[NFC][AArch64] Precommit test to optimize instruction selection for aarch64_neon_pmull64 intrinsic.

[mingmingl]

Summary

Add test cases when instruction selection for aarch64_neon_pmull64 intrinsic is sub-optimal.

This is to prepare for D131047, and makes diff clearer.

[mingmingl]

DO NOT SEND

Diffbase to show an alternative approach

[mingmingl]

add another test case test2 in llvm/test/CodeGen/AArch64/aarch64-pmull2.ll

[mingmingl]

simplify test case a little bit

[mingmingl]

simplify test case, by removing unnecessary xor, etc

[mingmingl]

Update revision title and summary. Use same comments for NFC and actual patch, so only assembly diff is shown in phabricator UI (not comment diff)

From discussion with colleagues I should probably ask for review for both pre-commit tests and the actual change in the first place to ensure the quality of new test cases (e.g., not overly sensitive to changes that might make it brittle/likely to fail for unrelated reasons, not insufficiently minimized so that it's hard to maintain/read/possibly and more likely to fail for unrelated reasons, have good readability in terms of naming, etc)

So do it now and will remember to ask reviews for both in the future. Hope it's not too late!

[dmgreen]

Oh yeah - LGTM

llvm/test/CodeGen/AArch64/aarch64-pmull2.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -verify-machineinstrs  -mtriple=aarch64-linux-gnu -mattr=+aes -o - %s| FileCheck %s --check-prefixes=CHECK
+
+; User code intends to execute {pmull, pmull2} instructions on {lower, higher} half of the same vector registers directly.
+; Test that PMULL2 are generated for higher-half operands.
+; The suboptimal code generation fails to use higher-half contents in place; instead, it moves higher-lane contents to lower lane
+; to make use of PMULL everywhere, and generates unnecessary moves.
+define void @test1(ptr %0, ptr %1) {
+; CHECK-LABEL: test1:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ldp q0, q1, [x1]
+; CHECK-NEXT:    mov w8, #56824
+; CHECK-NEXT:    mov w9, #61186
+; CHECK-NEXT:    movk w8, #40522, lsl #16
+; CHECK-NEXT:    movk w9, #29710, lsl #16
+; CHECK-NEXT:    mov x10, v0.d[1]
+; CHECK-NEXT:    fmov d2, x9
+; CHECK-NEXT:    mov x11, v1.d[1]
+; CHECK-NEXT:    fmov d3, x8
+; CHECK-NEXT:    fmov d4, x10
+; CHECK-NEXT:    pmull v0.1q, v0.1d, v2.1d
+; CHECK-NEXT:    fmov d5, x11
+; CHECK-NEXT:    pmull v1.1q, v1.1d, v2.1d
+; CHECK-NEXT:    pmull v2.1q, v4.1d, v3.1d
+; CHECK-NEXT:    pmull v3.1q, v5.1d, v3.1d
+; CHECK-NEXT:    eor v0.16b, v0.16b, v2.16b
+; CHECK-NEXT:    eor v1.16b, v1.16b, v3.16b
+; CHECK-NEXT:    stp q0, q1, [x1]
+; CHECK-NEXT:    ret
+  %3 = load <2 x i64>, ptr %1
+  %4 = getelementptr inbounds <2 x i64>, ptr %1, i64 1
+  %5 = load <2 x i64>, ptr %4
+  %6 = extractelement <2 x i64> %3, i64 1
+  %7 = tail call <16 x i8> @llvm.aarch64.neon.pmull64(i64 %6, i64 2655706616)
+  %8 = extractelement <2 x i64> %5, i64 1
+  %9 = tail call <16 x i8> @llvm.aarch64.neon.pmull64(i64 %8, i64 2655706616)
+  %10 = load <2 x i64>, ptr %0
+  %11 = getelementptr inbounds i8, ptr %0, i64 16
+  %12 = load <2 x i64>, ptr %11
+  %13 = extractelement <2 x i64> %3, i64 0
+  %14 = tail call <16 x i8> @llvm.aarch64.neon.pmull64(i64 %13, i64 1947135746)
+  %15 = extractelement <2 x i64> %5, i64 0
+  %16 = tail call <16 x i8> @llvm.aarch64.neon.pmull64(i64 %15, i64 1947135746)
+  %17 = xor <16 x i8> %14, %7
+  %18 = xor <16 x i8> %16, %9
+  store <16 x i8> %17, ptr %1
+  store <16 x i8> %18, ptr %4
+  ret void
+}
+
+; One operand is higher-half of SIMD register, and the other operand is lower-half of another SIMD register.
+; Tests that codegen doesn't generate unnecessary moves.
+define void @test2(ptr %0, <2 x i64> %1, <2 x i64> %2) {
+; CHECK-LABEL: test2:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov x8, v0.d[1]
+; CHECK-NEXT:    fmov d0, x8
+; CHECK-NEXT:    pmull v0.1q, v0.1d, v1.1d
+; CHECK-NEXT:    str q0, [x0]
+; CHECK-NEXT:    ret
+  %4 = extractelement <2 x i64> %1, i64 1
+  %5 = extractelement <2 x i64> %2, i64 0
+  %6 = tail call <16 x i8> @llvm.aarch64.neon.pmull64(i64 %4, i64 %5)
+  store <16 x i8> %6, ptr %0, align 16
+  ret void
+}
+
+declare <16 x i8> @llvm.aarch64.neon.pmull64(i64, i64)
 No newline at end of file

llvm/test/CodeGen/AArch64/pmull-ldr-merge.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -verify-machineinstrs  -mtriple=aarch64-linux-gnu -mattr=+aes -o - %s| FileCheck %s --check-prefixes=CHECK
 
-; Tests that scalar i64 arguments of llvm.aarch64.neon.pmull64 are
-; loaded into SIMD registers, as opposed to being loaded into GPR followed by a mov.
-
+; Two operands are in scalar form.
+; Tests that both operands are loaded into SIMD registers directly as opposed to being loaded into GPR followed by a fmov.
 define void @test1(ptr %0, i64 %1, i64 %2) {
 ; CHECK-LABEL: test1:
 ; CHECK:       // %bb.0:
 ; CHECK-NEXT:    add x8, x0, x2, lsl #4
 ; CHECK-NEXT:    add x9, x0, x1, lsl #4
 ; CHECK-NEXT:    ldr d0, [x8, #8]
 ; CHECK-NEXT:    ldr d1, [x9, #8]
 ; CHECK-NEXT:    pmull v0.1q, v1.1d, v0.1d
-; CHECK-NEXT:    str q0, [x9]
+; CHECK-NEXT:    str q0, [x0]
 ; CHECK-NEXT:    ret
   %4 = getelementptr inbounds <2 x i64>, ptr %0, i64 %1
   %5 = getelementptr inbounds <2 x i64>, ptr %0, i64 %1, i64 1
   %6 = load i64, ptr %5, align 8
   %7 = getelementptr inbounds <2 x i64>, ptr %0, i64 %2, i64 1
   %8 = load i64, ptr %7, align 8
   %9 = tail call <16 x i8> @llvm.aarch64.neon.pmull64(i64 %6, i64 %8)
-  store <16 x i8> %9, ptr %4, align 16
+  store <16 x i8> %9, ptr %0, align 16
   ret void
 }
 
+; Operand %8 is higher-half of v2i64, and operand %7 is a scalar load.
+; Tests that operand is loaded into SIMD registers directly as opposed to being loaded into GPR followed by a fmov.
 define void @test2(ptr %0, i64 %1, i64 %2, <2 x i64> %3) {
 ; CHECK-LABEL: test2:
 ; CHECK:       // %bb.0:
 ; CHECK-NEXT:    mov x9, v0.d[1]
 ; CHECK-NEXT:    add x8, x0, x1, lsl #4
 ; CHECK-NEXT:    ldr d0, [x8, #8]
 ; CHECK-NEXT:    fmov d1, x9
-; CHECK-NEXT:    pmull v0.1q, v0.1d, v1.1d
+; CHECK-NEXT:    pmull v0.1q, v1.1d, v0.1d
 ; CHECK-NEXT:    str q0, [x8]
 ; CHECK-NEXT:    ret
   %5 = getelementptr inbounds <2 x i64>, ptr %0, i64 %1
   %6 = getelementptr inbounds <2 x i64>, ptr %0, i64 %1, i64 1
   %7 = load i64, ptr %6, align 8
   %8 = extractelement <2 x i64> %3, i64 1
-  %9 = tail call <16 x i8> @llvm.aarch64.neon.pmull64(i64 %7, i64 %8)
+  %9 = tail call <16 x i8> @llvm.aarch64.neon.pmull64(i64 %8, i64 %7)
   store <16 x i8> %9, ptr %5, align 16
   ret void
 }
 
+; Operand %7 is a scalar load, and operand %3 is an input parameter of function `test4`.
+; Test that %7 is loaded into SIMD registers.
 define void @test3(ptr %0, i64 %1, i64 %2, i64 %3) {
 ; CHECK-LABEL: test3:
 ; CHECK:       // %bb.0:
 ; CHECK-NEXT:    add x8, x0, x1, lsl #4
 ; CHECK-NEXT:    fmov d0, x3
 ; CHECK-NEXT:    ldr d1, [x8, #8]
 ; CHECK-NEXT:    pmull v0.1q, v1.1d, v0.1d
 ; CHECK-NEXT:    str q0, [x8]
 ; CHECK-NEXT:    ret
   %5 = getelementptr inbounds <2 x i64>, ptr %0, i64 %1
   %6 = getelementptr inbounds <2 x i64>, ptr %0, i64 %1, i64 1
   %7 = load i64, ptr %6, align 8
   %8 = tail call <16 x i8> @llvm.aarch64.neon.pmull64(i64 %7, i64 %3)
   store <16 x i8> %8, ptr %5, align 16
   ret void
 }
 
+; Operand %4 is the higher-half of v2i64, and operand %2 is an input parameter of i64.
+; Test that %2 is duplicated into the proper lane of SIMD directly for optimal codegen.
+define void @test4(ptr %0, <2 x i64> %1, i64 %2) {
+; CHECK-LABEL: test4:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov x8, v0.d[1]
+; CHECK-NEXT:    fmov d0, x1
+; CHECK-NEXT:    fmov d1, x8
+; CHECK-NEXT:    pmull v0.1q, v1.1d, v0.1d
+; CHECK-NEXT:    str q0, [x0]
+; CHECK-NEXT:    ret
+  %4 = extractelement <2 x i64> %1, i64 1
+  %5 = tail call <16 x i8> @llvm.aarch64.neon.pmull64(i64 %4, i64 %2)
+  store <16 x i8> %5, ptr %0, align 16
+  ret void
+}
+
 declare <16 x i8> @llvm.aarch64.neon.pmull64(i64, i64)