[llvm] AtomicExpand: Preserve metadata when expanding partword RMW (PR #89769)
Matt Arsenault via llvm-commits
llvm-commits at lists.llvm.org
Tue May 7 10:12:13 PDT 2024
https://github.com/arsenm updated https://github.com/llvm/llvm-project/pull/89769
>From bc3284d533fc91e93d36726808189fceb911700f Mon Sep 17 00:00:00 2001
From: Matt Arsenault <Matthew.Arsenault at amd.com>
Date: Mon, 22 Apr 2024 23:35:11 +0200
Subject: [PATCH] AtomicExpand: Preserve metadata when expanding partword RMW
This will be important for AMDGPU in a future patch.
---
llvm/lib/CodeGen/AtomicExpandPass.cpp | 5 ++++-
.../AtomicExpand/AMDGPU/expand-atomic-i16.ll | 12 ++++++------
2 files changed, 10 insertions(+), 7 deletions(-)
diff --git a/llvm/lib/CodeGen/AtomicExpandPass.cpp b/llvm/lib/CodeGen/AtomicExpandPass.cpp
index ee44e9353d04..f1bc7262150b 100644
--- a/llvm/lib/CodeGen/AtomicExpandPass.cpp
+++ b/llvm/lib/CodeGen/AtomicExpandPass.cpp
@@ -965,7 +965,10 @@ AtomicRMWInst *AtomicExpandImpl::widenPartwordAtomicRMW(AtomicRMWInst *AI) {
AtomicRMWInst *NewAI = Builder.CreateAtomicRMW(
Op, PMV.AlignedAddr, NewOperand, PMV.AlignedAddrAlignment,
AI->getOrdering(), AI->getSyncScopeID());
- // TODO: Preserve metadata
+
+ // TODO: Do we need to drop noundef? We widened the operation and could be
+ // loading undefined bits.
+ NewAI->copyMetadata(*AI);
Value *FinalOldResult = extractMaskedValue(Builder, NewAI, PMV);
AI->replaceAllUsesWith(FinalOldResult);
diff --git a/llvm/test/Transforms/AtomicExpand/AMDGPU/expand-atomic-i16.ll b/llvm/test/Transforms/AtomicExpand/AMDGPU/expand-atomic-i16.ll
index ce8524c70af6..bdf5c4c6129a 100644
--- a/llvm/test/Transforms/AtomicExpand/AMDGPU/expand-atomic-i16.ll
+++ b/llvm/test/Transforms/AtomicExpand/AMDGPU/expand-atomic-i16.ll
@@ -189,7 +189,7 @@ define i16 @test_atomicrmw_and_i16_global_agent_preserve_md(ptr addrspace(1) %pt
; CHECK-NEXT: [[TMP3:%.*]] = zext i16 [[VALUE:%.*]] to i32
; CHECK-NEXT: [[VALOPERAND_SHIFTED:%.*]] = shl i32 [[TMP3]], [[SHIFTAMT]]
; CHECK-NEXT: [[ANDOPERAND:%.*]] = or i32 [[VALOPERAND_SHIFTED]], [[INV_MASK]]
-; CHECK-NEXT: [[TMP4:%.*]] = atomicrmw and ptr addrspace(1) [[ALIGNEDADDR]], i32 [[ANDOPERAND]] syncscope("agent") seq_cst, align 4
+; CHECK-NEXT: [[TMP4:%.*]] = atomicrmw and ptr addrspace(1) [[ALIGNEDADDR]], i32 [[ANDOPERAND]] syncscope("agent") seq_cst, align 4, !noundef [[META0:![0-9]+]], !some.unknown.md [[META0]]
; CHECK-NEXT: [[SHIFTED:%.*]] = lshr i32 [[TMP4]], [[SHIFTAMT]]
; CHECK-NEXT: [[EXTRACTED:%.*]] = trunc i32 [[SHIFTED]] to i16
; CHECK-NEXT: ret i16 [[EXTRACTED]]
@@ -203,7 +203,7 @@ define i16 @test_atomicrmw_and_i16_global_agent_align4_preserve_md(ptr addrspace
; CHECK-LABEL: @test_atomicrmw_and_i16_global_agent_align4_preserve_md(
; CHECK-NEXT: [[TMP1:%.*]] = zext i16 [[VALUE:%.*]] to i32
; CHECK-NEXT: [[ANDOPERAND:%.*]] = or i32 [[TMP1]], -65536
-; CHECK-NEXT: [[TMP2:%.*]] = atomicrmw and ptr addrspace(1) [[PTR:%.*]], i32 [[ANDOPERAND]] syncscope("agent") seq_cst, align 4
+; CHECK-NEXT: [[TMP2:%.*]] = atomicrmw and ptr addrspace(1) [[PTR:%.*]], i32 [[ANDOPERAND]] syncscope("agent") seq_cst, align 4, !noundef [[META0]], !some.unknown.md [[META0]]
; CHECK-NEXT: [[EXTRACTED:%.*]] = trunc i32 [[TMP2]] to i16
; CHECK-NEXT: ret i16 [[EXTRACTED]]
;
@@ -223,7 +223,7 @@ define i16 @test_atomicrmw_and_i16_global_agent__amdgpu_no_remote_memory(ptr add
; CHECK-NEXT: [[TMP3:%.*]] = zext i16 [[VALUE:%.*]] to i32
; CHECK-NEXT: [[VALOPERAND_SHIFTED:%.*]] = shl i32 [[TMP3]], [[SHIFTAMT]]
; CHECK-NEXT: [[ANDOPERAND:%.*]] = or i32 [[VALOPERAND_SHIFTED]], [[INV_MASK]]
-; CHECK-NEXT: [[TMP4:%.*]] = atomicrmw and ptr addrspace(1) [[ALIGNEDADDR]], i32 [[ANDOPERAND]] syncscope("agent") seq_cst, align 4
+; CHECK-NEXT: [[TMP4:%.*]] = atomicrmw and ptr addrspace(1) [[ALIGNEDADDR]], i32 [[ANDOPERAND]] syncscope("agent") seq_cst, align 4, !amdgpu.no.remote.memory [[META0]]
; CHECK-NEXT: [[SHIFTED:%.*]] = lshr i32 [[TMP4]], [[SHIFTAMT]]
; CHECK-NEXT: [[EXTRACTED:%.*]] = trunc i32 [[SHIFTED]] to i16
; CHECK-NEXT: ret i16 [[EXTRACTED]]
@@ -236,7 +236,7 @@ define i16 @test_atomicrmw_and_i16_global_agent_align4__amdgpu_no_remote_memory(
; CHECK-LABEL: @test_atomicrmw_and_i16_global_agent_align4__amdgpu_no_remote_memory(
; CHECK-NEXT: [[TMP1:%.*]] = zext i16 [[VALUE:%.*]] to i32
; CHECK-NEXT: [[ANDOPERAND:%.*]] = or i32 [[TMP1]], -65536
-; CHECK-NEXT: [[TMP2:%.*]] = atomicrmw and ptr addrspace(1) [[PTR:%.*]], i32 [[ANDOPERAND]] syncscope("agent") seq_cst, align 4
+; CHECK-NEXT: [[TMP2:%.*]] = atomicrmw and ptr addrspace(1) [[PTR:%.*]], i32 [[ANDOPERAND]] syncscope("agent") seq_cst, align 4, !amdgpu.no.remote.memory [[META0]]
; CHECK-NEXT: [[EXTRACTED:%.*]] = trunc i32 [[TMP2]] to i16
; CHECK-NEXT: ret i16 [[EXTRACTED]]
;
@@ -256,7 +256,7 @@ define i16 @test_atomicrmw_and_i16_global_agent__amdgpu_no_fine_grained_memory(p
; CHECK-NEXT: [[TMP3:%.*]] = zext i16 [[VALUE:%.*]] to i32
; CHECK-NEXT: [[VALOPERAND_SHIFTED:%.*]] = shl i32 [[TMP3]], [[SHIFTAMT]]
; CHECK-NEXT: [[ANDOPERAND:%.*]] = or i32 [[VALOPERAND_SHIFTED]], [[INV_MASK]]
-; CHECK-NEXT: [[TMP4:%.*]] = atomicrmw and ptr addrspace(1) [[ALIGNEDADDR]], i32 [[ANDOPERAND]] syncscope("agent") seq_cst, align 4
+; CHECK-NEXT: [[TMP4:%.*]] = atomicrmw and ptr addrspace(1) [[ALIGNEDADDR]], i32 [[ANDOPERAND]] syncscope("agent") seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]]
; CHECK-NEXT: [[SHIFTED:%.*]] = lshr i32 [[TMP4]], [[SHIFTAMT]]
; CHECK-NEXT: [[EXTRACTED:%.*]] = trunc i32 [[SHIFTED]] to i16
; CHECK-NEXT: ret i16 [[EXTRACTED]]
@@ -269,7 +269,7 @@ define i16 @test_atomicrmw_and_i16_global_agent_align4__amdgpu_no_fine_grained_m
; CHECK-LABEL: @test_atomicrmw_and_i16_global_agent_align4__amdgpu_no_fine_grained_memory(
; CHECK-NEXT: [[TMP1:%.*]] = zext i16 [[VALUE:%.*]] to i32
; CHECK-NEXT: [[ANDOPERAND:%.*]] = or i32 [[TMP1]], -65536
-; CHECK-NEXT: [[TMP2:%.*]] = atomicrmw and ptr addrspace(1) [[PTR:%.*]], i32 [[ANDOPERAND]] syncscope("agent") seq_cst, align 4
+; CHECK-NEXT: [[TMP2:%.*]] = atomicrmw and ptr addrspace(1) [[PTR:%.*]], i32 [[ANDOPERAND]] syncscope("agent") seq_cst, align 4, !amdgpu.no.fine.grained.memory [[META0]]
; CHECK-NEXT: [[EXTRACTED:%.*]] = trunc i32 [[TMP2]] to i16
; CHECK-NEXT: ret i16 [[EXTRACTED]]
;
More information about the llvm-commits
mailing list