[clang] [llvm] Adding splitdouble HLSL function (PR #109331)

[Tex Riddell via llvm-commits]

================
@@ -0,0 +1,98 @@
+// RUN: %clang_cc1 -finclude-default-header -x hlsl -triple dxil-pc-shadermodel6.3-library %s -fnative-half-type -emit-llvm -O1 -o - | FileCheck %s
+// RUN: %clang_cc1 -finclude-default-header -x hlsl -triple spirv-vulkan-library %s -fnative-half-type -emit-llvm -O0 -o - | FileCheck %s --check-prefix=SPIRV
+
+
+
+// CHECK: define {{.*}} i32 {{.*}}test_scalar{{.*}}(double {{.*}} [[VALD:%.*]])
+// CHECK: [[VALRET:%.*]] = {{.*}} call { i32, i32 } @llvm.dx.splitdouble.i32(double [[VALD]])
+// CHECK-NEXT: extractvalue { i32, i32 } [[VALRET]], 0
+// CHECK-NEXT: extractvalue { i32, i32 } [[VALRET]], 1
+// SPIRV: define spir_func {{.*}} i32 {{.*}}test_scalar{{.*}}(double {{.*}} [[VALD:%.*]])
+// SPIRV-NOT: @llvm.dx.splitdouble.i32
+// SPIRV: [[REG:%.*]] = load double, ptr [[VALD]].addr, align 8
+// SPIRV-NEXT: [[CAST:%.*]] = bitcast double [[REG]] to <2 x i32>
+// SPIRV-NEXT: extractelement <2 x i32> [[CAST]], i64 0
+// SPIRV-NEXT: extractelement <2 x i32> [[CAST]], i64 1
+uint test_scalar(double D) {
+  uint A, B;
+  asuint(D, A, B);
+  return A + B;
+}
+
+// CHECK: define {{.*}} i32 {{.*}}test_double1{{.*}}(<1 x double> {{.*}} [[VALD:%.*]])
+// CHECK: [[TRUNC:%.*]] = extractelement <1 x double> %D, i64 0
+// CHECK-NEXT: [[VALRET:%.*]] = {{.*}} call { i32, i32 } @llvm.dx.splitdouble.i32(double [[TRUNC]])
+// CHECK-NEXT: extractvalue { i32, i32 } [[VALRET]], 0
+// CHECK-NEXT: extractvalue { i32, i32 } [[VALRET]], 1
+// SPIRV: define spir_func {{.*}} i32 {{.*}}test_double1{{.*}}(<1 x double> {{.*}} [[VALD:%.*]])
+// SPIRV-NOT: @llvm.dx.splitdouble.i32
+// SPIRV: [[REG:%.*]] = load <1 x double>, ptr [[VALD]].addr, align 8
+// SPIRV-NEXT: [[TRUNC:%.*]] = extractelement <1 x double> %1, i64 0
----------------
tex3d wrote:

@farzonl I wasn't suggesting that we run `spirv-val` in *this* test.  I was suggesting that this generated IR should also be tested in the other test (`llvm/test/CodeGen/SPIRV/hlsl-intrinsics/splitdouble.ll`) where we test how the generated IR patterns lower to SPIR-V and *do* run `spirv-val`.

In reality, I'm expecting that the vector size 1 case is problematic for SPIR-V target. I still consider this comment resolved because CGBuiltin is doing it's best to operate with scalars, but the incoming type is `<1 x double>` which will generate an illegal type in SPIR-V.

Ultimately, I think we need to sort out our vector size 1 plan more thoroughly, rather than generate and test something speculative that won't lower to legal code, and/or won't be how we want to represent vector1 in IR in the first place.

https://github.com/llvm/llvm-project/pull/109331