1) codegen a shift of a register as a shift, not an LEA.
2) teach the RA to convert a shift to an LEA instruction if it wants something
in three-address form.
This gives us asm diffs like:
- leal (,%eax,4), %eax
+ shll $2, %eax
which is faster on some processors and smaller on all of them.
and, more interestingly:
- movl 24(%esi), %eax
- leal (,%eax,4), %edi
+ movl 24(%esi), %edi
+ shll $2, %edi
Without #2, #1 was a significant pessimization in some cases.
This implements CodeGen/X86/shift-codegen.ll
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35204 91177308-0d34-0410-b5e6-96231b3b80d8
actually *removes* one of the operands, instead of just assigning both operands
the same register. This make reasoning about instructions unnecessarily complex,
because you need to know if you are before or after register allocation to match
up operand #'s with the target description file.
Changing this also gets rid of a bunch of hacky code in various places.
This patch also includes changes to fold loads into cmp/test instructions in
the X86 backend, along with a significant simplification to the X86 spill
folding code.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30108 91177308-0d34-0410-b5e6-96231b3b80d8
movw. That is we promote the destination operand to r16. So
%CH = TRUNC_R16_R8 %BP
is emitted as
movw %bp, %cx.
This is incorrect. If %cl is live, it would be clobbered.
Ideally we want to do the opposite, that is emitted it as
movb ??, %ch
But this is not possible since %bp does not have a r8 sub-register.
We are now defining a new register class R16_ which is a subclass of R16
containing only those 16-bit registers that have r8 sub-registers (i.e.
AX - DX). We isel the truncate to two instructions, a MOV16to16_ to copy the
value to the R16_ class, followed by a TRUNC_R16_R8.
Due to bug 770, the register colaescer is not going to coalesce between R16 and
R16_. That will be fixed later so we can eliminate the MOV16to16_. Right now, it
can only be eliminated if we are lucky that source and destination registers are
the same.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28164 91177308-0d34-0410-b5e6-96231b3b80d8
- Handle FR32 to VR128:v4f32 and FR64 to VR128:v2f64 with aliases of MOVAPS
and MOVAPD. Mark them as move instructions and *hope* they will be deleted.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26919 91177308-0d34-0410-b5e6-96231b3b80d8
proves to be worth 20% on Ptrdist/ks. Might be related to dependency
breaking support.
2. Added FsMOVAPSrr and FsMOVAPDrr as aliases to MOVAPSrr and MOVAPDrr. These
are used for FR32 / FR64 reg-to-reg copies.
3. Tell reg-allocator to generate MOVSSrm / MOVSDrm and MOVSSmr / MOVSDmr to
spill / restore FsMOVAPSrr and FsMOVAPDrr.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26241 91177308-0d34-0410-b5e6-96231b3b80d8
XMM registers. There are many known deficiencies and fixmes, which will be
addressed ASAP. The major benefit of this work is that it will allow the
LLVM register allocator to allocate FP registers across basic blocks.
The x86 backend will still default to x87 style FP. To enable this work,
you must pass -enable-sse-scalar-fp and either -sse2 or -sse3 to llc.
An example before and after would be for:
double foo(double *P) { double Sum = 0; int i; for (i = 0; i < 1000; ++i)
Sum += P[i]; return Sum; }
The inner loop looks like the following:
x87:
.LBB_foo_1: # no_exit
fldl (%esp)
faddl (%eax,%ecx,8)
fstpl (%esp)
incl %ecx
cmpl $1000, %ecx
#FP_REG_KILL
jne .LBB_foo_1 # no_exit
SSE2:
addsd (%eax,%ecx,8), %xmm0
incl %ecx
cmpl $1000, %ecx
#FP_REG_KILL
jne .LBB_foo_1 # no_exit
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22340 91177308-0d34-0410-b5e6-96231b3b80d8
