x86-32: missed zext folding optimization for cvtsi2ss

[jroelofs]

This function:

float foo(unsigned char x) {
  return x;
}

compiles to (x86-32):

define float @foo(i8 zeroext  %x) nounwind  {
entry:
	%tmp12 = uitofp i8 %x to float		; <float> [#uses=1]
	ret float %tmp12
}

compiles to:

_foo:
	subl	$4, %esp
	movzbl	8(%esp), %eax
	cvtsi2ss	%eax, %xmm0
	movss	%xmm0, (%esp)
	flds	(%esp)
	addl	$4, %esp
	ret

We should be able to use:
cvtsi2ss 8($esp), %xmm0
since we know the stack slot is already zext'd.

rdar://5907648

[jroelofs]

Added a test referencing this in af6c1dbcdd15df92e9c902a28987301a18df8c91 3e0cdf3

[llvmbot]

@llvm/issue-subscribers-backend-x86

[RKSimon]

Added a test referencing this in af6c1dbcdd15df92e9c902a28987301a18df8c91

@jroelofs I think this is actually 3e0cdf3 ?

[jroelofs]

Added a test referencing this in af6c1dbcdd15df92e9c902a28987301a18df8c91

@jroelofs I think this is actually 3e0cdf3 ?

oops, yes. I think I might have copied the sha before rebasing.