[X86] Fix bug in legalize vector types - Split large loads

[aymanmus]

Summary

When splitting a large load to smaller legally-typed loads, the last load should be padded to reach the size of the previous one so a CONCAT_VECTORS node could reunite them again.
The code currently pads the last load to reach the size of the first load (instead of the previous).

[RKSimon]

Does this fix PR27708 as well?

[aymanmus]

The bug was filed for Clang's 3.8 version, it passed with clang-4.0 and clang-ToT (top of trunc).
Seems like the issue was fixed while ago.

[RKSimon]

In D38495#891561, @aymanmus wrote:

The bug was filed for Clang's 3.8 version, it passed with clang-4.0 and clang-ToT (top of trunc).
Seems like the issue was fixed while ago.

Confirmed I added a test case at rL315186

[RKSimon]

Possibly add a test case for PR34657 as well? @zvi has already reduced much of it

test/CodeGen/X86/pr34653.ll

1

Add a triple, also I realise its an assertion test but it doesn't usually hurt to try and create FileChecks for a test

[aymanmus]

Adding triple & a test case for PR23657

[RKSimon]

test/CodeGen/X86/pr34653-pr34657.ll

314 ↗

(On Diff #119845)

I have a feeling this will reduce further - bugpoint typically stops a little early.

[aymanmus]

Manually reducing "bugpointed" test.

[RKSimon]

test/CodeGen/X86/pr34653-pr34657.ll

3 ↗

(On Diff #120740)

Why the difference in attributes?

[aymanmus]

test/CodeGen/X86/pr34653-pr34657.ll

3 ↗

(On Diff #120740)

To reproduce both of the bugs as they were filed.

[RKSimon]

LGTM once the tests have been split up

test/CodeGen/X86/pr34653-pr34657.ll

3 ↗

(On Diff #120740)

OK - please can you just split these into 2 test files. pr34653.ll and pr34657.ll

[aymanmus]

test/CodeGen/X86/pr34653-pr34657.ll

3 ↗

(On Diff #120740)

Done. Thanks

lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp

while (LdWidth > 0) {
     } else {
       L = DAG.getLoad(NewVT, dl, Chain, BasePtr,
                       LD->getPointerInfo().getWithOffset(Offset),
                       MinAlign(Align, Increment), MMOFlags, AAInfo);
       LdChain.push_back(L.getValue(1));
     }
 
     LdOps.push_back(L);
-
+    LdOp = L;
 
     LdWidth -= NewVTWidth;
   }
 
   // Build the vector from the load operations.
   unsigned End = LdOps.size();
   if (!LdOps[0].getValueType().isVector())
     // All the loads are scalar loads.

test/CodeGen/X86/pr34653.ll

+; RUN: llc %s -O0 -mattr=+avx512f
+
+declare fastcc <38 x double> @test() 
+ 
+define void @main() {
+entry:
+  %v = call fastcc <38 x double> @test()
+  %v.0 = extractelement <38 x double> %v, i32 0
+  ret void
+}