[llvm-branch-commits] [llvm] IR/AMDGPU: Autoupgrade amdgpu-unsafe-fp-atomics attribute (PR #101698)
Matt Arsenault via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Fri Aug 2 08:54:48 PDT 2024
https://github.com/arsenm created https://github.com/llvm/llvm-project/pull/101698
Delete the attribute and annotate any atomicrmw instructions in the
function with new metadata.
>From f19c1c2205115215567b7860a1a9fda2489eb114 Mon Sep 17 00:00:00 2001
From: Matt Arsenault <Matthew.Arsenault at amd.com>
Date: Thu, 18 Apr 2024 15:14:24 +0200
Subject: [PATCH] IR/AMDGPU: Autoupgrade amdgpu-unsafe-fp-atomics attribute
Delete the attribute and annotate any atomicrmw instructions in the
function with new metadata.
---
llvm/lib/IR/AutoUpgrade.cpp | 34 ++++++++
.../amdgpu-unsafe-fp-atomics-upgrade.ll | 80 +++++++++++++++++++
2 files changed, 114 insertions(+)
create mode 100644 llvm/test/Bitcode/amdgpu-unsafe-fp-atomics-upgrade.ll
diff --git a/llvm/lib/IR/AutoUpgrade.cpp b/llvm/lib/IR/AutoUpgrade.cpp
index 0e95c6df670dc..eda5c0f66349f 100644
--- a/llvm/lib/IR/AutoUpgrade.cpp
+++ b/llvm/lib/IR/AutoUpgrade.cpp
@@ -5260,6 +5260,22 @@ struct StrictFPUpgradeVisitor : public InstVisitor<StrictFPUpgradeVisitor> {
Call.addFnAttr(Attribute::NoBuiltin);
}
};
+
+/// Replace "amdgpu-unsafe-fp-atomics" metadata with atomicrmw metadata
+struct AMDGPUUnsafeFPAtomicsUpgradeVisitor
+ : public InstVisitor<AMDGPUUnsafeFPAtomicsUpgradeVisitor> {
+ AMDGPUUnsafeFPAtomicsUpgradeVisitor() = default;
+
+ void visitAtomicRMWInst(AtomicRMWInst &RMW) {
+ if (!RMW.isFloatingPointOperation())
+ return;
+
+ MDNode *Empty = MDNode::get(RMW.getContext(), {});
+ RMW.setMetadata("amdgpu.no.fine.grained.host.memory", Empty);
+ RMW.setMetadata("amdgpu.no.remote.memory.access", Empty);
+ RMW.setMetadata("amdgpu.ignore.denormal.mode", Empty);
+ }
+};
} // namespace
void llvm::UpgradeFunctionAttributes(Function &F) {
@@ -5282,6 +5298,24 @@ void llvm::UpgradeFunctionAttributes(Function &F) {
F.setSection(A.getValueAsString());
F.removeFnAttr("implicit-section-name");
}
+
+ if (!F.empty()) {
+ // For some reason this is called twice, and the first time is before any
+ // instructions are loaded into the body.
+
+ if (Attribute A = F.getFnAttribute("amdgpu-unsafe-fp-atomics");
+ A.isValid()) {
+
+ if (A.getValueAsBool()) {
+ AMDGPUUnsafeFPAtomicsUpgradeVisitor Visitor;
+ Visitor.visit(F);
+ }
+
+ // We will leave behind dead attribute uses on external declarations, but
+ // clang never added these to declarations anyway.
+ F.removeFnAttr("amdgpu-unsafe-fp-atomics");
+ }
+ }
}
static bool isOldLoopArgument(Metadata *MD) {
diff --git a/llvm/test/Bitcode/amdgpu-unsafe-fp-atomics-upgrade.ll b/llvm/test/Bitcode/amdgpu-unsafe-fp-atomics-upgrade.ll
new file mode 100644
index 0000000000000..ceac496d1e8dd
--- /dev/null
+++ b/llvm/test/Bitcode/amdgpu-unsafe-fp-atomics-upgrade.ll
@@ -0,0 +1,80 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals all --version 4
+; RUN: llvm-as < %s | llvm-dis | FileCheck %s
+
+; amdgpu-unsafe-fp-atomics attribute should be removed and replaced
+; with metadata attached to any atomicrmw with floating-point
+; operations.
+
+; Maybe the attribute should be dropped from declarations, but it
+; didn't do anything on one and clang never added it.
+declare void @unsafe_fp_atomics_true_decl() "amdgpu-unsafe-fp-atomics"="true"
+declare void @unsafe_fp_atomics_false_decl() "amdgpu-unsafe-fp-atomics"="false"
+
+; Delete the attribute and replace with the most aggressive metadata possible
+define void @unsafe_fp_atomics_true(ptr addrspace(1) %ptr, float %val, i32 %ival, <2 x half> %vval) "amdgpu-unsafe-fp-atomics"="true" {
+; CHECK-LABEL: define void @unsafe_fp_atomics_true(
+; CHECK-SAME: ptr addrspace(1) [[PTR:%.*]], float [[VAL:%.*]], i32 [[IVAL:%.*]], <2 x half> [[VVAL:%.*]]) {
+; CHECK-NEXT: [[RMW_FADD:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], float [[VAL]] syncscope("one-as") seq_cst, align 4, !amdgpu.no.fine.grained.host.memory [[META0:![0-9]+]], !amdgpu.no.remote.memory.access [[META0]], !amdgpu.ignore.denormal.mode [[META0]]
+; CHECK-NEXT: [[RMW_FSUB:%.*]] = atomicrmw fsub ptr addrspace(1) [[PTR]], float [[VAL]] syncscope("one-as") seq_cst, align 4, !amdgpu.no.fine.grained.host.memory [[META0]], !amdgpu.no.remote.memory.access [[META0]], !amdgpu.ignore.denormal.mode [[META0]]
+; CHECK-NEXT: [[RMW_FMIN:%.*]] = atomicrmw fmin ptr addrspace(1) [[PTR]], float [[VAL]] syncscope("one-as") seq_cst, align 4, !amdgpu.no.fine.grained.host.memory [[META0]], !amdgpu.no.remote.memory.access [[META0]], !amdgpu.ignore.denormal.mode [[META0]]
+; CHECK-NEXT: [[RMW_FMAX:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR]], float [[VAL]] syncscope("one-as") seq_cst, align 4, !amdgpu.no.fine.grained.host.memory [[META0]], !amdgpu.no.remote.memory.access [[META0]], !amdgpu.ignore.denormal.mode [[META0]]
+; CHECK-NEXT: [[RMW_XOR:%.*]] = atomicrmw xor ptr addrspace(1) [[PTR]], i32 [[IVAL]] syncscope("one-as") seq_cst, align 4
+; CHECK-NEXT: [[RMW_FADD_VECTOR:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], <2 x half> [[VVAL]] syncscope("one-as") seq_cst, align 4, !amdgpu.no.fine.grained.host.memory [[META0]], !amdgpu.no.remote.memory.access [[META0]], !amdgpu.ignore.denormal.mode [[META0]]
+; CHECK-NEXT: [[RMW_FSUB_VECTOR:%.*]] = atomicrmw fsub ptr addrspace(1) [[PTR]], <2 x half> [[VVAL]] syncscope("one-as") seq_cst, align 4, !amdgpu.no.fine.grained.host.memory [[META0]], !amdgpu.no.remote.memory.access [[META0]], !amdgpu.ignore.denormal.mode [[META0]]
+; CHECK-NEXT: [[RMW_FMIN_VECTOR:%.*]] = atomicrmw fmin ptr addrspace(1) [[PTR]], <2 x half> [[VVAL]] syncscope("one-as") seq_cst, align 4, !amdgpu.no.fine.grained.host.memory [[META0]], !amdgpu.no.remote.memory.access [[META0]], !amdgpu.ignore.denormal.mode [[META0]]
+; CHECK-NEXT: [[RMW_FMAX_VECTOR:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR]], <2 x half> [[VVAL]] syncscope("one-as") seq_cst, align 4, !amdgpu.no.fine.grained.host.memory [[META0]], !amdgpu.no.remote.memory.access [[META0]], !amdgpu.ignore.denormal.mode [[META0]]
+; CHECK-NEXT: [[RMW_XCHG:%.*]] = atomicrmw xchg ptr addrspace(1) [[PTR]], float [[VAL]] syncscope("one-as") seq_cst, align 4
+; CHECK-NEXT: ret void
+;
+ %rmw.fadd = atomicrmw fadd ptr addrspace(1) %ptr, float %val syncscope("one-as") seq_cst
+ %rmw.fsub = atomicrmw fsub ptr addrspace(1) %ptr, float %val syncscope("one-as") seq_cst
+ %rmw.fmin = atomicrmw fmin ptr addrspace(1) %ptr, float %val syncscope("one-as") seq_cst
+ %rmw.fmax = atomicrmw fmax ptr addrspace(1) %ptr, float %val syncscope("one-as") seq_cst
+ %rmw.xor = atomicrmw xor ptr addrspace(1) %ptr, i32 %ival syncscope("one-as") seq_cst
+ %rmw.fadd.vector = atomicrmw fadd ptr addrspace(1) %ptr, <2 x half> %vval syncscope("one-as") seq_cst
+ %rmw.fsub.vector = atomicrmw fsub ptr addrspace(1) %ptr, <2 x half> %vval syncscope("one-as") seq_cst
+ %rmw.fmin.vector = atomicrmw fmin ptr addrspace(1) %ptr, <2 x half> %vval syncscope("one-as") seq_cst
+ %rmw.fmax.vector = atomicrmw fmax ptr addrspace(1) %ptr, <2 x half> %vval syncscope("one-as") seq_cst
+
+ ; xchg doesn't need any metadata
+ %rmw.xchg = atomicrmw xchg ptr addrspace(1) %ptr, float %val syncscope("one-as") seq_cst
+ ret void
+}
+
+; Should just delete the effectless attribute if it exists
+define void @unsafe_fp_atomics_false(ptr addrspace(1) %ptr, float %val, i32 %ival, <2 x half> %vval) "amdgpu-unsafe-fp-atomics"="false" {
+; CHECK-LABEL: define void @unsafe_fp_atomics_false(
+; CHECK-SAME: ptr addrspace(1) [[PTR:%.*]], float [[VAL:%.*]], i32 [[IVAL:%.*]], <2 x half> [[VVAL:%.*]]) {
+; CHECK-NEXT: [[RMW_FADD:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], float [[VAL]] syncscope("one-as") seq_cst, align 4
+; CHECK-NEXT: [[RMW_FSUB:%.*]] = atomicrmw fsub ptr addrspace(1) [[PTR]], float [[VAL]] syncscope("one-as") seq_cst, align 4
+; CHECK-NEXT: [[RMW_FMIN:%.*]] = atomicrmw fmin ptr addrspace(1) [[PTR]], float [[VAL]] syncscope("one-as") seq_cst, align 4
+; CHECK-NEXT: [[RMW_FMAX:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR]], float [[VAL]] syncscope("one-as") seq_cst, align 4
+; CHECK-NEXT: [[RMW_XOR:%.*]] = atomicrmw xor ptr addrspace(1) [[PTR]], i32 [[IVAL]] syncscope("one-as") seq_cst, align 4
+; CHECK-NEXT: [[RMW_FADD_VECTOR:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], <2 x half> [[VVAL]] syncscope("one-as") seq_cst, align 4
+; CHECK-NEXT: [[RMW_FSUB_VECTOR:%.*]] = atomicrmw fsub ptr addrspace(1) [[PTR]], <2 x half> [[VVAL]] syncscope("one-as") seq_cst, align 4
+; CHECK-NEXT: [[RMW_FMIN_VECTOR:%.*]] = atomicrmw fmin ptr addrspace(1) [[PTR]], <2 x half> [[VVAL]] syncscope("one-as") seq_cst, align 4
+; CHECK-NEXT: [[RMW_FMAX_VECTOR:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR]], <2 x half> [[VVAL]] syncscope("one-as") seq_cst, align 4
+; CHECK-NEXT: [[RMW_XCHG:%.*]] = atomicrmw xchg ptr addrspace(1) [[PTR]], float [[VAL]] syncscope("one-as") seq_cst, align 4
+; CHECK-NEXT: ret void
+;
+ %rmw.fadd = atomicrmw fadd ptr addrspace(1) %ptr, float %val syncscope("one-as") seq_cst
+ %rmw.fsub = atomicrmw fsub ptr addrspace(1) %ptr, float %val syncscope("one-as") seq_cst
+ %rmw.fmin = atomicrmw fmin ptr addrspace(1) %ptr, float %val syncscope("one-as") seq_cst
+ %rmw.fmax = atomicrmw fmax ptr addrspace(1) %ptr, float %val syncscope("one-as") seq_cst
+ %rmw.xor = atomicrmw xor ptr addrspace(1) %ptr, i32 %ival syncscope("one-as") seq_cst
+ %rmw.fadd.vector = atomicrmw fadd ptr addrspace(1) %ptr, <2 x half> %vval syncscope("one-as") seq_cst
+ %rmw.fsub.vector = atomicrmw fsub ptr addrspace(1) %ptr, <2 x half> %vval syncscope("one-as") seq_cst
+ %rmw.fmin.vector = atomicrmw fmin ptr addrspace(1) %ptr, <2 x half> %vval syncscope("one-as") seq_cst
+ %rmw.fmax.vector = atomicrmw fmax ptr addrspace(1) %ptr, <2 x half> %vval syncscope("one-as") seq_cst
+
+ ; xchg doesn't need any metadata
+ %rmw.xchg = atomicrmw xchg ptr addrspace(1) %ptr, float %val syncscope("one-as") seq_cst
+ ret void
+}
+
+;.
+; CHECK: attributes #[[ATTR0:[0-9]+]] = { "amdgpu-unsafe-fp-atomics"="true" }
+; CHECK: attributes #[[ATTR1:[0-9]+]] = { "amdgpu-unsafe-fp-atomics"="false" }
+;.
+; CHECK: [[META0]] = !{}
+;.
More information about the llvm-branch-commits
mailing list