[llvm] [X86] Make test more robust against different builds

[aartbik]

Summary

Rationale:
Using the --debug-only flag requires a debug build. Also, the debug output is not always consistent over different builds.
This change avoids all problems by just testing the generated assembly for AVX.

[mehdi_amini]

I don't really understand why? Can you add the motivation in the commit/revision description?

[craig.topper]

In D78609#1995999, @mehdi_amini wrote:

I don't really understand why? Can you add the motivation in the commit/revision description?

I think because the test is checking the -debug output.

Can we just check the assembly output and trust that asserts builds will trigger the assert that found this originally?

[aartbik]

In D78609#1996030, @craig.topper wrote:

In D78609#1995999, @mehdi_amini wrote:

I don't really understand why? Can you add the motivation in the commit/revision description?

I think because the test is checking the -debug output.

Can we just check the assembly output and trust that asserts builds will trigger the assert that found this originally?

I just found out that this is needed for debug builds. Otherwise the test will fail for optimized builds.
I took inspiration from other files in the same directory, see e.g. merge-store-partially-alias-loads.ll
In this case, testing the assembly would of course work, but it would not be a strict as I wanted it to be.
Given that other tests do exactly this, I figured this was the best-practice for debug output testing.

If there are other best-practices for this, happy to comply...

[aartbik]

Hmm, given that some builds now break on using "extract_subvector" instead of "vselect",
I am actually really start to like the suggestion to just test the assembly and rely on
not hitting asserts in builds that support these, but not let the test rely on debug output.

CHECK: v4f64 = vselect

<stdin>:8:40: note: scanning from here
Initial selection DAG: %bb.0 'bug45563:'

<stdin>:41:7: note: possible intended match here
t54: v2f64 = extract_subvector t52, Constant:i64<0

[aartbik]

more robust test

[aartbik]

PTAL

[craig.topper]

llvm/test/CodeGen/X86/pr45563.ll

1

Drop the stderr redirect?

[aartbik]

dropped the stderr redirect, no longer needed

[craig.topper]

LGTM

llvm/test/CodeGen/X86/pr45563.ll

-; RUN: llc < %s -debug-only=isel -O3 -mattr=avx  2>&1 | FileCheck %s
+; RUN: llc < %s -O3 -mattr=avx  2>&1 | FileCheck %s
 
 ; Bug 45563:
 ; The LowerMLOAD() method AVX masked load branch should
 ; use the operand vector type rather than the mask type.
 ; Given, for example:
 ;   v4f64,ch = masked_load ..
 ; The select should be:
 ;   v4f64 = vselect ..
 ; instead of:
 ;   v4i64 = vselect ..
 
 define <16 x double> @bug45563(<16 x double>* %addr, <16 x double> %dst, <16 x i64> %e, <16 x i64> %f) {
 ; CHECK-LABEL: bug45563:
-; CHECK:       v4f64 = vselect
+; CHECK:       vblendvpd
+; CHECK:       vmaskmovpd
   %mask = icmp slt <16 x i64> %e, %f
   %res = call <16 x double> @llvm.masked.load.v16f64.p0v16f64(<16 x double>* %addr, i32 4, <16 x i1>%mask, <16 x double> %dst)
   ret <16 x double> %res
 }
 
 declare <16 x double> @llvm.masked.load.v16f64.p0v16f64(<16 x double>* %addr, i32 %align, <16 x i1> %mask, <16 x double> %dst)