[llvm] r363406 - AMDGPU: Fold readlane intrinsics of constants
Matt Arsenault via llvm-commits
llvm-commits at lists.llvm.org
Fri Jun 14 07:51:26 PDT 2019
Author: arsenm
Date: Fri Jun 14 07:51:26 2019
New Revision: 363406
URL: http://llvm.org/viewvc/llvm-project?rev=363406&view=rev
Log:
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.
Modified:
llvm/trunk/lib/Transforms/InstCombine/InstCombineCalls.cpp
llvm/trunk/test/Transforms/InstCombine/AMDGPU/amdgcn-intrinsics.ll
Modified: llvm/trunk/lib/Transforms/InstCombine/InstCombineCalls.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/InstCombine/InstCombineCalls.cpp?rev=363406&r1=363405&r2=363406&view=diff
==============================================================================
--- llvm/trunk/lib/Transforms/InstCombine/InstCombineCalls.cpp (original)
+++ llvm/trunk/lib/Transforms/InstCombine/InstCombineCalls.cpp Fri Jun 14 07:51:26 2019
@@ -3776,6 +3776,13 @@ Instruction *InstCombiner::visitCallInst
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.
Modified: llvm/trunk/test/Transforms/InstCombine/AMDGPU/amdgcn-intrinsics.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/InstCombine/AMDGPU/amdgcn-intrinsics.ll?rev=363406&r1=363405&r2=363406&view=diff
==============================================================================
--- llvm/trunk/test/Transforms/InstCombine/AMDGPU/amdgcn-intrinsics.ll (original)
+++ llvm/trunk/test/Transforms/InstCombine/AMDGPU/amdgcn-intrinsics.ll Fri Jun 14 07:51:26 2019
@@ -2432,6 +2432,62 @@ define void @kill_true() {
}
; --------------------------------------------------------------------
+; llvm.amdgcn.readfirstlane
+; --------------------------------------------------------------------
+
+declare i32 @llvm.amdgcn.readfirstlane(i32)
+
+ at 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
; --------------------------------------------------------------------
More information about the llvm-commits
mailing list