[RISCV] Expand multiplication by 2^N * 3/5/9 + 1 with SHL_ADD
#166933
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) ChangesFull diff: https://github.com/llvm/llvm-project/pull/166933.diff 2 Files Affected:
diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
index c1d38419992b1..637a46c508c1f 100644
--- a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
+++ b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
@@ -16544,12 +16544,23 @@ static SDValue expandMulToShlAddShlAdd(SDNode *N, SelectionDAG &DAG,
break;
}
- // 2/4/8 * 3/5/9 + 1 -> (shXadd (shYadd X, X), X)
int ShX;
if (int ShY = isShifted359(MulAmt - 1, ShX)) {
assert(ShX != 0 && "MulAmt=4,6,10 handled before");
+ // 2/4/8 * 3/5/9 + 1 -> (shXadd (shYadd X, X), X)
if (ShX <= 3)
return getShlAddShlAdd(N, DAG, ShX, ShY, /*AddX=*/true, Shift);
+ // 2^N * 3/5/9 + 1 -> (add (shYadd (shl X, N), (shl X, N)), X)
+ if (Shift == 0) {
+ SDLoc DL(N);
+ EVT VT = N->getValueType(0);
+ SDValue X = N->getOperand(0);
+ SDValue Shl =
+ DAG.getNode(ISD::SHL, DL, VT, X, DAG.getConstant(ShX, DL, VT));
+ SDValue ShlAdd = DAG.getNode(RISCVISD::SHL_ADD, DL, VT, Shl,
+ DAG.getTargetConstant(ShY, DL, VT), Shl);
+ return DAG.getNode(ISD::ADD, DL, VT, ShlAdd, X);
+ }
}
return SDValue();
}
@@ -16610,7 +16621,7 @@ static SDValue expandMul(SDNode *N, SelectionDAG &DAG,
DAG.getTargetConstant(Shift, DL, VT), Shift1);
}
- // TODO: 2^(C1>3) * 3,5,9 +/- 1
+ // TODO: 2^(C1>3) * 3/5/9 - 1
// 2^n + 2/4/8 + 1 -> (add (shl X, C1), (shXadd X, X))
if (MulAmt > 2 && isPowerOf2_64((MulAmt - 1) & (MulAmt - 2))) {
diff --git a/llvm/test/CodeGen/RISCV/rv64zba.ll b/llvm/test/CodeGen/RISCV/rv64zba.ll
index e56c7b41d43ce..156599fb72877 100644
--- a/llvm/test/CodeGen/RISCV/rv64zba.ll
+++ b/llvm/test/CodeGen/RISCV/rv64zba.ll
@@ -944,6 +944,58 @@ define i64 @addmul146(i64 %a, i64 %b) {
ret i64 %d
}
+define i64 @mul49(i64 %a) {
+; RV64I-LABEL: mul49:
+; RV64I: # %bb.0:
+; RV64I-NEXT: li a1, 49
+; RV64I-NEXT: mul a0, a0, a1
+; RV64I-NEXT: ret
+;
+; RV64ZBA-LABEL: mul49:
+; RV64ZBA: # %bb.0:
+; RV64ZBA-NEXT: slli a1, a0, 4
+; RV64ZBA-NEXT: sh1add a1, a1, a1
+; RV64ZBA-NEXT: add a0, a1, a0
+; RV64ZBA-NEXT: ret
+;
+; RV64XANDESPERF-LABEL: mul49:
+; RV64XANDESPERF: # %bb.0:
+; RV64XANDESPERF-NEXT: slli a1, a0, 4
+; RV64XANDESPERF-NEXT: nds.lea.h a1, a1, a1
+; RV64XANDESPERF-NEXT: add a0, a1, a0
+; RV64XANDESPERF-NEXT: ret
+ %c = mul i64 %a, 49
+ ret i64 %c
+}
+
+define i64 @zext_mul49(i32 signext %a) {
+; RV64I-LABEL: zext_mul49:
+; RV64I: # %bb.0:
+; RV64I-NEXT: li a1, 49
+; RV64I-NEXT: slli a1, a1, 32
+; RV64I-NEXT: slli a0, a0, 32
+; RV64I-NEXT: mulhu a0, a0, a1
+; RV64I-NEXT: ret
+;
+; RV64ZBA-LABEL: zext_mul49:
+; RV64ZBA: # %bb.0:
+; RV64ZBA-NEXT: slli.uw a1, a0, 4
+; RV64ZBA-NEXT: sh1add a1, a1, a1
+; RV64ZBA-NEXT: add.uw a0, a0, a1
+; RV64ZBA-NEXT: ret
+;
+; RV64XANDESPERF-LABEL: zext_mul49:
+; RV64XANDESPERF: # %bb.0:
+; RV64XANDESPERF-NEXT: slli a1, a0, 32
+; RV64XANDESPERF-NEXT: srli a1, a1, 28
+; RV64XANDESPERF-NEXT: nds.lea.h a1, a1, a1
+; RV64XANDESPERF-NEXT: nds.lea.b.ze a0, a1, a0
+; RV64XANDESPERF-NEXT: ret
+ %b = zext i32 %a to i64
+ %c = mul i64 %b, 49
+ ret i64 %c
+}
+
define i64 @mul50(i64 %a) {
; RV64I-LABEL: mul50:
; RV64I: # %bb.0:
@@ -1044,6 +1096,54 @@ define i64 @addmul100(i64 %a, i64 %b) {
ret i64 %d
}
+define i64 @mul145(i64 %a) {
+; RV64I-LABEL: mul145:
+; RV64I: # %bb.0:
+; RV64I-NEXT: li a1, 145
+; RV64I-NEXT: mul a0, a0, a1
+; RV64I-NEXT: ret
+;
+; RV64ZBA-LABEL: mul145:
+; RV64ZBA: # %bb.0:
+; RV64ZBA-NEXT: slli a1, a0, 4
+; RV64ZBA-NEXT: sh3add a1, a1, a1
+; RV64ZBA-NEXT: add a0, a1, a0
+; RV64ZBA-NEXT: ret
+;
+; RV64XANDESPERF-LABEL: mul145:
+; RV64XANDESPERF: # %bb.0:
+; RV64XANDESPERF-NEXT: slli a1, a0, 4
+; RV64XANDESPERF-NEXT: nds.lea.d a1, a1, a1
+; RV64XANDESPERF-NEXT: add a0, a1, a0
+; RV64XANDESPERF-NEXT: ret
+ %c = mul i64 %a, 145
+ ret i64 %c
+}
+
+define i64 @mul161(i64 %a) {
+; RV64I-LABEL: mul161:
+; RV64I: # %bb.0:
+; RV64I-NEXT: li a1, 161
+; RV64I-NEXT: mul a0, a0, a1
+; RV64I-NEXT: ret
+;
+; RV64ZBA-LABEL: mul161:
+; RV64ZBA: # %bb.0:
+; RV64ZBA-NEXT: slli a1, a0, 5
+; RV64ZBA-NEXT: sh2add a1, a1, a1
+; RV64ZBA-NEXT: add a0, a1, a0
+; RV64ZBA-NEXT: ret
+;
+; RV64XANDESPERF-LABEL: mul161:
+; RV64XANDESPERF: # %bb.0:
+; RV64XANDESPERF-NEXT: slli a1, a0, 5
+; RV64XANDESPERF-NEXT: nds.lea.w a1, a1, a1
+; RV64XANDESPERF-NEXT: add a0, a1, a0
+; RV64XANDESPERF-NEXT: ret
+ %c = mul i64 %a, 161
+ ret i64 %c
+}
+
define i64 @mul162(i64 %a) {
; RV64I-LABEL: mul162:
; RV64I: # %bb.0:
|
topperc
reviewed
Nov 9, 2025
| ; RV64ZBA: # %bb.0: | ||
| ; RV64ZBA-NEXT: slli a1, a0, 4 | ||
| ; RV64ZBA-NEXT: sh1add a1, a1, a1 | ||
| ; RV64ZBA-NEXT: add a0, a1, a0 |
Collaborator
There was a problem hiding this comment.
If my CPU has a 2 cycle multiply latency like SiFiveP400/600/800, this seems worse. Should I consider adding a tuning flag to disable this for my CPUs?
Contributor
Author
There was a problem hiding this comment.
This PR expands mul into 3 instructions, with a possible fold of zext into slli.uw.
Most of the transforms in expandMul expand into three instructions, so I assumed it's okay.
If not (for reasons other than hasMinSize() which is handled), I think we should handle that together with the other 3-instruction expansions.
Collaborator
There was a problem hiding this comment.
Ok, I agree with that.
Some of the transforms have ILP between 2 of the 3 instructions. This lowering is serial. There are probably other serial lowerings. On a core with more shift/add resources than multiply, a serial implementation this could still be beneficial if it prevents the multiplier from being a bottleneck.
On a single issue in order CPU with a pipelined multiplier, this may just be increasing the code size for no benefit. Freeing up the multiplier doesn't help if you can't issue any more instructions.
Contributor
Author
There was a problem hiding this comment.
There are probably other serial lowerings.
- 3/5/9 * 3/5/9 * 2^N
- (2/4/8 * 3/5/9 + 1) * 2^N
git-crd
pushed a commit
to git-crd/crd-llvm-project
that referenced
this pull request
Nov 13, 2025
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.
No description provided.