Skip to content

Conversation

@nikic
Copy link
Contributor

@nikic nikic commented Oct 31, 2025

Currently, out-of-range branch fixups will be silently miscompiled. GNU as will instead print an "operand out of range" error instead.

Check that the branch target fixups fit into the proper range and have low zero bits. The implementation is inspired by SystemZ:

auto checkFixupInRange = [&](int64_t Min, int64_t Max) -> bool {

I'm not sure if there is any way to test the absolute cases. It looks like something like beqa 0, 0x10000 already gets rejected during parsing.

(My actual interest here is not actually assembly usage, but LLVM failing to use the correct branch kind for reasons I've not tracked down yet. Currently this just silently truncates the branch target instead of producing an error.)

@llvmbot
Copy link
Member

llvmbot commented Oct 31, 2025

@llvm/pr-subscribers-backend-powerpc

Author: Nikita Popov (nikic)

Changes

Currently, out-of-range branch fixups will be silently miscompiled. GNU as will instead print an "operand out of range" error instead.

Check that the branch target fixups fit into the proper range and have low zero bits. The implementation is inspired by SystemZ:

auto checkFixupInRange = [&](int64_t Min, int64_t Max) -> bool {

I'm not sure if there is any way to test the absolute cases. It looks like something like beqa 0, 0x10000 already gets rejected during parsing.

(My actual interest here is not actually assembly usage, but LLVM failing to use the correct branch kind for reasons I've not tracked down yet. Currently this just silently truncates the branch target instead of producing an error.)


Full diff: https://github.com/llvm/llvm-project/pull/165859.diff

2 Files Affected:

  • (modified) llvm/lib/Target/PowerPC/MCTargetDesc/PPCAsmBackend.cpp (+23-2)
  • (added) llvm/test/MC/PowerPC/fixup-out-of-range.s (+43)
diff --git a/llvm/lib/Target/PowerPC/MCTargetDesc/PPCAsmBackend.cpp b/llvm/lib/Target/PowerPC/MCTargetDesc/PPCAsmBackend.cpp
index 04b886ae74993..7b69e73ea65be 100644
--- a/llvm/lib/Target/PowerPC/MCTargetDesc/PPCAsmBackend.cpp
+++ b/llvm/lib/Target/PowerPC/MCTargetDesc/PPCAsmBackend.cpp
@@ -25,7 +25,26 @@
 #include "llvm/Support/ErrorHandling.h"
 using namespace llvm;
 
-static uint64_t adjustFixupValue(unsigned Kind, uint64_t Value) {
+static uint64_t adjustFixupValue(MCContext &Ctx, const MCFixup &Fixup,
+                                 unsigned Kind, uint64_t Value) {
+  auto checkBrFixup = [&](unsigned Bits) {
+    int64_t SVal = int64_t(Value);
+    if ((Value & 3) != 0) {
+      Ctx.reportError(Fixup.getLoc(), "branch target not a multiple of four (" +
+                                          Twine(SVal) + ")");
+      return;
+    }
+
+    // Low two bits are not encoded.
+    Bits += 2;
+    if (!isIntN(Bits, Value)) {
+      Ctx.reportError(Fixup.getLoc(), "branch target out of range (" +
+                                          Twine(SVal) + " not between " +
+                                          Twine(minIntN(Bits)) + " and " +
+                                          Twine(maxIntN(Bits)) + ")");
+    }
+  };
+
   switch (Kind) {
   default:
     llvm_unreachable("Unknown fixup kind!");
@@ -37,10 +56,12 @@ static uint64_t adjustFixupValue(unsigned Kind, uint64_t Value) {
     return Value;
   case PPC::fixup_ppc_brcond14:
   case PPC::fixup_ppc_brcond14abs:
+    checkBrFixup(14);
     return Value & 0xfffc;
   case PPC::fixup_ppc_br24:
   case PPC::fixup_ppc_br24abs:
   case PPC::fixup_ppc_br24_notoc:
+    checkBrFixup(24);
     return Value & 0x3fffffc;
   case PPC::fixup_ppc_half16:
     return Value & 0xffff;
@@ -202,7 +223,7 @@ void PPCAsmBackend::applyFixup(const MCFragment &F, const MCFixup &Fixup,
   MCFixupKind Kind = Fixup.getKind();
   if (mc::isRelocation(Kind))
     return;
-  Value = adjustFixupValue(Kind, Value);
+  Value = adjustFixupValue(getContext(), Fixup, Kind, Value);
   if (!Value)
     return; // Doesn't change encoding.
 
diff --git a/llvm/test/MC/PowerPC/fixup-out-of-range.s b/llvm/test/MC/PowerPC/fixup-out-of-range.s
new file mode 100644
index 0000000000000..e9c4f513cceda
--- /dev/null
+++ b/llvm/test/MC/PowerPC/fixup-out-of-range.s
@@ -0,0 +1,43 @@
+# RUN: not llvm-mc -triple powerpc64le-unknown-unknown -filetype=obj %s 2>&1 >/dev/null | FileCheck %s
+
+; CHECK: error: branch target out of range (32772 not between -32768 and 32767)
+brcond14_out_of_range_hi:
+    beq 0, brcond14_target
+    .space 0x8000
+
+brcond14_target:
+    blr
+
+; CHECK: error: branch target out of range (-32772 not between -32768 and 32767)
+brcond14_out_of_range_lo:
+    .space 0x8004
+    beq 0, brcond14_out_of_range_lo
+
+; CHECK: error: branch target not a multiple of four (5)
+brcond14_misaligned:
+    beq 0, brcond14_misaligned_target
+    .byte 0
+
+brcond14_misaligned_target:
+    blr
+
+; CHECK: error: branch target out of range (33554436 not between -33554432 and 33554431)
+br24_out_of_range_hi:
+    b br24_target
+    .space 0x2000000
+
+br24_target:
+    blr
+
+; CHECK: error: branch target out of range (-33554436 not between -33554432 and 33554431)
+br24_out_of_range_lo:
+    .space 0x2000004
+    b br24_out_of_range_lo
+
+; CHECK: error: branch target not a multiple of four (5)
+br24_misaligned:
+    b br24_misaligned_target
+    .byte 0
+
+br24_misaligned_target:
+    blr

Currently, out-of-range branch fixups will be silently miscompiled.
GNU as will instead print a message like:

> Error: operand out of range (0x0000000000010004 is not between 0xffffffffffff8000 and 0x0000000000007ffc)

Check that the branch target fixups fit into the proper range and
have low zero bits.

I'm not sure if there is any way to test the absolute cases. It
looks like something like `beqa 0, 0x10000` already gets rejected
during parsing.

(My actual interest here is not actually assembly usage, but LLVM
producing invalid branches for reasons I've not tracked down yet.
Currently this just silently truncates the branch target, and I'd
rather make it produce an error instead.)
@RolandF77
Copy link
Collaborator

I'm not sure if there is any way to test the absolute cases. It looks like something like beqa 0, 0x10000 already gets rejected during parsing.

Could you use an expression, like beqa 0,lab-. ?

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants