-
Notifications
You must be signed in to change notification settings - Fork 12.9k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[RISCV] Avoid matching 3/5/9 * 2^N as 2^N + 2/4/8 (e.g. 24) #88937
Conversation
The former is better as a zero extend can be folded into the sll, whereas the later currently produces a seperate zext.w due to bad interactions with other combines.
@llvm/pr-subscribers-backend-risc-v Author: Philip Reames (preames) ChangesThe former is better as a zero extend can be folded into the sll, whereas the later currently produces a seperate zext.w due to bad interactions with other combines. Full diff: https://github.com/llvm/llvm-project/pull/88937.diff 3 Files Affected:
diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
index 765838aafb58d2..de2ad639f0d6c8 100644
--- a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
+++ b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
@@ -13416,6 +13416,12 @@ static SDValue expandMul(SDNode *N, SelectionDAG &DAG,
return SDValue();
uint64_t MulAmt = CNode->getZExtValue();
+ // 3/5/9 * 2^N -> shXadd (sll X, C), (sll X, C)
+ // Matched in tablegen, avoid perturbing patterns.
+ for (uint64_t Divisor : {3, 5, 9})
+ if (MulAmt % Divisor == 0 && isPowerOf2_64(MulAmt / Divisor))
+ return SDValue();
+
// If this is a power 2 + 2/4/8, we can use a shift followed by a single
// shXadd. First check if this a sum of two power of 2s because that's
// easy. Then count how many zeros are up to the first bit.
diff --git a/llvm/test/CodeGen/RISCV/addimm-mulimm.ll b/llvm/test/CodeGen/RISCV/addimm-mulimm.ll
index 10103f071462c5..48fa69e1045656 100644
--- a/llvm/test/CodeGen/RISCV/addimm-mulimm.ll
+++ b/llvm/test/CodeGen/RISCV/addimm-mulimm.ll
@@ -551,9 +551,8 @@ define i64 @add_mul_combine_infinite_loop(i64 %x) {
; RV32IMB-NEXT: sh3add a1, a1, a2
; RV32IMB-NEXT: sh1add a0, a0, a0
; RV32IMB-NEXT: slli a2, a0, 3
-; RV32IMB-NEXT: li a3, 1
-; RV32IMB-NEXT: slli a3, a3, 11
-; RV32IMB-NEXT: sh3add a0, a0, a3
+; RV32IMB-NEXT: addi a0, a2, 2047
+; RV32IMB-NEXT: addi a0, a0, 1
; RV32IMB-NEXT: sltu a2, a0, a2
; RV32IMB-NEXT: add a1, a1, a2
; RV32IMB-NEXT: ret
@@ -562,8 +561,8 @@ define i64 @add_mul_combine_infinite_loop(i64 %x) {
; RV64IMB: # %bb.0:
; RV64IMB-NEXT: addi a0, a0, 86
; RV64IMB-NEXT: sh1add a0, a0, a0
-; RV64IMB-NEXT: slli a0, a0, 3
-; RV64IMB-NEXT: addi a0, a0, -16
+; RV64IMB-NEXT: li a1, -16
+; RV64IMB-NEXT: sh3add a0, a0, a1
; RV64IMB-NEXT: ret
%tmp0 = mul i64 %x, 24
%tmp1 = add i64 %tmp0, 2048
diff --git a/llvm/test/CodeGen/RISCV/rv64zba.ll b/llvm/test/CodeGen/RISCV/rv64zba.ll
index a84b9e5e7962f6..c3c757656be933 100644
--- a/llvm/test/CodeGen/RISCV/rv64zba.ll
+++ b/llvm/test/CodeGen/RISCV/rv64zba.ll
@@ -2490,3 +2490,25 @@ define ptr @test_gep_gep_dont_crash(ptr %p, i64 %a1, i64 %a2) {
%gep2 = getelementptr i64, ptr %gep1, i64 %a1
ret ptr %gep2
}
+
+define i64 @regression(i32 signext %x, i32 signext %y) {
+; RV64I-LABEL: regression:
+; RV64I: # %bb.0:
+; RV64I-NEXT: subw a0, a0, a1
+; RV64I-NEXT: slli a0, a0, 32
+; RV64I-NEXT: li a1, 3
+; RV64I-NEXT: slli a1, a1, 35
+; RV64I-NEXT: mulhu a0, a0, a1
+; RV64I-NEXT: ret
+;
+; RV64ZBA-LABEL: regression:
+; RV64ZBA: # %bb.0:
+; RV64ZBA-NEXT: subw a0, a0, a1
+; RV64ZBA-NEXT: slli.uw a0, a0, 3
+; RV64ZBA-NEXT: sh1add a0, a0, a0
+; RV64ZBA-NEXT: ret
+ %sub = sub i32 %x, %y
+ %ext = zext i32 %sub to i64
+ %res = mul nuw nsw i64 %ext, 24
+ ret i64 %res
+}
|
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
LGTM
The former is better as a zero extend can be folded into the sll, whereas the later currently produces a seperate zext.w due to bad interactions with other combines.