[SystemZ] Implement .machine (push|pop) directives #137302

dominik-steenken · 2025-04-25T09:47:36Z

The .machine push and .machine pop directives were missing from the SystemZ Backend Asm Parser. This PR adds them, and expands the corresponding test to test proper operation.

Internally, this is modeled as a simple stack implemented on a SmallVector<StringRef>.

llvmbot · 2025-04-25T09:48:08Z

@llvm/pr-subscribers-mc

@llvm/pr-subscribers-backend-systemz

Author: Dominik Steenken (dominik-steenken)

Changes

The .machine push and .machine pop directives were missing from the SystemZ Backend Asm Parser. This PR adds them, and expands the corresponding test to test proper operation.

Internally, this is modeled as a simple stack implemented on a SmallVector<StringRef>.

This implementation makes the assumption that, without any .machine directive, the "current" machine is undefined, rather than trying to derive it based on the host machine. Thus, in this implementation, any .machine push must be preceded by at least one .machine <cpu>. Otherwise, a parsing error is recognized.

Full diff: https://github.com/llvm/llvm-project/pull/137302.diff

3 Files Affected:

(modified) llvm/lib/Target/SystemZ/AsmParser/SystemZAsmParser.cpp (+32-6)
(modified) llvm/test/MC/SystemZ/machine-directive-invalid.s (+10-3)
(modified) llvm/test/MC/SystemZ/machine-directive.s (+2-1)

diff --git a/llvm/lib/Target/SystemZ/AsmParser/SystemZAsmParser.cpp b/llvm/lib/Target/SystemZ/AsmParser/SystemZAsmParser.cpp
index 6d9a7a73f72db..a7612ecfd954f 100644
--- a/llvm/lib/Target/SystemZ/AsmParser/SystemZAsmParser.cpp
+++ b/llvm/lib/Target/SystemZ/AsmParser/SystemZAsmParser.cpp
@@ -410,6 +410,12 @@ class SystemZAsmParser : public MCTargetAsmParser {
 
 private:
   MCAsmParser &Parser;
+
+  // A vector to contain the stack of machine specs created by `.machine push`
+  SmallVector<StringRef> MachineStack;
+  // Specifies the current CPU
+  StringRef CurrentCPU = "";
+
   enum RegisterGroup {
     RegGR,
     RegFP,
@@ -1382,17 +1388,37 @@ bool SystemZAsmParser::parseDirectiveMachine(SMLoc L) {
       Parser.getTok().isNot(AsmToken::String))
     return TokError("unexpected token in '.machine' directive");
 
-  StringRef CPU = Parser.getTok().getIdentifier();
+  StringRef Id = Parser.getTok().getIdentifier();
+
+  // Parse push and pop directives first
+  StringRef CPU = "";
+  if (Id == "push") {
+    if (CurrentCPU.empty())
+      return TokError(
+          "push without preceding cpu spec in '.machine' directive");
+    MachineStack.push_back(CurrentCPU);
+  } else if (Id == "pop") {
+    if (MachineStack.empty())
+      return TokError("pop without corresponding push in '.machine' directive");
+    CPU = MachineStack.back();
+    MachineStack.pop_back();
+  } else {
+    CPU = Id;
+  }
+
   Parser.Lex();
   if (parseEOL())
     return true;
 
-  MCSubtargetInfo &STI = copySTI();
-  STI.setDefaultFeatures(CPU, /*TuneCPU*/ CPU, "");
-  setAvailableFeatures(ComputeAvailableFeatures(STI.getFeatureBits()));
-
-  getTargetStreamer().emitMachine(CPU);
+  // check if we need to change cpu
+  if (!CPU.empty()) {
+    MCSubtargetInfo &STI = copySTI();
+    STI.setDefaultFeatures(CPU, /*TuneCPU*/ CPU, "");
+    setAvailableFeatures(ComputeAvailableFeatures(STI.getFeatureBits()));
+    CurrentCPU = CPU;
 
+    getTargetStreamer().emitMachine(CPU);
+  }
   return false;
 }
 
diff --git a/llvm/test/MC/SystemZ/machine-directive-invalid.s b/llvm/test/MC/SystemZ/machine-directive-invalid.s
index 8b3147ecc62d4..f9e6aea4f2604 100644
--- a/llvm/test/MC/SystemZ/machine-directive-invalid.s
+++ b/llvm/test/MC/SystemZ/machine-directive-invalid.s
@@ -1,10 +1,17 @@
+// NOTE: Assertions have been autogenerated by utils/update_mc_test_checks.py UTC_ARGS: --version 5
 # RUN: not llvm-mc -triple=s390x %s 2>&1 | FileCheck %s
 
-# CHECK: [[#@LINE+1]]:9: error: unexpected token in '.machine' directive
 .machine
+// CHECK: :[[@LINE-1]]:9: error: unexpected token in '.machine' directive
+
+.machine push
+// CHECK: :[[@LINE-1]]:10: error: push without preceding cpu spec in '.machine' directive
+
+.machine pop
+// CHECK: :[[@LINE-1]]:10: error: pop without corresponding push in '.machine' directive
 
-# CHECK: [[#@LINE+1]]:10: error: unexpected token in '.machine' directive
 .machine 42
+// CHECK: :[[@LINE-1]]:10: error: unexpected token in '.machine' directive
 
-# CHECK: [[#@LINE+1]]:13: error: expected newline
 .machine z13+
+// CHECK: :[[@LINE-1]]:13: error: expected newline
diff --git a/llvm/test/MC/SystemZ/machine-directive.s b/llvm/test/MC/SystemZ/machine-directive.s
index aa71c6aea5572..7861c56420cd3 100644
--- a/llvm/test/MC/SystemZ/machine-directive.s
+++ b/llvm/test/MC/SystemZ/machine-directive.s
@@ -12,9 +12,10 @@
 # CHECK: vgbm	%v0, 3
 
 .machine z13
+.machine push
 vgbm    %v0, 0
 .machine zEC12
 vgbm    %v0, 1
-.machine z13
+.machine pop
 vgbm    %v0, 3

dominik-steenken · 2025-04-25T10:03:01Z

@uweigand FYI

dominik-steenken · 2025-04-25T10:34:05Z

investigating ci fail

uweigand · 2025-04-25T13:31:24Z

This implementation makes the assumption that, without any .machine directive, the "current" machine is undefined, rather than trying to derive it based on the host machine. Thus, in this implementation, any .machine push must be preceded by at least one .machine <cpu>. Otherwise, a parsing error is recognized.

I don't think this matches GAS behavior. Why is this necessary? Can't we just push the feature bits instead of the name onto the stack? There's the default set of feature bits that's already detected (and used before any explicit machine directive); a first push would just push this; subsequent pops would restore whatever bits are set without having to re-compute them from a name.

uweigand · 2025-04-28T11:23:05Z

llvm/lib/Target/SystemZ/AsmParser/SystemZAsmParser.cpp

+  // it always reflects the current FeatureBitset used to determine which
+  // machine features are available, i.e. either the default or the result
+  // of the most reecent `.machine` directive.
+  FeatureBitset CurrentFeatures;


Why do you need this copy? You should be able to just use STI.getAvailableFeatures() ...

Good point, i have removed the copy and use getAvailableFeatures() instead.

uweigand · 2025-04-28T11:24:41Z

llvm/lib/Target/SystemZ/AsmParser/SystemZAsmParser.cpp

+      return TokError("pop without corresponding push in '.machine' directive");
+    CurrentFeatures = MachineStack.back();
+    MachineStack.pop_back();
+    setAvailableFeatures(CurrentFeatures);


There seems to be an inconsistency w.r.t. calling ComputeAvailableFeatures. This needs to be called somewhere, either before setting up CurrentFeatures or else when using it here. (If you do push STI.getAvailableFeatures() instead as mentioned above, this issue would go away as well.)

Thanks, i was a bit confused as to what ComputeAvailableFeatures did since it takes and produces a FeatureBitset. Since i took your advice above, i think this is now addressed as well.

uweigand · 2025-04-28T11:25:19Z

llvm/lib/Target/SystemZ/AsmParser/SystemZAsmParser.cpp

+    STI.setDefaultFeatures(Id, /*TuneCPU*/ Id, "");
+    CurrentFeatures = STI.getFeatureBits();
+    setAvailableFeatures(ComputeAvailableFeatures(CurrentFeatures));
+    getTargetStreamer().emitMachine(Id);


Minor nit: looks like you do the same emitMachine call in all three branches, so it probably should be moved after the if.

Yes, those were originally different but over the various edits they became the same. I've moved them as you suggested.

uweigand · 2025-04-28T11:25:59Z

llvm/lib/Target/SystemZ/AsmParser/SystemZAsmParser.cpp

+    setAvailableFeatures(ComputeAvailableFeatures(CurrentFeatures));
+    getTargetStreamer().emitMachine(Id);
+  }
+
  Parser.Lex();
  if (parseEOL())
    return true;


Probably better to leave this error check before performing any action.

Yes. This introduced a minor compilcation where i had to separate the reporting of the TokError from the existing if block, but i think it is now in the right place. Thanks for the feedback :)

dominik-steenken · 2025-04-28T12:26:14Z

This implementation makes the assumption that, without any .machine directive, the "current" machine is undefined, rather than trying to derive it based on the host machine. Thus, in this implementation, any .machine push must be preceded by at least one .machine <cpu>. Otherwise, a parsing error is recognized.

I don't think this matches GAS behavior. Why is this necessary? Can't we just push the feature bits instead of the name onto the stack? There's the default set of feature bits that's already detected (and used before any explicit machine directive); a first push would just push this; subsequent pops would restore whatever bits are set without having to re-compute them from a name.

Thanks, that was very helpful. I was somewhat unhappy with this solution as well. It originated from me having difficulties obtaining a string description of the CPU from the SubTargetInfo - it seemed no matter where i sourced this from, it would always be an empty string. But as you point out, it is also possible to just store the bits and restore them once the .machine stack is empty. I have updated the commit to provide an implementation that does this. .machine push without a prior .machine <cpu> is thus now legal and i have modified the tests accordingly. I have also tried to incorporate your additional feedback. I am now waiting for the CI to complete.

uweigand · 2025-04-28T12:38:32Z

llvm/lib/Target/SystemZ/AsmParser/SystemZAsmParser.cpp

+  // Do error check before lexing the next token so that TokError
+  // points at the correct token.
+  if ((Id == "pop") && (MachineStack.empty()))
+    return TokError("pop without corresponding push in '.machine' directive");


Hmm. I do think it would be better to have the syntax check before the semantics check. E.g. if you do a .machine pop z15 or something, you should preferably see the error that there's an unexpected token at the end, over the error that there was no corresponding push.

To get the location correct, I think it would be better to use an Error with an explicit location (e.g. by calling Parser.getTok().getLoc() at the place where you do the getIdentifier call, and then using that location with Error further down).

makes sense to me, i have reincorporated the check for the semantic error and added an additional line to the error test to check that the .machine pop <cpu> situation you outlined does in fact point at the syntax issue and not the semantic one.

actually hang on i used a push instead of a pop...

and it does report the semantic issue first :/
GIve me a second please

resolved. It works now, at least in my local testing.

uweigand

LGTM now assuming CI passes.

dominik-steenken · 2025-04-30T12:12:55Z

From what i can tell, this lldb check fail exists on main as well. Not sure what to do about this, for now i'll just rebase.

The `.machine push` and `.machine pop` directives were missing from the SystemZ Backend Asm Parser. This commit adds them, and expands the corresponding test to test proper operation.

uweigand · 2025-04-30T13:53:54Z

Thanks. The Windows failure looks unrelated, so I think this is good to go now.

The `.machine push` and `.machine pop` directives were missing from the SystemZ Backend Asm Parser. This PR adds them, and expands the corresponding test to test proper operation. Internally, this is modeled as a simple stack implemented on a `SmallVector<StringRef>`.

llvmbot added backend:SystemZ mc Machine (object) code labels Apr 25, 2025

dominik-steenken force-pushed the add-machine-push-pop branch 2 times, most recently from 35949c3 to 0834e7a Compare April 28, 2025 11:04

uweigand reviewed Apr 28, 2025

View reviewed changes

dominik-steenken force-pushed the add-machine-push-pop branch from 0834e7a to 8b0af95 Compare April 28, 2025 12:22

uweigand reviewed Apr 28, 2025

View reviewed changes

dominik-steenken force-pushed the add-machine-push-pop branch 2 times, most recently from 4c94500 to 17df235 Compare April 28, 2025 13:00

uweigand approved these changes Apr 28, 2025

View reviewed changes

dominik-steenken force-pushed the add-machine-push-pop branch from 17df235 to 1ace0c7 Compare April 28, 2025 14:36

[SystemZ] Implement .machine (push|pop) directives

4774136

The `.machine push` and `.machine pop` directives were missing from the SystemZ Backend Asm Parser. This commit adds them, and expands the corresponding test to test proper operation.

dominik-steenken force-pushed the add-machine-push-pop branch from 1ace0c7 to 4774136 Compare April 30, 2025 12:15

uweigand merged commit c626ef9 into llvm:main Apr 30, 2025
9 of 11 checks passed

uweigand mentioned this pull request Apr 30, 2025

s390x: support .machine push and .machine pop assembler directives #129053

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[SystemZ] Implement .machine (push|pop) directives #137302

[SystemZ] Implement .machine (push|pop) directives #137302

dominik-steenken commented Apr 25, 2025 •

edited

Loading

llvmbot commented Apr 25, 2025 •

edited

Loading

dominik-steenken commented Apr 25, 2025

dominik-steenken commented Apr 25, 2025 •

edited

Loading

uweigand commented Apr 25, 2025

uweigand Apr 28, 2025

dominik-steenken Apr 28, 2025

uweigand Apr 28, 2025

dominik-steenken Apr 28, 2025

uweigand Apr 28, 2025

dominik-steenken Apr 28, 2025

uweigand Apr 28, 2025

dominik-steenken Apr 28, 2025 •

edited

Loading

dominik-steenken commented Apr 28, 2025

uweigand Apr 28, 2025

dominik-steenken Apr 28, 2025

dominik-steenken Apr 28, 2025

dominik-steenken Apr 28, 2025

dominik-steenken Apr 28, 2025

uweigand left a comment

dominik-steenken commented Apr 30, 2025

uweigand commented Apr 30, 2025

[SystemZ] Implement .machine (push|pop) directives #137302

[SystemZ] Implement .machine (push|pop) directives #137302

Conversation

dominik-steenken commented Apr 25, 2025 • edited Loading

llvmbot commented Apr 25, 2025 • edited Loading

dominik-steenken commented Apr 25, 2025

dominik-steenken commented Apr 25, 2025 • edited Loading

uweigand commented Apr 25, 2025

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

dominik-steenken Apr 28, 2025 • edited Loading

Choose a reason for hiding this comment

dominik-steenken commented Apr 28, 2025

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

uweigand left a comment

Choose a reason for hiding this comment

dominik-steenken commented Apr 30, 2025

uweigand commented Apr 30, 2025

dominik-steenken commented Apr 25, 2025 •

edited

Loading

llvmbot commented Apr 25, 2025 •

edited

Loading

dominik-steenken commented Apr 25, 2025 •

edited

Loading

dominik-steenken Apr 28, 2025 •

edited

Loading