[llvm] AMDGPU: Validate processor and features in TargetID parsing (PR #209304)

[Matt Arsenault via llvm-commits]

================
@@ -2828,6 +2828,155 @@ TEST(TargetParserTest, testAMDGPUgetIsaVersionFromSubArch) {
             (AMDGPU::IsaVersion{0, 0, 0}));
 }
 
+TEST(TargetParserTest, testAMDGPUParseTargetIDString) {
+  using AMDGPU::TargetID;
+  using AMDGPU::TargetIDSetting;
+
+  // A well-formed target id parses, canonicalizing the processor and features.
+  {
+    std::optional<TargetID> TID =
+        TargetID::parseTargetIDString("amdgcn-amd-amdhsa-unknown-gfx90a");
+    ASSERT_TRUE(TID.has_value());
+    EXPECT_EQ(TID->getGPUKind(), AMDGPU::GK_GFX90A);
+    EXPECT_EQ(TID->getXnackSetting(), TargetIDSetting::Any);
+    EXPECT_EQ(TID->getSramEccSetting(), TargetIDSetting::Any);
+  }
+
+  // Explicit feature modifiers are applied.
+  {
+    std::optional<TargetID> TID = TargetID::parseTargetIDString(
+        "amdgcn-amd-amdhsa-unknown-gfx90a:xnack+:sramecc-");
+    ASSERT_TRUE(TID.has_value());
+    EXPECT_EQ(TID->getXnackSetting(), TargetIDSetting::On);
+    EXPECT_EQ(TID->getSramEccSetting(), TargetIDSetting::Off);
+  }
+
+  // The processor+features field may be empty; the ISA is taken from the
+  // triple subarch.
+  EXPECT_TRUE(TargetID::parseTargetIDString("amdgpu9.0a-amd-amdhsa-unknown-")
+                  .has_value());
+
+  // Structurally malformed strings (missing the processor+features field or a
+  // non-AMDGCN triple) are rejected.
+  EXPECT_FALSE(TargetID::parseTargetIDString("not-a-valid-target-id"));
+  EXPECT_FALSE(
+      TargetID::parseTargetIDString("x86_64-unknown-linux-gnu-gfx90a"));
+
+  // An unrecognized processor is rejected.
+  EXPECT_FALSE(
+      TargetID::parseTargetIDString("amdgcn-amd-amdhsa-unknown-gfxbogus"));
+
+  // A feature the processor does not support is rejected: gfx600 has neither
+  // xnack nor sramecc.
+  EXPECT_FALSE(
+      TargetID::parseTargetIDString("amdgcn-amd-amdhsa-unknown-gfx600:xnack+"));
+  EXPECT_FALSE(TargetID::parseTargetIDString(
+      "amdgcn-amd-amdhsa-unknown-gfx900:sramecc+"));
----------------
arsenm wrote:

There isn't a need for a new test, this would already be covered. There isn't an encoding change here. The xnack always targets always emit EF_AMDGPU_FEATURE_XNACK_ANY_V4 (though I'd consider that a bug, they shouldn't emit any xnack flag)

https://github.com/llvm/llvm-project/pull/209304