[NVPTX] expand extload/truncstore for vectors of floats

Summary:
According to PTX ISA:

For convenience, ld, st, and cvt instructions permit source and destination data operands to be wider than the instruction-type size, so that narrow values may be loaded, stored, and converted using regular-width registers. For example, 8-bit or 16-bit values may be held directly in 32-bit or 64-bit registers when being loaded, stored, or converted to other types and sizes. The operand type checking rules are relaxed for bit-size and integer (signed and unsigned) instruction types; floating-point instruction types still require that the operand type-size matches exactly, unless the operand is of bit-size type.

So, the ISA does not support load with extending/store with truncatation for floating numbers. This is reflected in setting the loadext/truncstore actions to expand in the code for floating numbers, but vectors of floating numbers are not taken care of.

As a result, loading a vector of floats followed by a fp_extend may be combined by DAGCombiner to a extload, and the extload may be lowered to NVPTXISD::LoadV2 with extending information. However, NVPTXISD::LoadV2 does not perform extending, and no extending instructions are inserted. Finally, PTX instructions with mismatched types are generated, like
ld.v2.f32 {%fd3, %fd4}, [%rd2]

This patch adds the correct actions for vectors of floats, so DAGCombiner would not create loads with extending, and correct code is generated.

Patched by Gang Hu.

Test Plan: Test case attached.

Reviewers: jingyue

Reviewed By: jingyue

Subscribers: llvm-commits, jholewinski

Differential Revision: http://reviews.llvm.org/D10876

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@241191 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/NVPTX/NVPTXISelLowering.cpp b/lib/Target/NVPTX/NVPTXISelLowering.cpp
index b5af72ab855aa43cfb975d229b401a5d292d0286..09e0bd5d3d8839044eb47821c1b6c2365347e3ff 100644 (file)
--- a/lib/Target/NVPTX/NVPTXISelLowering.cpp
+++ b/lib/Target/NVPTX/NVPTXISelLowering.cpp
@@ -206,7 +206,14 @@ NVPTXTargetLowering::NVPTXTargetLowering(const NVPTXTargetMachine &TM,
   setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
   setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
   setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f32, Expand);
+  setLoadExtAction(ISD::EXTLOAD, MVT::v2f32, MVT::v2f16, Expand);
+  setLoadExtAction(ISD::EXTLOAD, MVT::v2f64, MVT::v2f16, Expand);
+  setLoadExtAction(ISD::EXTLOAD, MVT::v2f64, MVT::v2f32, Expand);
+  setLoadExtAction(ISD::EXTLOAD, MVT::v4f32, MVT::v4f16, Expand);
+  setLoadExtAction(ISD::EXTLOAD, MVT::v4f64, MVT::v4f16, Expand);
+  setLoadExtAction(ISD::EXTLOAD, MVT::v4f64, MVT::v4f32, Expand);
   // Turn FP truncstore into trunc + store.
+  // FIXME: vector types should also be expanded
   setTruncStoreAction(MVT::f32, MVT::f16, Expand);
   setTruncStoreAction(MVT::f64, MVT::f16, Expand);
   setTruncStoreAction(MVT::f64, MVT::f32, Expand);

diff --git a/test/CodeGen/NVPTX/extloadv.ll b/test/CodeGen/NVPTX/extloadv.ll
new file mode 100644 (file)
index 0000000..8c264ae
--- /dev/null
+++ b/test/CodeGen/NVPTX/extloadv.ll
@@ -0,0 +1,15 @@
+; RUN: llc < %s -march=nvptx64 -mcpu=sm_35 | FileCheck %s
+
+define void @foo(float* nocapture readonly %x_value, double* nocapture %output) #0 {
+  %1 = bitcast float* %x_value to <4 x float>*
+  %2 = load <4 x float>, <4 x float>* %1, align 16
+  %3 = fpext <4 x float> %2 to <4 x double>
+; CHECK-NOT: ld.v2.f32 {%fd{{[0-9]+}}, %fd{{[0-9]+}}}, [%rd{{[0-9]+}}];
+; CHECK:  cvt.f64.f32
+; CHECK:  cvt.f64.f32
+; CHECK:  cvt.f64.f32
+; CHECK:  cvt.f64.f32
+  %4 = bitcast double* %output to <4 x double>*
+  store <4 x double> %3, <4 x double>* %4
+  ret void
+}