[llvm] 65e0574 - [RISCV] Expand mul X, C where C=2^N*(3, 5, 9)*(3, 5, 9) (#108100)
via llvm-commits
llvm-commits at lists.llvm.org
Wed Sep 11 07:33:10 PDT 2024
Author: Philip Reames
Date: 2024-09-11T07:33:06-07:00
New Revision: 65e0574076b86fd5adb432387c23ce9954fb95c4
URL: https://github.com/llvm/llvm-project/commit/65e0574076b86fd5adb432387c23ce9954fb95c4
DIFF: https://github.com/llvm/llvm-project/commit/65e0574076b86fd5adb432387c23ce9954fb95c4.diff
LOG: [RISCV] Expand mul X, C where C=2^N*(3,5,9)*(3,5,9) (#108100)
This is a three deep expression which is deeper than we've otherwise
gone for multiple expansions, but I think it's reasonable to do so. This
covers mul by 50, 100, and 200 which are reasonably common naturally
arising numbers.
Added:
Modified:
llvm/lib/Target/RISCV/RISCVISelLowering.cpp
llvm/test/CodeGen/RISCV/rv64zba.ll
Removed:
################################################################################
diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
index 23f2b0e96495e9..d2e62579bbddfc 100644
--- a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
+++ b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
@@ -14131,6 +14131,32 @@ static SDValue expandMul(SDNode *N, SelectionDAG &DAG,
return DAG.getNode(ISD::SUB, DL, VT, Shift1, Shift2);
}
+ if (HasShlAdd) {
+ for (uint64_t Divisor : {3, 5, 9}) {
+ if (MulAmt % Divisor != 0)
+ continue;
+ uint64_t MulAmt2 = MulAmt / Divisor;
+ // 3/5/9 * 3/5/9 * 2^N - In particular, this covers multiples
+ // of 25 which happen to be quite common.
+ for (uint64_t Divisor2 : {3, 5, 9}) {
+ if (MulAmt2 % Divisor2 != 0)
+ continue;
+ uint64_t MulAmt3 = MulAmt2 / Divisor2;
+ if (isPowerOf2_64(MulAmt3)) {
+ SDLoc DL(N);
+ SDValue Mul359A =
+ DAG.getNode(RISCVISD::SHL_ADD, DL, VT, X,
+ DAG.getConstant(Log2_64(Divisor - 1), DL, VT), X);
+ SDValue Mul359B = DAG.getNode(
+ RISCVISD::SHL_ADD, DL, VT, Mul359A,
+ DAG.getConstant(Log2_64(Divisor2 - 1), DL, VT), Mul359A);
+ return DAG.getNode(ISD::SHL, DL, VT, Mul359B,
+ DAG.getConstant(Log2_64(MulAmt3), DL, VT));
+ }
+ }
+ }
+ }
+
return SDValue();
}
diff --git a/llvm/test/CodeGen/RISCV/rv64zba.ll b/llvm/test/CodeGen/RISCV/rv64zba.ll
index 19d7324eeff4a1..a381ee67354b32 100644
--- a/llvm/test/CodeGen/RISCV/rv64zba.ll
+++ b/llvm/test/CodeGen/RISCV/rv64zba.ll
@@ -570,122 +570,200 @@ define i64 @addmul72(i64 %a, i64 %b) {
}
define i64 @mul50(i64 %a) {
-; CHECK-LABEL: mul50:
-; CHECK: # %bb.0:
-; CHECK-NEXT: li a1, 50
-; CHECK-NEXT: mul a0, a0, a1
-; CHECK-NEXT: ret
+; RV64I-LABEL: mul50:
+; RV64I: # %bb.0:
+; RV64I-NEXT: li a1, 50
+; RV64I-NEXT: mul a0, a0, a1
+; RV64I-NEXT: ret
+;
+; RV64ZBA-LABEL: mul50:
+; RV64ZBA: # %bb.0:
+; RV64ZBA-NEXT: sh2add a0, a0, a0
+; RV64ZBA-NEXT: sh2add a0, a0, a0
+; RV64ZBA-NEXT: slli a0, a0, 1
+; RV64ZBA-NEXT: ret
%c = mul i64 %a, 50
ret i64 %c
}
define i64 @addmul50(i64 %a, i64 %b) {
-; CHECK-LABEL: addmul50:
-; CHECK: # %bb.0:
-; CHECK-NEXT: li a2, 50
-; CHECK-NEXT: mul a0, a0, a2
-; CHECK-NEXT: add a0, a0, a1
-; CHECK-NEXT: ret
+; RV64I-LABEL: addmul50:
+; RV64I: # %bb.0:
+; RV64I-NEXT: li a2, 50
+; RV64I-NEXT: mul a0, a0, a2
+; RV64I-NEXT: add a0, a0, a1
+; RV64I-NEXT: ret
+;
+; RV64ZBA-LABEL: addmul50:
+; RV64ZBA: # %bb.0:
+; RV64ZBA-NEXT: sh2add a0, a0, a0
+; RV64ZBA-NEXT: sh2add a0, a0, a0
+; RV64ZBA-NEXT: sh1add a0, a0, a1
+; RV64ZBA-NEXT: ret
%c = mul i64 %a, 50
%d = add i64 %c, %b
ret i64 %d
}
define i64 @mul100(i64 %a) {
-; CHECK-LABEL: mul100:
-; CHECK: # %bb.0:
-; CHECK-NEXT: li a1, 100
-; CHECK-NEXT: mul a0, a0, a1
-; CHECK-NEXT: ret
+; RV64I-LABEL: mul100:
+; RV64I: # %bb.0:
+; RV64I-NEXT: li a1, 100
+; RV64I-NEXT: mul a0, a0, a1
+; RV64I-NEXT: ret
+;
+; RV64ZBA-LABEL: mul100:
+; RV64ZBA: # %bb.0:
+; RV64ZBA-NEXT: sh2add a0, a0, a0
+; RV64ZBA-NEXT: sh2add a0, a0, a0
+; RV64ZBA-NEXT: slli a0, a0, 2
+; RV64ZBA-NEXT: ret
%c = mul i64 %a, 100
ret i64 %c
}
define i64 @addmul100(i64 %a, i64 %b) {
-; CHECK-LABEL: addmul100:
-; CHECK: # %bb.0:
-; CHECK-NEXT: li a2, 100
-; CHECK-NEXT: mul a0, a0, a2
-; CHECK-NEXT: add a0, a0, a1
-; CHECK-NEXT: ret
+; RV64I-LABEL: addmul100:
+; RV64I: # %bb.0:
+; RV64I-NEXT: li a2, 100
+; RV64I-NEXT: mul a0, a0, a2
+; RV64I-NEXT: add a0, a0, a1
+; RV64I-NEXT: ret
+;
+; RV64ZBA-LABEL: addmul100:
+; RV64ZBA: # %bb.0:
+; RV64ZBA-NEXT: sh2add a0, a0, a0
+; RV64ZBA-NEXT: sh2add a0, a0, a0
+; RV64ZBA-NEXT: sh2add a0, a0, a1
+; RV64ZBA-NEXT: ret
%c = mul i64 %a, 100
%d = add i64 %c, %b
ret i64 %d
}
define i64 @mul162(i64 %a) {
-; CHECK-LABEL: mul162:
-; CHECK: # %bb.0:
-; CHECK-NEXT: li a1, 162
-; CHECK-NEXT: mul a0, a0, a1
-; CHECK-NEXT: ret
+; RV64I-LABEL: mul162:
+; RV64I: # %bb.0:
+; RV64I-NEXT: li a1, 162
+; RV64I-NEXT: mul a0, a0, a1
+; RV64I-NEXT: ret
+;
+; RV64ZBA-LABEL: mul162:
+; RV64ZBA: # %bb.0:
+; RV64ZBA-NEXT: sh3add a0, a0, a0
+; RV64ZBA-NEXT: sh3add a0, a0, a0
+; RV64ZBA-NEXT: slli a0, a0, 1
+; RV64ZBA-NEXT: ret
%c = mul i64 %a, 162
ret i64 %c
}
define i64 @addmul162(i64 %a, i64 %b) {
-; CHECK-LABEL: addmul162:
-; CHECK: # %bb.0:
-; CHECK-NEXT: li a2, 162
-; CHECK-NEXT: mul a0, a0, a2
-; CHECK-NEXT: add a0, a0, a1
-; CHECK-NEXT: ret
+; RV64I-LABEL: addmul162:
+; RV64I: # %bb.0:
+; RV64I-NEXT: li a2, 162
+; RV64I-NEXT: mul a0, a0, a2
+; RV64I-NEXT: add a0, a0, a1
+; RV64I-NEXT: ret
+;
+; RV64ZBA-LABEL: addmul162:
+; RV64ZBA: # %bb.0:
+; RV64ZBA-NEXT: sh3add a0, a0, a0
+; RV64ZBA-NEXT: sh3add a0, a0, a0
+; RV64ZBA-NEXT: sh1add a0, a0, a1
+; RV64ZBA-NEXT: ret
%c = mul i64 %a, 162
%d = add i64 %c, %b
ret i64 %d
}
define i64 @mul180(i64 %a) {
-; CHECK-LABEL: mul180:
-; CHECK: # %bb.0:
-; CHECK-NEXT: li a1, 180
-; CHECK-NEXT: mul a0, a0, a1
-; CHECK-NEXT: ret
+; RV64I-LABEL: mul180:
+; RV64I: # %bb.0:
+; RV64I-NEXT: li a1, 180
+; RV64I-NEXT: mul a0, a0, a1
+; RV64I-NEXT: ret
+;
+; RV64ZBA-LABEL: mul180:
+; RV64ZBA: # %bb.0:
+; RV64ZBA-NEXT: sh2add a0, a0, a0
+; RV64ZBA-NEXT: sh3add a0, a0, a0
+; RV64ZBA-NEXT: slli a0, a0, 2
+; RV64ZBA-NEXT: ret
%c = mul i64 %a, 180
ret i64 %c
}
define i64 @addmul180(i64 %a, i64 %b) {
-; CHECK-LABEL: addmul180:
-; CHECK: # %bb.0:
-; CHECK-NEXT: li a2, 180
-; CHECK-NEXT: mul a0, a0, a2
-; CHECK-NEXT: add a0, a0, a1
-; CHECK-NEXT: ret
+; RV64I-LABEL: addmul180:
+; RV64I: # %bb.0:
+; RV64I-NEXT: li a2, 180
+; RV64I-NEXT: mul a0, a0, a2
+; RV64I-NEXT: add a0, a0, a1
+; RV64I-NEXT: ret
+;
+; RV64ZBA-LABEL: addmul180:
+; RV64ZBA: # %bb.0:
+; RV64ZBA-NEXT: sh2add a0, a0, a0
+; RV64ZBA-NEXT: sh3add a0, a0, a0
+; RV64ZBA-NEXT: sh2add a0, a0, a1
+; RV64ZBA-NEXT: ret
%c = mul i64 %a, 180
%d = add i64 %c, %b
ret i64 %d
}
define i64 @add255mul180(i64 %a) {
-; CHECK-LABEL: add255mul180:
-; CHECK: # %bb.0:
-; CHECK-NEXT: li a1, 180
-; CHECK-NEXT: mul a0, a0, a1
-; CHECK-NEXT: addi a0, a0, 255
-; CHECK-NEXT: ret
+; RV64I-LABEL: add255mul180:
+; RV64I: # %bb.0:
+; RV64I-NEXT: li a1, 180
+; RV64I-NEXT: mul a0, a0, a1
+; RV64I-NEXT: addi a0, a0, 255
+; RV64I-NEXT: ret
+;
+; RV64ZBA-LABEL: add255mul180:
+; RV64ZBA: # %bb.0:
+; RV64ZBA-NEXT: sh2add a0, a0, a0
+; RV64ZBA-NEXT: sh3add a0, a0, a0
+; RV64ZBA-NEXT: slli a0, a0, 2
+; RV64ZBA-NEXT: addi a0, a0, 255
+; RV64ZBA-NEXT: ret
%c = mul i64 %a, 180
%d = add i64 %c, 255
ret i64 %d
}
define i64 @mul200(i64 %a) {
-; CHECK-LABEL: mul200:
-; CHECK: # %bb.0:
-; CHECK-NEXT: li a1, 200
-; CHECK-NEXT: mul a0, a0, a1
-; CHECK-NEXT: ret
+; RV64I-LABEL: mul200:
+; RV64I: # %bb.0:
+; RV64I-NEXT: li a1, 200
+; RV64I-NEXT: mul a0, a0, a1
+; RV64I-NEXT: ret
+;
+; RV64ZBA-LABEL: mul200:
+; RV64ZBA: # %bb.0:
+; RV64ZBA-NEXT: sh2add a0, a0, a0
+; RV64ZBA-NEXT: sh2add a0, a0, a0
+; RV64ZBA-NEXT: slli a0, a0, 3
+; RV64ZBA-NEXT: ret
%c = mul i64 %a, 200
ret i64 %c
}
define i64 @addmul200(i64 %a, i64 %b) {
-; CHECK-LABEL: addmul200:
-; CHECK: # %bb.0:
-; CHECK-NEXT: li a2, 200
-; CHECK-NEXT: mul a0, a0, a2
-; CHECK-NEXT: add a0, a0, a1
-; CHECK-NEXT: ret
+; RV64I-LABEL: addmul200:
+; RV64I: # %bb.0:
+; RV64I-NEXT: li a2, 200
+; RV64I-NEXT: mul a0, a0, a2
+; RV64I-NEXT: add a0, a0, a1
+; RV64I-NEXT: ret
+;
+; RV64ZBA-LABEL: addmul200:
+; RV64ZBA: # %bb.0:
+; RV64ZBA-NEXT: sh2add a0, a0, a0
+; RV64ZBA-NEXT: sh2add a0, a0, a0
+; RV64ZBA-NEXT: sh3add a0, a0, a1
+; RV64ZBA-NEXT: ret
%c = mul i64 %a, 200
%d = add i64 %c, %b
ret i64 %d
More information about the llvm-commits
mailing list