Go to the documentation of this file.
14 static Target TheX86_32Target;
15 return TheX86_32Target;
18 static Target TheX86_64Target;
19 return TheX86_64Target;
Move duplicate certain instructions close to their use
This might compile to this xmm1 xorps xmm0 movss xmm0 ret Now consider if the code caused xmm1 to get spilled This might produce this xmm1 movaps xmm0 movaps xmm1 movss xmm0 ret since the reload is only used by these we could fold it into the producing something like xmm1 movaps xmm0 ret saving two instructions The basic idea is that a reload from a spill if only one byte chunk is bring in zeros the one element instead of elements This can be used to simplify a variety of shuffle where the elements are fixed zeros This code generates ugly probably due to costs being off or< 4 x float > eax movaps(%eax)
We currently generate a but we really shouldn eax ecx xorl edx divl ecx eax divl ecx movl eax ret A similar code sequence works for division We currently compile i32 v2 eax eax jo LBB1_2 atomic and others It is also currently not done for read modify write instructions It is also current not done if the OF or CF flags are needed The shift operators have the complication that when the shift count is EFLAGS is not set
amdgpu propagate attributes Late propagate attributes from kernels to functions
zeros_impl< sizeof(T)> zeros(const T &)
Generic address nodes are lowered to some combination of target independent and machine specific ABI
should just be implemented with a CLZ instruction Since there are other e that share this it would be best to implement this in a target independent as zero is the default value for the binary encoder e add r0 add r5 Register operands should be distinct That is
compiles conv shl5 shl ret i32 or10 it would be better as
Instead of the following for memset char edx edx edx It might be better to generate eax movl edx movl edx movw edx when we can spare a register It reduces code size Evaluate what the best way to codegen sdiv C is For we currently get ret i32 Y eax movl ecx ecx ecx addl ecx
RUN< i32 >< i1 > br i1 label then
should just be implemented with a CLZ instruction Since there are other targets
the custom lowered code happens to be right
into xmm2 addss xmm2 xmm1 xmm3 addss xmm3 movaps xmm0 unpcklps xmm0 ret seems silly when it could just be one addps Expand libm rounding functions main should enable SSE DAZ mode and other fast SSE modes Think about doing i64 math in SSE regs on x86 This testcase should have no SSE instructions in and only one load from a constant double ret double C the select is being which prevents the dag combiner from turning load CPI2 load(select CPI1, CPI2)' The pattern isel got this one right. Lower memcpy/memset to a series of SSE 128 bit move instructions when it 's feasible. Codegen
This might compile to this xmm1 xorps xmm0 movss xmm0 ret Now consider if the code caused xmm1 to get spilled This might produce this xmm1 movaps xmm0 movaps xmm1 movss xmm0 ret since the reload is only used by these we could fold it into the producing something like xmm1 movaps xmm0 ret saving two instructions The basic idea is that a reload from a spill if only one byte chunk is bring in zeros the one element instead of elements This can be used to simplify a variety of shuffle where the elements are fixed zeros This code generates ugly probably due to costs being off or< 4 x float > eax xmm0 pxor xmm1 movaps xmm2 xmm2 xmm0 movaps eax ret Would it be better to generate
we should consider alternate ways to model stack dependencies Lots of things could be done in WebAssemblyTargetTransformInfo cpp there are numerous optimization related hooks that can be overridden in WebAssemblyTargetLowering Instead of the OptimizeReturned which should consider preserving the returned attribute through to MachineInstrs and extending the MemIntrinsicResults pass to do this optimization on calls too That would also let the WebAssemblyPeephole pass clean up dead defs for such as it does for stores Consider implementing and or getMachineCombinerPatterns Find a clean way to fix the problem which leads to the Shrink Wrapping pass being run after the WebAssembly PEI pass When setting multiple variables to the same constant
This is an optimization pass for GlobalISel generic memory operations.
Add support for conditional and other related patterns Instead eax eax je LBB16_2 eax edi eax movl eax
This might compile to this xmm1 xorps xmm0 movss xmm0 ret Now consider if the code caused xmm1 to get spilled This might produce this xmm1 movaps c2(%esp) ... xorps %xmm0
into xmm2 addss xmm2 xmm1 xmm3 addss xmm3 movaps xmm0 unpcklps xmm0 ret seems silly when it could just be one addps Expand libm rounding functions main should enable SSE DAZ mode and other fast SSE modes Think about doing i64 math in SSE regs on x86 This testcase should have no SSE instructions in it
StringRef extension(StringRef path, Style style=Style::native)
Get extension.
< float * > store float float *tmp5 ret void Compiles rax shrq
Add support for conditional and other related patterns Instead eax cmpl
We currently generate a but we really shouldn eax ecx xorl edx divl ecx eax divl ecx movl eax ret A similar code sequence works for division We currently compile i32 v2 eax eax jo LBB1_2 atomic and others It is also currently not done for read modify write instructions It is also current not done if the OF or CF flags are needed The shift operators have the complication that when the shift count is zero
We have fiadd patterns now but the followings have the same cost and complexity We need a way to specify the later is more profitable def def The FP stackifier should handle simple permutates to reduce number of shuffle e g ceil
@ Custom
The target wants to do something special with this combination of operand and type.
The object format emitted by the WebAssembly backed is documented see the home and packaging for producing WebAssembly applications that can run in browsers and other environments wasi sdk provides a more minimal C C SDK based on llvm and a libc based on for producing WebAssemmbly applictions that use the WASI ABI Rust provides WebAssembly support integrated into Cargo There are two main which provides a relatively minimal environment that has an emphasis on being native wasm32 unknown which uses Emscripten internally and provides standard C C filesystem GL and SDL bindings For more see
Replace within non kernel function use of LDS with pointer
This currently compiles esp xmm0 movsd esp eax eax esp ret We should use not the dag combiner This is because dagcombine2 needs to be able to see through the X86ISD::Wrapper which DAGCombine can t really do The code for turning x load into a single vector load is target independent and should be moved to the dag combiner The code for turning 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
We have fiadd patterns now but the followings have the same cost and complexity We need a way to specify the later is more profitable def RFP
Target & getTheX86_64Target()
_bar mov r0 mov r1 fldd LCPI1_0 fmrrd d0 bl _foo fmdrr r5 fmsr r0 fsitod s2 faddd d0 fmrrd d0 ldmfd r0 mov r1 the copys to callee save registers and the fact they are only being used by the fmdrr instruction It would have been better had the fmdrr been scheduled before the call and place the result in a callee save DPR register The two mov ops would not have been necessary Calling convention related stuff
return AArch64::GPR64RegClass contains(Reg)
Target - Wrapper for Target specific information.
into xmm2 addss xmm2 xmm1 xmm3 addss xmm3 movaps xmm0 unpcklps xmm0 ret seems silly when it could just be one addps Expand libm rounding functions main should enable SSE DAZ mode and other fast SSE modes Think about doing i64 math in SSE regs on x86 This testcase should have no SSE instructions in and only one load from a constant double
This might compile to this xmm1 xorps xmm0 movss xmm0 ret Now consider if the code caused xmm1 to get spilled This might produce this xmm1 movaps xmm0 movaps xmm1 movss xmm0 ret since the reload is only used by these we could fold it into the uses
SSE has instructions for doing operations on complex we should pattern match them For this should turn into a horizontal add
amdgpu aa AMDGPU Address space based Alias Analysis Wrapper
Since we know that Vector is byte aligned and we know the element offset of we should change the load into a lve *x instead of doing a load store lve *x sequence Implement passing vectors by value into calls and receiving them as arguments GCC apparently tries to codegen
This currently compiles esp xmm0 movsd esp eax eax addl
The object format emitted by the WebAssembly backed is documented see the home and packaging for producing WebAssembly applications that can run in browsers and other environments wasi sdk provides a more minimal C C SDK based on llvm and a libc based on for producing WebAssemmbly applictions that use the WASI ABI Rust provides WebAssembly support integrated into Cargo There are two main which provides a relatively minimal environment that has an emphasis on being native wasm32 unknown which uses Emscripten internally and provides standard C C filesystem GL and SDL bindings For more information
AMD64 Optimization Manual has some nice information about optimizing integer multiplication by a constant How much of it applies to Intel s X86 implementation There are definite trade offs to xmm0 cvttss2siq rdx jb L3 subss xmm0 rax cvttss2siq rdx xorq rdx rax ret instead of xmm1 cvttss2siq rcx movaps xmm2 subss xmm2 cvttss2siq rax rdx xorq rax ucomiss xmm0 cmovb rax ret Seems like the jb branch has high likelihood of being taken It would have saved a few instructions It s not possible to reference CH
This might compile to this xmm1 xorps xmm0 movss xmm0 ret Now consider if the code caused xmm1 to get spilled This might produce this xmm1 movaps xmm0 movaps xmm1 movss xmm0 ret since the reload is only used by these we could fold it into the producing something like xmm1 movaps xmm0 ret saving two instructions The basic idea is that a reload from a spill if only one byte chunk is bring in zeros the one element instead of elements This can be used to simplify a variety of shuffle where the elements are fixed zeros This code generates ugly probably due to costs being off or< 4 x float > eax xmm0 pxor xmm1 movaps xmm2 xmm2 xmm0 movaps eax ret Would it be better to ecx xmm0 xor eax xmm0 xmm0 movaps ecx ret Some useful information in the Apple Altivec SSE Migration or Various SSE compare translations Add hooks to commute some CMPP operations Apply the same transformation that merged four float into a single bit load to loads from constant pool Floating point max min are commutable when enable unsafe fp path is specified We should turn int_x86_sse_max_ss and X86ISD::FMIN etc into other nodes which are selected to max min instructions that are marked commutable We should materialize vector constants like all ones and signbit with code like
into eax xorps xmm0 xmm0 eax xmm0 eax xmm0 ret esp eax movdqa xmm0 xmm0 esp const ret align it should be movdqa xmm0 xmm0 We should transform a shuffle of two vectors of constants into a single vector of constants insertelement of a constant into a vector of constants should also result in a vector of constants e g VecISelBug ll We compiled it to something globl _t xmm0 movhps xmm0 movss LCPI1_1
@ FP_TO_SINT
FP_TO_[US]INT - Convert a floating point value to a signed or unsigned integer.
const_iterator end(StringRef path)
Get end iterator over path.
PowerPC MI Peephole Optimization
the custom lowered code happens to be but we shouldn t have to custom lower anything This is probably related to< 2 x i64 > ops being so bad LLVM currently generates stack realignment when it is not necessary needed The problem is that we need to know about stack alignment too before RA runs At that point we don t know
into eax xorps xmm0 xmm0 eax xmm0 eax xmm0 ret esp eax movdqa xmm0 xmm0 esp const ret align it should be movdqa xmm0 pinsrb
hexagon Hexagon specific predictive commoning for HVX vectors
to esp esp setne al movzbw ax esp setg cl movzbw cx cmove cx cl jne LBB1_2 esp ret(also really horrible code on ppc). This is due to the expand code for 64-bit compares. GCC produces multiple branches
compiles ldr LCPI1_0 ldr ldr mov lsr tst moveq r1 ldr LCPI1_1 and r0 bx lr It would be better to do something like to fold the shift into the conditional ldr LCPI1_0 ldr ldr tst movne lsr ldr LCPI1_1 and r0 bx lr it saves an instruction and a register It might be profitable to cse MOVi16 if there are lots of bit immediates with the same bottom half Robert Muth started working on an alternate jump table implementation that does not put the tables in line in the text This is more like the llvm default jump table implementation This might be useful sometime Several revisions of patches are on the mailing beginning at
add sub stmia L5 ldr r0 bl L_printf $stub Instead of a and a wouldn t it be better to do three moves *Return an aggregate type is even worse
This currently compiles esp xmm0 movsd esp eax eax esp ret We should use movmskp
Expected< ExpressionValue > max(const ExpressionValue &Lhs, const ExpressionValue &Rhs)
into eax xorps xmm0 xmm0 eax xmm0 movl
we should consider alternate ways to model stack dependencies Lots of things could be done in WebAssemblyTargetTransformInfo cpp there are numerous optimization related hooks that can be overridden in WebAssemblyTargetLowering Instead of the OptimizeReturned which should consider preserving the returned attribute through to MachineInstrs and extending the MemIntrinsicResults pass to do this optimization on calls too That would also let the WebAssemblyPeephole pass clean up dead defs for such as it does for stores Consider implementing and or getMachineCombinerPatterns Find a clean way to fix the problem which leads to the Shrink Wrapping pass being run after the WebAssembly PEI pass When setting multiple variables to the same we currently get code like const It could be done with a smaller encoding like local tee $pop5 local $pop6 WebAssembly registers are implicitly initialized to zero Explicit zeroing is therefore often redundant and could be optimized away Small indices may use smaller encodings than large indices WebAssemblyRegColoring and or WebAssemblyRegRenumbering should sort registers according to their usage frequency to maximize the usage of smaller encodings Many cases of irreducible control flow could be transformed more optimally than via the transform in WebAssemblyFixIrreducibleControlFlow cpp It may also be worthwhile to do transforms before register particularly when duplicating to allow register coloring to be aware of the duplication WebAssemblyRegStackify could use AliasAnalysis to reorder loads and stores more aggressively WebAssemblyRegStackify is currently a greedy algorithm This means that
alloca< 16 x float >, align 16 %tmp2=alloca< 16 x float >, align 16 store< 16 x float > %A,< 16 x float > *%tmp %s=bitcast< 16 x float > *%tmp to i8 *%s2=bitcast< 16 x float > *%tmp2 to i8 *call void @llvm.memcpy.i64(i8 *%s, i8 *%s2, i64 64, i32 16) %R=load< 16 x float > *%tmp2 ret< 16 x float > %R } declare void @llvm.memcpy.i64(i8 *nocapture, i8 *nocapture, i64, i32) nounwind which compiles to:_foo:subl $140, %esp movaps %xmm3, 112(%esp) movaps %xmm2, 96(%esp) movaps %xmm1, 80(%esp) movaps %xmm0, 64(%esp) movl 60(%esp), %eax movl %eax, 124(%esp) movl 56(%esp), %eax movl %eax, 120(%esp) movl 52(%esp), %eax< many many more 32-bit copies > movaps(%esp), %xmm0 movaps 16(%esp), %xmm1 movaps 32(%esp), %xmm2 movaps 48(%esp), %xmm3 addl $140, %esp ret On Nehalem, it may even be cheaper to just use movups when unaligned than to fall back to lower-granularity chunks. Implement processor-specific optimizations for parity with GCC on these processors. GCC does two optimizations:1. ix86_pad_returns inserts a noop before ret instructions if immediately preceded by a conditional branch or is the target of a jump. 2. ix86_avoid_jump_misspredicts inserts noops in cases where a 16-byte block of code contains more than 3 branches. The first one is done for all AMDs, Core2, and "Generic" The second one is done for:Atom, Pentium Pro, all AMDs, Pentium 4, Nocona, Core 2, and "Generic" Testcase:int x(int a) { return(a &0xf0)> >4 tmp
then ret i32 result Tail recursion elimination should handle
gets compiled into this on rsp movaps rsp movaps rsp movaps rsp movaps rsp movaps rsp movaps rsp movaps rsp movaps rsp movq rsp movq rsp movq rsp movq rsp movq rsp rax movq rsp rax movq rsp rsp rsp eax eax jbe LBB1_3 rcx rax movq rsp eax rsp ret ecx eax rcx movl rsp jmp LBB1_2 gcc generates
We currently generate a but we really shouldn eax ecx xorl edx divl ecx eax divl ecx movl eax ret A similar code sequence works for division We currently compile i32 v2 eax eax jo LBB1_2 and
Note that only the low bits of effective_addr2 are used On bit we don t eliminate the computation of the top half of effective_addr2 because we don t have whole function selection dags On x86
include(LLVM-Build) add_subdirectory(IR) add_subdirectory(FuzzMutate) add_subdirectory(FileCheck) add_subdirectory(InterfaceStub) add_subdirectory(IRPrinter) add_subdirectory(IRReader) add_subdirectory(CodeGen) add_subdirectory(BinaryFormat) add_subdirectory(Bitcode) add_subdirectory(Bitstream) add_subdirectory(DWARFLinker) add_subdirectory(Extensions) add_subdirectory(Frontend) add_subdirectory(Transforms) add_subdirectory(Linker) add_subdirectory(Analysis) add_subdirectory(LTO) add_subdirectory(MC) add_subdirectory(MCA) add_subdirectory(ObjCopy) add_subdirectory(Object) add_subdirectory(ObjectYAML) add_subdirectory(Option) add_subdirectory(Remarks) add_subdirectory(Debuginfod) add_subdirectory(DebugInfo) add_subdirectory(DWP) add_subdirectory(ExecutionEngine) add_subdirectory(Target) add_subdirectory(AsmParser) add_subdirectory(LineEditor) add_subdirectory(ProfileData) add_subdirectory(Passes) add_subdirectory(TextAPI) add_subdirectory(ToolDrivers) add_subdirectory(XRay) if(LLVM_INCLUDE_TESTS) add_subdirectory(Testing) endif() add_subdirectory(WindowsDriver) add_subdirectory(WindowsManifest) set(LLVMCONFIGLIBRARYDEPENDENCIESINC "$
We have fiadd patterns now but the followings have the same cost and complexity We need a way to specify the later is more profitable def FpADD32m
The following code is currently eax eax ecx jb LBB1_2 eax movzbl eax ret eax ret We could change the eax into movzwl(%esp)
Predicate any(Predicate P0, Predicate P1)
True iff P0 or P1 are true.
This currently compiles esp xmm0 movsd esp eax eax esp ret We should use not the dag combiner This is because dagcombine2 needs to be able to see through the X86ISD::Wrapper which DAGCombine can t really do The code for turning x load into a single vector load is target independent and should be moved to the dag combiner The code for turning 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 where P can be anything The alignment inference code cannot handle loads from globals in static non mode because it doesn t look through the extra dyld stub load If you try vec_align ll without relocation you ll see what I mean We should lower store(fneg(load p), q) into an integer load+xor+store
Clang compiles this i1 i64 store i64 i64 store i64 i64 store i64 i64 store i64 align Which gets codegen d xmm0 movaps rbp movaps rbp movaps rbp movaps rbp rbp rbp rbp rbp It would be better to have movq s of instead of the movaps s http
Should compile to something like
we should consider alternate ways to model stack dependencies Lots of things could be done in WebAssemblyTargetTransformInfo cpp there are numerous optimization related hooks that can be overridden in WebAssemblyTargetLowering Instead of the OptimizeReturned which should consider preserving the returned attribute through to MachineInstrs and extending the MemIntrinsicResults pass to do this optimization on calls too That would also let the WebAssemblyPeephole pass clean up dead defs for such as it does for stores Consider implementing and or getMachineCombinerPatterns Find a clean way to fix the problem which leads to the Shrink Wrapping pass being run after the WebAssembly PEI pass When setting multiple variables to the same we currently get code like const It could be done with a smaller encoding like local tee $pop5 local $pop6 WebAssembly registers are implicitly initialized to zero Explicit zeroing is therefore often redundant and could be optimized away Small indices may use smaller encodings than large indices WebAssemblyRegColoring and or WebAssemblyRegRenumbering should sort registers according to their usage frequency to maximize the usage of smaller encodings Many cases of irreducible control flow could be transformed more optimally than via the transform in WebAssemblyFixIrreducibleControlFlow cpp It may also be worthwhile to do transforms before register particularly when duplicating to allow register coloring to be aware of the duplication WebAssemblyRegStackify could use AliasAnalysis to reorder loads and stores more aggressively WebAssemblyRegStackify is currently a greedy algorithm This means for a binary however wasm doesn t actually require this WebAssemblyRegStackify could be extended
=0.0 ? 0.0 :(a > 0.0 ? 1.0 :-1.0) a
We have fiadd patterns now but the followings have the same cost and complexity We need a way to specify the later is more profitable def def The FP stackifier should handle simple permutates to reduce number of shuffle e g floor
We have fiadd patterns now but the followings have the same cost and complexity We need a way to specify the later is more profitable def def i32mem
should just be implemented with a CLZ instruction Since there are other e that share this it would be best to implement this in a target independent way
It looks like we only need to define PPCfmarto for these because according to these instructions perform RTO on fma s result
In x86 we generate this spiffy xmm0 xmm0 ret in x86 we generate this which could be xmm1 movss xmm1 xmm0 ret In sse4 we could use insertps to make both better Here s another testcase that could use x3
< i1 > br i1 label label bb bb
We have fiadd patterns now but the followings have the same cost and complexity We need a way to specify the later is more profitable def def The FP stackifier should handle simple permutates to reduce number of shuffle e g turning
AMD64 Optimization Manual has some nice information about optimizing integer multiplication by a constant How much of it applies to Intel s X86 implementation There are definite trade offs to xmm0 cvttss2siq rdx jb L3 subss xmm0 rax cvttss2siq rdx xorq rdx rax ret instead of xmm1 cvttss2siq rcx movaps xmm2 subss xmm2 cvttss2siq rax rdx xorq rax ucomiss xmm0 cmovb rcx
into xmm2 addss xmm2 xmm1 xmm3 addss xmm3 movaps xmm0 unpcklps xmm0 ret seems silly when it could just be one addps Expand libm rounding functions main should enable SSE DAZ mode and other fast SSE modes Think about doing i64 math in SSE regs on x86 This testcase should have no SSE instructions in and only one load from a constant double ret double C the select is being which prevents the dag combiner from turning select(load CPI1)
bool match(Val *V, const Pattern &P)
In x86 we generate this spiffy xmm0 xmm0 ret in x86 we generate this which could be xmm1 movss xmm1 xmm0 ret In sse4 we could use insertps to make both better Here s another testcase that could use insertps[mem]
This might compile to this code
@ ABS
ABS - Determine the unsigned absolute value of a signed integer value of the same bitwidth.
Common register allocation spilling lr str ldr sxth r3 ldr mla r4 can lr mov lr str ldr sxth r3 mla r4 and then merge mul and lr str ldr sxth r3 mla r4 It also increase the likelihood the store may become dead bb27 Successors according to LLVM ID Predecessors according to mbb< bb27, 0x8b0a7c0 > Note ADDri is not a two address instruction its result reg1037 is an operand of the PHI node in bb76 and its operand reg1039 is the result of the PHI node We should treat it as a two address code and make sure the ADDri is scheduled after any node that reads reg1039 Use info(i.e. register scavenger) to assign it a free register to allow reuse the collector could move the objects and invalidate the derived pointer This is bad enough in the first but safe points can crop up unpredictably **array_addr i32 n y store obj obj **nth_el If the i64 division is lowered to a then a safe point array and nth_el no longer point into the correct object The fix for this is to copy address calculations so that dependent pointers are never live across safe point boundaries But the loads cannot be copied like this if there was an intervening so may be hard to get right Only a concurrent mutator can trigger a collection at the libcall safe point So single threaded programs do not have this even with a copying collector LLVM optimizations would probably undo a front end s careful work The ocaml frametable structure supports liveness information It would be good to support it The FIXME in ComputeCommonTailLength in BranchFolding cpp needs to be revisited The check is there to work around a misuse of directives in assembly It would be good to detect collector target compatibility instead of silently doing the wrong thing It would be really nice to be able to write patterns in td files for which would eliminate a bunch of predicates on them(e.g. no side effects). Once this is in place
Itanium Name Demangler i e convert the string _Z1fv into f()". You can also use the CRTP base ManglingParser to perform some simple analysis on the mangled name
This might compile to this xmm1 xorps xmm0 movss xmm0 ret Now consider if the code caused xmm1 to get spilled This might produce this xmm1 movaps xmm0 movaps xmm1 movss xmm0 ret since the reload is only used by these we could fold it into the producing something like xmm1 movaps xmm0 ret saving two instructions The basic idea is that a reload from a spill if only one byte chunk is bring in zeros the one element instead of elements This can be used to simplify a variety of shuffle where the elements are fixed zeros This code generates ugly probably due to costs being off or< 4 x float > * P2
(vector float) vec_cmpeq(*A, *B) C
This might compile to this xmm1 xorps xmm0 movss xmm0 ret Now consider if the code caused xmm1 to get spilled This might produce this xmm1 movaps xmm0 movaps xmm1 movss xmm0 ret since the reload is only used by these we could fold it into the producing something like xmm1 movaps xmm0 ret saving two instructions The basic idea is that a reload from a spill if only one byte chunk is bring in zeros the one element instead of elements This can be used to simplify a variety of shuffle where the elements are fixed zeros This code generates ugly probably due to costs being off or< 4 x float > eax xmm0 pxor xmm1 movaps xmm2 xmm2 xmm0 movaps eax ret Would it be better to ecx xmm0 xor eax xmm0 xmm0 movaps ecx ret Some useful information in the Apple Altivec SSE Migration andnot
Clang compiles this i1 i64 store i64 i64 store i64 i64 store i64 i64 store i64 align Which gets codegen d xmm0 movaps rbp movaps rbp movaps rbp movaps rbp rbp rbp rbp rbp It would be better to have movq s of instead of the movaps s LLVM produces ret int
AMD64 Optimization Manual has some nice information about optimizing integer multiplication by a constant How much of it applies to Intel s X86 implementation There are definite trade offs to xmm0 cvttss2siq rdx jb L3 subss xmm0 rax cvttss2siq rdx xorq rax
static GCMetadataPrinterRegistry::Add< OcamlGCMetadataPrinter > Y("ocaml", "ocaml 3.10-compatible collector")
http eax xorl edx cl sete al setne dl sall eax sall edx But that requires good bit subreg support this might be better It s an extra but it s one instruction shorter
bitcast float %x to i32 %s=and i32 %t, 2147483647 %d=bitcast i32 %s to float ret float %d } declare float @fabsf(float %n) define float @bar(float %x) nounwind { %d=call float @fabsf(float %x) ret float %d } This IR(from PR6194):target datalayout="e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128" target triple="x86_64-apple-darwin10.0.0" %0=type { double, double } %struct.float3=type { float, float, float } define void @test(%0, %struct.float3 *nocapture %res) nounwind noinline ssp { entry:%tmp18=extractvalue %0 %0, 0 t
the resulting code requires compare and branches when and if the revised code is with conditional branches instead of More there is a byte word extend before each where there should be only and the condition codes are not remembered when the same two values are compared twice More LSR enhancements i8 and i32 load store addressing modes are identical int b
We have fiadd patterns now but the followings have the same cost and complexity We need a way to specify the later is more profitable def f32mem
This requires reassociating to forms of expressions that are already something that reassoc doesn t think about yet These two functions should generate the same code on big endian int * l
AMD64 Optimization Manual has some nice information about optimizing integer multiplication by a constant How much of it applies to Intel s X86 implementation There are definite trade offs to xmm0 cvttss2siq rdx jb L3 subss xmm0 rax cvttss2siq rdx xorq rdx rax ret instead of xmm1 cvttss2siq rcx movaps xmm2 subss xmm2 cvttss2siq rax rdx xorq rax ucomiss xmm0 cmovb rax ret Seems like the jb branch has high likelihood of being taken It would have saved a few instructions It s not possible to reference and DH registers in an instruction requiring REX prefix divb and mulb both produce results in AH If isel emits a CopyFromReg which gets turned into a movb and that can be allocated a r8b r15b To get around isel emits a CopyFromReg from AX and then right shift it down by and truncate it It s not pretty but it works We need some register allocation magic to make the hack go which would often require a callee saved register Callees usually need to keep this value live for most of their body so it doesn t add a significant burden on them We currently implement this in codegen
the custom lowered code happens to be but we shouldn t have to custom lower anything This is probably related to< 2 x i64 > ops being so bad LLVM currently generates stack realignment when it is not necessary needed The problem is that we need to know about stack alignment too before RA runs At that point we don t whether there will be vector or not Stack realignment logic is overly conservative here
The object format emitted by the WebAssembly backed is documented in
Common register allocation spilling lr str ldr sxth r3 ldr mla r4 can be
into xmm2 addss xmm2 xmm1 xmm3 addss xmm3 movaps xmm0 unpcklps xmm0 ret seems silly when it could just be one addps Expand libm rounding functions main should enable SSE DAZ mode and other fast SSE modes Think about doing i64 math in SSE regs on x86 This testcase should have no SSE instructions in and only one load from a constant pool
@ SINT_TO_FP
[SU]INT_TO_FP - These operators convert integers (whose interpreted sign depends on the first letter)...
_foo edi jbe LBB1_3 LBB1_1
To do *Keep the address of the constant pool in a register instead of forming its address all of the time *We can fold small constant offsets into the hi lo references to constant pool addresses as well *When in V9 register allocate icc *[0-3] Add support for isel ing UMUL_LOHI instead of marking it as Expand *Emit the Branch on Integer Register with Prediction instructions It s not clear how to write a pattern for this int o6 subcc l0 bne LBBt1_2 ! F nop l0 st g0 retl nop should be replaced with a brz in V9 mode *Same as above
We have fiadd patterns now but the followings have the same cost and complexity We need a way to specify the later is more profitable def def FpIADD32m
Statically lint checks LLVM IR
*Add support for compiling functions in both ARM and Thumb then taking the smallest *Add support for compiling individual basic blocks in thumb when in a larger ARM function This can be used for presumed cold like paths to EH handling etc *Thumb doesn t have normal pre post increment addressing modes
the resulting code requires compare and branches when and if the revised code is with conditional branches instead of More there is a byte word extend before each where there should be only and the condition codes are not remembered when the same two values are compared twice More LSR enhancements i8 and i32 load store addressing modes are identical int int c
Instead we get xmm1 addss xmm1 pshufd
@ CopyFromReg
CopyFromReg - This node indicates that the input value is a virtual or physical register that is defi...
Common register allocation spilling lr str ldr sxth r3 ldr mla r4 can lr mov lr str ldr sxth r3 mla r4 and then merge mul and lr str ldr sxth r3 mla r4 It also increase the likelihood the store may become dead bb27 Successors according to LLVM ID Predecessors according to mbb< bb27, 0x8b0a7c0 > Note ADDri is not a two address instruction its result reg1037 is an operand of the PHI node in bb76 and its operand reg1039 is the result of the PHI node We should treat it as a two address code and make sure the ADDri is scheduled after any node that reads reg1039 Use info(i.e. register scavenger) to assign it a free register to allow reuse the collector could move the objects and invalidate the derived pointer This is bad enough in the first place
return _mm_set_ps(0.0, 0.0, 0.0, b)
static sys::TimePoint< std::chrono::seconds > now(bool Deterministic)
static GCMetadataPrinterRegistry::Add< ErlangGCPrinter > X("erlang", "erlang-compatible garbage collector")
In fpstack this compiles esp eax movl esp fildl(%esp) fmuls LCPI1_0 addl $4
The initial backend is deliberately restricted to z10 We should add support for later architectures at some point If an asm ties an i32 r result to an i64 the input will be treated as an leaving the upper bits uninitialised For i64 store i32 i32 *dst ret void from CodeGen SystemZ asm ll will use LHI rather than LGHI to load This seems to be a general target independent problem The tuning of the choice between LOAD XC and CLC for constant length block operations We could extend them to variable length operations too
This might compile to this xmm1 xorps xmm0 movss xmm0 ret Now consider if the code caused xmm1 to get spilled This might produce this xmm1 movaps xmm0 movaps xmm1 movss xmm0 ret since the reload is only used by these we could fold it into the producing something like xmm1 movaps xmm0 ret saving two instructions The basic idea is that a reload from a spill if only one byte chunk is bring in zeros the one element instead of elements This can be used to simplify a variety of shuffle where the elements are fixed zeros This code generates ugly probably due to costs being off or< 4 x float > eax xmm0 pxor xmm1 movaps xmm2 xmm2 xmm0 movaps eax ret Would it be better to ecx xmm0 xor eax pinsrw
Common register allocation spilling lr str ldr sxth r3 ldr mla r4 can lr mov lr str ldr sxth r3 mla r4 and then merge mul and lr str ldr sxth r3 mla r4 It also increase the likelihood the store may become dead bb27 Successors according to LLVM ID Predecessors according to mbb< bb27, 0x8b0a7c0 > Note ADDri is not a two address instruction its result reg1037 is an operand of the PHI node in bb76 and its operand reg1039 is the result of the PHI node We should treat it as a two address code and make sure the ADDri is scheduled after any node that reads reg1039 Use info(i.e. register scavenger) to assign it a free register to allow reuse the collector could move the objects and invalidate the derived pointer This is bad enough in the first but safe points can crop up unpredictably **array_addr i32 n y store obj obj **nth_el If the i64 division is lowered to a then a safe point will(must) appear for the call site. If a collection occurs
the resulting code requires compare and branches when and if the revised code is with conditional branches instead of More there is a byte word extend before each where there should be only one
Current eax eax eax ret Ideal eax eax ret Re implement atomic builtins x86 does not have to use add to implement these Instead
Common register allocation spilling lr str ldr sxth r3 ldr mla r4 can lr mov lr str ldr sxth r3 mla r4 and then merge mul and lr str ldr sxth r3 mla r4 It also increase the likelihood the store may become dead bb27 Successors according to LLVM ID Predecessors according to mbb< bb27, 0x8b0a7c0 > Note ADDri is not a two address instruction its result reg1037 is an operand of the PHI node in bb76 and its operand reg1039 is the result of the PHI node We should treat it as a two address code and make sure the ADDri is scheduled after any node that reads reg1039 Use info(i.e. register scavenger) to assign it a free register to allow reuse the collector could move the objects and invalidate the derived pointer This is bad enough in the first but safe points can crop up unpredictably **array_addr i32
__m128 foo1(float x1, float x4)
into eax xorps xmm0 xmm0 eax xmm0 eax xmm0 ret esp eax movdqa xmm0 xmm0 esp const ret align it should be movdqa xmm0 xmm0 We should transform a shuffle of two vectors of constants into a single vector of constants insertelement of a constant into a vector of constants should also result in a vector of constants e g VecISelBug ll We compiled it to something globl _t _t
AMD64 Optimization Manual has some nice information about optimizing integer multiplication by a constant How much of it applies to Intel s X86 implementation There are definite trade offs to xmm0 cvttss2siq rdx jb L3 subss xmm0 rax cvttss2siq rdx xorq rdx rax ret instead of xmm1 cvttss2siq rcx movaps xmm2 subss xmm2 cvttss2siq rax rdx xorq rax ucomiss xmm0 cmovb rax ret Seems like the jb branch has high likelihood of being taken It would have saved a few instructions It s not possible to reference and DH registers in an instruction requiring REX prefix divb and mulb both produce results in AH If isel emits a CopyFromReg which gets turned into a movb and that can be allocated a r8b r15b To get around this
This currently compiles esp movsd(%esp)
We currently generate an sqrtsd and divsd instructions This is bad
We have fiadd patterns now but the followings have the same cost and complexity We need a way to specify the later is more profitable def def The FP stackifier should handle simple permutates to reduce number of shuffle e g which prevents the spiller from folding spill code into the instructions Currently the x86 codegen isn t very good at mixing SSE and FPStack code
Error backend(const Config &C, AddStreamFn AddStream, unsigned ParallelCodeGenParallelismLevel, Module &M, ModuleSummaryIndex &CombinedIndex)
Runs a regular LTO backend.
Target & getTheX86_32Target()
We should do a little better with eliminating dead stores The stores to the stack are dead since a and b are not needed
gcc mainline compiles it x2(%rip)
AMD64 Optimization Manual has some nice information about optimizing integer multiplication by a constant How much of it applies to Intel s X86 implementation There are definite trade offs to xmm0 cvttss2siq rdx jb L3 subss xmm0 rax cvttss2siq rdx xorq rdx rax ret instead of xmm1 cvttss2siq rcx movaps xmm2 subss xmm2 cvttss2siq rax rdx xorq rax ucomiss xmm0 cmovb rax ret Seems like the jb branch has high likelihood of being taken It would have saved a few instructions It s not possible to reference and DH registers in an instruction requiring REX prefix divb and mulb both produce results in AH If isel emits a CopyFromReg which gets turned into a movb and that can be allocated a r8b r15b To get around isel emits a CopyFromReg from AX and then right shift it down by and truncate it It s not pretty but it works We need some register allocation magic to make the hack go which would often require a callee saved register Callees usually need to keep this value live for most of their body so it doesn t add a significant burden on them We currently implement this in however this is suboptimal because it means that it would be quite awkward to implement the optimization for callers A better implementation would be to relax the LLVM IR rules for sret arguments to allow a function with an sret argument to have a non void return type
Unify divergent function exit nodes
This is equivalent to the following
#define LLVM_EXTERNAL_VISIBILITY
gcc mainline compiles it xmm0 xmm1 movaps xmm2 movlhps xmm2 movaps xmm0 ret We compile vector multiply by constant into poor< i32 10, i32 10, i32 10, i32 10 > ret< 4 x i32 > A On targets without SSE4
AMD64 Optimization Manual has some nice information about optimizing integer multiplication by a constant How much of it applies to Intel s X86 implementation There are definite trade offs to xmm0 cvttss2siq rdx jb L3 subss xmm0 rax cvttss2siq rdx xorq rdx rax ret instead of xmm1 cvttss2siq rcx movaps xmm2 subss xmm2 cvttss2siq rax rdx xorq rax ucomiss xmm0 cmovb rax ret Seems like the jb branch has high likelihood of being taken It would have saved a few instructions It s not possible to reference AH
This might compile to this xmm1 xorps xmm0 movss xmm0 ret Now consider if the code caused xmm1 to get spilled This might produce this xmm1 movaps xmm0 movaps xmm1 movss xmm0 ret since the reload is only used by these we could fold it into the producing something like xmm1 movaps xmm0 ret saving two instructions The basic idea is that a reload from a spill slot
Predicate all(Predicate P0, Predicate P1)
True iff P0 and P1 are true.
multiplies can be turned into SHL s
compiles ldr LCPI1_0 ldr ldr mov lsr tst moveq r1 ldr LCPI1_1 and r0 bx lr It would be better to do something like to fold the shift into the conditional move
entry stw r5 blr GCC r3 srawi xor r4 subf r0 stw r5 blr which is much nicer This theoretically may help improve twolf slightly(used in dimbox.c:142?).
This compiles xmm1 mulss xmm1 xorps xmm0 movss xmm0 ret Because mulss doesn t modify the top the top elements of xmm1 are already zero d We could compile this to
This might compile to this xmm1 xorps xmm0 movss xmm0 ret Now consider if the code caused xmm1 to get spilled This might produce this xmm1 movaps xmm0 movaps xmm1 movss xmm0 ret since the reload is only used by these we could fold it into the producing something like xmm1 movaps xmm0 ret saving two instructions The basic idea is that a reload from a spill if only one byte chunk is bring in zeros the one element instead of elements This can be used to simplify a variety of shuffle where the elements are fixed zeros This code generates ugly probably due to costs being off or something
S is passed via registers r2 But gcc stores them to the and then reload them to and r3 before issuing the call(r0 contains the address of the format string)
This currently compiles esp xmm0 movsd esp eax eax esp ret We should use not the dag combiner This is because dagcombine2 needs to be able to see through the X86ISD::Wrapper node
The following code is currently generated
SSE has instructions for doing operations on complex numbers
> ldr r0, pc, #((LCPI1_0-(LPCRELL0+4))&0xfffffffc) We compile the following:define i16 @func_entry_2E_ce(i32 %i) { switch i32 %i, label %bb12.exitStub[i32 0, label %bb4.exitStub i32 1, label %bb9.exitStub i32 2, label %bb4.exitStub i32 3, label %bb4.exitStub i32 7, label %bb9.exitStub i32 8, label %bb.exitStub i32 9, label %bb9.exitStub] bb12.exitStub:ret i16 0 bb4.exitStub:ret i16 1 bb9.exitStub:ret i16 2 bb.exitStub:ret i16 3 } into:_func_entry_2E_ce:mov r2, #1 lsl r2, r0 cmp r0, #9 bhi LBB1_4 @bb12.exitStub LBB1_1:@newFuncRoot mov r1, #13 tst r2, r1 bne LBB1_5 @bb4.exitStub LBB1_2:@newFuncRoot ldr r1, LCPI1_0 tst r2, r1 bne LBB1_6 @bb9.exitStub LBB1_3:@newFuncRoot mov r1, #1 lsl r1, r1, #8 tst r2, r1 bne LBB1_7 @bb.exitStub LBB1_4:@bb12.exitStub mov r0, #0 bx lr LBB1_5:@bb4.exitStub mov r0, #1 bx lr LBB1_6:@bb9.exitStub mov r0, #2 bx lr LBB1_7:@bb.exitStub mov r0, #3 bx lr LBB1_8:.align 2 LCPI1_0:.long 642 gcc compiles to:cmp r0, #9 @ lr needed for prologue bhi L2 ldr r3, L11 mov r2, #1 mov r1, r2, asl r0 ands r0, r3, r2, asl r0 movne r0, #2 bxne lr tst r1, #13 beq L9 L3:mov r0, r2 bx lr L9:tst r1, #256 movne r0, #3 bxne lr L2:mov r0, #0 bx lr L12:.align 2 L11:.long 642 GCC is doing a couple of clever things here:1. It is predicating one of the returns. This isn 't a clear win though:in cases where that return isn 't taken, it is replacing one condbranch with two 'ne' predicated instructions. 2. It is sinking the shift of "1 << i" into the tst, and using ands instead of tst. This will probably require whole function isel. 3. GCC emits:tst r1, #256 we emit:mov r1, #1 lsl r1, r1, #8 tst r2, r1 When spilling in thumb mode and the sp offset is too large to fit in the ldr/str offset field, we load the offset from a constpool entry and add it to sp:ldr r2, LCPI add r2, sp ldr r2,[r2] These instructions preserve the condition code which is important if the spill is between a cmp and a bcc instruction. However, we can use the(potentially) cheaper sequence if we know it 's ok to clobber the condition register. add r2, sp, #255 *4 add r2, #132 ldr r2,[r2, #7 *4] This is especially bad when dynamic alloca is used. The all fixed size stack objects are referenced off the frame pointer with negative offsets. See oggenc for an example. Poor codegen test/CodeGen/ARM/select.ll f7:ldr r5, LCPI1_0 LPC0:add r5, pc ldr r6, LCPI1_1 ldr r2, LCPI1_2 mov r3, r6 mov lr, pc bx r5 Make register allocator/spiller smarter so we can re-materialize "mov r, imm", etc. Almost all Thumb instructions clobber condition code. Thumb load/store address mode offsets are scaled. The values kept in the instruction operands are pre-scale values. This probably ought to be changed to avoid extra work when we convert Thumb2 instructions to Thumb1 instructions. We need to make(some of the) Thumb1 instructions predicable. That will allow shrinking of predicated Thumb2 instructions. To allow this, we need to be able to toggle the 's' bit since they do not set CPSR when they are inside IT blocks. Make use of hi register variants of cmp:tCMPhir/tCMPZhir. Thumb1 immediate field sometimes keep pre-scaled values. See ThumbRegisterInfo::eliminateFrameIndex. This is inconsistent from ARM and Thumb2. Rather than having tBR_JTr print a ".align 2" and constant island pass pad it, add a target specific ALIGN instruction instead. That way, getInstSizeInBytes won 't have to over-estimate. It can also be used for loop alignment pass. We generate conditional code for icmp when we don 't need to. This code:int foo(int s) { return s==1 LCPI1_0
void shuffle(Iterator first, Iterator last, RNG &&g)
instcombine should handle this transform
This might compile to this xmm1 xorps xmm0 movss xmm0 ret Now consider if the code caused xmm1 to get spilled This might produce this xmm1 movaps xmm0 movaps xmm1 movss xmm0 ret since the reload is only used by these we could fold it into the producing something like xmm1 movaps xmm0 ret saving two instructions The basic idea is that a reload from a spill if only one byte chunk is bring in zeros the one element instead of elements This can be used to simplify a variety of shuffle where the elements are fixed zeros This code generates ugly probably due to costs being off or< 4 x float > eax xmm0 pxor xmm1 movaps xmm2 xmm2 xmm0 movaps eax ret Would it be better to ecx xmm0 xor eax xmm0 xmm0 movaps ecx ret Some useful information in the Apple Altivec SSE Migration Guide
<%struct.s * > cast struct s *S to sbyte *< sbyte * > sbyte uint cast struct s *agg result to sbyte *< sbyte * > sbyte uint cast struct s *memtmp to sbyte *< sbyte * > sbyte uint ret void llc ends up issuing two memcpy or custom lower memcpy(of small size) to be ldmia/stmia. I think option 2 is better but the current register allocator cannot allocate a chunk of registers at a time. A feasible temporary solution is to use specific physical registers at the lowering time for small(<
Instead we get xmm1 addss xmm1 xmm2 movhlps xmm0 movaps xmm3 addss xmm3 movdqa xmm0 addss xmm0 ret there are cases where some simple SLP would improve codegen a bit compiling _Complex float B
int compare(DigitsT LDigits, int16_t LScale, DigitsT RDigits, int16_t RScale)
Compare two scaled numbers.
Instead of the following for memset char edx edx edx It might be better to generate eax movl edx movl edx movw edx when we can spare a register It reduces code size Evaluate what the best way to codegen sdiv C is For we currently get ret i32 Y eax movl ecx ecx ecx addl eax eax ret GCC knows several different ways to codegen one of which is eax eax ecx cmovle eax eax ret which is probably slower
into eax xorps xmm0 xmm0 eax xmm0 eax xmm0 ret esp eax movdqa xmm0 xmm0 esp const ret align it should be movdqa xmm0 edi
Clang compiles this i1 i64 store i64 i64 store i64 i64 store i64 i64 store i64 align Which gets codegen d xmm0 movaps rbp movaps rbp movaps rbp movaps rbp movq
print Print MemDeps of function
In fpstack this compiles esp eax movl esp esp ret in SSE it compiles into significantly slower esp cvtsi2sd(%esp)
So that lo16() _a(r2) lbz r2
SI optimize exec mask operations pre RA
This might compile to this xmm1 xorps xmm0 movss xmm0 ret Now consider if the code caused xmm1 to get spilled This might produce this xmm1 movaps xmm0 movaps xmm1 movss xmm0 ret since the reload is only used by these we could fold it into the producing something like xmm1 movaps xmm0 ret saving two instructions The basic idea is that a reload from a spill if only one byte chunk is bring in zeros the one element instead of elements This can be used to simplify a variety of shuffle where the elements are fixed zeros This code generates ugly probably due to costs being off or< 4 x float > eax xmm0 pxor xmm1 movaps xmm2 shufps
Doing so could allow SROA of the destination pointers See also
void norm(double x, double y, double z)
into eax xorps xmm0 xmm0 eax xmm0 eax xmm0 ret gcc
into eax xorps xmm0 xmm0 eax xmm0 eax xmm0 ret esp eax movdqa xmm0 xmm0 esp const ret align it should be movdqa xmm0 xmm0 We should transform a shuffle of two vectors of constants into a single vector of constants insertelement of a constant into a vector of constants should also result in a vector of constants e g VecISelBug ll We compiled it to something globl _t xmm0 movhps xmm0 movss xmm1 movaps xmm2 xmm2 xmm0 movaps rdar
Expected< ExpressionValue > min(const ExpressionValue &Lhs, const ExpressionValue &Rhs)
This compiles xmm1 mulss xmm1 xorps xmm0 movss xmm0 ret Because mulss doesn t modify the top elements
_foo edi jbe LBB1_3 eax LBB1_2
Instead we get xmm1 addss xmm1 xmm2 movhlps xmm0 movaps xmm3 addss xmm3 movdqa xmm2
Since we know that Vector is byte aligned and we know the element offset of we should change the load into a lve *x instruction
into xmm2 addss xmm2 xmm1 xmm3 addss xmm3 movaps xmm0 unpcklps xmm0 ret seems silly when it could just be one addps Expand libm rounding functions main should enable SSE DAZ mode and other fast SSE modes Think about doing i64 math in SSE regs on x86 This testcase should have no SSE instructions in and only one load from a constant double ret double C the select is being lowered
static uint64_t scale(uint64_t Num, uint32_t N, uint32_t D)
We have fiadd patterns now but the followings have the same cost and complexity We need a way to specify the later is more profitable def def The FP stackifier should handle simple permutates to reduce number of shuffle e g trunc
In fpstack this compiles esp eax movl esp esp ret in SSE it compiles into significantly slower esp xmm0 mulsd xmm0 movsd esp fldl(%esp) addl $12
For the entry BB esp pxor xmm0 xmm1 ucomisd xmm1 setnp al sete cl testb al jne LBB1_5 xmm2 cvtss2sd xmm3 ucomisd xmm0 ja LBB1_3 xmm2 LBB1_3
Common register allocation spilling lr str ldr sxth r3 ldr mla r4 can lr mov lr str ldr sxth r3 mla r4 and then merge mul and lr str ldr sxth r3 mla r4 It also increase the likelihood the store may become dead bb27 Successors according to LLVM ID Predecessors according to mbb< bb27, 0x8b0a7c0 > Note ADDri is not a two address instruction its result reg1037 is an operand of the PHI node in bb76 and its operand reg1039 is the result of the PHI node We should treat it as a two address code and make sure the ADDri is scheduled after any node that reads reg1039 Use info(i.e. register scavenger) to assign it a free register to allow reuse the collector could move the objects and invalidate the derived pointer This is bad enough in the first but safe points can crop up unpredictably Consider
We currently generate a but we really shouldn eax ecx xorl edx divl ecx eax divl ecx movl eax ret A similar code sequence works for division We currently compile i32 v2 eax eax jo LBB1_2 xor
this could be done in SelectionDAGISel along with other special for
This currently compiles esp xmm0 movsd esp eax eax esp ret We should use not the dag combiner This is because dagcombine2 needs to be able to see through the X86ISD::Wrapper which DAGCombine can t really do The code for turning x load into a single vector load is target independent and should be moved to the dag combiner The code for turning 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 where P can be anything The alignment inference code cannot handle loads from globals in static non mode because it doesn t look through the extra dyld stub load If you try vec_align ll without relocation you ll see what I mean We should lower which eliminates a constant pool load For consider
We have fiadd patterns now but the followings have the same cost and complexity We need a way to specify the later is more profitable def def The FP stackifier should handle simple permutates to reduce number of shuffle instructions
AMD64 Optimization Manual has some nice information about optimizing integer multiplication by a constant How much of it applies to Intel s X86 implementation There are definite trade offs to xmm0 cvttss2siq rdx jb L3 subss xmm0 rax cvttss2siq rdx xorq rdx rax ret instead of xmm1 cvttss2siq rcx movaps xmm2 subss xmm2 cvttss2siq rax rdx xorq rax ucomiss xmm0 cmovb rax ret Seems like the jb branch has high likelihood of being taken It would have saved a few instructions It s not possible to reference and DH registers in an instruction requiring REX prefix However
Should compile r2 movcc movcs str strb mov lr r1 movcs movcc mov lr not
into eax xorps xmm0 xmm0 movzbl(%esp)
<%struct.bf ** > define void t1() nounwind ssp
This might compile to this xmm1 xorps xmm0 movss xmm0 ret Now consider if the code caused xmm1 to get spilled This might produce this xmm1 movaps xmm0 movaps xmm1 movss xmm0 ret since the reload is only used by these instructions
Common register allocation spilling lr str ldr sxth r3 ldr mla r4 can lr mov lr str ldr sxth r3 mla r4 and then merge mul and lr str ldr sxth r3 mla r4 It also increase the likelihood the store may become dead bb27 Successors according to CFG
the custom lowered code happens to be but we shouldn t have to custom lower anything This is probably related to< 2 x i64 > ops being so bad LLVM currently generates stack realignment when it is not necessary needed The problem is that we need to know about stack alignment too before RA runs At that point we don t whether there will be vector or not Stack realignment logic is overly conservative but otherwise we can produce unaligned loads stores Fixing this will require some huge RA changes Testcase
LLVM_EXTERNAL_VISIBILITY void LLVMInitializeX86TargetInfo()
*Add support for compiling functions in both ARM and Thumb mode
the resulting code requires compare and branches when and if the revised code is with conditional branches instead of More there is a byte word extend before each where there should be only and the condition codes are not remembered when the same two values are compared twice More LSR enhancements possible
We have fiadd patterns now but the followings have the same cost and complexity We need a way to specify the later is more profitable def OneArgFPRW
< float * > store float float *tmp5 ret void Compiles rax rax movl rdi ret This would be better kept in the SSE unit by treating XMM0 as a and doing a shuffle from v[1] to v[0] then a float store[UNSAFE FP] void foo(double, double, double)
compiles ldr LCPI1_0 ldr ldr mov lsr tst moveq r1 ldr LCPI1_1 and r0 bx lr It would be better to do something like to fold the shift into the conditional ldr LCPI1_0 ldr ldr tst movne lsr ldr LCPI1_1 and r0 bx lr it saves an instruction and a register It might be profitable to cse MOVi16 if there are lots of bit immediates with the same bottom half Robert Muth started working on an alternate jump table implementation that does not put the tables in line in the text This is more like the llvm default jump table implementation This might be useful sometime Several revisions of patches are on the mailing beginning while CMP sets them like a subtract Therefore to be able to use CMN for comparisons other than the Z bit
*Add support for compiling functions in both ARM and Thumb then taking the smallest *Add support for compiling individual basic blocks in thumb when in a larger ARM function This can be used for presumed cold code
This might compile to this xmm1 xorps xmm0 movss xmm0 ret Now consider if the code caused xmm1 to get spilled This might produce this xmm1 movaps xmm0 movaps xmm1 movss xmm0 ret since the reload is only used by these we could fold it into the producing something like xmm1 movaps xmm0 ret saving two instructions The basic idea is that a reload from a spill can
Should compile to something r4 addze r3 instead we get
float space text globl _test align _test
Some targets(e.g. athlons) prefer freep to fstp ST(0)
In x86 we generate this spiffy xmm0 xmm0 ret in x86 we generate this which could be better
RegisterTarget - Helper template for registering a target, for use in the target's initialization fun...
into eax xorps xmm0 xmm0 eax xmm0 eax xmm0 ret esp eax movdqa LC0
Generic address nodes are lowered to some combination of target independent and machine specific and compilation options The choice of specific instructions that are to be used is delegated to ISel which in turn relies on TableGen patterns to choose subtarget specific instructions For in the pseudo code generated got(sym))
gcc mainline compiles it xmm0 xmm1 movaps xmm2 movlhps xmm2 movaps xmm0 ret We compile vector multiply by constant into poor< i32 10, i32 10, i32 10, i32 10 > ret< 4 x i32 > A On targets without this compiles globl _f xmm1 movd eax imull LCPI1_0
So that lo16() r2 stb r3 blr Becomes r3 they should compile to something better than
SSE Variable shift can be custom lowered to something like which uses a small table unaligned load shuffle instead of going through memory __m128i shift_right(__m128i value, unsigned long offset)
esp eax movl ecx ecx cvtsi2ss xmm0 eax cvtsi2ss xmm1 xmm0 addss xmm0 movss flds(%esp, 1) 0000002d addl $0x04
we compile this esp call L1 $pb L1 esp je LBB1_2 esp ret but is currently always computed in the entry block It would be better to sink the picbase computation down into the block for the as it is the only one that uses it This happens for a lot of code with early outs Another example is loads of arguments
into eax xorps xmm0 xmm0 eax xmm0 eax xmm0 ret esp eax movdqa xmm0 xmm0 esp const ret align it should be movdqa xmm0 xmm0 We should transform a shuffle of two vectors of constants into a single vector of constants insertelement of a constant into a vector of constants should also result in a vector of constants e g VecISelBug ll We compiled it to something horrible
http eax xorl edx cl sete al setne dl sall eax sall edx But that requires good bit subreg support this might be better It s an extra but it s one instruction and doesn t stress bit subreg eax eax movl edx edx sall eax sall cl edx bit shifts(in general) expand to really bad code. Instead of using cmovs
the resulting code requires compare and branches when and if the revised code is with conditional branches instead of More there is a byte word extend before each where there should be only and the condition codes are not remembered when the same two values are compared twice More LSR enhancements i8 and i32 load store addressing modes are identical This
AMD64 Optimization Manual has some nice information about optimizing integer multiplication by a constant How much of it applies to Intel s X86 implementation There are definite trade offs to xmm0 cvttss2siq rdx jb L3 subss xmm0 rax cvttss2siq rdx xorq rdx rax ret instead of xmm1 cvttss2siq rcx movaps xmm2 subss xmm2 cvttss2siq rax rdx xorq rax ucomiss xmm0 cmovb rax ret Seems like the jb branch has high likelihood of being taken It would have saved a few instructions It s not possible to reference BH
gcc mainline compiles it xmm0 xmm1 movaps xmm2 movlhps xmm2 movaps xmm0 ret We compile vector multiply by constant into poor< i32 10, i32 10, i32 10, i32 10 > ret< 4 x i32 > A On targets without this compiles globl _f _f
Instead we get xmm1 addss xmm1 xmm2 movhlps xmm0 movaps xmm3 addss xmm3 movdqa xmm0 addss xmm3
into xmm2 addss xmm2 xmm1 xmm3 addss xmm3 movaps xmm0 unpcklps xmm0 ret seems silly when it could just be one addps Expand libm rounding functions main should enable SSE DAZ mode and other fast SSE modes Think about doing i64 math in SSE regs on x86 This testcase should have no SSE instructions in and only one load from a constant double ret double C Currently
Instead we get xmm1 addss xmm1 xmm2 movhlps xmm0 movaps xmm3 addss xmm3 movdqa xmm0 addss xmm0 ret Also
This might compile to this xmm1 xorps xmm0 movss xmm0 ret Now consider if the code caused xmm1 to get spilled This might produce this xmm1 movaps xmm0 movaps xmm1 movss xmm0 ret since the reload is only used by these we could fold it into the producing something like xmm1 movaps xmm0 ret saving two instructions The basic idea is that a reload from a spill if only one byte chunk is used
http eax xorl edx cl sete al setne dl sall eax sall edx But that requires good bit subreg support this might be better It s an extra shift
MIPS Relocation Principles In LLVM
S is passed via registers r2 But gcc stores them to the stack
This might compile to this xmm1 xorps xmm0 movss xmm0 ret Now consider if the code caused xmm1 to get spilled This might produce this xmm1 movaps xmm0 movaps xmm1 movss xmm0 ret However
the custom lowered code happens to be but we shouldn t have to custom lower anything This is probably related to< 2 x i64 > ops being so bad LLVM currently generates stack realignment when it is not necessary needed The problem is that we need to know about stack alignment too early
esp eax movl ecx ecx cvtsi2ss xmm0 andl
the multiplication has a latency of four as opposed to two cycles for the movl lea variant It appears gcc place string data with linkonce linkage in section coalesced instead of section coalesced Take a look at darwin h
Common register allocation spilling lr str ldr sxth r3 ldr mla r4 can lr mov lr str ldr sxth r3 mla r4 and then merge mul and lr str ldr sxth r3 mla r4 It also increase the likelihood the store may become dead bb27 Successors according to LLVM BB
Common register allocation spilling problem
this lets us change the cmpl into a which is and eliminate the shift We compile this i32 i32 i8 zeroext d nounwind
Vector Shift Left don t map to llvm shl and lshr
the custom lowered code happens to be but we shouldn t have to custom lower anything This is probably related to< 2 x i64 > ops being so bad LLVM currently generates stack realignment when it is not necessary needed The problem is that we need to know about stack alignment too before RA runs At that point we don t whether there will be vector spill
QP Compare Ordered outs ins xscmpudp No builtin are required Or llvm fcmp order unorder compare DP QP Compare builtin are required DP xscmp *dp write to VSX register Use int_ppc_vsx_xscmpeqdp f64
the resulting code requires compare and branches when and if the revised code is with conditional branches instead of More there is a byte word extend before each where there should be only and the condition codes are not remembered when the same two values are compared twice More LSR enhancements i8 and i32 load store addressing modes are identical int int int d
The same transformation can work with an even modulo with the addition of a and shrink the compare RHS by the same amount Unless the target supports that transformation probably isn t worthwhile The transformation can also easily be made to work with non zero equality for n
Implement PPCInstrInfo::isLoadFromStackSlot isStoreToStackSlot for vector registers
This currently compiles esp xmm0 movsd esp eax eax esp ret We should use not the dag combiner This is because dagcombine2 needs to be able to see through the X86ISD::Wrapper which DAGCombine can t really do The code for turning x load into a single vector load is target independent and should be moved to the dag combiner The code for turning 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 where P can be anything The alignment inference code cannot handle loads from globals in static non mode because it doesn t look through the extra dyld stub load If you try vec_align ll without relocation model
Unrolling by would eliminate the &in both leading to a net reduction in code size The resultant code would then also be suitable for exit value computation We miss a bunch of rotate opportunities on various including etc On X86
This currently compiles esp xmm0 movsd esp eax shrl
So we should use XX3Form_Rcr to implement intrinsic Convert DP outs ins xscvdpsp No builtin are required Round &Convert QP DP(dword[1] is set to zero) No builtin are required Round to Quad Precision because you need to assign rounding mode in instruction Provide builtin(set f128:$vT,(int_ppc_vsx_xsrqpi f128:$vB))(set f128 yields< n x< ty > >< result > yields< ty >< result > No builtin are required Load Store load store see def memrix16 in PPCInstrInfo td Load Store Vector load store outs ins lxsdx set load store with conversion from to outs ins lxsspx set load store outs ins lxsiwzx set PPClfiwzx ins stxsiwx $dst
bb420 i The CBE manages to produce
Common register allocation spilling lr str ldr sxth r3 ldr mla r4 can lr mov lr str ldr sxth r3 mla r4 and then merge mul and lr str ldr sxth r3 mla r4 It also increase the likelihood the store may become dead bb27 Successors according to LLVM ID Predecessors according to mbb< bb27, 0x8b0a7c0 > Note ADDri is not a two address instruction its result reg1037 is an operand of the PHI node in bb76 and its operand reg1039 is the result of the PHI node We should treat it as a two address code and make sure the ADDri is scheduled after any node that reads reg1039 Use info(i.e. register scavenger) to assign it a free register to allow reuse the collector could move the objects and invalidate the derived pointer This is bad enough in the first but safe points can crop up unpredictably **array_addr i32 n y store obj obj **nth_el If the i64 division is lowered to a libcall
hexagon cext Hexagon constant extender optimization
print Instructions which execute on loop entry
This might compile to this xmm1 xorps xmm0 movss xmm0 ret Now consider if the code caused xmm1 to get spilled This might produce this xmm1 movaps xmm0 movaps xmm1 movss xmm0 ret since the reload is only used by these we could fold it into the producing something like xmm1 movaps xmm0 ret saving two instructions The basic idea is that a reload from a spill if only one byte chunk is bring in zeros the one element instead of elements This can be used to simplify a variety of shuffle operations
This might compile to this xmm1 xorps xmm0 movss xmm0 ret Now consider if the code caused xmm1 to get spilled This might produce this xmm1 movaps xmm0 movaps xmm1 movss xmm0 ret since the reload is only used by these we could fold it into the producing something like xmm1 movaps xmm0 ret saving two instructions The basic idea is that a reload from a spill if only one byte chunk is bring in zeros the one element instead of elements This can be used to simplify a variety of shuffle where the elements are fixed zeros This code generates ugly probably due to costs being off or< 4 x float > eax xmm0 pxor xmm1 movaps xmm2 xmm2 xmm0 movaps eax ret Would it be better to ecx xmm0 xor eax
Add support for conditional and other related patterns Instead of
Instead we get xmm1 addss xmm1 xmm2 movhlps xmm0 movaps xmm3 addss xmm1
Code Generation Notes for reduce the size of the ISel and reduce repetition in the implementation In a small number of cases
SSE has instructions for doing operations on complex we should pattern match them For example
