[llvm-branch-commits] [llvm] release/19.x: FastISel: Fix incorrectly using getPointerTy (#110465) (PR #110490)
via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Mon Sep 30 04:02:25 PDT 2024
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-backend-x86
Author: None (llvmbot)
<details>
<summary>Changes</summary>
Backport 81ba95cefe1b5a12f0a7d8e6a383bcce9e95b785
Requested by: @<!-- -->arsenm
---
Full diff: https://github.com/llvm/llvm-project/pull/110490.diff
4 Files Affected:
- (modified) llvm/include/llvm/CodeGen/FastISel.h (+1-1)
- (modified) llvm/lib/CodeGen/SelectionDAG/FastISel.cpp (+4-4)
- (modified) llvm/lib/Target/X86/X86FastISel.cpp (+3-1)
- (added) llvm/test/CodeGen/X86/issue56055.ll (+81)
``````````diff
diff --git a/llvm/include/llvm/CodeGen/FastISel.h b/llvm/include/llvm/CodeGen/FastISel.h
index 3cbc35400181dd..f3c4cc8d0511d4 100644
--- a/llvm/include/llvm/CodeGen/FastISel.h
+++ b/llvm/include/llvm/CodeGen/FastISel.h
@@ -275,7 +275,7 @@ class FastISel {
/// This is a wrapper around getRegForValue that also takes care of
/// truncating or sign-extending the given getelementptr index value.
- Register getRegForGEPIndex(const Value *Idx);
+ Register getRegForGEPIndex(MVT PtrVT, const Value *Idx);
/// We're checking to see if we can fold \p LI into \p FoldInst. Note
/// that we could have a sequence where multiple LLVM IR instructions are
diff --git a/llvm/lib/CodeGen/SelectionDAG/FastISel.cpp b/llvm/lib/CodeGen/SelectionDAG/FastISel.cpp
index ef9f7833551905..246acc7f405837 100644
--- a/llvm/lib/CodeGen/SelectionDAG/FastISel.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/FastISel.cpp
@@ -380,14 +380,13 @@ void FastISel::updateValueMap(const Value *I, Register Reg, unsigned NumRegs) {
}
}
-Register FastISel::getRegForGEPIndex(const Value *Idx) {
+Register FastISel::getRegForGEPIndex(MVT PtrVT, const Value *Idx) {
Register IdxN = getRegForValue(Idx);
if (!IdxN)
// Unhandled operand. Halt "fast" selection and bail.
return Register();
// If the index is smaller or larger than intptr_t, truncate or extend it.
- MVT PtrVT = TLI.getPointerTy(DL);
EVT IdxVT = EVT::getEVT(Idx->getType(), /*HandleUnknown=*/false);
if (IdxVT.bitsLT(PtrVT)) {
IdxN = fastEmit_r(IdxVT.getSimpleVT(), PtrVT, ISD::SIGN_EXTEND, IdxN);
@@ -543,7 +542,8 @@ bool FastISel::selectGetElementPtr(const User *I) {
uint64_t TotalOffs = 0;
// FIXME: What's a good SWAG number for MaxOffs?
uint64_t MaxOffs = 2048;
- MVT VT = TLI.getPointerTy(DL);
+ MVT VT = TLI.getValueType(DL, I->getType()).getSimpleVT();
+
for (gep_type_iterator GTI = gep_type_begin(I), E = gep_type_end(I);
GTI != E; ++GTI) {
const Value *Idx = GTI.getOperand();
@@ -584,7 +584,7 @@ bool FastISel::selectGetElementPtr(const User *I) {
// N = N + Idx * ElementSize;
uint64_t ElementSize = GTI.getSequentialElementStride(DL);
- Register IdxN = getRegForGEPIndex(Idx);
+ Register IdxN = getRegForGEPIndex(VT, Idx);
if (!IdxN) // Unhandled operand. Halt "fast" selection and bail.
return false;
diff --git a/llvm/lib/Target/X86/X86FastISel.cpp b/llvm/lib/Target/X86/X86FastISel.cpp
index 2eae155956368f..5d594bd54fbfc4 100644
--- a/llvm/lib/Target/X86/X86FastISel.cpp
+++ b/llvm/lib/Target/X86/X86FastISel.cpp
@@ -902,6 +902,8 @@ bool X86FastISel::X86SelectAddress(const Value *V, X86AddressMode &AM) {
uint64_t Disp = (int32_t)AM.Disp;
unsigned IndexReg = AM.IndexReg;
unsigned Scale = AM.Scale;
+ MVT PtrVT = TLI.getValueType(DL, U->getType()).getSimpleVT();
+
gep_type_iterator GTI = gep_type_begin(U);
// Iterate through the indices, folding what we can. Constants can be
// folded, and one dynamic index can be handled, if the scale is supported.
@@ -937,7 +939,7 @@ bool X86FastISel::X86SelectAddress(const Value *V, X86AddressMode &AM) {
(S == 1 || S == 2 || S == 4 || S == 8)) {
// Scaled-index addressing.
Scale = S;
- IndexReg = getRegForGEPIndex(Op);
+ IndexReg = getRegForGEPIndex(PtrVT, Op);
if (IndexReg == 0)
return false;
break;
diff --git a/llvm/test/CodeGen/X86/issue56055.ll b/llvm/test/CodeGen/X86/issue56055.ll
new file mode 100644
index 00000000000000..27eaf13e3b00be
--- /dev/null
+++ b/llvm/test/CodeGen/X86/issue56055.ll
@@ -0,0 +1,81 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -fast-isel < %s | FileCheck -check-prefixes=CHECK,FASTISEL %s
+; RUN: llc < %s | FileCheck -check-prefixes=CHECK,SDAG %s
+
+target datalayout = "e-m:w-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-windows-msvc"
+
+define void @issue56055(ptr addrspace(270) %ptr, ptr %out) {
+; CHECK-LABEL: issue56055:
+; CHECK: # %bb.0:
+; CHECK-NEXT: addl $2, %ecx
+; CHECK-NEXT: movl %ecx, (%rdx)
+; CHECK-NEXT: retq
+ %add.ptr = getelementptr inbounds i8, ptr addrspace(270) %ptr, i32 2
+ store ptr addrspace(270) %add.ptr, ptr %out
+ ret void
+}
+
+define void @issue56055_vector(<2 x ptr addrspace(270)> %ptr, ptr %out) {
+; CHECK-LABEL: issue56055_vector:
+; CHECK: # %bb.0:
+; CHECK-NEXT: movdqa (%rcx), %xmm0
+; CHECK-NEXT: paddd __xmm at 00000000000000000000000200000002(%rip), %xmm0
+; CHECK-NEXT: movq %xmm0, (%rdx)
+; CHECK-NEXT: retq
+ %add.ptr = getelementptr inbounds i8, <2 x ptr addrspace(270)> %ptr, <2 x i32> <i32 2, i32 2>
+ store <2 x ptr addrspace(270)> %add.ptr, ptr %out
+ ret void
+}
+
+define void @issue56055_small_idx(ptr addrspace(270) %ptr, ptr %out, i16 %idx) {
+; CHECK-LABEL: issue56055_small_idx:
+; CHECK: # %bb.0:
+; CHECK-NEXT: movswl %r8w, %eax
+; CHECK-NEXT: addl %ecx, %eax
+; CHECK-NEXT: movl %eax, (%rdx)
+; CHECK-NEXT: retq
+ %add.ptr = getelementptr inbounds i8, ptr addrspace(270) %ptr, i16 %idx
+ store ptr addrspace(270) %add.ptr, ptr %out
+ ret void
+}
+
+define void @issue56055_small_idx_vector(<2 x ptr addrspace(270)> %ptr, ptr %out, <2 x i16> %idx) {
+; CHECK-LABEL: issue56055_small_idx_vector:
+; CHECK: # %bb.0:
+; CHECK-NEXT: pshuflw {{.*#+}} xmm0 = mem[0,0,2,1,4,5,6,7]
+; CHECK-NEXT: psrad $16, %xmm0
+; CHECK-NEXT: paddd (%rcx), %xmm0
+; CHECK-NEXT: movq %xmm0, (%rdx)
+; CHECK-NEXT: retq
+ %add.ptr = getelementptr inbounds i8, <2 x ptr addrspace(270)> %ptr, <2 x i16> %idx
+ store <2 x ptr addrspace(270)> %add.ptr, ptr %out
+ ret void
+}
+
+define void @issue56055_large_idx(ptr addrspace(270) %ptr, ptr %out, i64 %idx) {
+; CHECK-LABEL: issue56055_large_idx:
+; CHECK: # %bb.0:
+; CHECK-NEXT: addl %ecx, %r8d
+; CHECK-NEXT: movl %r8d, (%rdx)
+; CHECK-NEXT: retq
+ %add.ptr = getelementptr inbounds i8, ptr addrspace(270) %ptr, i64 %idx
+ store ptr addrspace(270) %add.ptr, ptr %out
+ ret void
+}
+
+define void @issue56055_large_idx_vector(<2 x ptr addrspace(270)> %ptr, ptr %out, <2 x i64> %idx) {
+; CHECK-LABEL: issue56055_large_idx_vector:
+; CHECK: # %bb.0:
+; CHECK-NEXT: pshufd {{.*#+}} xmm0 = mem[0,2,2,3]
+; CHECK-NEXT: paddd (%rcx), %xmm0
+; CHECK-NEXT: movq %xmm0, (%rdx)
+; CHECK-NEXT: retq
+ %add.ptr = getelementptr inbounds i8, <2 x ptr addrspace(270)> %ptr, <2 x i64> %idx
+ store <2 x ptr addrspace(270)> %add.ptr, ptr %out
+ ret void
+}
+
+;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+; FASTISEL: {{.*}}
+; SDAG: {{.*}}
``````````
</details>
https://github.com/llvm/llvm-project/pull/110490
More information about the llvm-branch-commits
mailing list