AMDGPU: Fold readlane intrinsics of constants

I'm not 100% sure about this, since I'm worried about IR transforms
that might end up introducing divergence downstream once replaced with
a constant, but I haven't come up with an example yet.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363406 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Transforms/InstCombine/InstCombineCalls.cpp b/lib/Transforms/InstCombine/InstCombineCalls.cpp
index f167762b60203226e4a7e80db1d573f888348a51..347aeb6989a7c45d2068bd021e22f001b21cf74f 100644 (file)
--- a/lib/Transforms/InstCombine/InstCombineCalls.cpp
+++ b/lib/Transforms/InstCombine/InstCombineCalls.cpp
@@ -3776,6 +3776,13 @@ Instruction *InstCombiner::visitCallInst(CallInst &CI) {
     II->setOperand(0, UndefValue::get(Old->getType()));
     return II;
   }
+  case Intrinsic::amdgcn_readfirstlane:
+  case Intrinsic::amdgcn_readlane: {
+    // A constant value is trivially uniform.
+    if (Constant *C = dyn_cast<Constant>(II->getArgOperand(0)))
+      return replaceInstUsesWith(*II, C);
+    break;
+  }
   case Intrinsic::stackrestore: {
     // If the save is right next to the restore, remove the restore.  This can
     // happen when variable allocas are DCE'd.

diff --git a/test/Transforms/InstCombine/AMDGPU/amdgcn-intrinsics.ll b/test/Transforms/InstCombine/AMDGPU/amdgcn-intrinsics.ll
index 6c5c02569e588ed5b904bdbe19a96d7d28ae1af5..1d54d78c57dad01497a6232414eb0553daf5371f 100644 (file)
--- a/test/Transforms/InstCombine/AMDGPU/amdgcn-intrinsics.ll
+++ b/test/Transforms/InstCombine/AMDGPU/amdgcn-intrinsics.ll
@@ -2431,6 +2431,62 @@ define void @kill_true() {
   ret void
 }
 
+; --------------------------------------------------------------------
+; llvm.amdgcn.readfirstlane
+; --------------------------------------------------------------------
+
+declare i32 @llvm.amdgcn.readfirstlane(i32)
+
+@gv = constant i32 0
+
+define amdgpu_kernel void @readfirstlane_constant(i32 %arg) {
+; CHECK-LABEL: @readfirstlane_constant(
+; CHECK-NEXT: %var = call i32 @llvm.amdgcn.readfirstlane(i32 %arg)
+; CHECK-NEXT: store volatile i32 %var, i32* undef, align 4
+; CHECK-NEXT: store volatile i32 0, i32* undef, align 4
+; CHECK-NEXT: store volatile i32 123, i32* undef, align 4
+; CHECK-NEXT: store volatile i32 ptrtoint (i32* @gv to i32), i32* undef, align 4
+; CHECK-NEXT: store volatile i32 undef, i32* undef, align 4
+  %var = call i32 @llvm.amdgcn.readfirstlane(i32 %arg)
+  %zero = call i32 @llvm.amdgcn.readfirstlane(i32 0)
+  %imm = call i32 @llvm.amdgcn.readfirstlane(i32 123)
+  %constexpr = call i32 @llvm.amdgcn.readfirstlane(i32 ptrtoint (i32* @gv to i32))
+  %undef = call i32 @llvm.amdgcn.readfirstlane(i32 undef)
+  store volatile i32 %var, i32* undef
+  store volatile i32 %zero, i32* undef
+  store volatile i32 %imm, i32* undef
+  store volatile i32 %constexpr, i32* undef
+  store volatile i32 %undef, i32* undef
+  ret void
+}
+
+; --------------------------------------------------------------------
+; llvm.amdgcn.readlane
+; --------------------------------------------------------------------
+
+declare i32 @llvm.amdgcn.readlane(i32, i32)
+
+define amdgpu_kernel void @readlane_constant(i32 %arg, i32 %lane) {
+; CHECK-LABEL: @readlane_constant(
+; CHECK-NEXT: %var = call i32 @llvm.amdgcn.readlane(i32 %arg, i32 7)
+; CHECK-NEXT: store volatile i32 %var, i32* undef, align 4
+; CHECK-NEXT: store volatile i32 0, i32* undef, align 4
+; CHECK-NEXT: store volatile i32 123, i32* undef, align 4
+; CHECK-NEXT: store volatile i32 ptrtoint (i32* @gv to i32), i32* undef, align 4
+; CHECK-NEXT: store volatile i32 undef, i32* undef, align 4
+  %var = call i32 @llvm.amdgcn.readlane(i32 %arg, i32 7)
+  %zero = call i32 @llvm.amdgcn.readlane(i32 0, i32 %lane)
+  %imm = call i32 @llvm.amdgcn.readlane(i32 123, i32 %lane)
+  %constexpr = call i32 @llvm.amdgcn.readlane(i32 ptrtoint (i32* @gv to i32), i32 %lane)
+  %undef = call i32 @llvm.amdgcn.readlane(i32 undef, i32 %lane)
+  store volatile i32 %var, i32* undef
+  store volatile i32 %zero, i32* undef
+  store volatile i32 %imm, i32* undef
+  store volatile i32 %constexpr, i32* undef
+  store volatile i32 %undef, i32* undef
+  ret void
+}
+
 ; --------------------------------------------------------------------
 ; llvm.amdgcn.update.dpp.i32
 ; --------------------------------------------------------------------