diet/include/llvm/Use.h Mon Sep 15 15:01:53 2008
@@ -161,7 +161,6 @@
if (isStop(Next))
assert((isStop(*StrippedPrev) || (StrippedNext ? isStop(StrippedNext->Next) : true)) && "joining digits?");
*StrippedPrev = Next;
- // Use *StrippedNext(getNext());
if (StrippedNext) StrippedNext->setPrev(StrippedPrev);
}
Modified: llvm/branches/ggreif/use-diet/lib/VMCore/getValue.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/branches/ggreif/use-diet/lib/VMCore/getValue.cpp?rev=56227&r1=56226&r2=56227&view=diff
==============================================================================
--- llvm/branches/ggreif/use-diet/lib/VMCore/getValue.cpp (original)
+++ llvm/branches/ggreif/use-diet/lib/VMCore/getValue.cpp Mon Sep 15 15:01:53 2008
@@ -12,8 +12,6 @@
//===----------------------------------------------------------------------===//
#include "llvm/User.h"
-#include
-//#include
// this can later be removed:
#include
@@ -23,7 +21,7 @@
friend class Use;
-enum { requiredSteps = sizeof(Value*) * 8 - 2 };
+enum { spareBits = 2, requiredSteps = sizeof(Value*) * 8 - spareBits };
/// repaintByCopying -- given a pattern and a
/// junk tagspace, copy the former's tags into
@@ -46,7 +44,7 @@
/// junk tagspace, compute tags into the latter
///
static inline void repaintByCalculating(unsigned long Tags, Use *Junk) {
- Tags >>= 2;
+ Tags >>= spareBits;
for (int I = requiredSteps - 1; I >= 0; --I) {
Use::NextPtrTag Tag(Tags & (1 << I) ? Use::oneDigitTagN : Use::zeroDigitTagN);
@@ -62,12 +60,11 @@
/// punchAwayDigits -- ensure that repainted area
/// begins with a stop
///
-static inline void punchAwayDigits(Use *PrevU, Use **Uprev) {
+static inline void punchAwayDigits(Use **Uprev) {
Uprev = stripTag(Uprev);
- if (PrevU)
- assert(&PrevU->Next == Uprev && "U->Prev differs from PrevU?");
-
- *Uprev = stripTag(*Uprev);
+ // if (PrevU)
+ // assert(&PrevU->Next == Uprev && "U->Prev differs from PrevU?");
+ *Uprev = stripTag(*Uprev);
}
@@ -79,6 +76,7 @@
/// overpainted.
/// In any case this routine is invoked with U being
/// pointed at from a Use with a stop tag.
+///
static inline Value *gatherAndPotentiallyRepaint(Use *U) {
int Cushion = requiredSteps;
@@ -94,23 +92,14 @@
if (Cushion <= 0) {
assert((Tag == Use::fullStopTagN || Tag == Use::stopTagN)
&& "More digits?");
- return reinterpret_cast(Acc << 2);
+ return reinterpret_cast(Acc << spareBits);
}
switch (Tag) {
case Use::fullStopTagN:
return reinterpret_cast(Next);
case Use::stopTagN: {
- // for efficieny:
- // goto efficiency
- /* THIS PESSIMIZES THE ddsd30S case !!!
- Next = Next->Next;
- // __builtin_prefetch(Next);
- --Cushion;
- Tag = extractTag(Next);
- Next = stripTag(Next);
- */
- break;
+ goto efficiency;
}
default:
Acc = (Acc << 1) | (Tag & 1);
@@ -124,12 +113,6 @@
break;
}
- // FIXME: We track prev this way, in the real
- // Use struct there is a (tagged) Prev.
- // It is needed to punch over junk
- // digit sequence just before repaint.
- Use *PrevU = NULL;
-
while (Cushion > 0) {
switch (Tag) {
case Use::fullStopTagN:
@@ -142,7 +125,7 @@
while (1) {
if (!digits)
- return reinterpret_cast(Acc << 2);
+ return reinterpret_cast(Acc << spareBits);
Next = Next->Next;
// __builtin_prefetch(Next);
@@ -152,13 +135,12 @@
switch (Tag) {
case Use::fullStopTagN:
if (Cushion <= 0) {
- punchAwayDigits(PrevU, U->Prev);
+ punchAwayDigits(U->Prev);
repaintByCalculating(reinterpret_cast(Next), U);
}
return reinterpret_cast(Next);
case Use::stopTagN: {
if (Cushion <= 0) {
- PrevU = U;
U = stripTag(U->Next);
}
goto efficiency;
@@ -167,7 +149,6 @@
--digits;
Acc = (Acc << 1) | (Tag & 1);
if (Cushion <= 0) {
- PrevU = U;
U = stripTag(U->Next);
}
continue;
@@ -185,16 +166,14 @@
} // switch
} // while
- // Now we know that we have a nice cushion between
- // U and Next. Do the same thing as above, but
- // don't decrement Cushion any more, instead
- // push U forward. After the value is found,
- // repaint beginning at U.
+ // Now we know that we have a nice cushion between U and Next. Do the same
+ // thing as above, but don't decrement Cushion any more, instead push U
+ // forward. After the value is found, repaint beginning at U.
while (1) {
switch (Tag) {
case Use::fullStopTagN: {
- punchAwayDigits(PrevU, U->Prev);
+ punchAwayDigits(U->Prev);
repaintByCalculating(reinterpret_cast(Next), U);
return reinterpret_cast(Next);
}
@@ -206,20 +185,19 @@
while (1) {
if (!digits) {
- punchAwayDigits(PrevU, U->Prev);
+ punchAwayDigits(U->Prev);
repaintByCopying(Tagspace, U);
- return reinterpret_cast(Acc << 2);
+ return reinterpret_cast(Acc << spareBits);
}
Next = Next->Next;
// __builtin_prefetch(Next);
- PrevU = U;
U = stripTag(U->Next);
Tag = extractTag(Next);
Next = stripTag(Next);
switch (Tag) {
case Use::fullStopTagN: {
- punchAwayDigits(PrevU, U->Prev);
+ punchAwayDigits(U->Prev);
repaintByCalculating(reinterpret_cast(Next), U);
return reinterpret_cast(Next);
}
@@ -238,7 +216,6 @@
default:
Next = Next->Next;
// __builtin_prefetch(Next);
- PrevU = U;
U = stripTag(U->Next);
Tag = extractTag(Next);
Next = stripTag(Next);
@@ -251,12 +228,13 @@
/// - picking up exactly ToGo digits
/// - or finding a stop which marks the beginning
/// of a repaintable area
+///
static inline Value *skipPotentiallyGathering(Use *U,
unsigned long Acc,
int ToGo) {
while (1) {
if (!ToGo)
- return reinterpret_cast(Acc << 2);
+ return reinterpret_cast(Acc << spareBits);
Use *Next(U->Next);
// __builtin_prefetch(Next);
From bruno.cardoso at gmail.com Mon Sep 15 16:06:58 2008
From: bruno.cardoso at gmail.com (Bruno Cardoso Lopes)
Date: Mon, 15 Sep 2008 21:06:58 -0000
Subject: [llvm-commits] [llvm] r56228 - in /llvm/trunk/lib/Target/Mips:
MipsSubtarget.cpp MipsTargetMachine.cpp
Message-ID: <200809152106.m8FL6xPU014011@zion.cs.uiuc.edu>
Author: bruno
Date: Mon Sep 15 16:06:55 2008
New Revision: 56228
URL: http://llvm.org/viewvc/llvm-project?rev=56228&view=rev
Log:
Fixed Bug 2751
http://llvm.org/bugs/show_bug.cgi?id=2751
Abicall was enabled even when static code model was provided
in the command line.
The correct behavior is to disable abicall when static is
specified.
Modified:
llvm/trunk/lib/Target/Mips/MipsSubtarget.cpp
llvm/trunk/lib/Target/Mips/MipsTargetMachine.cpp
Modified: llvm/trunk/lib/Target/Mips/MipsSubtarget.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/Mips/MipsSubtarget.cpp?rev=56228&r1=56227&r2=56228&view=diff
==============================================================================
--- llvm/trunk/lib/Target/Mips/MipsSubtarget.cpp (original)
+++ llvm/trunk/lib/Target/Mips/MipsSubtarget.cpp Mon Sep 15 16:06:55 2008
@@ -65,9 +65,9 @@
HasCondMov = true;
}
- // Abicall is the default for O32 ABI and is ignored
- // for EABI.
- if (NotABICall || isABI_EABI())
+ // Abicall is the default for O32 ABI, but is disabled within EABI and in
+ // static code.
+ if (NotABICall || isABI_EABI() || (TM.getRelocationModel() == Reloc::Static))
HasABICall = false;
// TODO: disable when handling 64 bit symbols in the future.
Modified: llvm/trunk/lib/Target/Mips/MipsTargetMachine.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/Mips/MipsTargetMachine.cpp?rev=56228&r1=56227&r2=56228&view=diff
==============================================================================
--- llvm/trunk/lib/Target/Mips/MipsTargetMachine.cpp (original)
+++ llvm/trunk/lib/Target/Mips/MipsTargetMachine.cpp Mon Sep 15 16:06:55 2008
@@ -46,7 +46,7 @@
TLInfo(*this)
{
// Abicall enables PIC by default
- if (Subtarget.hasABICall() && (getRelocationModel() != Reloc::Static))
+ if (Subtarget.hasABICall())
setRelocationModel(Reloc::PIC_);
// TODO: create an option to enable long calls, like -mlong-calls,
From dpatel at apple.com Mon Sep 15 16:13:46 2008
From: dpatel at apple.com (Devang Patel)
Date: Mon, 15 Sep 2008 21:13:46 -0000
Subject: [llvm-commits] [llvm] r56229 - in /llvm/trunk:
include/llvm/PassManagerUtils.h utils/PassManagerUtils.cpp
Message-ID: <200809152113.m8FLDkQM014227@zion.cs.uiuc.edu>
Author: dpatel
Date: Mon Sep 15 16:13:42 2008
New Revision: 56229
URL: http://llvm.org/viewvc/llvm-project?rev=56229&view=rev
Log:
Extract optimization pass selection code from llvm-gcc into a separate routine.
This can be used by other stand alone tools, such as 'opt'.
Added:
llvm/trunk/include/llvm/PassManagerUtils.h
llvm/trunk/utils/PassManagerUtils.cpp
Added: llvm/trunk/include/llvm/PassManagerUtils.h
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/include/llvm/PassManagerUtils.h?rev=56229&view=auto
==============================================================================
--- llvm/trunk/include/llvm/PassManagerUtils.h (added)
+++ llvm/trunk/include/llvm/PassManagerUtils.h Mon Sep 15 16:13:42 2008
@@ -0,0 +1,36 @@
+//===-- llvm/Support/PassManagerUtils.h -------------------------*- C++ -*-===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file provides interface to pass manager utilities.
+//
+//===----------------------------------------------------------------------===//
+
+namespace llvm {
+
+class FunctionPassManager;
+class PassManager;
+
+/// AddOptimizationPasses - This routine adds optimization passes
+/// based on selected optimization level, OptLevel. This routine is
+/// used by llvm-gcc and other tools.
+///
+/// OptLevel - Optimization Level
+/// EnableIPO - Enables IPO passes. llvm-gcc enables this when
+/// flag_unit_at_a_time is set.
+/// InlinerSelection - 1 : Add function inliner.
+/// - 2 : Add AlwaysInliner.
+/// OptLibCalls - Simplify lib calls, if set.
+/// PruneEH - Add PruneEHPass, if set.
+/// UnrollLoop - Unroll loops, if set.
+void AddOptimizationPasses(FunctionPassManager &FPM, PassManager &MPM,
+ unsigned OptLevel, bool EnableIPO,
+ unsigned InlinerSelection, bool OptLibCalls,
+ bool PruneEH, bool UnrollLoop);
+
+}
Added: llvm/trunk/utils/PassManagerUtils.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/utils/PassManagerUtils.cpp?rev=56229&view=auto
==============================================================================
--- llvm/trunk/utils/PassManagerUtils.cpp (added)
+++ llvm/trunk/utils/PassManagerUtils.cpp Mon Sep 15 16:13:42 2008
@@ -0,0 +1,108 @@
+//===-- PassManagerUtils.cpp - --------------------------------------------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements pass manager utiliy routines.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/PassManagerUtils.h"
+#include "llvm/PassManagers.h"
+#include "llvm/Transforms/Scalar.h"
+#include "llvm/Transforms/IPO.h"
+#include "llvm/Analysis/LoopPass.h"
+
+/// AddOptimizationPasses - This routine adds optimization passes
+/// based on selected optimization level, OptLevel. This routine is
+/// used by llvm-gcc and other tools.
+///
+/// OptLevel - Optimization Level
+/// EnableIPO - Enables IPO passes. llvm-gcc enables this when
+/// flag_unit_at_a_time is set.
+/// InlinerSelection - 1 : Add function inliner.
+/// - 2 : Add AlwaysInliner.
+/// OptLibCalls - Simplify lib calls, if set.
+/// PruneEH - Add PruneEHPass, if set.
+/// UnrollLoop - Unroll loops, if set.
+void llvm::AddOptimizationPasses(FunctionPassManager &FPM, PassManager &MPM,
+ unsigned OptLevel, bool EnableIPO,
+ unsigned InlinerSelection, bool OptLibCalls,
+ bool PruneEH, bool UnrollLoop) {
+ if (OptLevel == 0)
+ return;
+
+ FPM.add(createCFGSimplificationPass());
+ if (OptLevel == 1)
+ FPM.add(createPromoteMemoryToRegisterPass());
+ else
+ FPM.add(createScalarReplAggregatesPass());
+ FPM.add(createInstructionCombiningPass());
+
+ if (EnableIPO)
+ MPM.add(createRaiseAllocationsPass()); // call %malloc -> malloc inst
+ MPM.add(createCFGSimplificationPass()); // Clean up disgusting code
+ MPM.add(createPromoteMemoryToRegisterPass()); // Kill useless allocas
+ if (EnableIPO) {
+ MPM.add(createGlobalOptimizerPass()); // OptLevel out global vars
+ MPM.add(createGlobalDCEPass()); // Remove unused fns and globs
+ MPM.add(createIPConstantPropagationPass()); // IP Constant Propagation
+ MPM.add(createDeadArgEliminationPass()); // Dead argument elimination
+ }
+ MPM.add(createInstructionCombiningPass()); // Clean up after IPCP & DAE
+ MPM.add(createCFGSimplificationPass()); // Clean up after IPCP & DAE
+ if (EnableIPO && PruneEH)
+ MPM.add(createPruneEHPass()); // Remove dead EH info
+ if (InlinerSelection == 1) // respect -fno-inline-functions
+ MPM.add(createFunctionInliningPass()); // Inline small functions
+ else if (InlinerSelection == 2)
+ MPM.add(createAlwaysInlinerPass()); // Inline always_inline functions
+ if (OptLevel > 2)
+ MPM.add(createArgumentPromotionPass()); // Scalarize uninlined fn args
+ if (OptLibCalls)
+ MPM.add(createSimplifyLibCallsPass()); // Library Call Optimizations
+ MPM.add(createInstructionCombiningPass()); // Cleanup for scalarrepl.
+ MPM.add(createJumpThreadingPass()); // Thread jumps.
+ MPM.add(createCFGSimplificationPass()); // Merge & remove BBs
+ MPM.add(createScalarReplAggregatesPass()); // Break up aggregate allocas
+ MPM.add(createInstructionCombiningPass()); // Combine silly seq's
+ MPM.add(createCondPropagationPass()); // Propagate conditionals
+ MPM.add(createTailCallEliminationPass()); // Eliminate tail calls
+ MPM.add(createCFGSimplificationPass()); // Merge & remove BBs
+ MPM.add(createReassociatePass()); // Reassociate expressions
+ MPM.add(createLoopRotatePass()); // Rotate Loop
+ MPM.add(createLICMPass()); // Hoist loop invariants
+ MPM.add(createLoopUnswitchPass());
+ MPM.add(createLoopIndexSplitPass()); // Split loop index
+ MPM.add(createInstructionCombiningPass());
+ MPM.add(createIndVarSimplifyPass()); // Canonicalize indvars
+ MPM.add(createLoopDeletionPass()); // Delete dead loops
+ if (UnrollLoop)
+ MPM.add(createLoopUnrollPass()); // Unroll small loops
+ MPM.add(createInstructionCombiningPass()); // Clean up after the unroller
+ MPM.add(createGVNPass()); // Remove redundancies
+ MPM.add(createMemCpyOptPass()); // Remove memcpy / form memset
+ MPM.add(createSCCPPass()); // Constant prop with SCCP
+
+ // Run instcombine after redundancy elimination to exploit opportunities
+ // opened up by them.
+ MPM.add(createInstructionCombiningPass());
+ MPM.add(createCondPropagationPass()); // Propagate conditionals
+ MPM.add(createDeadStoreEliminationPass()); // Delete dead stores
+ MPM.add(createAggressiveDCEPass()); // Delete dead instructions
+ MPM.add(createCFGSimplificationPass()); // Merge & remove BBs
+
+ if (EnableIPO) {
+ MPM.add(createStripDeadPrototypesPass()); // Get rid of dead prototypes
+ MPM.add(createDeadTypeEliminationPass()); // Eliminate dead types
+ }
+
+ if (OptLevel > 1 && EnableIPO)
+ MPM.add(createConstantMergePass()); // Merge dup global constants
+
+ return;
+}
From gohman at apple.com Mon Sep 15 16:22:08 2008
From: gohman at apple.com (Dan Gohman)
Date: Mon, 15 Sep 2008 21:22:08 -0000
Subject: [llvm-commits] [llvm] r56230 - in /llvm/trunk:
lib/Transforms/Scalar/LoopStrengthReduce.cpp
test/CodeGen/X86/optimize-smax.ll
Message-ID: <200809152122.m8FLM8iT014496@zion.cs.uiuc.edu>
Author: djg
Date: Mon Sep 15 16:22:06 2008
New Revision: 56230
URL: http://llvm.org/viewvc/llvm-project?rev=56230&view=rev
Log:
Teach LSR to optimize away SMAX operations for tripcounts in common
cases. See the comment above OptimizeSMax for the full story, and
the testcase for an example. This cancels out a pessimization
commonly attributed to indvars, and will allow us to lift some of
the artificial throttles in indvars, rather than add new ones.
Added:
llvm/trunk/test/CodeGen/X86/optimize-smax.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=56230&r1=56229&r2=56230&view=diff
==============================================================================
--- llvm/trunk/lib/Transforms/Scalar/LoopStrengthReduce.cpp (original)
+++ llvm/trunk/lib/Transforms/Scalar/LoopStrengthReduce.cpp Mon Sep 15 16:22:06 2008
@@ -184,6 +184,11 @@
/// inside the loop then try to eliminate the cast opeation.
void OptimizeShadowIV(Loop *L);
+ /// OptimizeSMax - Rewrite the loop's terminating condition
+ /// if it uses an smax computation.
+ ICmpInst *OptimizeSMax(Loop *L, ICmpInst *Cond,
+ IVStrideUse* &CondUse);
+
bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
const SCEVHandle *&CondStride);
bool RequiresTypeConversion(const Type *Ty, const Type *NewTy);
@@ -1695,6 +1700,123 @@
return Cond;
}
+/// OptimizeSMax - Rewrite the loop's terminating condition if it uses
+/// an smax computation.
+///
+/// This is a narrow solution to a specific, but acute, problem. For loops
+/// like this:
+///
+/// i = 0;
+/// do {
+/// p[i] = 0.0;
+/// } while (++i < n);
+///
+/// where the comparison is signed, the trip count isn't just 'n', because
+/// 'n' could be negative. And unfortunately this can come up even for loops
+/// where the user didn't use a C do-while loop. For example, seemingly
+/// well-behaved top-test loops will commonly be lowered like this:
+//
+/// if (n > 0) {
+/// i = 0;
+/// do {
+/// p[i] = 0.0;
+/// } while (++i < n);
+/// }
+///
+/// and then it's possible for subsequent optimization to obscure the if
+/// test in such a way that indvars can't find it.
+///
+/// When indvars can't find the if test in loops like this, it creates a
+/// signed-max expression, which allows it to give the loop a canonical
+/// induction variable:
+///
+/// i = 0;
+/// smax = n < 1 ? 1 : n;
+/// do {
+/// p[i] = 0.0;
+/// } while (++i != smax);
+///
+/// Canonical induction variables are necessary because the loop passes
+/// are designed around them. The most obvious example of this is the
+/// LoopInfo analysis, which doesn't remember trip count values. It
+/// expects to be able to rediscover the trip count each time it is
+/// needed, and it does this using a simple analyis that only succeeds if
+/// the loop has a canonical induction variable.
+///
+/// However, when it comes time to generate code, the maximum operation
+/// can be quite costly, especially if it's inside of an outer loop.
+///
+/// This function solves this problem by detecting this type of loop and
+/// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
+/// the instructions for the maximum computation.
+///
+ICmpInst *LoopStrengthReduce::OptimizeSMax(Loop *L, ICmpInst *Cond,
+ IVStrideUse* &CondUse) {
+ // Check that the loop matches the pattern we're looking for.
+ if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
+ Cond->getPredicate() != CmpInst::ICMP_NE)
+ return Cond;
+
+ SelectInst *Sel = dyn_cast(Cond->getOperand(1));
+ if (!Sel || !Sel->hasOneUse()) return Cond;
+
+ SCEVHandle IterationCount = SE->getIterationCount(L);
+ if (isa(IterationCount))
+ return Cond;
+ SCEVHandle One = SE->getIntegerSCEV(1, IterationCount->getType());
+
+ // Adjust for an annoying getIterationCount quirk.
+ IterationCount = SE->getAddExpr(IterationCount, One);
+
+ // Check for a max calculation that matches the pattern.
+ SCEVSMaxExpr *SMax = dyn_cast(IterationCount);
+ if (!SMax || SMax != SE->getSCEV(Sel)) return Cond;
+
+ SCEVHandle SMaxLHS = SMax->getOperand(0);
+ SCEVHandle SMaxRHS = SMax->getOperand(1);
+ if (!SMaxLHS || SMaxLHS != One) return Cond;
+
+ // Check the relevant induction variable for conformance to
+ // the pattern.
+ SCEVHandle IV = SE->getSCEV(Cond->getOperand(0));
+ SCEVAddRecExpr *AR = dyn_cast(IV);
+ if (!AR || !AR->isAffine() ||
+ AR->getStart() != One ||
+ AR->getStepRecurrence(*SE) != One)
+ return Cond;
+
+ // Check the right operand of the select, and remember it, as it will
+ // be used in the new comparison instruction.
+ Value *NewRHS = 0;
+ if (SE->getSCEV(Sel->getOperand(1)) == SMaxRHS)
+ NewRHS = Sel->getOperand(1);
+ else if (SE->getSCEV(Sel->getOperand(2)) == SMaxRHS)
+ NewRHS = Sel->getOperand(2);
+ if (!NewRHS) return Cond;
+
+ // Ok, everything looks ok to change the condition into an SLT or SGE and
+ // delete the max calculation.
+ ICmpInst *NewCond =
+ new ICmpInst(Cond->getPredicate() == CmpInst::ICMP_NE ?
+ CmpInst::ICMP_SLT :
+ CmpInst::ICMP_SGE,
+ Cond->getOperand(0), NewRHS, "scmp", Cond);
+
+ // Delete the max calculation instructions.
+ Cond->replaceAllUsesWith(NewCond);
+ Cond->eraseFromParent();
+ SE->deleteValueFromRecords(Cond);
+ Instruction *Cmp = cast(Sel->getOperand(0));
+ Sel->eraseFromParent();
+ SE->deleteValueFromRecords(Sel);
+ if (Cmp->use_empty()) {
+ Cmp->eraseFromParent();
+ SE->deleteValueFromRecords(Cmp);
+ }
+ CondUse->User = NewCond;
+ return NewCond;
+}
+
/// OptimizeShadowIV - If IV is used in a int-to-float cast
/// inside the loop then try to eliminate the cast opeation.
void LoopStrengthReduce::OptimizeShadowIV(Loop *L) {
@@ -1836,6 +1958,11 @@
if (!FindIVUserForCond(Cond, CondUse, CondStride))
return; // setcc doesn't use the IV.
+ // If the trip count is computed in terms of an smax (due to ScalarEvolution
+ // being unable to find a sufficient guard, for example), change the loop
+ // comparison to use SLT instead of NE.
+ Cond = OptimizeSMax(L, Cond, CondUse);
+
// If possible, change stride and operands of the compare instruction to
// eliminate one stride.
Cond = ChangeCompareStride(L, Cond, CondUse, CondStride);
Added: llvm/trunk/test/CodeGen/X86/optimize-smax.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/optimize-smax.ll?rev=56230&view=auto
==============================================================================
--- llvm/trunk/test/CodeGen/X86/optimize-smax.ll (added)
+++ llvm/trunk/test/CodeGen/X86/optimize-smax.ll Mon Sep 15 16:22:06 2008
@@ -0,0 +1,236 @@
+; RUN: llvm-as < %s | llc -march=x86 | not grep cmov
+
+; LSR should be able to eliminate the smax computations by
+; making the loops use slt comparisons instead of ne comparisons.
+
+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 = "i386-apple-darwin9"
+
+define void @foo(i8* %r, i32 %s, i32 %w, i32 %x, i8* %j, i32 %d) nounwind {
+entry:
+ %0 = mul i32 %x, %w ; [#uses=2]
+ %1 = mul i32 %x, %w ; [#uses=1]
+ %2 = sdiv i32 %1, 4 ; [#uses=1]
+ %.sum2 = add i32 %2, %0 ; [#uses=2]
+ %cond = icmp eq i32 %d, 1 ; [#uses=1]
+ br i1 %cond, label %bb29, label %bb10.preheader
+
+bb10.preheader: ; preds = %entry
+ %3 = icmp sgt i32 %x, 0 ; [#uses=1]
+ br i1 %3, label %bb.nph9, label %bb18.loopexit
+
+bb.nph7: ; preds = %bb7.preheader
+ %4 = mul i32 %y.08, %w ; [#uses=1]
+ %5 = mul i32 %y.08, %s ; [#uses=1]
+ %6 = add i32 %5, 1 ; [#uses=1]
+ %tmp8 = icmp sgt i32 1, %w ; [#uses=1]
+ %smax9 = select i1 %tmp8, i32 1, i32 %w ; [#uses=1]
+ br label %bb6
+
+bb6: ; preds = %bb7, %bb.nph7
+ %x.06 = phi i32 [ 0, %bb.nph7 ], [ %indvar.next7, %bb7 ] ; [#uses=3]
+ %7 = add i32 %x.06, %4 ; [#uses=1]
+ %8 = shl i32 %x.06, 1 ; [#uses=1]
+ %9 = add i32 %6, %8 ; [#uses=1]
+ %10 = getelementptr i8* %r, i32 %9 ; [#uses=1]
+ %11 = load i8* %10, align 1 ; [#uses=1]
+ %12 = getelementptr i8* %j, i32 %7 ; [#uses=1]
+ store i8 %11, i8* %12, align 1
+ br label %bb7
+
+bb7: ; preds = %bb6
+ %indvar.next7 = add i32 %x.06, 1 ; [#uses=2]
+ %exitcond10 = icmp ne i32 %indvar.next7, %smax9 ; [#uses=1]
+ br i1 %exitcond10, label %bb6, label %bb7.bb9_crit_edge
+
+bb7.bb9_crit_edge: ; preds = %bb7
+ br label %bb9
+
+bb9: ; preds = %bb7.preheader, %bb7.bb9_crit_edge
+ br label %bb10
+
+bb10: ; preds = %bb9
+ %indvar.next11 = add i32 %y.08, 1 ; [#uses=2]
+ %exitcond12 = icmp ne i32 %indvar.next11, %x ; [#uses=1]
+ br i1 %exitcond12, label %bb7.preheader, label %bb10.bb18.loopexit_crit_edge
+
+bb10.bb18.loopexit_crit_edge: ; preds = %bb10
+ br label %bb10.bb18.loopexit_crit_edge.split
+
+bb10.bb18.loopexit_crit_edge.split: ; preds = %bb.nph9, %bb10.bb18.loopexit_crit_edge
+ br label %bb18.loopexit
+
+bb.nph9: ; preds = %bb10.preheader
+ %13 = icmp sgt i32 %w, 0 ; [#uses=1]
+ br i1 %13, label %bb.nph9.split, label %bb10.bb18.loopexit_crit_edge.split
+
+bb.nph9.split: ; preds = %bb.nph9
+ br label %bb7.preheader
+
+bb7.preheader: ; preds = %bb.nph9.split, %bb10
+ %y.08 = phi i32 [ 0, %bb.nph9.split ], [ %indvar.next11, %bb10 ] ; [#uses=3]
+ br i1 true, label %bb.nph7, label %bb9
+
+bb.nph5: ; preds = %bb18.loopexit
+ %14 = sdiv i32 %w, 2 ; [#uses=1]
+ %15 = icmp slt i32 %w, 2 ; [#uses=1]
+ %16 = sdiv i32 %x, 2 ; [#uses=2]
+ br i1 %15, label %bb18.bb20_crit_edge.split, label %bb.nph5.split
+
+bb.nph5.split: ; preds = %bb.nph5
+ %tmp2 = icmp sgt i32 1, %16 ; [#uses=1]
+ %smax3 = select i1 %tmp2, i32 1, i32 %16 ; [#uses=1]
+ br label %bb13
+
+bb13: ; preds = %bb18, %bb.nph5.split
+ %y.14 = phi i32 [ 0, %bb.nph5.split ], [ %indvar.next1, %bb18 ] ; [#uses=4]
+ %17 = mul i32 %14, %y.14 ; [#uses=2]
+ %18 = shl i32 %y.14, 1 ; [#uses=1]
+ %19 = srem i32 %y.14, 2 ; [#uses=1]
+ %20 = add i32 %19, %18 ; [#uses=1]
+ %21 = mul i32 %20, %s ; [#uses=2]
+ br i1 true, label %bb.nph3, label %bb17
+
+bb.nph3: ; preds = %bb13
+ %22 = add i32 %17, %0 ; [#uses=1]
+ %23 = add i32 %17, %.sum2 ; [#uses=1]
+ %24 = sdiv i32 %w, 2 ; [#uses=2]
+ %tmp = icmp sgt i32 1, %24 ; [#uses=1]
+ %smax = select i1 %tmp, i32 1, i32 %24 ; [#uses=1]
+ br label %bb14
+
+bb14: ; preds = %bb15, %bb.nph3
+ %x.12 = phi i32 [ 0, %bb.nph3 ], [ %indvar.next, %bb15 ] ; [#uses=5]
+ %25 = shl i32 %x.12, 2 ; [#uses=1]
+ %26 = add i32 %25, %21 ; [#uses=1]
+ %27 = getelementptr i8* %r, i32 %26 ; [#uses=1]
+ %28 = load i8* %27, align 1 ; [#uses=1]
+ %.sum = add i32 %22, %x.12 ; [#uses=1]
+ %29 = getelementptr i8* %j, i32 %.sum ; [#uses=1]
+ store i8 %28, i8* %29, align 1
+ %30 = shl i32 %x.12, 2 ; [#uses=1]
+ %31 = or i32 %30, 2 ; [#uses=1]
+ %32 = add i32 %31, %21 ; [#uses=1]
+ %33 = getelementptr i8* %r, i32 %32 ; [#uses=1]
+ %34 = load i8* %33, align 1 ; [#uses=1]
+ %.sum6 = add i32 %23, %x.12 ; [#uses=1]
+ %35 = getelementptr i8* %j, i32 %.sum6 ; [#uses=1]
+ store i8 %34, i8* %35, align 1
+ br label %bb15
+
+bb15: ; preds = %bb14
+ %indvar.next = add i32 %x.12, 1 ; [#uses=2]
+ %exitcond = icmp ne i32 %indvar.next, %smax ; [#uses=1]
+ br i1 %exitcond, label %bb14, label %bb15.bb17_crit_edge
+
+bb15.bb17_crit_edge: ; preds = %bb15
+ br label %bb17
+
+bb17: ; preds = %bb15.bb17_crit_edge, %bb13
+ br label %bb18
+
+bb18.loopexit: ; preds = %bb10.bb18.loopexit_crit_edge.split, %bb10.preheader
+ %36 = icmp slt i32 %x, 2 ; [#uses=1]
+ br i1 %36, label %bb20, label %bb.nph5
+
+bb18: ; preds = %bb17
+ %indvar.next1 = add i32 %y.14, 1 ; [#uses=2]
+ %exitcond4 = icmp ne i32 %indvar.next1, %smax3 ; [#uses=1]
+ br i1 %exitcond4, label %bb13, label %bb18.bb20_crit_edge
+
+bb18.bb20_crit_edge: ; preds = %bb18
+ br label %bb18.bb20_crit_edge.split
+
+bb18.bb20_crit_edge.split: ; preds = %bb18.bb20_crit_edge, %bb.nph5
+ br label %bb20
+
+bb20: ; preds = %bb18.bb20_crit_edge.split, %bb18.loopexit
+ switch i32 %d, label %return [
+ i32 3, label %bb22
+ i32 1, label %bb29
+ ]
+
+bb22: ; preds = %bb20
+ %37 = mul i32 %x, %w ; [#uses=1]
+ %38 = sdiv i32 %37, 4 ; [#uses=1]
+ %.sum3 = add i32 %38, %.sum2 ; [#uses=2]
+ %39 = add i32 %x, 15 ; [#uses=1]
+ %40 = and i32 %39, -16 ; [#uses=1]
+ %41 = add i32 %w, 15 ; [#uses=1]
+ %42 = and i32 %41, -16 ; [#uses=1]
+ %43 = mul i32 %40, %s ; [#uses=1]
+ %44 = icmp sgt i32 %x, 0 ; [#uses=1]
+ br i1 %44, label %bb.nph, label %bb26
+
+bb.nph: ; preds = %bb22
+ br label %bb23
+
+bb23: ; preds = %bb24, %bb.nph
+ %y.21 = phi i32 [ 0, %bb.nph ], [ %indvar.next5, %bb24 ] ; [#uses=3]
+ %45 = mul i32 %y.21, %42 ; [#uses=1]
+ %.sum1 = add i32 %45, %43 ; [#uses=1]
+ %46 = getelementptr i8* %r, i32 %.sum1 ; [#uses=1]
+ %47 = mul i32 %y.21, %w ; [#uses=1]
+ %.sum5 = add i32 %47, %.sum3 ; [#uses=1]
+ %48 = getelementptr i8* %j, i32 %.sum5 ; [#uses=1]
+ tail call void @llvm.memcpy.i32(i8* %48, i8* %46, i32 %w, i32 1)
+ br label %bb24
+
+bb24: ; preds = %bb23
+ %indvar.next5 = add i32 %y.21, 1 ; [#uses=2]
+ %exitcond6 = icmp ne i32 %indvar.next5, %x ; [#uses=1]
+ br i1 %exitcond6, label %bb23, label %bb24.bb26_crit_edge
+
+bb24.bb26_crit_edge: ; preds = %bb24
+ br label %bb26
+
+bb26: ; preds = %bb24.bb26_crit_edge, %bb22
+ %49 = mul i32 %x, %w ; [#uses=1]
+ %.sum4 = add i32 %.sum3, %49 ; [#uses=1]
+ %50 = getelementptr i8* %j, i32 %.sum4 ; [#uses=1]
+ %51 = mul i32 %x, %w ; [#uses=1]
+ %52 = sdiv i32 %51, 2 ; [#uses=1]
+ tail call void @llvm.memset.i32(i8* %50, i8 -128, i32 %52, i32 1)
+ ret void
+
+bb29: ; preds = %bb20, %entry
+ %53 = add i32 %w, 15 ; [#uses=1]
+ %54 = and i32 %53, -16 ; [#uses=1]
+ %55 = icmp sgt i32 %x, 0 ; [#uses=1]
+ br i1 %55, label %bb.nph11, label %bb33
+
+bb.nph11: ; preds = %bb29
+ br label %bb30
+
+bb30: ; preds = %bb31, %bb.nph11
+ %y.310 = phi i32 [ 0, %bb.nph11 ], [ %indvar.next13, %bb31 ] ; [#uses=3]
+ %56 = mul i32 %y.310, %54 ; [#uses=1]
+ %57 = getelementptr i8* %r, i32 %56 ; [#uses=1]
+ %58 = mul i32 %y.310, %w ; [#uses=1]
+ %59 = getelementptr i8* %j, i32 %58 ; [#uses=1]
+ tail call void @llvm.memcpy.i32(i8* %59, i8* %57, i32 %w, i32 1)
+ br label %bb31
+
+bb31: ; preds = %bb30
+ %indvar.next13 = add i32 %y.310, 1 ; [#uses=2]
+ %exitcond14 = icmp ne i32 %indvar.next13, %x ; [#uses=1]
+ br i1 %exitcond14, label %bb30, label %bb31.bb33_crit_edge
+
+bb31.bb33_crit_edge: ; preds = %bb31
+ br label %bb33
+
+bb33: ; preds = %bb31.bb33_crit_edge, %bb29
+ %60 = mul i32 %x, %w ; [#uses=1]
+ %61 = getelementptr i8* %j, i32 %60 ; [#uses=1]
+ %62 = mul i32 %x, %w ; [#uses=1]
+ %63 = sdiv i32 %62, 2 ; [#uses=1]
+ tail call void @llvm.memset.i32(i8* %61, i8 -128, i32 %63, i32 1)
+ ret void
+
+return: ; preds = %bb20
+ ret void
+}
+
+declare void @llvm.memcpy.i32(i8*, i8*, i32, i32) nounwind
+
+declare void @llvm.memset.i32(i8*, i8, i32, i32) nounwind
From resistor at mac.com Mon Sep 15 17:00:39 2008
From: resistor at mac.com (Owen Anderson)
Date: Mon, 15 Sep 2008 22:00:39 -0000
Subject: [llvm-commits] [llvm] r56231 -
/llvm/trunk/lib/CodeGen/LiveIntervalAnalysis.cpp
Message-ID: <200809152200.m8FM0dR5015709@zion.cs.uiuc.edu>
Author: resistor
Date: Mon Sep 15 17:00:38 2008
New Revision: 56231
URL: http://llvm.org/viewvc/llvm-project?rev=56231&view=rev
Log:
Live intervals for live-in registers should begin at the beginning of a basic block, not at the first
instruction. Also, their valno's should have an unknown def. This has no effect currently, but was
causing issues when StrongPHIElimination was enabled.
Modified:
llvm/trunk/lib/CodeGen/LiveIntervalAnalysis.cpp
Modified: llvm/trunk/lib/CodeGen/LiveIntervalAnalysis.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/CodeGen/LiveIntervalAnalysis.cpp?rev=56231&r1=56230&r2=56231&view=diff
==============================================================================
--- llvm/trunk/lib/CodeGen/LiveIntervalAnalysis.cpp (original)
+++ llvm/trunk/lib/CodeGen/LiveIntervalAnalysis.cpp Mon Sep 15 17:00:38 2008
@@ -624,7 +624,11 @@
MachineBasicBlock::iterator mi = MBB->begin();
unsigned baseIndex = MIIdx;
unsigned start = baseIndex;
- unsigned end = start;
+ while (baseIndex / InstrSlots::NUM < i2miMap_.size() &&
+ getInstructionFromIndex(baseIndex) == 0)
+ baseIndex += InstrSlots::NUM;
+ unsigned end = baseIndex;
+
while (mi != MBB->end()) {
if (mi->killsRegister(interval.reg, tri_)) {
DOUT << " killed";
@@ -659,7 +663,7 @@
}
}
- LiveRange LR(start, end, interval.getNextValue(start, 0, VNInfoAllocator));
+ LiveRange LR(start, end, interval.getNextValue(~0U, 0, VNInfoAllocator));
interval.addRange(LR);
interval.addKill(LR.valno, end);
DOUT << " +" << LR << '\n';
@@ -676,11 +680,6 @@
// Track the index of the current machine instr.
unsigned MIIndex = 0;
- // Skip over empty initial indices.
- while (MIIndex / InstrSlots::NUM < i2miMap_.size() &&
- getInstructionFromIndex(MIIndex) == 0)
- MIIndex += InstrSlots::NUM;
-
for (MachineFunction::iterator MBBI = mf_->begin(), E = mf_->end();
MBBI != E; ++MBBI) {
MachineBasicBlock *MBB = MBBI;
@@ -699,6 +698,11 @@
true);
}
+ // Skip over empty initial indices.
+ while (MIIndex / InstrSlots::NUM < i2miMap_.size() &&
+ getInstructionFromIndex(MIIndex) == 0)
+ MIIndex += InstrSlots::NUM;
+
for (; MI != miEnd; ++MI) {
DOUT << MIIndex << "\t" << *MI;
From gohman at apple.com Mon Sep 15 17:18:04 2008
From: gohman at apple.com (Dan Gohman)
Date: Mon, 15 Sep 2008 22:18:04 -0000
Subject: [llvm-commits] [llvm] r56232 - in /llvm/trunk:
lib/Analysis/ScalarEvolution.cpp
test/Analysis/ScalarEvolution/avoid-smax-0.ll
test/Analysis/ScalarEvolution/avoid-smax-1.ll
test/Analysis/ScalarEvolution/avoid-smax.ll
Message-ID: <200809152218.m8FMI4oD016186@zion.cs.uiuc.edu>
Author: djg
Date: Mon Sep 15 17:18:04 2008
New Revision: 56232
URL: http://llvm.org/viewvc/llvm-project?rev=56232&view=rev
Log:
Teach ScalarEvolution to consider loop preheaders in the search for
an if statement that guards a loop, to allow indvars to avoid smax
operations in more situations.
Added:
llvm/trunk/test/Analysis/ScalarEvolution/avoid-smax-0.ll
- copied unchanged from r56221, llvm/trunk/test/Analysis/ScalarEvolution/avoid-smax.ll
llvm/trunk/test/Analysis/ScalarEvolution/avoid-smax-1.ll
Removed:
llvm/trunk/test/Analysis/ScalarEvolution/avoid-smax.ll
Modified:
llvm/trunk/lib/Analysis/ScalarEvolution.cpp
Modified: llvm/trunk/lib/Analysis/ScalarEvolution.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Analysis/ScalarEvolution.cpp?rev=56232&r1=56231&r2=56232&view=diff
==============================================================================
--- llvm/trunk/lib/Analysis/ScalarEvolution.cpp (original)
+++ llvm/trunk/lib/Analysis/ScalarEvolution.cpp Mon Sep 15 17:18:04 2008
@@ -1480,6 +1480,12 @@
SCEVHandle HowManyLessThans(SCEV *LHS, SCEV *RHS, const Loop *L,
bool isSigned);
+ /// getPredecessorWithUniqueSuccessorForBB - Return a predecessor of BB
+ /// (which may not be an immediate predecessor) which has exactly one
+ /// successor from which BB is reachable, or null if no such block is
+ /// found.
+ BasicBlock* getPredecessorWithUniqueSuccessorForBB(BasicBlock *BB);
+
/// executesAtLeastOnce - Test whether entry to the loop is protected by
/// a conditional between LHS and RHS.
bool executesAtLeastOnce(const Loop *L, bool isSigned, SCEV *LHS, SCEV *RHS);
@@ -1825,6 +1831,11 @@
case Instruction::Select:
// This could be a smax or umax that was lowered earlier.
// Try to recover it.
+ //
+ // FIXME: This doesn't recognize code like this:
+ // %t = icmp sgt i32 %n, -1
+ // %max = select i1 %t, i32 %n, i32 0
+ //
if (ICmpInst *ICI = dyn_cast(U->getOperand(0))) {
Value *LHS = ICI->getOperand(0);
Value *RHS = ICI->getOperand(1);
@@ -2703,6 +2714,28 @@
return UnknownValue;
}
+/// getPredecessorWithUniqueSuccessorForBB - Return a predecessor of BB
+/// (which may not be an immediate predecessor) which has exactly one
+/// successor from which BB is reachable, or null if no such block is
+/// found.
+///
+BasicBlock *
+ScalarEvolutionsImpl::getPredecessorWithUniqueSuccessorForBB(BasicBlock *BB) {
+ // If the block has a unique predecessor, the predecessor must have
+ // no other successors from which BB is reachable.
+ if (BasicBlock *Pred = BB->getSinglePredecessor())
+ return Pred;
+
+ // A loop's header is defined to be a block that dominates the loop.
+ // If the loop has a preheader, it must be a block that has exactly
+ // one successor that can reach BB. This is slightly more strict
+ // than necessary, but works if critical edges are split.
+ if (Loop *L = LI.getLoopFor(BB))
+ return L->getLoopPreheader();
+
+ return 0;
+}
+
/// executesAtLeastOnce - Test whether entry to the loop is protected by
/// a conditional between LHS and RHS.
bool ScalarEvolutionsImpl::executesAtLeastOnce(const Loop *L, bool isSigned,
@@ -2711,14 +2744,11 @@
BasicBlock *PreheaderDest = L->getHeader();
// Starting at the preheader, climb up the predecessor chain, as long as
- // there are unique predecessors, looking for a conditional branch that
- // protects the loop.
- //
- // This is a conservative apporoximation of a climb of the
- // control-dependence predecessors.
-
- for (; Preheader; PreheaderDest = Preheader,
- Preheader = Preheader->getSinglePredecessor()) {
+ // there are predecessors that can be found that have unique successors
+ // leading to the original header.
+ for (; Preheader;
+ PreheaderDest = Preheader,
+ Preheader = getPredecessorWithUniqueSuccessorForBB(Preheader)) {
BranchInst *LoopEntryPredicate =
dyn_cast(Preheader->getTerminator());
Added: llvm/trunk/test/Analysis/ScalarEvolution/avoid-smax-1.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Analysis/ScalarEvolution/avoid-smax-1.ll?rev=56232&view=auto
==============================================================================
--- llvm/trunk/test/Analysis/ScalarEvolution/avoid-smax-1.ll (added)
+++ llvm/trunk/test/Analysis/ScalarEvolution/avoid-smax-1.ll Mon Sep 15 17:18:04 2008
@@ -0,0 +1,234 @@
+; RUN: llvm-as < %s | opt -indvars | llvm-dis > %t
+; RUN: not grep select %t
+; RUN: grep {icmp ne i32.\* %w } %t
+
+; Indvars should be able to insert a canonical induction variable
+; for the bb6 loop without using a maximum calculation (icmp, select)
+; because it should be able to prove that the comparison is guarded
+; by an appropriate conditional branch.
+
+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 = "i386-apple-darwin9"
+
+define void @foo(i8* %r, i32 %s, i32 %w, i32 %x, i8* %j, i32 %d) nounwind {
+entry:
+ %0 = mul i32 %x, %w ; [#uses=2]
+ %1 = mul i32 %x, %w ; [#uses=1]
+ %2 = sdiv i32 %1, 4 ; [#uses=1]
+ %.sum2 = add i32 %2, %0 ; [#uses=2]
+ %cond = icmp eq i32 %d, 1 ; [#uses=1]
+ br i1 %cond, label %bb29, label %bb10.preheader
+
+bb10.preheader: ; preds = %entry
+ %3 = icmp sgt i32 %x, 0 ; [#uses=1]
+ br i1 %3, label %bb.nph9, label %bb18.loopexit
+
+bb.nph7: ; preds = %bb7.preheader
+ %4 = mul i32 %y.08, %w ; [#uses=1]
+ %5 = mul i32 %y.08, %s ; [#uses=1]
+ %6 = add i32 %5, 1 ; [#uses=1]
+ br label %bb6
+
+bb6: ; preds = %bb7, %bb.nph7
+ %x.06 = phi i32 [ %13, %bb7 ], [ 0, %bb.nph7 ] ; [#uses=3]
+ %7 = add i32 %x.06, %4 ; [#uses=1]
+ %8 = shl i32 %x.06, 1 ; [#uses=1]
+ %9 = add i32 %6, %8 ; [#uses=1]
+ %10 = getelementptr i8* %r, i32 %9 ; [#uses=1]
+ %11 = load i8* %10, align 1 ; [#uses=1]
+ %12 = getelementptr i8* %j, i32 %7 ; [#uses=1]
+ store i8 %11, i8* %12, align 1
+ %13 = add i32 %x.06, 1 ; [#uses=2]
+ br label %bb7
+
+bb7: ; preds = %bb6
+ %14 = icmp slt i32 %13, %w ; [#uses=1]
+ br i1 %14, label %bb6, label %bb7.bb9_crit_edge
+
+bb7.bb9_crit_edge: ; preds = %bb7
+ br label %bb9
+
+bb9: ; preds = %bb7.preheader, %bb7.bb9_crit_edge
+ %15 = add i32 %y.08, 1 ; [#uses=2]
+ br label %bb10
+
+bb10: ; preds = %bb9
+ %16 = icmp slt i32 %15, %x ; [#uses=1]
+ br i1 %16, label %bb7.preheader, label %bb10.bb18.loopexit_crit_edge
+
+bb10.bb18.loopexit_crit_edge: ; preds = %bb10
+ br label %bb10.bb18.loopexit_crit_edge.split
+
+bb10.bb18.loopexit_crit_edge.split: ; preds = %bb.nph9, %bb10.bb18.loopexit_crit_edge
+ br label %bb18.loopexit
+
+bb.nph9: ; preds = %bb10.preheader
+ %17 = icmp sgt i32 %w, 0 ; [#uses=1]
+ br i1 %17, label %bb.nph9.split, label %bb10.bb18.loopexit_crit_edge.split
+
+bb.nph9.split: ; preds = %bb.nph9
+ br label %bb7.preheader
+
+bb7.preheader: ; preds = %bb.nph9.split, %bb10
+ %y.08 = phi i32 [ %15, %bb10 ], [ 0, %bb.nph9.split ] ; [#uses=3]
+ br i1 true, label %bb.nph7, label %bb9
+
+bb.nph5: ; preds = %bb18.loopexit
+ %18 = sdiv i32 %w, 2 ; [#uses=1]
+ %19 = icmp slt i32 %w, 2 ; [#uses=1]
+ %20 = sdiv i32 %x, 2 ; [#uses=1]
+ br i1 %19, label %bb18.bb20_crit_edge.split, label %bb.nph5.split
+
+bb.nph5.split: ; preds = %bb.nph5
+ br label %bb13
+
+bb13: ; preds = %bb18, %bb.nph5.split
+ %y.14 = phi i32 [ %42, %bb18 ], [ 0, %bb.nph5.split ] ; [#uses=4]
+ %21 = mul i32 %18, %y.14 ; [#uses=2]
+ %22 = shl i32 %y.14, 1 ; [#uses=1]
+ %23 = srem i32 %y.14, 2 ; [#uses=1]
+ %24 = add i32 %23, %22 ; [#uses=1]
+ %25 = mul i32 %24, %s ; [#uses=2]
+ br i1 true, label %bb.nph3, label %bb17
+
+bb.nph3: ; preds = %bb13
+ %26 = add i32 %21, %0 ; [#uses=1]
+ %27 = add i32 %21, %.sum2 ; [#uses=1]
+ %28 = sdiv i32 %w, 2 ; [#uses=1]
+ br label %bb14
+
+bb14: ; preds = %bb15, %bb.nph3
+ %x.12 = phi i32 [ %40, %bb15 ], [ 0, %bb.nph3 ] ; [#uses=5]
+ %29 = shl i32 %x.12, 2 ; [#uses=1]
+ %30 = add i32 %29, %25 ; [#uses=1]
+ %31 = getelementptr i8* %r, i32 %30 ; [#uses=1]
+ %32 = load i8* %31, align 1 ; [#uses=1]
+ %.sum = add i32 %26, %x.12 ; [#uses=1]
+ %33 = getelementptr i8* %j, i32 %.sum ; [#uses=1]
+ store i8 %32, i8* %33, align 1
+ %34 = shl i32 %x.12, 2 ; [#uses=1]
+ %35 = or i32 %34, 2 ; [#uses=1]
+ %36 = add i32 %35, %25 ; [#uses=1]
+ %37 = getelementptr i8* %r, i32 %36 ; [#uses=1]
+ %38 = load i8* %37, align 1 ; [#uses=1]
+ %.sum6 = add i32 %27, %x.12 ; [#uses=1]
+ %39 = getelementptr i8* %j, i32 %.sum6 ; [#uses=1]
+ store i8 %38, i8* %39, align 1
+ %40 = add i32 %x.12, 1 ; [#uses=2]
+ br label %bb15
+
+bb15: ; preds = %bb14
+ %41 = icmp sgt i32 %28, %40 ; [#uses=1]
+ br i1 %41, label %bb14, label %bb15.bb17_crit_edge
+
+bb15.bb17_crit_edge: ; preds = %bb15
+ br label %bb17
+
+bb17: ; preds = %bb15.bb17_crit_edge, %bb13
+ %42 = add i32 %y.14, 1 ; [#uses=2]
+ br label %bb18
+
+bb18.loopexit: ; preds = %bb10.bb18.loopexit_crit_edge.split, %bb10.preheader
+ %43 = icmp slt i32 %x, 2 ; [#uses=1]
+ br i1 %43, label %bb20, label %bb.nph5
+
+bb18: ; preds = %bb17
+ %44 = icmp sgt i32 %20, %42 ; [#uses=1]
+ br i1 %44, label %bb13, label %bb18.bb20_crit_edge
+
+bb18.bb20_crit_edge: ; preds = %bb18
+ br label %bb18.bb20_crit_edge.split
+
+bb18.bb20_crit_edge.split: ; preds = %bb18.bb20_crit_edge, %bb.nph5
+ br label %bb20
+
+bb20: ; preds = %bb18.bb20_crit_edge.split, %bb18.loopexit
+ switch i32 %d, label %return [
+ i32 3, label %bb22
+ i32 1, label %bb29
+ ]
+
+bb22: ; preds = %bb20
+ %45 = mul i32 %x, %w ; [#uses=1]
+ %46 = sdiv i32 %45, 4 ; [#uses=1]
+ %.sum3 = add i32 %46, %.sum2 ; [#uses=2]
+ %47 = add i32 %x, 15 ; [#uses=1]
+ %48 = and i32 %47, -16 ; [#uses=1]
+ %49 = add i32 %w, 15 ; [#uses=1]
+ %50 = and i32 %49, -16 ; [#uses=1]
+ %51 = mul i32 %48, %s ; [#uses=1]
+ %52 = icmp sgt i32 %x, 0 ; [#uses=1]
+ br i1 %52, label %bb.nph, label %bb26
+
+bb.nph: ; preds = %bb22
+ br label %bb23
+
+bb23: ; preds = %bb24, %bb.nph
+ %y.21 = phi i32 [ %57, %bb24 ], [ 0, %bb.nph ] ; [#uses=3]
+ %53 = mul i32 %y.21, %50 ; [#uses=1]
+ %.sum1 = add i32 %53, %51 ; [#uses=1]
+ %54 = getelementptr i8* %r, i32 %.sum1 ; [#uses=1]
+ %55 = mul i32 %y.21, %w ; [#uses=1]
+ %.sum5 = add i32 %55, %.sum3 ; [#uses=1]
+ %56 = getelementptr i8* %j, i32 %.sum5 ; [#uses=1]
+ tail call void @llvm.memcpy.i32(i8* %56, i8* %54, i32 %w, i32 1)
+ %57 = add i32 %y.21, 1 ; [#uses=2]
+ br label %bb24
+
+bb24: ; preds = %bb23
+ %58 = icmp slt i32 %57, %x ; [#uses=1]
+ br i1 %58, label %bb23, label %bb24.bb26_crit_edge
+
+bb24.bb26_crit_edge: ; preds = %bb24
+ br label %bb26
+
+bb26: ; preds = %bb24.bb26_crit_edge, %bb22
+ %59 = mul i32 %x, %w ; [#uses=1]
+ %.sum4 = add i32 %.sum3, %59 ;