Using CallInst directly actually works

Author: arsenm

llvm/lib/Target/AMDGPU/AMDGPUInstCombineIntrinsic.cpp

@@ -1134,18 +1134,19 @@ GCNTTIImpl::instCombineIntrinsic(InstCombiner &IC, IntrinsicInst &II) const {
 
       // TODO: Handle this for update_dpp, mov_ddp8, and all permlane variants.
       if (isTypeLegal(BCSrc->getType())) {
+        Module *M = IC.Builder.GetInsertBlock()->getModule();
+        Function *Remangled =
+            Intrinsic::getOrInsertDeclaration(M, IID, {BCSrc->getType()});
+
         // Make sure convergence tokens are preserved.
         // TODO: CreateIntrinsic should allow directly copying bundles
         SmallVector<OperandBundleDef, 2> OpBundles;
         II.getOperandBundlesAsDefs(OpBundles);
 
-        IRBuilderBase::OperandBundlesGuard Guard(IC.Builder);
-        IC.Builder.setDefaultOperandBundles(OpBundles);
-
         SmallVector<Value *, 3> Args(II.args());
         Args[0] = BCSrc;
-        CallInst *NewCall = IC.Builder.CreateIntrinsic(
-            II.getIntrinsicID(), {BCSrc->getType()}, Args);
+
+        CallInst *NewCall = IC.Builder.CreateCall(Remangled, Args, OpBundles);
         NewCall->takeName(&II);
         return new BitCastInst(NewCall, II.getType());
       }

llvm/test/Transforms/InstCombine/AMDGPU/bitcast-fold-lane-ops.ll

@@ -288,8 +288,8 @@ define i32 @test_bitcast_f32_to_i32_readfirstlane_convergencetoken(float %val) c
 ; CHECK-LABEL: define i32 @test_bitcast_f32_to_i32_readfirstlane_convergencetoken(
 ; CHECK-SAME: float [[VAL:%.*]]) #[[ATTR1]] {
 ; CHECK-NEXT:    [[T:%.*]] = call token @llvm.experimental.convergence.entry()
-; CHECK-NEXT:    [[BITCAST:%.*]] = bitcast float [[VAL]] to i32
-; CHECK-NEXT:    [[RESULT:%.*]] = call i32 @llvm.amdgcn.readfirstlane.i32(i32 [[BITCAST]]) [ "convergencectrl"(token [[T]]) ]
+; CHECK-NEXT:    [[RESULT1:%.*]] = call float @llvm.amdgcn.readfirstlane.f32(float [[VAL]]) [ "convergencectrl"(token [[T]]) ]
+; CHECK-NEXT:    [[RESULT:%.*]] = bitcast float [[RESULT1]] to i32
 ; CHECK-NEXT:    ret i32 [[RESULT]]
 ;
   %t = call token @llvm.experimental.convergence.entry()
@@ -302,8 +302,8 @@ define i32 @test_bitcast_f32_to_i32_readlane_convergencetoken(float %val, i32 in
 ; CHECK-LABEL: define i32 @test_bitcast_f32_to_i32_readlane_convergencetoken(
 ; CHECK-SAME: float [[VAL:%.*]], i32 inreg [[LANE_INDEX:%.*]]) #[[ATTR1]] {
 ; CHECK-NEXT:    [[T:%.*]] = call token @llvm.experimental.convergence.entry()
-; CHECK-NEXT:    [[BITCAST:%.*]] = bitcast float [[VAL]] to i32
-; CHECK-NEXT:    [[RESULT:%.*]] = call i32 @llvm.amdgcn.readlane.i32(i32 [[BITCAST]], i32 [[LANE_INDEX]]) [ "convergencectrl"(token [[T]]) ]
+; CHECK-NEXT:    [[RESULT1:%.*]] = call float @llvm.amdgcn.readlane.f32(float [[VAL]], i32 [[LANE_INDEX]]) [ "convergencectrl"(token [[T]]) ]
+; CHECK-NEXT:    [[RESULT:%.*]] = bitcast float [[RESULT1]] to i32
 ; CHECK-NEXT:    ret i32 [[RESULT]]
 ;
   %t = call token @llvm.experimental.convergence.entry()

llvm/test/Transforms/InstCombine/AMDGPU/permlane64.ll

@@ -24,9 +24,9 @@ define i32 @test_bitcast_f32_to_i32_permlane64(float %val) {
 define i32 @test_bitcast_f32_to_i32_permlane64_convergencetokenn(float %val) convergent {
 ; CHECK-LABEL: @test_bitcast_f32_to_i32_permlane64_convergencetokenn(
 ; CHECK-NEXT:    [[T:%.*]] = call token @llvm.experimental.convergence.entry()
-; CHECK-NEXT:    [[BITCAST:%.*]] = bitcast float [[VAL:%.*]] to i32
-; CHECK-NEXT:    [[RESULT:%.*]] = call i32 @llvm.amdgcn.permlane64.i32(i32 [[BITCAST]]) [ "convergencectrl"(token [[T]]) ]
-; CHECK-NEXT:    ret i32 [[RESULT]]
+; CHECK-NEXT:    [[VAL:%.*]] = call float @llvm.amdgcn.permlane64.f32(float [[VAL1:%.*]]) [ "convergencectrl"(token [[T]]) ]
+; CHECK-NEXT:    [[BITCAST:%.*]] = bitcast float [[VAL]] to i32
+; CHECK-NEXT:    ret i32 [[BITCAST]]
 ;
   %t = call token @llvm.experimental.convergence.entry()
   %bitcast = bitcast float %val to i32