[RISCV] Fold LI 1 / SLLI into BSETI during i64 materialization #142348
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
Member
|
@llvm/pr-subscribers-backend-risc-v Author: Piotr Fusik (pfusik) ChangesMy first approach was to avoid emitting LI 1 / SLLI in the first place. Another possible approach would be to keep LI 1 / SLLI if it is to be Full diff: https://github.com/llvm/llvm-project/pull/142348.diff 2 Files Affected:
diff --git a/llvm/lib/Target/RISCV/MCTargetDesc/RISCVMatInt.cpp b/llvm/lib/Target/RISCV/MCTargetDesc/RISCVMatInt.cpp
index 8ea2548258fdb..f239129617b97 100644
--- a/llvm/lib/Target/RISCV/MCTargetDesc/RISCVMatInt.cpp
+++ b/llvm/lib/Target/RISCV/MCTargetDesc/RISCVMatInt.cpp
@@ -353,6 +353,13 @@ InstSeq generateInstSeq(int64_t Val, const MCSubtargetInfo &STI) {
} while (Hi != 0);
Res = TmpSeq;
}
+
+ // Fold LI 1 + SLLI into BSETI.
+ if (Res[0].getOpcode() == RISCV::ADDI && Res[0].getImm() == 1 &&
+ Res[1].getOpcode() == RISCV::SLLI) {
+ Res.erase(Res.begin()); // Remove ADDI.
+ Res.front() = Inst(RISCV::BSETI, Res.front().getImm()); // Patch SLLI.
+ }
}
// Perform optimization with BCLRI in the Zbs extension.
diff --git a/llvm/test/CodeGen/RISCV/imm.ll b/llvm/test/CodeGen/RISCV/imm.ll
index f324a9bc120ef..418407d9b7cd6 100644
--- a/llvm/test/CodeGen/RISCV/imm.ll
+++ b/llvm/test/CodeGen/RISCV/imm.ll
@@ -4637,3 +4637,127 @@ define i64 @imm64_0xFF7FFFFF7FFFFFFE() {
; RV64-REMAT-NEXT: ret
ret i64 -36028799166447617 ; 0xFF7FFFFF7FFFFFFE
}
+
+define i64 @imm64_0xFFFFFFFF0() {
+; RV32I-LABEL: imm64_0xFFFFFFFF0:
+; RV32I: # %bb.0:
+; RV32I-NEXT: li a0, -16
+; RV32I-NEXT: li a1, 15
+; RV32I-NEXT: ret
+;
+; RV32IXQCILI-LABEL: imm64_0xFFFFFFFF0:
+; RV32IXQCILI: # %bb.0:
+; RV32IXQCILI-NEXT: li a0, -16
+; RV32IXQCILI-NEXT: li a1, 15
+; RV32IXQCILI-NEXT: ret
+;
+; RV64I-LABEL: imm64_0xFFFFFFFF0:
+; RV64I: # %bb.0:
+; RV64I-NEXT: li a0, 1
+; RV64I-NEXT: slli a0, a0, 36
+; RV64I-NEXT: addi a0, a0, -16
+; RV64I-NEXT: ret
+;
+; RV64IZBA-LABEL: imm64_0xFFFFFFFF0:
+; RV64IZBA: # %bb.0:
+; RV64IZBA-NEXT: li a0, 1
+; RV64IZBA-NEXT: slli a0, a0, 36
+; RV64IZBA-NEXT: addi a0, a0, -16
+; RV64IZBA-NEXT: ret
+;
+; RV64IZBB-LABEL: imm64_0xFFFFFFFF0:
+; RV64IZBB: # %bb.0:
+; RV64IZBB-NEXT: li a0, 1
+; RV64IZBB-NEXT: slli a0, a0, 36
+; RV64IZBB-NEXT: addi a0, a0, -16
+; RV64IZBB-NEXT: ret
+;
+; RV64IZBS-LABEL: imm64_0xFFFFFFFF0:
+; RV64IZBS: # %bb.0:
+; RV64IZBS-NEXT: bseti a0, zero, 36
+; RV64IZBS-NEXT: addi a0, a0, -16
+; RV64IZBS-NEXT: ret
+;
+; RV64IXTHEADBB-LABEL: imm64_0xFFFFFFFF0:
+; RV64IXTHEADBB: # %bb.0:
+; RV64IXTHEADBB-NEXT: li a0, 1
+; RV64IXTHEADBB-NEXT: slli a0, a0, 36
+; RV64IXTHEADBB-NEXT: addi a0, a0, -16
+; RV64IXTHEADBB-NEXT: ret
+;
+; RV32-REMAT-LABEL: imm64_0xFFFFFFFF0:
+; RV32-REMAT: # %bb.0:
+; RV32-REMAT-NEXT: li a0, -16
+; RV32-REMAT-NEXT: li a1, 15
+; RV32-REMAT-NEXT: ret
+;
+; RV64-REMAT-LABEL: imm64_0xFFFFFFFF0:
+; RV64-REMAT: # %bb.0:
+; RV64-REMAT-NEXT: li a0, 1
+; RV64-REMAT-NEXT: slli a0, a0, 36
+; RV64-REMAT-NEXT: addi a0, a0, -16
+; RV64-REMAT-NEXT: ret
+ ret i64 68719476720 ; 0xFFFFFFFF0
+}
+
+define i64 @imm64_0x1FFFFFF08() {
+; RV32I-LABEL: imm64_0x1FFFFFF08:
+; RV32I: # %bb.0:
+; RV32I-NEXT: li a0, -248
+; RV32I-NEXT: li a1, 1
+; RV32I-NEXT: ret
+;
+; RV32IXQCILI-LABEL: imm64_0x1FFFFFF08:
+; RV32IXQCILI: # %bb.0:
+; RV32IXQCILI-NEXT: li a0, -248
+; RV32IXQCILI-NEXT: li a1, 1
+; RV32IXQCILI-NEXT: ret
+;
+; RV64I-LABEL: imm64_0x1FFFFFF08:
+; RV64I: # %bb.0:
+; RV64I-NEXT: li a0, 1
+; RV64I-NEXT: slli a0, a0, 33
+; RV64I-NEXT: addi a0, a0, -248
+; RV64I-NEXT: ret
+;
+; RV64IZBA-LABEL: imm64_0x1FFFFFF08:
+; RV64IZBA: # %bb.0:
+; RV64IZBA-NEXT: li a0, 1
+; RV64IZBA-NEXT: slli a0, a0, 33
+; RV64IZBA-NEXT: addi a0, a0, -248
+; RV64IZBA-NEXT: ret
+;
+; RV64IZBB-LABEL: imm64_0x1FFFFFF08:
+; RV64IZBB: # %bb.0:
+; RV64IZBB-NEXT: li a0, 1
+; RV64IZBB-NEXT: slli a0, a0, 33
+; RV64IZBB-NEXT: addi a0, a0, -248
+; RV64IZBB-NEXT: ret
+;
+; RV64IZBS-LABEL: imm64_0x1FFFFFF08:
+; RV64IZBS: # %bb.0:
+; RV64IZBS-NEXT: bseti a0, zero, 33
+; RV64IZBS-NEXT: addi a0, a0, -248
+; RV64IZBS-NEXT: ret
+;
+; RV64IXTHEADBB-LABEL: imm64_0x1FFFFFF08:
+; RV64IXTHEADBB: # %bb.0:
+; RV64IXTHEADBB-NEXT: li a0, 1
+; RV64IXTHEADBB-NEXT: slli a0, a0, 33
+; RV64IXTHEADBB-NEXT: addi a0, a0, -248
+; RV64IXTHEADBB-NEXT: ret
+;
+; RV32-REMAT-LABEL: imm64_0x1FFFFFF08:
+; RV32-REMAT: # %bb.0:
+; RV32-REMAT-NEXT: li a0, -248
+; RV32-REMAT-NEXT: li a1, 1
+; RV32-REMAT-NEXT: ret
+;
+; RV64-REMAT-LABEL: imm64_0x1FFFFFF08:
+; RV64-REMAT: # %bb.0:
+; RV64-REMAT-NEXT: li a0, 1
+; RV64-REMAT-NEXT: slli a0, a0, 33
+; RV64-REMAT-NEXT: addi a0, a0, -248
+; RV64-REMAT-NEXT: ret
+ ret i64 8589934344 ; 0x1FFFFFF08
+}
|
pfusik
commented
Jun 2, 2025
Contributor
Author
There was a problem hiding this comment.
Another option would be to add setOpcode().
|
✅ With the latest revision this PR passed the C/C++ code formatter. |
pfusik
commented
Jun 2, 2025
Contributor
Author
There was a problem hiding this comment.
My first approach was to avoid emitting LI 1 / SLLI in the first place. Unfortunately, that favors BSETI C / ADDI -1 over LI -1 / SRLI 64-C even though the latter has both instructions compressible. This is because the code assumes in several places that a two-instruction sequence (here: BSETI / ADDI) cannot be improved. Another possible approach would be to keep LI 1 / SLLI if it is to be later replaced with SRLI. This would be harder to grasp than simply patching LI 1 / SLLI with BSETI.
Contributor
Author
|
Tests merged into main branch as 405c31f. PR rebased. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
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 first approach was to avoid emitting LI 1 / SLLI in the first place.
Unfortunately, that favors BSETI C / ADDI -1 over LI -1 / SRLI 64-C
even though the latter has both instructions compressible.
This is because the code assumes in several places that a two-instruction
sequence (here: BSETI / ADDI) cannot be improved.
Another possible approach would be to keep LI 1 / SLLI if it is to be
later replaced with SRLI. This would be harder to grasp than eventually
patching LI 1 / SLLI with BSETI.