[llvm] 27d1910 - [ARM][ParallelDSP] Handle squaring multiplies
Sam Parker via llvm-commits
llvm-commits at lists.llvm.org
Tue Apr 21 00:40:18 PDT 2020
Author: Sam Parker
Date: 2020-04-21T08:39:56+01:00
New Revision: 27d19101e9a3bf6cabcc2b481847aa7b86087cc4
URL: https://github.com/llvm/llvm-project/commit/27d19101e9a3bf6cabcc2b481847aa7b86087cc4
DIFF: https://github.com/llvm/llvm-project/commit/27d19101e9a3bf6cabcc2b481847aa7b86087cc4.diff
LOG: [ARM][ParallelDSP] Handle squaring multiplies
The logic in ARMParallelDSP is setup to merge two 16-bits loads into
a 32-bit load and feed them into the smlads. This requires that four
loads are combined for the four inputs, but there wasn't actually a
check for this.
Differential Revision: https://reviews.llvm.org/D78492
Added:
llvm/test/CodeGen/ARM/ParallelDSP/squaring.ll
Modified:
llvm/lib/Target/ARM/ARMParallelDSP.cpp
Removed:
################################################################################
diff --git a/llvm/lib/Target/ARM/ARMParallelDSP.cpp b/llvm/lib/Target/ARM/ARMParallelDSP.cpp
index bf90abc2a0e3..c916d17c375d 100644
--- a/llvm/lib/Target/ARM/ARMParallelDSP.cpp
+++ b/llvm/lib/Target/ARM/ARMParallelDSP.cpp
@@ -570,6 +570,10 @@ bool ARMParallelDSP::CreateParallelPairs(Reduction &R) {
auto Ld2 = static_cast<LoadInst*>(PMul0->RHS);
auto Ld3 = static_cast<LoadInst*>(PMul1->RHS);
+ // Check that each mul is operating on two
diff erent loads.
+ if (Ld0 == Ld2 || Ld1 == Ld3)
+ return false;
+
if (AreSequentialLoads(Ld0, Ld1, PMul0->VecLd)) {
if (AreSequentialLoads(Ld2, Ld3, PMul1->VecLd)) {
LLVM_DEBUG(dbgs() << "OK: found two pairs of parallel loads!\n");
diff --git a/llvm/test/CodeGen/ARM/ParallelDSP/squaring.ll b/llvm/test/CodeGen/ARM/ParallelDSP/squaring.ll
new file mode 100644
index 000000000000..661597975f10
--- /dev/null
+++ b/llvm/test/CodeGen/ARM/ParallelDSP/squaring.ll
@@ -0,0 +1,275 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -mtriple=armv8-a-linux-gnueabihf -arm-parallel-dsp -dce --verify %s -S -o - | FileCheck %s
+
+define dso_local void @a() align 2 {
+; CHECK-LABEL: @a(
+; CHECK-NEXT: for.end:
+; CHECK-NEXT: [[B:%.*]] = alloca i32, align 4
+; CHECK-NEXT: [[TMP0:%.*]] = load i16, i16* bitcast (void ()* @a to i16*), align 2
+; CHECK-NEXT: [[CONV:%.*]] = sext i16 [[TMP0]] to i32
+; CHECK-NEXT: [[MUL:%.*]] = mul nsw i32 [[CONV]], [[CONV]]
+; CHECK-NEXT: [[TMP1:%.*]] = load i16, i16* getelementptr (i16, i16* bitcast (void ()* @a to i16*), i32 1), align 2
+; CHECK-NEXT: [[CONV3:%.*]] = sext i16 [[TMP1]] to i32
+; CHECK-NEXT: [[MUL6:%.*]] = mul nsw i32 [[CONV3]], [[CONV3]]
+; CHECK-NEXT: [[ADD:%.*]] = add nuw nsw i32 [[MUL6]], [[MUL]]
+; CHECK-NEXT: [[TMP2:%.*]] = load i16, i16* getelementptr (i16, i16* bitcast (void ()* @a to i16*), i32 2), align 2
+; CHECK-NEXT: [[CONV11:%.*]] = sext i16 [[TMP2]] to i32
+; CHECK-NEXT: [[MUL12:%.*]] = mul nsw i32 [[CONV11]], [[CONV3]]
+; CHECK-NEXT: [[ADD14:%.*]] = add nsw i32 [[MUL12]], [[ADD]]
+; CHECK-NEXT: [[TMP3:%.*]] = load i16, i16* getelementptr (i16, i16* bitcast (void ()* @a to i16*), i32 3), align 2
+; CHECK-NEXT: [[CONV17:%.*]] = sext i16 [[TMP3]] to i32
+; CHECK-NEXT: [[ADD19:%.*]] = add nsw i32 [[ADD14]], [[CONV17]]
+; CHECK-NEXT: store i32 [[ADD19]], i32* [[B]], align 4
+; CHECK-NEXT: [[TMP4:%.*]] = load i16, i16* getelementptr (i16, i16* bitcast (void ()* @a to i16*), i32 4), align 2
+; CHECK-NEXT: [[CONV21:%.*]] = sext i16 [[TMP4]] to i32
+; CHECK-NEXT: [[ADD_PTR:%.*]] = getelementptr inbounds i32, i32* [[B]], i32 [[CONV21]]
+; CHECK-NEXT: [[ARRAYIDX22:%.*]] = getelementptr inbounds i32, i32* [[ADD_PTR]], i32 9
+; CHECK-NEXT: [[TMP5:%.*]] = load i32, i32* [[ARRAYIDX22]], align 4
+; CHECK-NEXT: [[SHL:%.*]] = shl i32 [[TMP5]], 1
+; CHECK-NEXT: store i32 [[SHL]], i32* [[ARRAYIDX22]], align 4
+; CHECK-NEXT: br label [[FOR_COND23:%.*]]
+; CHECK: for.cond23:
+; CHECK-NEXT: br label [[FOR_COND23]]
+;
+for.end:
+ %b = alloca i32, align 4
+ %0 = bitcast i32* %b to i8*
+ %1 = load i16, i16* bitcast (void ()* @a to i16*), align 2
+ %conv = sext i16 %1 to i32
+ %mul = mul nsw i32 %conv, %conv
+ %2 = load i16, i16* getelementptr (i16, i16* bitcast (void ()* @a to i16*), i32 1), align 2
+ %conv3 = sext i16 %2 to i32
+ %mul6 = mul nsw i32 %conv3, %conv3
+ %add = add nuw nsw i32 %mul6, %mul
+ %3 = load i16, i16* getelementptr (i16, i16* bitcast (void ()* @a to i16*), i32 2), align 2
+ %conv11 = sext i16 %3 to i32
+ %mul12 = mul nsw i32 %conv11, %conv3
+ %add14 = add nsw i32 %mul12, %add
+ %4 = load i16, i16* getelementptr (i16, i16* bitcast (void ()* @a to i16*), i32 3), align 2
+ %conv17 = sext i16 %4 to i32
+ %add19 = add nsw i32 %add14, %conv17
+ store i32 %add19, i32* %b, align 4
+ %5 = load i16, i16* getelementptr (i16, i16* bitcast (void ()* @a to i16*), i32 4), align 2
+ %conv21 = sext i16 %5 to i32
+ %add.ptr = getelementptr inbounds i32, i32* %b, i32 %conv21
+ %arrayidx22 = getelementptr inbounds i32, i32* %add.ptr, i32 9
+ %6 = load i32, i32* %arrayidx22, align 4
+ %shl = shl i32 %6, 1
+ store i32 %shl, i32* %arrayidx22, align 4
+ br label %for.cond23
+
+for.cond23: ; preds = %for.cond23, %for.end
+ br label %for.cond23
+}
+
+define i32 @accumulate_square_a0(i16* %a, i16* %b, i32 %acc) {
+; CHECK-LABEL: @accumulate_square_a0(
+; CHECK-NEXT: entry:
+; CHECK-NEXT: [[ADDR_A_1:%.*]] = getelementptr i16, i16* [[A:%.*]], i32 1
+; CHECK-NEXT: [[ADDR_B_1:%.*]] = getelementptr i16, i16* [[B:%.*]], i32 1
+; CHECK-NEXT: [[LD_A_0:%.*]] = load i16, i16* [[A]]
+; CHECK-NEXT: [[SEXT_A_0:%.*]] = sext i16 [[LD_A_0]] to i32
+; CHECK-NEXT: [[TMP0:%.*]] = bitcast i16* [[ADDR_A_1]] to i32*
+; CHECK-NEXT: [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
+; CHECK-NEXT: [[TMP2:%.*]] = trunc i32 [[TMP1]] to i16
+; CHECK-NEXT: [[TMP3:%.*]] = sext i16 [[TMP2]] to i32
+; CHECK-NEXT: [[TMP4:%.*]] = bitcast i16* [[ADDR_B_1]] to i32*
+; CHECK-NEXT: [[TMP5:%.*]] = load i32, i32* [[TMP4]], align 4
+; CHECK-NEXT: [[TMP6:%.*]] = trunc i32 [[TMP5]] to i16
+; CHECK-NEXT: [[TMP7:%.*]] = sext i16 [[TMP6]] to i32
+; CHECK-NEXT: [[MUL_0:%.*]] = mul i32 [[SEXT_A_0]], [[SEXT_A_0]]
+; CHECK-NEXT: [[TMP8:%.*]] = add i32 [[MUL_0]], [[ACC:%.*]]
+; CHECK-NEXT: [[MUL_1:%.*]] = mul i32 [[TMP3]], [[TMP7]]
+; CHECK-NEXT: [[TMP9:%.*]] = add i32 [[MUL_1]], [[TMP8]]
+; CHECK-NEXT: [[TMP10:%.*]] = call i32 @llvm.arm.smlad(i32 [[TMP1]], i32 [[TMP5]], i32 [[TMP9]])
+; CHECK-NEXT: ret i32 [[TMP10]]
+;
+entry:
+ %addr.a.1 = getelementptr i16, i16* %a, i32 1
+ %addr.b.1 = getelementptr i16, i16* %b, i32 1
+ %ld.a.0 = load i16, i16* %a
+ %sext.a.0 = sext i16 %ld.a.0 to i32
+ %ld.b.0 = load i16, i16* %b
+ %ld.a.1 = load i16, i16* %addr.a.1
+ %ld.b.1 = load i16, i16* %addr.b.1
+ %sext.a.1 = sext i16 %ld.a.1 to i32
+ %sext.b.1 = sext i16 %ld.b.1 to i32
+ %sext.b.0 = sext i16 %ld.b.0 to i32
+ %mul.0 = mul i32 %sext.a.0, %sext.a.0
+ %mul.1 = mul i32 %sext.a.1, %sext.b.1
+ %addr.a.2 = getelementptr i16, i16* %a, i32 2
+ %addr.b.2 = getelementptr i16, i16* %b, i32 2
+ %ld.a.2 = load i16, i16* %addr.a.2
+ %ld.b.2 = load i16, i16* %addr.b.2
+ %sext.a.2 = sext i16 %ld.a.2 to i32
+ %sext.b.2 = sext i16 %ld.b.2 to i32
+ %mul.2 = mul i32 %sext.a.2, %sext.b.2
+ %add = add i32 %mul.0, %mul.1
+ %add.1 = add i32 %mul.1, %mul.2
+ %add.2 = add i32 %add.1, %add
+ %res = add i32 %add.2, %acc
+ ret i32 %res
+}
+
+define i32 @accumulate_square_a2(i16* %a, i16* %b, i32 %acc) {
+; CHECK-LABEL: @accumulate_square_a2(
+; CHECK-NEXT: entry:
+; CHECK-NEXT: [[TMP0:%.*]] = bitcast i16* [[A:%.*]] to i32*
+; CHECK-NEXT: [[TMP1:%.*]] = load i32, i32* [[TMP0]], align 4
+; CHECK-NEXT: [[TMP2:%.*]] = lshr i32 [[TMP1]], 16
+; CHECK-NEXT: [[TMP3:%.*]] = trunc i32 [[TMP2]] to i16
+; CHECK-NEXT: [[TMP4:%.*]] = sext i16 [[TMP3]] to i32
+; CHECK-NEXT: [[TMP5:%.*]] = bitcast i16* [[B:%.*]] to i32*
+; CHECK-NEXT: [[TMP6:%.*]] = load i32, i32* [[TMP5]], align 4
+; CHECK-NEXT: [[TMP7:%.*]] = lshr i32 [[TMP6]], 16
+; CHECK-NEXT: [[TMP8:%.*]] = trunc i32 [[TMP7]] to i16
+; CHECK-NEXT: [[TMP9:%.*]] = sext i16 [[TMP8]] to i32
+; CHECK-NEXT: [[MUL_1:%.*]] = mul i32 [[TMP4]], [[TMP9]]
+; CHECK-NEXT: [[ADDR_A_2:%.*]] = getelementptr i16, i16* [[A]], i32 2
+; CHECK-NEXT: [[ADDR_B_2:%.*]] = getelementptr i16, i16* [[B]], i32 2
+; CHECK-NEXT: [[LD_A_2:%.*]] = load i16, i16* [[ADDR_A_2]]
+; CHECK-NEXT: [[LD_B_2:%.*]] = load i16, i16* [[ADDR_B_2]]
+; CHECK-NEXT: [[SEXT_A_2:%.*]] = sext i16 [[LD_A_2]] to i32
+; CHECK-NEXT: [[SEXT_B_2:%.*]] = sext i16 [[LD_B_2]] to i32
+; CHECK-NEXT: [[MUL_2:%.*]] = mul i32 [[SEXT_A_2]], [[SEXT_A_2]]
+; CHECK-NEXT: [[TMP10:%.*]] = add i32 [[MUL_2]], [[ACC:%.*]]
+; CHECK-NEXT: [[TMP11:%.*]] = add i32 [[MUL_1]], [[TMP10]]
+; CHECK-NEXT: [[TMP12:%.*]] = call i32 @llvm.arm.smlad(i32 [[TMP1]], i32 [[TMP6]], i32 [[TMP11]])
+; CHECK-NEXT: [[RES:%.*]] = add i32 [[TMP12]], [[SEXT_B_2]]
+; CHECK-NEXT: ret i32 [[RES]]
+;
+entry:
+ %addr.a.1 = getelementptr i16, i16* %a, i32 1
+ %addr.b.1 = getelementptr i16, i16* %b, i32 1
+ %ld.a.0 = load i16, i16* %a
+ %sext.a.0 = sext i16 %ld.a.0 to i32
+ %ld.b.0 = load i16, i16* %b
+ %ld.a.1 = load i16, i16* %addr.a.1
+ %ld.b.1 = load i16, i16* %addr.b.1
+ %sext.a.1 = sext i16 %ld.a.1 to i32
+ %sext.b.1 = sext i16 %ld.b.1 to i32
+ %sext.b.0 = sext i16 %ld.b.0 to i32
+ %mul.0 = mul i32 %sext.a.0, %sext.b.0
+ %mul.1 = mul i32 %sext.a.1, %sext.b.1
+ %addr.a.2 = getelementptr i16, i16* %a, i32 2
+ %addr.b.2 = getelementptr i16, i16* %b, i32 2
+ %ld.a.2 = load i16, i16* %addr.a.2
+ %ld.b.2 = load i16, i16* %addr.b.2
+ %sext.a.2 = sext i16 %ld.a.2 to i32
+ %sext.b.2 = sext i16 %ld.b.2 to i32
+ %mul.2 = mul i32 %sext.a.2, %sext.a.2
+ %add = add i32 %mul.0, %mul.1
+ %add.1 = add i32 %mul.1, %mul.2
+ %add.2 = add i32 %add.1, %add
+ %add.3 = add i32 %add.2, %acc
+ %res = add i32 %add.3, %sext.b.2
+ ret i32 %res
+}
+
+define i32 @accumulate_square_b2(i16* %a, i16* %b, i32 %acc) {
+; CHECK-LABEL: @accumulate_square_b2(
+; CHECK-NEXT: entry:
+; CHECK-NEXT: [[ADDR_A_1:%.*]] = getelementptr i16, i16* [[A:%.*]], i32 1
+; CHECK-NEXT: [[ADDR_B_1:%.*]] = getelementptr i16, i16* [[B:%.*]], i32 1
+; CHECK-NEXT: [[LD_A_0:%.*]] = load i16, i16* [[A]]
+; CHECK-NEXT: [[SEXT_A_0:%.*]] = sext i16 [[LD_A_0]] to i32
+; CHECK-NEXT: [[LD_A_1:%.*]] = load i16, i16* [[ADDR_A_1]]
+; CHECK-NEXT: [[LD_B_1:%.*]] = load i16, i16* [[ADDR_B_1]]
+; CHECK-NEXT: [[SEXT_A_1:%.*]] = sext i16 [[LD_A_1]] to i32
+; CHECK-NEXT: [[SEXT_B_1:%.*]] = sext i16 [[LD_B_1]] to i32
+; CHECK-NEXT: [[MUL_0:%.*]] = mul i32 [[SEXT_A_0]], [[SEXT_A_0]]
+; CHECK-NEXT: [[MUL_1:%.*]] = mul i32 [[SEXT_A_1]], [[SEXT_B_1]]
+; CHECK-NEXT: [[ADDR_B_2:%.*]] = getelementptr i16, i16* [[B]], i32 2
+; CHECK-NEXT: [[LD_B_2:%.*]] = load i16, i16* [[ADDR_B_2]]
+; CHECK-NEXT: [[SEXT_B_2:%.*]] = sext i16 [[LD_B_2]] to i32
+; CHECK-NEXT: [[MUL_2:%.*]] = mul i32 [[SEXT_B_2]], [[SEXT_B_2]]
+; CHECK-NEXT: [[ADD:%.*]] = add i32 [[MUL_0]], [[MUL_1]]
+; CHECK-NEXT: [[ADD_1:%.*]] = add i32 [[MUL_1]], [[MUL_2]]
+; CHECK-NEXT: [[ADD_2:%.*]] = add i32 [[ADD_1]], [[ADD]]
+; CHECK-NEXT: [[RES:%.*]] = add i32 [[ADD_2]], [[ACC:%.*]]
+; CHECK-NEXT: ret i32 [[RES]]
+;
+entry:
+ %addr.a.1 = getelementptr i16, i16* %a, i32 1
+ %addr.b.1 = getelementptr i16, i16* %b, i32 1
+ %ld.a.0 = load i16, i16* %a
+ %sext.a.0 = sext i16 %ld.a.0 to i32
+ %ld.b.0 = load i16, i16* %b
+ %ld.a.1 = load i16, i16* %addr.a.1
+ %ld.b.1 = load i16, i16* %addr.b.1
+ %sext.a.1 = sext i16 %ld.a.1 to i32
+ %sext.b.1 = sext i16 %ld.b.1 to i32
+ %sext.b.0 = sext i16 %ld.b.0 to i32
+ %mul.0 = mul i32 %sext.a.0, %sext.a.0
+ %mul.1 = mul i32 %sext.a.1, %sext.b.1
+ %addr.a.2 = getelementptr i16, i16* %a, i32 2
+ %addr.b.2 = getelementptr i16, i16* %b, i32 2
+ %ld.a.2 = load i16, i16* %addr.a.2
+ %ld.b.2 = load i16, i16* %addr.b.2
+ %sext.a.2 = sext i16 %ld.a.2 to i32
+ %sext.b.2 = sext i16 %ld.b.2 to i32
+ %mul.2 = mul i32 %sext.b.2, %sext.b.2
+ %add = add i32 %mul.0, %mul.1
+ %add.1 = add i32 %mul.1, %mul.2
+ %add.2 = add i32 %add.1, %add
+ %add.3 = add i32 %add.2, %sext.a.2
+ %res = add i32 %add.2, %acc
+ ret i32 %res
+}
+
+define i32 @accumulate_square_a1(i16* %a, i16* %b, i32 %acc) {
+; CHECK-LABEL: @accumulate_square_a1(
+; CHECK-NEXT: entry:
+; CHECK-NEXT: [[ADDR_A_1:%.*]] = getelementptr i16, i16* [[A:%.*]], i32 1
+; CHECK-NEXT: [[ADDR_B_1:%.*]] = getelementptr i16, i16* [[B:%.*]], i32 1
+; CHECK-NEXT: [[LD_A_0:%.*]] = load i16, i16* [[A]]
+; CHECK-NEXT: [[SEXT_A_0:%.*]] = sext i16 [[LD_A_0]] to i32
+; CHECK-NEXT: [[LD_A_1:%.*]] = load i16, i16* [[ADDR_A_1]]
+; CHECK-NEXT: [[LD_B_1:%.*]] = load i16, i16* [[ADDR_B_1]]
+; CHECK-NEXT: [[SEXT_A_1:%.*]] = sext i16 [[LD_A_1]] to i32
+; CHECK-NEXT: [[SEXT_B_1:%.*]] = sext i16 [[LD_B_1]] to i32
+; CHECK-NEXT: [[MUL_0:%.*]] = mul i32 [[SEXT_A_0]], [[SEXT_A_0]]
+; CHECK-NEXT: [[MUL_1:%.*]] = mul i32 [[SEXT_A_1]], [[SEXT_A_1]]
+; CHECK-NEXT: [[ADDR_A_2:%.*]] = getelementptr i16, i16* [[A]], i32 2
+; CHECK-NEXT: [[ADDR_B_2:%.*]] = getelementptr i16, i16* [[B]], i32 2
+; CHECK-NEXT: [[LD_A_2:%.*]] = load i16, i16* [[ADDR_A_2]]
+; CHECK-NEXT: [[LD_B_2:%.*]] = load i16, i16* [[ADDR_B_2]]
+; CHECK-NEXT: [[SEXT_A_2:%.*]] = sext i16 [[LD_A_2]] to i32
+; CHECK-NEXT: [[SEXT_B_2:%.*]] = sext i16 [[LD_B_2]] to i32
+; CHECK-NEXT: [[MUL_2:%.*]] = mul i32 [[SEXT_A_2]], [[SEXT_B_2]]
+; CHECK-NEXT: [[ADD:%.*]] = add i32 [[MUL_1]], [[SEXT_B_1]]
+; CHECK-NEXT: [[ADD_1:%.*]] = add i32 [[MUL_0]], [[ADD]]
+; CHECK-NEXT: [[ADD_2:%.*]] = add i32 [[MUL_1]], [[MUL_2]]
+; CHECK-NEXT: [[ADD_3:%.*]] = add i32 [[ADD_2]], [[ADD_1]]
+; CHECK-NEXT: [[ADD_4:%.*]] = add i32 [[ADD_3]], [[SEXT_A_2]]
+; CHECK-NEXT: [[RES:%.*]] = add i32 [[ADD_4]], [[ACC:%.*]]
+; CHECK-NEXT: ret i32 [[RES]]
+;
+entry:
+ %addr.a.1 = getelementptr i16, i16* %a, i32 1
+ %addr.b.1 = getelementptr i16, i16* %b, i32 1
+ %ld.a.0 = load i16, i16* %a
+ %sext.a.0 = sext i16 %ld.a.0 to i32
+ %ld.b.0 = load i16, i16* %b
+ %ld.a.1 = load i16, i16* %addr.a.1
+ %ld.b.1 = load i16, i16* %addr.b.1
+ %sext.a.1 = sext i16 %ld.a.1 to i32
+ %sext.b.1 = sext i16 %ld.b.1 to i32
+ %sext.b.0 = sext i16 %ld.b.0 to i32
+ %mul.0 = mul i32 %sext.a.0, %sext.a.0
+ %mul.1 = mul i32 %sext.a.1, %sext.a.1
+ %addr.a.2 = getelementptr i16, i16* %a, i32 2
+ %addr.b.2 = getelementptr i16, i16* %b, i32 2
+ %ld.a.2 = load i16, i16* %addr.a.2
+ %ld.b.2 = load i16, i16* %addr.b.2
+ %sext.a.2 = sext i16 %ld.a.2 to i32
+ %sext.b.2 = sext i16 %ld.b.2 to i32
+ %mul.2 = mul i32 %sext.a.2, %sext.b.2
+ %add = add i32 %mul.1, %sext.b.1
+ %add.1 = add i32 %mul.0, %add
+ %add.2 = add i32 %mul.1, %mul.2
+ %add.3 = add i32 %add.2, %add.1
+ %add.4 = add i32 %add.3, %sext.a.2
+ %res = add i32 %add.4, %acc
+ ret i32 %res
+}
More information about the llvm-commits
mailing list