CompCertTSO: A Verified Compiler for Relaxed-Memory Concurrency

• p. 3 §1.2

  Moreover, for four phases the memory accesses are identical except that some values that are undefined in the source take particular values in the target

• p. 18 Fig. 3

  (undefined values can become defined)

Only the authors can say what is right here.

Judge's note: I counted the 17 phases in Fig. 3 (p.18). Only three carry the undefined-values annotation: Cstacked→Cminor, LTLin→Linear and MachConcr→Asm. §5.5 (p.34) names only the Cstacked-Cminor and spilling/reloading phases. Taking the intro's 2 whole-system phases, 3 fence phases and 1 register-allocation phase, Fig. 3 leaves 8 exact 1:1 phases and 3 concretizing phases, where the intro gives 7 and 4. Some unannotated phase may also concretize undefined values, but I found no statement saying so. Only the authors can say which phase is the fourth.

Second check: another model agreed

• p. 19 Proof outline after Theorem 5.1

  between each of the non-fence-elimination phases from Cminor to MachAbstr

• p. 18 Fig. 3

  Simplification (undefined values can become defined)

• p. 34 §5.5

  the Cstacked-Cminor and spilling/reloading phases may change some of the undefined values

• p. 3 §1.2, second contribution

  (7 of 17) the memory accesses of source and target are in exact 1:1 correspondence. Moreover, for four phases the memory accesses are identical except that some values that are undefined in the source take particular values in the target

Only the authors can say what is right here.

Judge's note: In Fig. 3 (p.18), Cstacked→Cminor is a single downward arrow inside the §5.5 lifted span. §5.5 (p.34) says this phase uses the concretizing variants of Theorems 4.5/4.7. The outline on p.19 covers Cstacked→Csharpminor and MachAbstr→MachConcr as whole-system proofs, and its threadwise range starts at Cminor, so Cstacked→Cminor appears nowhere. The range presumably should start at Cstacked. Reviewer note 2 in this cluster (four vs three concretizing phases) duplicates C001.

Second check: another model agreed

• p. 4 §1.2 Contributions

  abstracting from the 18 000 or so lines of underlying proof script of these two phases

• p. 43 Fig. 15

  Csharpminor to Cstacked (§5.3)

• p. 43 Fig. 15, rows Csharpminor to Cstacked and MachAbstr to MachConcr

  3421 8163

• p. 4 Section 1 (contributions), third paragraph

  abstracting from the 18 000 or so lines of underlying proof

Only the authors can say what is right here.

Judge's note: Fig. 15 (p.43) lists Csharpminor to Cstacked (§5.3) as Specs 3421, Proof 8163 and MachAbstr to MachConcr (§5.4) as Specs 2724, Proof 6525. Neither the proof-only total nor the specs-plus-proof total is about 18 000, and I found no other count of these phases in the text. The figure may come from an earlier version of the development or a different count. Because it is hedged ('or so'), I treat this as a query to the authors.

Second check: another model agreed

• p. 44 §9, paragraph on CompCert 1.12.1

  There are also more optimization passes, including function inlining, constant propagation, and redundant reload elimination.

• p. 18 Fig. 3

  Constant propagation

Only the authors can say what is right here.

Judge's note: Fig. 3 (p.18) shows a 'Constant propagation' RTL pass, and the text at §5.5 says constant propagation is performed. §9 presents it as a newer pass without saying how the 1.12.1 version differs; the const-locations remark suggests an extended version is meant. The wording is ambiguous.

Second check: another model agreed

• p. 33 §5.4.2 first paragraph

  which points to the stack frame in MachAbstr, but in MachConcr points to the frame

Only the authors can say what is right here.

Judge's note: The page 33 image confirms the wording. As printed, the sentence names no difference between the two pointers, presumably a frame block versus the frame's location in MachConcr stack memory. Only the authors can supply the intended contrast.

Second check: another model agreed

• p. 34 Footnote 4

  the instruction ADD %eax,%ebx corresponds to the assignment %eax <- %eax + %ebx

Only the authors can say what is right here.

Judge's note: Rests on outside knowledge: in AT&T syntax, which uses %-prefixed registers, the destination is the last operand, so ADD %eax,%ebx means %ebx <- %ebx + %eax. The footnote image matches the quote. The notation and the stated operand order do not agree; the authors should say which syntax they intend.

Second check: another model agreed

• p. 12 §4.2 last paragraph

  In CompcertTSO, we represent pointers again as block-offset pairs

• p. 4 §1.2

  including the optimizations present in Compcert 1.5

• p. 4 §1.2, third contribution

  including the optimizations present in Compcert 1.5 that are safe under TSO

• p. 33 §5.5

  We have enabled all the CompCert 1.5 optimizations that are sound under the TSO semantics except tail call optimization.

Judge's note: I confirmed the mixed capitalisation on the page images of p.4 and p.12; the text layer has further cases on lines 665-688 and 862. This is cosmetic. Reviewer note 2's tail-call point is not a defect: §5.5 (p.34) says plainly that tail call optimization is sound but not enabled, and the intro's 'safe under TSO' wording is only loose.

Second check: none (judge only)

• p. 40 §6.3.1

  Take > to be the empty relation (which is trivially well-founded)

• p. 15 Definition 4.6

  equipped with a well-founded order < on States(T)

Judge's note: Both images confirm the symbols. Because the relation is empty the meaning is unaffected, so this is a cosmetic notation inconsistency.

Second check: none (judge only)

• p. 31 Fig. 10 panel labels

  MachAbs Thread State

• p. 31 Fig. 10 caption

  MachConcr and MachAbstr thread states

Judge's note: I confirmed the labels on the p.31 page image. They are abbreviated names, inconsistent with the rest of the paper. This is cosmetic.

Second check: none (judge only)

• p. 35 §5.6 pseudo-instructions list

  do not include single-point floating point values

Judge's note: The image confirms 'single-point'. The context is truncating moves that drop precision, so 'single-precision' is the evident term.

Second check: none (judge only)

• p. 42 §6.3.3

  whatever analysis was used to used to decide to perform the

• p. 42 Section 6.3.3, first paragraph

  irrespective of whatever analysis was used to used to decide to perform the insertion

Judge's note: Confirmed in the page 42 image.

Second check: none (judge only)

• p. 45 §9 Related work

  since Sasha and Snir’s seminal paper

• p. 45 §9, second paragraph

  since Sasha and Snir’s seminal pa-

• p. 45 Sec. 9, memory barriers paragraph

  Sasha and Snir’s seminal paper

• p. 45 Section 9 Related Work, memory barrier paragraph

  Sasha and Snir’s seminal pa-per on delay set analysis [Shasha and Snir 1988]

Judge's note: The page 45 image confirms 'Sasha' in the prose and 'Shasha' in the citation that follows.

Second check: none (judge only)

• p. 46 References, Abadi and Lamport 1991

  Martn Abadi and Leslie Lamport

• p. 46 References, first entry

  Martn Abadi and Leslie Lamport. 1991.

Only the authors can say what is right here.

Judge's note: Rests on outside knowledge: the author's first name is spelled 'Martín'. I looked at the crop of the typeset reference list on PDF page 46. It shows 'Martn Abadi' with no i between 't' and 'n', so this is not an extraction artefact; the text layer (line 2767) matches. The name appears nowhere else in the paper; the in-text citation is surname-only, '[Abadi and Lamport 1991]'. This is cosmetic, most likely an accented character that was lost in BibTeX processing.

Second check: none (judge only)