[GISEL] G_SPLAT_VECTOR can take a splat that is larger than the vector element

[michaelmaitland]

This is what SelectionDAG does. We'd like to reuse SelectionDAG patterns.

[llvmbot]

@llvm/pr-subscribers-llvm-globalisel

Author: Michael Maitland (michaelmaitland)

Changes

This is what SelectionDAG does. We'd like to reuse SelectionDAG patterns.

---

Full diff: https://github.com/llvm/llvm-project/pull/86974.diff

3 Files Affected:

• (modified) llvm/docs/GlobalISel/GenericOpcode.rst (+4)
• (modified) llvm/lib/CodeGen/MachineVerifier.cpp (+4-3)
• (modified) llvm/test/MachineVerifier/test_g_splat_vector.mir (+2-2)

diff --git a/llvm/docs/GlobalISel/GenericOpcode.rst b/llvm/docs/GlobalISel/GenericOpcode.rst
index cae2c21b80d7e7..e3ec5272c1e167 100644
--- a/llvm/docs/GlobalISel/GenericOpcode.rst
+++ b/llvm/docs/GlobalISel/GenericOpcode.rst
@@ -690,6 +690,10 @@ G_SPLAT_VECTOR
 
 Create a vector where all elements are the scalar from the source operand.
 
+The type of the operand must be equal to or smaller than the vector element
+type. If the operand is smaller than the vector element type, the scalar is
+implicitly truncated to the vector element type.
+
 Vector Reduction Operations
 ---------------------------
 
diff --git a/llvm/lib/CodeGen/MachineVerifier.cpp b/llvm/lib/CodeGen/MachineVerifier.cpp
index e4e05ce9278caf..e568a379b213e6 100644
--- a/llvm/lib/CodeGen/MachineVerifier.cpp
+++ b/llvm/lib/CodeGen/MachineVerifier.cpp
@@ -1774,9 +1774,10 @@ void MachineVerifier::verifyPreISelGenericInstruction(const MachineInstr *MI) {
     if (!SrcTy.isScalar())
       report("Source type must be a scalar", MI);
 
-    if (DstTy.getScalarType() != SrcTy)
-      report("Element type of the destination must be the same type as the "
-             "source type",
+    if (TypeSize::isKnownGT(DstTy.getScalarType().getSizeInBits(),
+                            SrcTy.getSizeInBits()))
+      report("Element type of the destination must be the same size or smaller "
+             "than the source type",
              MI);
 
     break;
diff --git a/llvm/test/MachineVerifier/test_g_splat_vector.mir b/llvm/test/MachineVerifier/test_g_splat_vector.mir
index 0d1d8a3e6dcc64..c5d0c1b61f3915 100644
--- a/llvm/test/MachineVerifier/test_g_splat_vector.mir
+++ b/llvm/test/MachineVerifier/test_g_splat_vector.mir
@@ -22,6 +22,6 @@ body:             |
     ; CHECK: Source type must be a scalar
     %6:_(<vscale x 2 x s32>) = G_SPLAT_VECTOR %2
 
-    ; CHECK: Element type of the destination must be the same type as the source type
-    %7:_(<vscale x 2 x s64>) = G_SPLAT_VECTOR %0
+    ; CHECK: Element type of the destination must be the same size or smaller than the source type
+    %7:_(<vscale x 2 x s128>) = G_SPLAT_VECTOR %7
 ...

[tschuett]

extended?

[michaelmaitland]

I got it all backwards 😆 . The operand should be allowed to be greater than the size of the vector element.

[tschuett]

The title still says smaller. Assuming that this is again for patterns, but G_SPLAT_VECTOR(G_TRUNC(s128)) should suffice during legalisation?

[michaelmaitland]

Assuming that this is again for patterns, but G_SPLAT_VECTOR(G_TRUNC(s128)) should suffice during legalisation?

This is for patterns and for equivalence to SelectionDAG ISD::SPLAT_VECTOR.

[tschuett]

I am again confused. You want to take i64 splats on 32-bit machines and truncate the i64 to i32? You can just state that i64 are not legal?!?

[michaelmaitland]

I am again confused. You want to take i64 splats on 32-bit machines and truncate the i64 to i32? You can just state that i64 are not legal?!?

From selection dag:

The type of the operand must match the vector element type, except when they are integer types. In this case the operand is allowed to be wider than the vector element type, and is implicitly truncated to it.

I want to do this since the patterns have an sXLen operand which can be s32 on RV32 and I should be able to create a 8b, 16b splat vector. And on RV64 the operand will be sXLen and I want to create 8b, 16b, and 32b splat vectors

[aemerson]

This seems reasonable to me if it allows us to import more patterns.