X86: Use LiveRegUnits in findDeadCallerSavedReg #156817
Open
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This stack of pull requests is managed by Graphite. Learn more about stacking. |
|
@llvm/pr-subscribers-backend-x86 Author: Matt Arsenault (arsenm) ChangesMy goal here was to eliminate the use of getGPRsForTailCall. There are a few things I find confusing about this function; the API There are a few codegen test changes. One of them I think is a correct Full diff: https://github.com/llvm/llvm-project/pull/156817.diff 5 Files Affected:
diff --git a/llvm/lib/Target/X86/X86RegisterInfo.cpp b/llvm/lib/Target/X86/X86RegisterInfo.cpp
index b79e508df7c97..3f4955f28e68b 100644
--- a/llvm/lib/Target/X86/X86RegisterInfo.cpp
+++ b/llvm/lib/Target/X86/X86RegisterInfo.cpp
@@ -1002,8 +1002,6 @@ unsigned X86RegisterInfo::findDeadCallerSavedReg(
if (MF->callsEHReturn())
return 0;
- const TargetRegisterClass &AvailableRegs = *getGPRsForTailCall(*MF);
-
if (MBBI == MBB.end())
return 0;
@@ -1025,20 +1023,20 @@ unsigned X86RegisterInfo::findDeadCallerSavedReg(
case X86::TCRETURNmi64:
case X86::EH_RETURN:
case X86::EH_RETURN64: {
- SmallSet<uint16_t, 8> Uses;
- for (MachineOperand &MO : MBBI->operands()) {
- if (!MO.isReg() || MO.isDef())
- continue;
- Register Reg = MO.getReg();
- if (!Reg)
- continue;
- for (MCRegAliasIterator AI(Reg, this, true); AI.isValid(); ++AI)
- Uses.insert(*AI);
+ LiveRegUnits LRU(*this);
+ LRU.addLiveOuts(MBB);
+ LRU.stepBackward(*MBBI);
+
+ // FIXME: Why do we need to special case this register? Is it missing from
+ // return implicit uses?
+ LRU.removeReg(X86::RIP);
+
+ const TargetRegisterClass &RC =
+ Is64Bit ? X86::GR64_NOSPRegClass : X86::GR32_NOSPRegClass;
+ for (MCRegister Reg : RC) {
+ if (LRU.available(Reg))
+ return Reg;
}
-
- for (auto CS : AvailableRegs)
- if (!Uses.count(CS) && CS != X86::RIP && CS != X86::RSP && CS != X86::ESP)
- return CS;
}
}
diff --git a/llvm/test/CodeGen/X86/apx/push2-pop2-cfi-seh.ll b/llvm/test/CodeGen/X86/apx/push2-pop2-cfi-seh.ll
index ad24608d338ad..d6d4db3509103 100644
--- a/llvm/test/CodeGen/X86/apx/push2-pop2-cfi-seh.ll
+++ b/llvm/test/CodeGen/X86/apx/push2-pop2-cfi-seh.ll
@@ -81,7 +81,7 @@ define i32 @csr6_alloc16(ptr %argv) {
; LIN-NEXT: .cfi_def_cfa_offset 32
; LIN-NEXT: pop2 %rbp, %r15
; LIN-NEXT: .cfi_def_cfa_offset 16
-; LIN-NEXT: popq %rcx
+; LIN-NEXT: popq %rax
; LIN-NEXT: .cfi_def_cfa_offset 8
; LIN-NEXT: retq
;
@@ -116,7 +116,7 @@ define i32 @csr6_alloc16(ptr %argv) {
; LIN-PPX-NEXT: .cfi_def_cfa_offset 32
; LIN-PPX-NEXT: pop2p %rbp, %r15
; LIN-PPX-NEXT: .cfi_def_cfa_offset 16
-; LIN-PPX-NEXT: popq %rcx
+; LIN-PPX-NEXT: popq %rax
; LIN-PPX-NEXT: .cfi_def_cfa_offset 8
; LIN-PPX-NEXT: retq
;
@@ -180,7 +180,7 @@ define i32 @csr6_alloc16(ptr %argv) {
; WIN-NEXT: pop2 %rbp, %rbx
; WIN-NEXT: pop2 %r13, %r12
; WIN-NEXT: pop2 %r15, %r14
-; WIN-NEXT: popq %rcx
+; WIN-NEXT: popq %rax
; WIN-NEXT: .seh_endepilogue
; WIN-NEXT: retq
; WIN-NEXT: .seh_endproc
@@ -211,7 +211,7 @@ define i32 @csr6_alloc16(ptr %argv) {
; WIN-PPX-NEXT: pop2p %rbp, %rbx
; WIN-PPX-NEXT: pop2p %r13, %r12
; WIN-PPX-NEXT: pop2p %r15, %r14
-; WIN-PPX-NEXT: popq %rcx
+; WIN-PPX-NEXT: popq %rax
; WIN-PPX-NEXT: .seh_endepilogue
; WIN-PPX-NEXT: retq
; WIN-PPX-NEXT: .seh_endproc
diff --git a/llvm/test/CodeGen/X86/lvi-hardening-ret.ll b/llvm/test/CodeGen/X86/lvi-hardening-ret.ll
index faa8bff8f0943..954985a3798b7 100644
--- a/llvm/test/CodeGen/X86/lvi-hardening-ret.ll
+++ b/llvm/test/CodeGen/X86/lvi-hardening-ret.ll
@@ -41,9 +41,9 @@ entry:
%add = add nsw i32 %0, %1
ret i32 %add
; CHECK-NOT: retq
-; CHECK: popq %rcx
+; CHECK: popq %rsi
; CHECK-NEXT: lfence
-; CHECK-NEXT: jmpq *%rcx
+; CHECK-NEXT: jmpq *%rsi
}
; Function Attrs: noinline nounwind optnone uwtable
@@ -52,9 +52,9 @@ define dso_local preserve_mostcc void @preserve_most() #0 {
entry:
ret void
; CHECK-NOT: retq
-; CHECK: popq %rax
+; CHECK: popq %r11
; CHECK-NEXT: lfence
-; CHECK-NEXT: jmpq *%rax
+; CHECK-NEXT: jmpq *%r11
}
; Function Attrs: noinline nounwind optnone uwtable
@@ -63,9 +63,9 @@ define dso_local preserve_allcc void @preserve_all() #0 {
entry:
ret void
; CHECK-NOT: retq
-; CHECK: popq %rax
+; CHECK: popq %r11
; CHECK-NEXT: lfence
-; CHECK-NEXT: jmpq *%rax
+; CHECK-NEXT: jmpq *%r11
}
define { i64, i128 } @ret_i64_i128() #0 {
diff --git a/llvm/test/CodeGen/X86/pr40289-64bit.ll b/llvm/test/CodeGen/X86/pr40289-64bit.ll
index 58da5258a6703..a83674ac81998 100644
--- a/llvm/test/CodeGen/X86/pr40289-64bit.ll
+++ b/llvm/test/CodeGen/X86/pr40289-64bit.ll
@@ -6,5 +6,5 @@ define cc 92 < 9 x i64 > @clobber() {
ret < 9 x i64 > undef
; CHECK-LABEL: clobber:
; CHECK-NOT: popq %rsp
- ; CHECK: addq $8, %rsp
+ ; CHECK: pushq %rax
}
diff --git a/llvm/test/CodeGen/X86/pr40289.ll b/llvm/test/CodeGen/X86/pr40289.ll
index 851b23c002bdb..21e50931b40f2 100644
--- a/llvm/test/CodeGen/X86/pr40289.ll
+++ b/llvm/test/CodeGen/X86/pr40289.ll
@@ -6,5 +6,5 @@ define < 3 x i32 > @clobber() {
ret < 3 x i32 > undef
; CHECK-LABEL: clobber:
; CHECK-NOT: popl %esp
- ; CHECK: addl $4, %esp
+ ; CHECK: popl %eax
}
|
phoebewang
reviewed
Sep 4, 2025
My goal here was to eliminate the use of getGPRsForTailCall. I don't think it makes sense to use here, considering this function handles other non-tail call cases. It seems to have been used to find a non-callee saved register in a roundabout way. We can be more precise by letting LiveRegUnits figure out the CSR set for the current function. There are a few things I find confusing about this function; the API isn't right. The callers ideally would be maintaining LiveRegUnits in the context where they need the free register. It also doesn't provide a RegisterClass to use. Also, this should work for any block and this shouldn't need to special case this set of return opcodes. There are a few codegen test changes. One of them I think is a correct improvement since the old code didn't consider undef uses as available. The others I think are just to allocation order changes, since we're now using the broader GPR_*NOSP classes.
73e2257 to
c831a2a
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
My goal here was to eliminate the use of getGPRsForTailCall.
I don't think it makes sense to use here, considering this function
handles other non-tail call cases. It seems to have been used to
find a non-callee saved register in a roundabout way. We can be more
precise by letting LiveRegUnits figure out the CSR set for the current
function.
There are a few things I find confusing about this function; the API
isn't right. The callers ideally would be maintaining LiveRegUnits
in the context where they need the free register. It also doesn't
provide a RegisterClass to use. Also, this should work for any
block and this shouldn't need to special case this set of return
opcodes. I'm also not sure why there is a special case for RIP. It was introduced in
fe5e5dc but doesn't appear to
be tested.
There are a few codegen test changes. One of them I think is a correct
improvement since the old code didn't consider undef uses as available.
The others I think are just to allocation order changes, since we're now
using the broader GPR_*NOSP classes.