[PATCH] D11393: [X86] Allow X86::COND_NE_OR_P and X86::COND_NP_OR_E to be reversed.

[Cong Hou via llvm-commits]

congh added inline comments.

================
Comment at: lib/Target/X86/X86InstrInfo.cpp:4062
@@ -4039,3 +4061,3 @@
       return true;
 
     // Update the MachineOperand.
----------------
davidxl wrote:
> Should these two JCCs be eliminated?
This is just an intermediate state, when AnalyzeBranch() is called. I believe they will be eliminated later.

================
Comment at: test/CodeGen/X86/fp-une-cmp.ll:49
@@ +48,3 @@
+; CHECK-NEXT:  jp  [[LABEL]]
+; CHECK:       jmp
+;
----------------
davidxl wrote:
> The branch sequence does imply  BB2 is reordered before BB1. Is it better to explicit test the label order?
> 
> Also why is the jump sequence not the optimal one:
> 
> jne bb2
> jnp bb1
> bb2:
> bb1:
OK, I will test the BB orders.

The generated assembly is shown below, which should be optimal.


```
func2:                                  # @func2
# BB#0:                                 # %entry
	pushl	%eax
	cvtss2sd	8(%esp), %xmm1
	cvtss2sd	12(%esp), %xmm0
	mulsd	%xmm1, %xmm0
	xorpd	%xmm1, %xmm1
	ucomisd	%xmm1, %xmm0
	jne	.LBB1_1
	jp	.LBB1_1
.LBB1_2:                                # %bb2
	cvtsd2ss	%xmm0, %xmm0
	movss	%xmm0, (%esp)
	flds	(%esp)
	popl	%eax
	retl
.LBB1_1:                                # %bb1
	addsd	.LCPI1_0, %xmm0
	jmp	.LBB1_2
.Lfunc_end1:
	.size	func2, .Lfunc_end1-func2

```


http://reviews.llvm.org/D11393