[llvm-commits] [llvm] r40736 - in /llvm/trunk: lib/Target/X86/X86InstrMMX.td lib/Target/X86/X86InstrSSE.td test/CodeGen/X86/constant-pool-remat-0.ll

[Chris Lattner]

> Mark the SSE and MMX load instructions that
> X86InstrInfo::isReallyTriviallyReMaterializable knows how to handle
> with the isReMaterializable flag so that it is given a chance to  
> handle
> them. Without hoisting constant-pool loads from loops this isn't very
> visible, though it does keep CodeGen/X86/constant-pool-remat-0.ll from
> making a copy of the constant pool on the stack.
>
> @@ -0,0 +1,10 @@
> +; RUN: llvm-as < %s | llc -march=x86-64 | grep LCPI | wc -l | grep 3
> +
> +declare float @qux(float %y)
> +
> +define float @array(float %a) {
> +  %n = mul float %a, 9.0
> +  %m = call float @qux(float %n)
> +  %o = mul float %m, 9.0
> +  ret float %o
> +}

Interesting testcase.  I'm now getting:

_array:
         subq    $8, %rsp
         movss   LCPI1_0(%rip), %xmm1
         mulss   %xmm1, %xmm0
         call    _qux
         movss   LCPI1_0(%rip), %xmm1
         mulss   %xmm1, %xmm0
         addq    $8, %rsp
         ret

for this.  Both loads should be folded into the mulss's.  Because the  
load is shared at isel time, this can't be done there, but it can be  
done after the remat.  Maybe the register allocator needs to call  
MRegisterInfo::foldMemoryOperand if the remat is a load that has a  
single use?

-Chris