[SystemZ] Use the EVT version of getVectorVT() in combineTruncateExtract().

[JonPsson1]

A test case showed up where the new vector type is v24i16, which is not a simple
MVT. In order to get an extended value type for cases like this, EVT::getVectorVT()
needs to be called instead of MVT::getVectorVT(), otherwise the following call
to getVectorElementType() in combineExtract() will fail.

(This showed up in Csmith testing and looks to have been a potential problem for at least 1.5 years.)

There are other places where MVT::getVectorVT() is called, but hopefully they are ok.

[llvmbot]

@llvm/pr-subscribers-backend-systemz

Author: Jonas Paulsson (JonPsson1)

Changes

A test case showed up where the new vector type is v24i16, which is not a simple
MVT. In order to get an extended value type for cases like this, EVT::getVectorVT()
needs to be called instead of MVT::getVectorVT(), otherwise the following call
to getVectorElementType() in combineExtract() will fail.

(This showed up in Csmith testing and looks to have been a potential problem for at least 1.5 years.)

There are other places where MVT::getVectorVT() is called, but hopefully they are ok.

---

Full diff: https://github.com/llvm/llvm-project/pull/100150.diff

2 Files Affected:

• (modified) llvm/lib/Target/SystemZ/SystemZISelLowering.cpp (+2-1)
• (modified) llvm/test/CodeGen/SystemZ/vec-combine-01.ll (+10)

diff --git a/llvm/lib/Target/SystemZ/SystemZISelLowering.cpp b/llvm/lib/Target/SystemZ/SystemZISelLowering.cpp
index b2b88143354a5..383393914a169 100644
--- a/llvm/lib/Target/SystemZ/SystemZISelLowering.cpp
+++ b/llvm/lib/Target/SystemZ/SystemZISelLowering.cpp
@@ -6653,7 +6653,8 @@ SDValue SystemZTargetLowering::combineTruncateExtract(
 
           // Defer the creation of the bitcast from X to combineExtract,
           // which might be able to optimize the extraction.
-          VecVT = MVT::getVectorVT(MVT::getIntegerVT(TruncBytes * 8),
+          VecVT = EVT::getVectorVT(*DCI.DAG.getContext(),
+                                   MVT::getIntegerVT(TruncBytes * 8),
                                    VecVT.getStoreSize() / TruncBytes);
           EVT ResVT = (TruncBytes < 4 ? MVT::i32 : TruncVT);
           return combineExtract(DL, ResVT, VecVT, Vec, NewIndex, DCI, true);
diff --git a/llvm/test/CodeGen/SystemZ/vec-combine-01.ll b/llvm/test/CodeGen/SystemZ/vec-combine-01.ll
index 6f0abd6ea5baf..16231b2d89526 100644
--- a/llvm/test/CodeGen/SystemZ/vec-combine-01.ll
+++ b/llvm/test/CodeGen/SystemZ/vec-combine-01.ll
@@ -153,3 +153,13 @@ define void @f7(ptr %ptr1, ptr %ptr2, ptr %ptr3, ptr %ptr4) {
   store i8 %trunc3, ptr %ptr4
   ret void
 }
+
+; Test that a truncating store with a non-simple VT can be handled.
+define void @f8(ptr %src, ptr %dst) {
+; CHECK-LABEL: f8:
+  %1 = load <12 x i32>, ptr %src, align 64
+  %2 = extractelement <12 x i32> %1, i64 11
+  %3 = trunc i32 %2 to i16
+  store i16 %3, ptr %dst, align 2
+  ret void
+}

[uweigand]

LGTM, thanks!

[tstellar]

/cherry-pick 22bc9db

[llvmbot]

Failed to create pull request for issue100150 https://github.com/llvm/llvm-project/actions/runs/11873454462