Skip to content

[GlobalISel] Port computeNumSignBits for G_MUL #167311

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.

Sign up for GitHub

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

Jump to bottom
Open
wants to merge 2 commits into
base: main
Choose a base branch
from
Open

[GlobalISel] Port computeNumSignBits for G_MUL #167311

wants to merge 2 commits into from

Conversation

@AnushaK6
Copy link

@AnushaK6 AnushaK6 commented Nov 10, 2025

This patch continues the effort in issue #150515 to port computeSignBits functionality from the SelectionDAG (SDAG) to GlobalISel (GISel).
Specifically, it introduces support for G_MUL and improves precision when analyzing known bits and sign bits.

Changes:

  • Implemented computeSignBits support for the G_MUL operation in GISel.
  • Added logic to distinguish between constant and non-constant operands. When the operand is a constant, we can determine the exact number of sign bits.
  • When the operand is a variable, conservatively compute an upper bound valid for all possible values.
  • Added special handling for constant cases such as multiplication by zero.

Added new tests to:

  • known-signbits.mir with targeted test cases for G_MUL
  • Also modified existing test checks that use mul

@github-actions
Copy link

github-actions bot commented Nov 10, 2025

Thank you for submitting a Pull Request (PR) to the LLVM Project!

This PR will be automatically labeled and the relevant teams will be notified.

If you wish to, you can add reviewers by using the "Reviewers" section on this page.

If this is not working for you, it is probably because you do not have write permissions for the repository. In which case you can instead tag reviewers by name in a comment by using @ followed by their GitHub username.

If you have received no comments on your PR for a week, you can request a review by "ping"ing the PR by adding a comment “Ping”. The common courtesy "ping" rate is once a week. Please remember that you are asking for valuable time from other developers.

If you have further questions, they may be answered by the LLVM GitHub User Guide.

You can also ask questions in a comment on this PR, on the LLVM Discord or on the forums.

@llvmbot
Copy link
Member

llvmbot commented Nov 10, 2025 •
edited
Loading

@llvm/pr-subscribers-backend-aarch64

@llvm/pr-subscribers-backend-risc-v

Author: None (AnushaK6)

Changes

This patch continues the effort in issue #150515 to port computeSignBits functionality from the SelectionDAG (SDAG) to GlobalISel (GISel).
Specifically, it introduces support for G_MUL and improves precision when analyzing known bits and sign bits.

Changes:

  • Implemented computeSignBits support for the G_MUL operation in GISel.
  • Added logic to distinguish between constant and non-constant operands. When the operand is a constant, we can determine the exact number of sign bits.
  • When the operand is a variable, conservatively compute an upper bound valid for all possible values.
  • Added special handling for constant cases such as multiplication by zero.

Added new tests to:

  • known-signbits.mir with targeted test cases for G_MUL
  • Also modified existing test checks that use mul

Patch is 151.36 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/167311.diff

6 Files Affected:

  • (modified) llvm/lib/CodeGen/GlobalISel/GISelValueTracking.cpp (+36)
  • (added) llvm/test/CodeGen/AArch64/GlobalISel/knownbits-mul.mir (+79)
  • (modified) llvm/test/CodeGen/AArch64/combine-sdiv.ll (+1-3)
  • (modified) llvm/test/CodeGen/AArch64/neon-dotreduce.ll (+1564-1564)
  • (modified) llvm/test/CodeGen/AArch64/rem-by-const.ll (+1-2)
  • (modified) llvm/test/CodeGen/RISCV/GlobalISel/div-by-constant.ll (+16-32)
diff --git a/llvm/lib/CodeGen/GlobalISel/GISelValueTracking.cpp b/llvm/lib/CodeGen/GlobalISel/GISelValueTracking.cpp
index c1fb8b6d78ff8..3b5c68f266fff 100644
--- a/llvm/lib/CodeGen/GlobalISel/GISelValueTracking.cpp
+++ b/llvm/lib/CodeGen/GlobalISel/GISelValueTracking.cpp
@@ -31,6 +31,7 @@
 #include "llvm/CodeGen/TargetLowering.h"
 #include "llvm/CodeGen/TargetOpcodes.h"
 #include "llvm/IR/ConstantRange.h"
+#include "llvm/IR/Constants.h"
 #include "llvm/IR/DerivedTypes.h"
 #include "llvm/IR/FMF.h"
 #include "llvm/MC/TargetRegistry.h"
@@ -2084,6 +2085,41 @@ unsigned GISelValueTracking::computeNumSignBits(Register R,
     FirstAnswer = std::min(Src1NumSignBits, Src2NumSignBits) - 1;
     break;
   }
+  case TargetOpcode::G_MUL: {
+    Register Src1 = MI.getOperand(1).getReg();
+    Register Src2 = MI.getOperand(2).getReg();
+
+    KnownBits Known1 = getKnownBits(Src1, DemandedElts, Depth + 1);
+    KnownBits Known2 = getKnownBits(Src2, DemandedElts, Depth + 1);
+
+    if (Known1.isZero() || Known2.isZero())
+      return TyBits;
+
+    auto C1 = getIConstantVRegValWithLookThrough(Src1, MRI);
+    auto C2 = getIConstantVRegValWithLookThrough(Src2, MRI);
+
+    if (C1 && C2) {
+      APInt Val1 = C1->Value;
+      APInt Val2 = C2->Value;
+      APInt Product = Val1 * Val2;
+      return Product.getNumSignBits();
+    }
+    unsigned Src1NumSignBits =
+        computeNumSignBits(Src1, DemandedElts, Depth + 1);
+    if (Src1NumSignBits == 1) {
+      return 1;
+    }
+    unsigned Src2NumSignBits =
+        computeNumSignBits(Src2, DemandedElts, Depth + 1);
+    if (Src2NumSignBits == 1) {
+      return 1;
+    }
+
+    unsigned OutValidBits =
+        (TyBits - Src1NumSignBits + 1) + (TyBits - Src2NumSignBits + 1);
+    FirstAnswer = OutValidBits > TyBits ? 1 : TyBits - OutValidBits + 1;
+    break;
+  }
   case TargetOpcode::G_FCMP:
   case TargetOpcode::G_ICMP: {
     bool IsFP = Opcode == TargetOpcode::G_FCMP;
diff --git a/llvm/test/CodeGen/AArch64/GlobalISel/knownbits-mul.mir b/llvm/test/CodeGen/AArch64/GlobalISel/knownbits-mul.mir
new file mode 100644
index 0000000000000..ea5281948a211
--- /dev/null
+++ b/llvm/test/CodeGen/AArch64/GlobalISel/knownbits-mul.mir
@@ -0,0 +1,79 @@
+# NOTE: Assertions have been autogenerated by utils/update_givaluetracking_test_checks.py UTC_ARGS: --version 5
+# RUN: llc -mtriple=aarch64 -passes="print<gisel-value-tracking>" -filetype=null %s 2>&1 | FileCheck %s
+
+---
+name:            ConstPositives
+body:             |
+  bb.1:
+  ; CHECK-LABEL: name: @ConstPositives
+  ; CHECK-NEXT: %0:_ KnownBits:00000011 SignBits:6
+  ; CHECK-NEXT: %1:_ KnownBits:00000101 SignBits:5
+  ; CHECK-NEXT: %2:_ KnownBits:00001111 SignBits:4
+    %0:_(s8) = G_CONSTANT i8 3
+    %1:_(s8) = G_CONSTANT i8 5
+    %2:_(s8) = G_MUL %0, %1
+...
+---
+name:            ConstZero
+body:             |
+  bb.1:
+  ; CHECK-LABEL: name: @ConstZero
+  ; CHECK-NEXT: %0:_ KnownBits:00000000 SignBits:8
+  ; CHECK-NEXT: %1:_ KnownBits:00000001 SignBits:7
+  ; CHECK-NEXT: %2:_ KnownBits:00000000 SignBits:8
+    %0:_(s8) = G_CONSTANT i8 0
+    %1:_(s8) = G_CONSTANT i8 1
+    %2:_(s8) = G_MUL %0, %1
+...
+---
+name:            ConstNegatives
+body:             |
+  bb.1:
+  ; CHECK-LABEL: name: @ConstNegatives
+  ; CHECK-NEXT: %0:_ KnownBits:11111110 SignBits:7
+  ; CHECK-NEXT: %1:_ KnownBits:11111100 SignBits:6
+  ; CHECK-NEXT: %2:_ KnownBits:00001000 SignBits:4
+    %0:_(s8) = G_CONSTANT i8 -2
+    %1:_(s8) = G_CONSTANT i8 -4
+    %2:_(s8) = G_MUL %0, %1
+...
+---
+name:            MixedSigns
+body:             |
+  bb.1:
+  ; CHECK-LABEL: name: @MixedSigns
+  ; CHECK-NEXT: %0:_ KnownBits:11111100 SignBits:6
+  ; CHECK-NEXT: %1:_ KnownBits:00000011 SignBits:6
+  ; CHECK-NEXT: %2:_ KnownBits:11110100 SignBits:4
+    %0:_(s8) = G_CONSTANT i8 -4
+    %1:_(s8) = G_CONSTANT i8 3
+    %2:_(s8) = G_MUL %0, %1
+...
+---
+name:            UnknownVar
+body:             |
+  bb.1:
+  ; CHECK-LABEL: name: @UnknownVar
+  ; CHECK-NEXT: %0:_ KnownBits:???????? SignBits:1
+  ; CHECK-NEXT: %1:_ KnownBits:00000010 SignBits:6
+  ; CHECK-NEXT: %2:_ KnownBits:???????0 SignBits:1
+    %0:_(s8) = COPY $b0
+    %1:_(s8) = G_CONSTANT i8 2
+    %2:_(s8) = G_MUL %0, %1
+...
+---
+name:            VectorMul
+body:             |
+  bb.1:
+  ; CHECK-LABEL: name: @VectorMul
+  ; CHECK-NEXT: %0:_ KnownBits:0000000000000001 SignBits:15
+  ; CHECK-NEXT: %1:_ KnownBits:0000000000000010 SignBits:14
+  ; CHECK-NEXT: %2:_ KnownBits:00000000000000?? SignBits:14
+  ; CHECK-NEXT: %3:_ KnownBits:00000000000000?? SignBits:14
+  ; CHECK-NEXT: %4:_ KnownBits:000000000000???? SignBits:12
+    %0:_(s16) = G_CONSTANT i16 1
+    %1:_(s16) = G_CONSTANT i16 2
+    %2:_(<4 x s16>) = G_BUILD_VECTOR %0, %1, %1, %0
+    %3:_(<4 x s16>) = G_BUILD_VECTOR %1, %0, %0, %1
+    %4:_(<4 x s16>) = G_MUL %2, %3
+...
diff --git a/llvm/test/CodeGen/AArch64/combine-sdiv.ll b/llvm/test/CodeGen/AArch64/combine-sdiv.ll
index cca190f08df2b..37ea64f700fd8 100644
--- a/llvm/test/CodeGen/AArch64/combine-sdiv.ll
+++ b/llvm/test/CodeGen/AArch64/combine-sdiv.ll
@@ -1510,7 +1510,6 @@ define i5 @combine_i5_sdiv_const7(i5 %x) {
 ; CHECK-GI-NEXT:    sbfx w9, w0, #0, #5
 ; CHECK-GI-NEXT:    sbfx w8, w8, #0, #5
 ; CHECK-GI-NEXT:    mul w8, w9, w8
-; CHECK-GI-NEXT:    sbfx w8, w8, #0, #10
 ; CHECK-GI-NEXT:    add w8, w0, w8, asr #5
 ; CHECK-GI-NEXT:    sbfx w8, w8, #0, #5
 ; CHECK-GI-NEXT:    asr w8, w8, #2
@@ -1560,7 +1559,6 @@ define i8 @combine_i8_sdiv_const7(i8 %x) {
 ; CHECK-GI-NEXT:    sxtb w8, w0
 ; CHECK-GI-NEXT:    mov w9, #-109 // =0xffffff93
 ; CHECK-GI-NEXT:    mul w8, w8, w9
-; CHECK-GI-NEXT:    sxth w8, w8
 ; CHECK-GI-NEXT:    add w8, w0, w8, asr #8
 ; CHECK-GI-NEXT:    sbfx w8, w8, #2, #6
 ; CHECK-GI-NEXT:    ubfx w9, w8, #7, #1
@@ -1585,7 +1583,7 @@ define i8 @combine_i8_sdiv_const100(i8 %x) {
 ; CHECK-GI-NEXT:    sxtb w8, w0
 ; CHECK-GI-NEXT:    mov w9, #41 // =0x29
 ; CHECK-GI-NEXT:    mul w8, w8, w9
-; CHECK-GI-NEXT:    sbfx w8, w8, #8, #8
+; CHECK-GI-NEXT:    asr w8, w8, #8
 ; CHECK-GI-NEXT:    asr w8, w8, #4
 ; CHECK-GI-NEXT:    ubfx w9, w8, #7, #1
 ; CHECK-GI-NEXT:    add w0, w8, w9
diff --git a/llvm/test/CodeGen/AArch64/neon-dotreduce.ll b/llvm/test/CodeGen/AArch64/neon-dotreduce.ll
index fb2a1fa697c26..581d8e116359b 100644
--- a/llvm/test/CodeGen/AArch64/neon-dotreduce.ll
+++ b/llvm/test/CodeGen/AArch64/neon-dotreduce.ll
@@ -2832,247 +2832,246 @@ define i32 @test_sdot_v25i8(ptr nocapture readonly %a, ptr nocapture readonly %b
 ; CHECK-GI-NEXT:    ldp q2, q1, [x1]
 ; CHECK-GI-NEXT:    movi d0, #0000000000000000
 ; CHECK-GI-NEXT:    str w2, [sp, #12] // 4-byte Folded Spill
-; CHECK-GI-NEXT:    mov b5, v2.b[2]
 ; CHECK-GI-NEXT:    mov b6, v2.b[3]
 ; CHECK-GI-NEXT:    mov b7, v2.b[4]
 ; CHECK-GI-NEXT:    mov b16, v2.b[5]
+; CHECK-GI-NEXT:    mov b19, v2.b[8]
+; CHECK-GI-NEXT:    mov b4, v2.b[1]
+; CHECK-GI-NEXT:    mov b5, v2.b[2]
 ; CHECK-GI-NEXT:    mov b17, v2.b[6]
 ; CHECK-GI-NEXT:    mov b18, v2.b[7]
-; CHECK-GI-NEXT:    mov b19, v2.b[8]
 ; CHECK-GI-NEXT:    mov b20, v2.b[9]
-; CHECK-GI-NEXT:    mov b21, v2.b[15]
-; CHECK-GI-NEXT:    mov b3, v2.b[1]
-; CHECK-GI-NEXT:    fmov w19, s2
-; CHECK-GI-NEXT:    mov b22, v1.b[6]
-; CHECK-GI-NEXT:    fmov w6, s5
-; CHECK-GI-NEXT:    mov b5, v2.b[10]
-; CHECK-GI-NEXT:    fmov w14, s6
-; CHECK-GI-NEXT:    mov b6, v2.b[11]
+; CHECK-GI-NEXT:    mov b21, v2.b[10]
+; CHECK-GI-NEXT:    mov b22, v2.b[11]
+; CHECK-GI-NEXT:    fmov w7, s2
+; CHECK-GI-NEXT:    fmov w13, s6
+; CHECK-GI-NEXT:    mov b6, v2.b[12]
 ; CHECK-GI-NEXT:    fmov w2, s7
-; CHECK-GI-NEXT:    stp s17, s18, [sp, #4] // 8-byte Folded Spill
-; CHECK-GI-NEXT:    mov b7, v2.b[12]
+; CHECK-GI-NEXT:    mov b7, v2.b[13]
 ; CHECK-GI-NEXT:    fmov w11, s16
-; CHECK-GI-NEXT:    sxtb w28, w19
-; CHECK-GI-NEXT:    mov b16, v2.b[13]
-; CHECK-GI-NEXT:    mov b18, v1.b[1]
-; CHECK-GI-NEXT:    sxtb w6, w6
-; CHECK-GI-NEXT:    mov b17, v2.b[14]
-; CHECK-GI-NEXT:    ldp q4, q2, [x0]
-; CHECK-GI-NEXT:    fmov w25, s19
+; CHECK-GI-NEXT:    mov b16, v2.b[14]
+; CHECK-GI-NEXT:    mov b23, v2.b[15]
+; CHECK-GI-NEXT:    ldp q3, q2, [x0]
+; CHECK-GI-NEXT:    fmov w26, s19
+; CHECK-GI-NEXT:    fmov w19, s4
+; CHECK-GI-NEXT:    stp s17, s18, [sp, #4] // 8-byte Folded Spill
+; CHECK-GI-NEXT:    fmov w29, s5
 ; CHECK-GI-NEXT:    fmov w24, s20
-; CHECK-GI-NEXT:    fmov w22, s5
-; CHECK-GI-NEXT:    mov b5, v1.b[2]
-; CHECK-GI-NEXT:    fmov w0, s6
-; CHECK-GI-NEXT:    sxtb w14, w14
-; CHECK-GI-NEXT:    mov b20, v1.b[3]
-; CHECK-GI-NEXT:    fmov w16, s7
-; CHECK-GI-NEXT:    mov b7, v1.b[4]
-; CHECK-GI-NEXT:    fmov w15, s16
-; CHECK-GI-NEXT:    sxtb w25, w25
+; CHECK-GI-NEXT:    sxtb w8, w7
+; CHECK-GI-NEXT:    mov b4, v3.b[2]
+; CHECK-GI-NEXT:    mov b5, v3.b[1]
+; CHECK-GI-NEXT:    sxtb w13, w13
+; CHECK-GI-NEXT:    mov b17, v1.b[1]
+; CHECK-GI-NEXT:    fmov w22, s21
+; CHECK-GI-NEXT:    sxtb w26, w26
+; CHECK-GI-NEXT:    mov b18, v1.b[2]
+; CHECK-GI-NEXT:    fmov w18, s22
 ; CHECK-GI-NEXT:    sxtb w24, w24
-; CHECK-GI-NEXT:    mov b16, v1.b[5]
-; CHECK-GI-NEXT:    fmov w13, s21
+; CHECK-GI-NEXT:    mov b19, v1.b[3]
+; CHECK-GI-NEXT:    fmov w16, s6
+; CHECK-GI-NEXT:    sxtb w19, w19
+; CHECK-GI-NEXT:    mov b21, v1.b[4]
+; CHECK-GI-NEXT:    fmov w15, s7
 ; CHECK-GI-NEXT:    sxtb w22, w22
-; CHECK-GI-NEXT:    mov b6, v4.b[2]
-; CHECK-GI-NEXT:    fmov w26, s18
-; CHECK-GI-NEXT:    sxtb w0, w0
-; CHECK-GI-NEXT:    mov b21, v1.b[7]
-; CHECK-GI-NEXT:    mov b18, v4.b[4]
-; CHECK-GI-NEXT:    fmov w7, s3
-; CHECK-GI-NEXT:    mov b3, v4.b[1]
-; CHECK-GI-NEXT:    fmov w12, s17
-; CHECK-GI-NEXT:    fmov w5, s5
-; CHECK-GI-NEXT:    mov b19, v4.b[3]
-; CHECK-GI-NEXT:    fmov w4, s20
-; CHECK-GI-NEXT:    fmov w3, s7
-; CHECK-GI-NEXT:    sxtb w29, w7
-; CHECK-GI-NEXT:    mov b17, v4.b[5]
-; CHECK-GI-NEXT:    fmov w1, s16
-; CHECK-GI-NEXT:    sxtb w5, w5
-; CHECK-GI-NEXT:    mov b16, v4.b[6]
-; CHECK-GI-NEXT:    fmov w18, s22
-; CHECK-GI-NEXT:    mov b7, v4.b[7]
-; CHECK-GI-NEXT:    fmov w17, s21
-; CHECK-GI-NEXT:    mov b5, v4.b[8]
-; CHECK-GI-NEXT:    mov b20, v4.b[9]
-; CHECK-GI-NEXT:    fmov w27, s6
-; CHECK-GI-NEXT:    mov b6, v4.b[10]
-; CHECK-GI-NEXT:    mov b21, v4.b[11]
-; CHECK-GI-NEXT:    fmov w21, s18
-; CHECK-GI-NEXT:    mov b18, v4.b[12]
-; CHECK-GI-NEXT:    mov b22, v4.b[13]
-; CHECK-GI-NEXT:    mov b23, v4.b[14]
-; CHECK-GI-NEXT:    fmov w10, s4
+; CHECK-GI-NEXT:    mov b7, v1.b[5]
+; CHECK-GI-NEXT:    mov b6, v3.b[3]
+; CHECK-GI-NEXT:    sxtb w11, w11
+; CHECK-GI-NEXT:    fmov w12, s23
+; CHECK-GI-NEXT:    mov b22, v1.b[6]
+; CHECK-GI-NEXT:    mov b23, v1.b[7]
+; CHECK-GI-NEXT:    mov b20, v3.b[4]
+; CHECK-GI-NEXT:    fmov w28, s4
+; CHECK-GI-NEXT:    fmov s4, w26
+; CHECK-GI-NEXT:    fmov w14, s16
+; CHECK-GI-NEXT:    fmov w27, s17
+; CHECK-GI-NEXT:    fmov w5, s18
+; CHECK-GI-NEXT:    sxtb w12, w12
+; CHECK-GI-NEXT:    fmov w4, s19
+; CHECK-GI-NEXT:    mov b19, v3.b[5]
+; CHECK-GI-NEXT:    sxtb w28, w28
+; CHECK-GI-NEXT:    fmov w3, s21
+; CHECK-GI-NEXT:    mov b18, v3.b[6]
 ; CHECK-GI-NEXT:    sxtb w27, w27
-; CHECK-GI-NEXT:    mov b24, v4.b[15]
-; CHECK-GI-NEXT:    fmov s4, w25
-; CHECK-GI-NEXT:    fmov w30, s3
-; CHECK-GI-NEXT:    fmov s3, w28
-; CHECK-GI-NEXT:    fmov w9, s5
-; CHECK-GI-NEXT:    sxtb w10, w10
-; CHECK-GI-NEXT:    fmov w7, s7
-; CHECK-GI-NEXT:    mov b7, v2.b[1]
+; CHECK-GI-NEXT:    sxtb w5, w5
+; CHECK-GI-NEXT:    fmov w1, s7
+; CHECK-GI-NEXT:    mov b16, v3.b[7]
+; CHECK-GI-NEXT:    fmov w0, s22
+; CHECK-GI-NEXT:    mov b17, v3.b[8]
+; CHECK-GI-NEXT:    fmov w17, s23
+; CHECK-GI-NEXT:    mov b7, v3.b[9]
+; CHECK-GI-NEXT:    fmov w30, s5
+; CHECK-GI-NEXT:    mov b5, v3.b[10]
+; CHECK-GI-NEXT:    mov b21, v3.b[11]
+; CHECK-GI-NEXT:    fmov w25, s6
+; CHECK-GI-NEXT:    mov b6, v3.b[12]
+; CHECK-GI-NEXT:    fmov w23, s20
+; CHECK-GI-NEXT:    mov b20, v3.b[13]
+; CHECK-GI-NEXT:    mov b22, v3.b[14]
+; CHECK-GI-NEXT:    fmov w6, s3
+; CHECK-GI-NEXT:    mov b23, v3.b[15]
+; CHECK-GI-NEXT:    fmov s3, w8
+; CHECK-GI-NEXT:    fmov w8, s1
 ; CHECK-GI-NEXT:    mov v4.h[1], w24
-; CHECK-GI-NEXT:    fmov w24, s1
-; CHECK-GI-NEXT:    fmov w8, s20
-; CHECK-GI-NEXT:    sxtb w9, w9
-; CHECK-GI-NEXT:    mov v3.h[1], w29
-; CHECK-GI-NEXT:    fmov w29, s6
-; CHECK-GI-NEXT:    fmov s6, w10
-; CHECK-GI-NEXT:    fmov w10, s2
-; CHECK-GI-NEXT:    fmov w19, s16
-; CHECK-GI-NEXT:    sxtb w24, w24
+; CHECK-GI-NEXT:    fmov w21, s19
+; CHECK-GI-NEXT:    mov b19, v2.b[1]
+; CHECK-GI-NEXT:    fmov w9, s17
+; CHECK-GI-NEXT:    fmov w24, s6
+; CHECK-GI-NEXT:    fmov w7, s16
+; CHECK-GI-NEXT:    mov b16, v2.b[2]
 ; CHECK-GI-NEXT:    sxtb w8, w8
-; CHECK-GI-NEXT:    mov b16, v2.b[3]
+; CHECK-GI-NEXT:    mov v3.h[1], w19
+; CHECK-GI-NEXT:    sxtb w19, w29
+; CHECK-GI-NEXT:    sxtb w9, w9
+; CHECK-GI-NEXT:    fmov w29, s5
+; CHECK-GI-NEXT:    mov v4.h[2], w22
+; CHECK-GI-NEXT:    sxtb w22, w6
+; CHECK-GI-NEXT:    fmov s5, w8
+; CHECK-GI-NEXT:    fmov w10, s7
+; CHECK-GI-NEXT:    fmov s7, w9
+; CHECK-GI-NEXT:    fmov w9, s16
+; CHECK-GI-NEXT:    fmov w20, s18
 ; CHECK-GI-NEXT:    sxtb w29, w29
-; CHECK-GI-NEXT:    fmov w23, s19
-; CHECK-GI-NEXT:    mov b19, v2.b[2]
+; CHECK-GI-NEXT:    fmov s6, w22
+; CHECK-GI-NEXT:    fmov w22, s2
 ; CHECK-GI-NEXT:    sxtb w10, w10
-; CHECK-GI-NEXT:    fmov s5, w24
-; CHECK-GI-NEXT:    sxtb w24, w30
-; CHECK-GI-NEXT:    mov v3.h[2], w6
-; CHECK-GI-NEXT:    sxtb w6, w26
-; CHECK-GI-NEXT:    fmov w28, s21
-; CHECK-GI-NEXT:    sxtb w23, w23
-; CHECK-GI-NEXT:    mov v6.h[1], w24
-; CHECK-GI-NEXT:    fmov w24, s7
-; CHECK-GI-NEXT:    fmov s7, w9
-; CHECK-GI-NEXT:    fmov w9, s19
-; CHECK-GI-NEXT:    mov v5.h[1], w6
-; CHECK-GI-NEXT:    mov v4.h[2], w22
-; CHECK-GI-NEXT:    fmov w20, s17
-; CHECK-GI-NEXT:    mov b17, v2.b[4]
-; CHECK-GI-NEXT:    sxtb w24, w24
-; CHECK-GI-NEXT:    mov v3.h[3], w14
-; CHECK-GI-NEXT:    sxtb w14, w2
+; CHECK-GI-NEXT:    mov v5.h[1], w27
+; CHECK-GI-NEXT:    sxtb w27, w30
 ; CHECK-GI-NEXT:    sxtb w9, w9
-; CHECK-GI-NEXT:    mov v7.h[1], w8
-; CHECK-GI-NEXT:    fmov w8, s16
-; CHECK-GI-NEXT:    fmov s16, w10
-; CHECK-GI-NEXT:    mov v6.h[2], w27
+; CHECK-GI-NEXT:    mov b18, v2.b[3]
+; CHECK-GI-NEXT:    mov v3.h[2], w19
+; CHECK-GI-NEXT:    sxtb w22, w22
+; CHECK-GI-NEXT:    mov v6.h[1], w27
+; CHECK-GI-NEXT:    fmov w27, s19
+; CHECK-GI-NEXT:    mov v7.h[1], w10
+; CHECK-GI-NEXT:    fmov w26, s21
+; CHECK-GI-NEXT:    mov b17, v2.b[4]
+; CHECK-GI-NEXT:    fmov s16, w22
 ; CHECK-GI-NEXT:    mov v5.h[2], w5
-; CHECK-GI-NEXT:    fmov w25, s18
-; CHECK-GI-NEXT:    mov v4.h[3], w0
-; CHECK-GI-NEXT:    sxtb w0, w4
-; CHECK-GI-NEXT:    sxtb w8, w8
-; CHECK-GI-NEXT:    mov b18, v2.b[5]
-; CHECK-GI-NEXT:    fmov w10, s17
-; CHECK-GI-NEXT:    mov v16.h[1], w24
+; CHECK-GI-NEXT:    sxtb w5, w25
+; CHECK-GI-NEXT:    sxtb w27, w27
+; CHECK-GI-NEXT:    fmov w10, s18
+; CHECK-GI-NEXT:    mov v3.h[3], w13
+; CHECK-GI-NEXT:    sxtb w13, w4
+; CHECK-GI-NEXT:    mov v6.h[2], w28
+; CHECK-GI-NEXT:    fmov w8, s20
+; CHECK-GI-NEXT:    mov v16.h[1], w27
 ; CHECK-GI-NEXT:    mov v7.h[2], w29
-; CHECK-GI-NEXT:    mov v3.h[4], w14
-; CHECK-GI-NEXT:    sxtb w14, w25
-; CHECK-GI-NEXT:    ldp x29, x30, [sp, #16] // 16-byte Folded Reload
+; CHECK-GI-NEXT:    mov b20, v2.b[5]
 ; CHECK-GI-NEXT:    sxtb w10, w10
-; CHECK-GI-NEXT:    mov v6.h[3], w23
-; CHECK-GI-NEXT:    mov v5.h[3], w0
-; CHECK-GI-NEXT:    fmov w26, s22
-; CHECK-GI-NEXT:    mov b19, v2.b[6]
-; CHECK-GI-NEXT:    fmov w27, s18
+; CHECK-GI-NEXT:    ldp x29, x30, [sp, #16] // 16-byte Folded Reload
+; CHECK-GI-NEXT:    sxtb w8, w8
+; CHECK-GI-NEXT:    fmov w22, s17
+; CHECK-GI-NEXT:    mov v5.h[3], w13
+; CHECK-GI-NEXT:    sxtb w13, w2
+; CHECK-GI-NEXT:    mov v6.h[3], w5
+; CHECK-GI-NEXT:    mov b21, v2.b[6]
 ; CHECK-GI-NEXT:    mov v16.h[2], w9
-; CHECK-GI-NEXT:    sxtb w9, w28
-; CHECK-GI-NEXT:    fmov w22, s23
+; CHECK-GI-NEXT:    sxtb w9, w18
+; CHECK-GI-NEXT:    sxtb w18, w23
+; CHECK-GI-NEXT:    mov v3.h[4], w13
+; CHECK-GI-NEXT:    sxtb w13, w24
+; CHECK-GI-NEXT:    fmov w27, s20
+; CHECK-GI-NEXT:    ldp x24, x23, [sp, #64] // 16-byte Folded Reload
+; CHECK-GI-NEXT:    mov v4.h[3], w9
+; CHECK-GI-NEXT:    sxtb w9, w26
+; CHECK-GI-NEXT:    ldp x26, x25, [sp, #48] // 16-byte Folded Reload
+; CHECK-GI-NEXT:    mov v16.h[3], w10
+; CHECK-GI-NEXT:    sxtb w10, w3
+; CHECK-GI-NEXT:    mov v6.h[4], w18
+; CHECK-GI-NEXT:    ldr w18, [sp, #4] // 4-byte Folded Reload
+; CHECK-GI-NEXT:    mov v7.h[3], w9
+; CHECK-GI-NEXT:    sxtb w9, w16
+; CHECK-GI-NEXT:    sxtb w16, w22
+; CHECK-GI-NEXT:    mov v5.h[4], w10
+; CHECK-GI-NEXT:    sxtb w10, w15
+; CHECK-GI-NEXT:    sxtb w18, w18
+; CHECK-GI-NEXT:    mov v4.h[4], w9
+; CHECK-GI-NEXT:    sxtb w9, w21
+; CHECK-GI-NEXT:    ldp x22, x21, [sp, #80] // 16-byte Folded Reload
+; CHECK-GI-NEXT:    mov v16.h[4], w16
+; CHECK-GI-NEXT:    mov v7.h[4], w13
+; CHECK-GI-NEXT:    ldr w13, [sp, #8] // 4-byte Folded Reload
+; CHECK-GI-NEXT:    mov v6.h[5], w9
+; CHECK-GI-NEXT:    sxtb w9, w1
+; CHECK-GI-NEXT:    mov v3.h[5], w11
+; CHECK-GI-NEXT:    sxtb w11, w27
+; CHECK-GI-NEXT:    fmov w19, s22
+; CHECK-GI-NEXT:    fmov w28, s21
+; CHECK-GI-NEXT:    sxtb w13, w13
 ; CHECK-GI-NEXT:    mov b17, v2.b[7]
-; CHECK-GI-NEXT:    fmov w6, s24
+; CHECK-GI-NEXT:    mov v5.h[5], w9
+; CHECK-GI-NEXT:    sxtb w9, w0
+; CHECK-GI-NEXT:    mov v4.h[5], w10
+; CHECK-GI-NEXT:    sxtb w10, w20
+; CHECK-GI-NEXT:    mov v7.h[5], w8
+; CHECK-GI-NEXT:    mov v16.h[5], w11
+; CHECK-GI-NEXT:    sxtb w8, w14
+; CHECK-GI-NEXT:    sxtb w11, w28
+; CHECK-GI-NEXT:    mov v6.h[6], w10
+; CHECK-GI-NEXT:    sxtb w10, w19
+; CHECK-GI-NEXT:    fmov w6, s23
+; CHECK-GI-NEXT:    mov v5.h[6], w9
+; CHECK-GI-NEXT:    fmov w9, s17
+; CHECK-GI-NEXT:    mov v3.h[6], w18
+; CHECK-GI-NEXT:    mov v4.h[6], w8
+; CHECK-GI-NEXT:    sxtb w8, w7
+; CHECK-GI-NEXT:    mov v7.h[6], w10
+; CHECK-GI-NEXT:    mov v16.h[6], w11
+; CHECK-GI-NEXT:    sxtb w10, w6
 ; CHECK-GI-NEXT:    mov v0.s[1], wzr
-; CHECK-GI-NEXT:    mov v7.h[3], w9
-; CHECK-GI-NEXT:    sxtb w9, w11
-; CHECK-GI-NEXT:    sxtb w11, w21
-; CHECK-GI-NEXT:    fmov w24, s19
-; CHECK-GI-NEXT:    mov v16.h[3], w8
-; CHECK-GI-NEXT:    sxtb w8, w16
-; CHECK-GI-NEXT:    sxtb w16, w3
-; CHECK-GI-NEXT:    mov v6.h[4], w11
-; CHECK-GI-NEXT:    ldr w11, [sp, #4] // 4-byte Folded Reload
-; CHECK-GI-NEXT:    mov v3.h[5], w9
-; CHECK-GI-NEXT:    sxtb w9, w15
-; CHECK-GI-NEXT:    sxtb w15, w27
-; CHECK-GI-NEXT:    mov v7.h[4], w14
-; CHECK-GI-NEXT:    sxtb w14, w1
-; CHECK-GI-NEXT:    sxtb w11, w11
-; CHECK-GI-NEXT:    mov v4.h[4], w8
-; CHECK-GI-NEXT:    sxtb w8, w20
+; CHECK-GI-NEXT:    mov v6.h[7], w8
+; CHECK-GI-NEXT:    sxtb w8, w17
+; CHECK-GI-NEXT:    sxtb w9, w9
+; CHECK-GI-NEXT:    mov v3.h[7], w13
 ; CHECK-GI-NEXT:    ldp x28, x27, [sp, #32] // 16-byte Folded Reload
-; CHECK-GI-NEXT:    mov v5.h[4], w16
-; CHECK-GI-NEXT:    mov v16.h[4], w10
-; CHECK-GI-NEXT:    sxtb w10, w26
-; CHECK-GI-NEXT:    ldp x26, x25, [sp, #48] // 16-byte Folded Reload
-; CHECK-GI-NEXT:    mov v6.h[5], w8
-; CHECK-GI-NEXT:    ldr w8, [sp, #8] // 4-byte Folded Reload
-; CHECK-GI-NEXT:    mov v7.h[5], w10
-; CHECK-GI-NEXT:    sxtb w10, w12
-; CHECK-GI-NEXT:    sxtb w12, w18
-; CHECK-GI-NEXT:    mov v4.h[5], w9
-; CHECK-GI-NEXT:    sxtb w9, w19
-; CHECK-GI-NEXT:    mov v5.h[5], w14
-; CHECK-GI-NEXT:    sxtb w8, w8
-; CHECK-GI-NEXT:    mov v16.h[5], w15
-; CHECK-GI-NEXT:    mov v3.h[6], w11
-; CHECK-GI-NEXT:    sxtb w11, w22
-; CHECK-GI-NEXT:    mov v6.h[6], w9
-; CHECK-GI-NEXT:    sxtb w9, w13
-; CHECK-GI-NEXT:    sxtb w13, w24
+; CHECK-GI-NEXT:    mov v4.h[7], w12
+; CHECK-GI-NEXT:    mov v5.h[7], w8
+; CHECK-GI-NEXT:    mov v7.h[7], w10
+; CHECK-GI-NEXT:    mov v16.h[7], w9
+; CHECK-GI-NEXT:    smov w8, v1.b[8]
+; CHECK-GI-NEXT:    smov w9, v2.b[8]
 ; CHECK-GI-NEXT:    mov v0.s[2], wzr
-; CHECK-GI-NEXT:    mov v7.h[6], w11
-; CHECK-GI-NEXT:    fmov w11, s17
-; CHECK-GI-NEXT:    mov v4.h[6], w10
-; CHECK-GI-NEXT:    sxtb w10, w7
-; CHECK-GI-NEXT:    mov v5.h[6], w12
-; CHECK-GI-NEXT:    mov v16.h[6], w13
-; CHECK-GI-NEXT:    mov v3.h[7], w8
-; CHECK-GI-NEXT:    sxtb w8, w6
-; CHECK-GI-NEXT:    smov w12, v1.b[8]
-; CHECK-GI-NEXT:    mov v6.h[7], w10
-; CHECK-GI-NEXT:    sxtb w10, w17
-; CHECK-GI-NEXT:    sxtb w11, w11
-; CHECK-GI-NEXT:    mov v4.h[7], w9
-; CHECK-GI-NEXT:    mov v7.h[7], w8
-; CHECK-GI-NEXT:    smov w8, v2.b[8]
-; CHECK-GI-NEXT:    mov v5.h[7], w10
-; CHECK-GI-NEXT:    ldp x22, x21, [sp, #80] // 16-byte Folded Reload
-; CHECK-GI-NEXT:    mov v16.h[7], w11
-; CHECK-GI-NEXT:    mov v...
[truncated]

@llvmbot
Copy link
Member

llvmbot commented Nov 10, 2025

@llvm/pr-subscribers-llvm-globalisel

Author: None (AnushaK6)

Changes

This patch continues the effort in issue #150515 to port computeSignBits functionality from the SelectionDAG (SDAG) to GlobalISel (GISel).
Specifically, it introduces support for G_MUL and improves precision when analyzing known bits and sign bits.

Changes:

  • Implemented computeSignBits support for the G_MUL operation in GISel.
  • Added logic to distinguish between constant and non-constant operands. When the operand is a constant, we can determine the exact number of sign bits.
  • When the operand is a variable, conservatively compute an upper bound valid for all possible values.
  • Added special handling for constant cases such as multiplication by zero.

Added new tests to:

  • known-signbits.mir with targeted test cases for G_MUL
  • Also modified existing test checks that use mul

Patch is 151.36 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/167311.diff

6 Files Affected:

  • (modified) llvm/lib/CodeGen/GlobalISel/GISelValueTracking.cpp (+36)
  • (added) llvm/test/CodeGen/AArch64/GlobalISel/knownbits-mul.mir (+79)
  • (modified) llvm/test/CodeGen/AArch64/combine-sdiv.ll (+1-3)
  • (modified) llvm/test/CodeGen/AArch64/neon-dotreduce.ll (+1564-1564)
  • (modified) llvm/test/CodeGen/AArch64/rem-by-const.ll (+1-2)
  • (modified) llvm/test/CodeGen/RISCV/GlobalISel/div-by-constant.ll (+16-32)
diff --git a/llvm/lib/CodeGen/GlobalISel/GISelValueTracking.cpp b/llvm/lib/CodeGen/GlobalISel/GISelValueTracking.cpp
index c1fb8b6d78ff8..3b5c68f266fff 100644
--- a/llvm/lib/CodeGen/GlobalISel/GISelValueTracking.cpp
+++ b/llvm/lib/CodeGen/GlobalISel/GISelValueTracking.cpp
@@ -31,6 +31,7 @@
 #include "llvm/CodeGen/TargetLowering.h"
 #include "llvm/CodeGen/TargetOpcodes.h"
 #include "llvm/IR/ConstantRange.h"
+#include "llvm/IR/Constants.h"
 #include "llvm/IR/DerivedTypes.h"
 #include "llvm/IR/FMF.h"
 #include "llvm/MC/TargetRegistry.h"
@@ -2084,6 +2085,41 @@ unsigned GISelValueTracking::computeNumSignBits(Register R,
     FirstAnswer = std::min(Src1NumSignBits, Src2NumSignBits) - 1;
     break;
   }
+  case TargetOpcode::G_MUL: {
+    Register Src1 = MI.getOperand(1).getReg();
+    Register Src2 = MI.getOperand(2).getReg();
+
+    KnownBits Known1 = getKnownBits(Src1, DemandedElts, Depth + 1);
+    KnownBits Known2 = getKnownBits(Src2, DemandedElts, Depth + 1);
+
+    if (Known1.isZero() || Known2.isZero())
+      return TyBits;
+
+    auto C1 = getIConstantVRegValWithLookThrough(Src1, MRI);
+    auto C2 = getIConstantVRegValWithLookThrough(Src2, MRI);
+
+    if (C1 && C2) {
+      APInt Val1 = C1->Value;
+      APInt Val2 = C2->Value;
+      APInt Product = Val1 * Val2;
+      return Product.getNumSignBits();
+    }
+    unsigned Src1NumSignBits =
+        computeNumSignBits(Src1, DemandedElts, Depth + 1);
+    if (Src1NumSignBits == 1) {
+      return 1;
+    }
+    unsigned Src2NumSignBits =
+        computeNumSignBits(Src2, DemandedElts, Depth + 1);
+    if (Src2NumSignBits == 1) {
+      return 1;
+    }
+
+    unsigned OutValidBits =
+        (TyBits - Src1NumSignBits + 1) + (TyBits - Src2NumSignBits + 1);
+    FirstAnswer = OutValidBits > TyBits ? 1 : TyBits - OutValidBits + 1;
+    break;
+  }
   case TargetOpcode::G_FCMP:
   case TargetOpcode::G_ICMP: {
     bool IsFP = Opcode == TargetOpcode::G_FCMP;
diff --git a/llvm/test/CodeGen/AArch64/GlobalISel/knownbits-mul.mir b/llvm/test/CodeGen/AArch64/GlobalISel/knownbits-mul.mir
new file mode 100644
index 0000000000000..ea5281948a211
--- /dev/null
+++ b/llvm/test/CodeGen/AArch64/GlobalISel/knownbits-mul.mir
@@ -0,0 +1,79 @@
+# NOTE: Assertions have been autogenerated by utils/update_givaluetracking_test_checks.py UTC_ARGS: --version 5
+# RUN: llc -mtriple=aarch64 -passes="print<gisel-value-tracking>" -filetype=null %s 2>&1 | FileCheck %s
+
+---
+name:            ConstPositives
+body:             |
+  bb.1:
+  ; CHECK-LABEL: name: @ConstPositives
+  ; CHECK-NEXT: %0:_ KnownBits:00000011 SignBits:6
+  ; CHECK-NEXT: %1:_ KnownBits:00000101 SignBits:5
+  ; CHECK-NEXT: %2:_ KnownBits:00001111 SignBits:4
+    %0:_(s8) = G_CONSTANT i8 3
+    %1:_(s8) = G_CONSTANT i8 5
+    %2:_(s8) = G_MUL %0, %1
+...
+---
+name:            ConstZero
+body:             |
+  bb.1:
+  ; CHECK-LABEL: name: @ConstZero
+  ; CHECK-NEXT: %0:_ KnownBits:00000000 SignBits:8
+  ; CHECK-NEXT: %1:_ KnownBits:00000001 SignBits:7
+  ; CHECK-NEXT: %2:_ KnownBits:00000000 SignBits:8
+    %0:_(s8) = G_CONSTANT i8 0
+    %1:_(s8) = G_CONSTANT i8 1
+    %2:_(s8) = G_MUL %0, %1
+...
+---
+name:            ConstNegatives
+body:             |
+  bb.1:
+  ; CHECK-LABEL: name: @ConstNegatives
+  ; CHECK-NEXT: %0:_ KnownBits:11111110 SignBits:7
+  ; CHECK-NEXT: %1:_ KnownBits:11111100 SignBits:6
+  ; CHECK-NEXT: %2:_ KnownBits:00001000 SignBits:4
+    %0:_(s8) = G_CONSTANT i8 -2
+    %1:_(s8) = G_CONSTANT i8 -4
+    %2:_(s8) = G_MUL %0, %1
+...
+---
+name:            MixedSigns
+body:             |
+  bb.1:
+  ; CHECK-LABEL: name: @MixedSigns
+  ; CHECK-NEXT: %0:_ KnownBits:11111100 SignBits:6
+  ; CHECK-NEXT: %1:_ KnownBits:00000011 SignBits:6
+  ; CHECK-NEXT: %2:_ KnownBits:11110100 SignBits:4
+    %0:_(s8) = G_CONSTANT i8 -4
+    %1:_(s8) = G_CONSTANT i8 3
+    %2:_(s8) = G_MUL %0, %1
+...
+---
+name:            UnknownVar
+body:             |
+  bb.1:
+  ; CHECK-LABEL: name: @UnknownVar
+  ; CHECK-NEXT: %0:_ KnownBits:???????? SignBits:1
+  ; CHECK-NEXT: %1:_ KnownBits:00000010 SignBits:6
+  ; CHECK-NEXT: %2:_ KnownBits:???????0 SignBits:1
+    %0:_(s8) = COPY $b0
+    %1:_(s8) = G_CONSTANT i8 2
+    %2:_(s8) = G_MUL %0, %1
+...
+---
+name:            VectorMul
+body:             |
+  bb.1:
+  ; CHECK-LABEL: name: @VectorMul
+  ; CHECK-NEXT: %0:_ KnownBits:0000000000000001 SignBits:15
+  ; CHECK-NEXT: %1:_ KnownBits:0000000000000010 SignBits:14
+  ; CHECK-NEXT: %2:_ KnownBits:00000000000000?? SignBits:14
+  ; CHECK-NEXT: %3:_ KnownBits:00000000000000?? SignBits:14
+  ; CHECK-NEXT: %4:_ KnownBits:000000000000???? SignBits:12
+    %0:_(s16) = G_CONSTANT i16 1
+    %1:_(s16) = G_CONSTANT i16 2
+    %2:_(<4 x s16>) = G_BUILD_VECTOR %0, %1, %1, %0
+    %3:_(<4 x s16>) = G_BUILD_VECTOR %1, %0, %0, %1
+    %4:_(<4 x s16>) = G_MUL %2, %3
+...
diff --git a/llvm/test/CodeGen/AArch64/combine-sdiv.ll b/llvm/test/CodeGen/AArch64/combine-sdiv.ll
index cca190f08df2b..37ea64f700fd8 100644
--- a/llvm/test/CodeGen/AArch64/combine-sdiv.ll
+++ b/llvm/test/CodeGen/AArch64/combine-sdiv.ll
@@ -1510,7 +1510,6 @@ define i5 @combine_i5_sdiv_const7(i5 %x) {
 ; CHECK-GI-NEXT:    sbfx w9, w0, #0, #5
 ; CHECK-GI-NEXT:    sbfx w8, w8, #0, #5
 ; CHECK-GI-NEXT:    mul w8, w9, w8
-; CHECK-GI-NEXT:    sbfx w8, w8, #0, #10
 ; CHECK-GI-NEXT:    add w8, w0, w8, asr #5
 ; CHECK-GI-NEXT:    sbfx w8, w8, #0, #5
 ; CHECK-GI-NEXT:    asr w8, w8, #2
@@ -1560,7 +1559,6 @@ define i8 @combine_i8_sdiv_const7(i8 %x) {
 ; CHECK-GI-NEXT:    sxtb w8, w0
 ; CHECK-GI-NEXT:    mov w9, #-109 // =0xffffff93
 ; CHECK-GI-NEXT:    mul w8, w8, w9
-; CHECK-GI-NEXT:    sxth w8, w8
 ; CHECK-GI-NEXT:    add w8, w0, w8, asr #8
 ; CHECK-GI-NEXT:    sbfx w8, w8, #2, #6
 ; CHECK-GI-NEXT:    ubfx w9, w8, #7, #1
@@ -1585,7 +1583,7 @@ define i8 @combine_i8_sdiv_const100(i8 %x) {
 ; CHECK-GI-NEXT:    sxtb w8, w0
 ; CHECK-GI-NEXT:    mov w9, #41 // =0x29
 ; CHECK-GI-NEXT:    mul w8, w8, w9
-; CHECK-GI-NEXT:    sbfx w8, w8, #8, #8
+; CHECK-GI-NEXT:    asr w8, w8, #8
 ; CHECK-GI-NEXT:    asr w8, w8, #4
 ; CHECK-GI-NEXT:    ubfx w9, w8, #7, #1
 ; CHECK-GI-NEXT:    add w0, w8, w9
diff --git a/llvm/test/CodeGen/AArch64/neon-dotreduce.ll b/llvm/test/CodeGen/AArch64/neon-dotreduce.ll
index fb2a1fa697c26..581d8e116359b 100644
--- a/llvm/test/CodeGen/AArch64/neon-dotreduce.ll
+++ b/llvm/test/CodeGen/AArch64/neon-dotreduce.ll
@@ -2832,247 +2832,246 @@ define i32 @test_sdot_v25i8(ptr nocapture readonly %a, ptr nocapture readonly %b
 ; CHECK-GI-NEXT:    ldp q2, q1, [x1]
 ; CHECK-GI-NEXT:    movi d0, #0000000000000000
 ; CHECK-GI-NEXT:    str w2, [sp, #12] // 4-byte Folded Spill
-; CHECK-GI-NEXT:    mov b5, v2.b[2]
 ; CHECK-GI-NEXT:    mov b6, v2.b[3]
 ; CHECK-GI-NEXT:    mov b7, v2.b[4]
 ; CHECK-GI-NEXT:    mov b16, v2.b[5]
+; CHECK-GI-NEXT:    mov b19, v2.b[8]
+; CHECK-GI-NEXT:    mov b4, v2.b[1]
+; CHECK-GI-NEXT:    mov b5, v2.b[2]
 ; CHECK-GI-NEXT:    mov b17, v2.b[6]
 ; CHECK-GI-NEXT:    mov b18, v2.b[7]
-; CHECK-GI-NEXT:    mov b19, v2.b[8]
 ; CHECK-GI-NEXT:    mov b20, v2.b[9]
-; CHECK-GI-NEXT:    mov b21, v2.b[15]
-; CHECK-GI-NEXT:    mov b3, v2.b[1]
-; CHECK-GI-NEXT:    fmov w19, s2
-; CHECK-GI-NEXT:    mov b22, v1.b[6]
-; CHECK-GI-NEXT:    fmov w6, s5
-; CHECK-GI-NEXT:    mov b5, v2.b[10]
-; CHECK-GI-NEXT:    fmov w14, s6
-; CHECK-GI-NEXT:    mov b6, v2.b[11]
+; CHECK-GI-NEXT:    mov b21, v2.b[10]
+; CHECK-GI-NEXT:    mov b22, v2.b[11]
+; CHECK-GI-NEXT:    fmov w7, s2
+; CHECK-GI-NEXT:    fmov w13, s6
+; CHECK-GI-NEXT:    mov b6, v2.b[12]
 ; CHECK-GI-NEXT:    fmov w2, s7
-; CHECK-GI-NEXT:    stp s17, s18, [sp, #4] // 8-byte Folded Spill
-; CHECK-GI-NEXT:    mov b7, v2.b[12]
+; CHECK-GI-NEXT:    mov b7, v2.b[13]
 ; CHECK-GI-NEXT:    fmov w11, s16
-; CHECK-GI-NEXT:    sxtb w28, w19
-; CHECK-GI-NEXT:    mov b16, v2.b[13]
-; CHECK-GI-NEXT:    mov b18, v1.b[1]
-; CHECK-GI-NEXT:    sxtb w6, w6
-; CHECK-GI-NEXT:    mov b17, v2.b[14]
-; CHECK-GI-NEXT:    ldp q4, q2, [x0]
-; CHECK-GI-NEXT:    fmov w25, s19
+; CHECK-GI-NEXT:    mov b16, v2.b[14]
+; CHECK-GI-NEXT:    mov b23, v2.b[15]
+; CHECK-GI-NEXT:    ldp q3, q2, [x0]
+; CHECK-GI-NEXT:    fmov w26, s19
+; CHECK-GI-NEXT:    fmov w19, s4
+; CHECK-GI-NEXT:    stp s17, s18, [sp, #4] // 8-byte Folded Spill
+; CHECK-GI-NEXT:    fmov w29, s5
 ; CHECK-GI-NEXT:    fmov w24, s20
-; CHECK-GI-NEXT:    fmov w22, s5
-; CHECK-GI-NEXT:    mov b5, v1.b[2]
-; CHECK-GI-NEXT:    fmov w0, s6
-; CHECK-GI-NEXT:    sxtb w14, w14
-; CHECK-GI-NEXT:    mov b20, v1.b[3]
-; CHECK-GI-NEXT:    fmov w16, s7
-; CHECK-GI-NEXT:    mov b7, v1.b[4]
-; CHECK-GI-NEXT:    fmov w15, s16
-; CHECK-GI-NEXT:    sxtb w25, w25
+; CHECK-GI-NEXT:    sxtb w8, w7
+; CHECK-GI-NEXT:    mov b4, v3.b[2]
+; CHECK-GI-NEXT:    mov b5, v3.b[1]
+; CHECK-GI-NEXT:    sxtb w13, w13
+; CHECK-GI-NEXT:    mov b17, v1.b[1]
+; CHECK-GI-NEXT:    fmov w22, s21
+; CHECK-GI-NEXT:    sxtb w26, w26
+; CHECK-GI-NEXT:    mov b18, v1.b[2]
+; CHECK-GI-NEXT:    fmov w18, s22
 ; CHECK-GI-NEXT:    sxtb w24, w24
-; CHECK-GI-NEXT:    mov b16, v1.b[5]
-; CHECK-GI-NEXT:    fmov w13, s21
+; CHECK-GI-NEXT:    mov b19, v1.b[3]
+; CHECK-GI-NEXT:    fmov w16, s6
+; CHECK-GI-NEXT:    sxtb w19, w19
+; CHECK-GI-NEXT:    mov b21, v1.b[4]
+; CHECK-GI-NEXT:    fmov w15, s7
 ; CHECK-GI-NEXT:    sxtb w22, w22
-; CHECK-GI-NEXT:    mov b6, v4.b[2]
-; CHECK-GI-NEXT:    fmov w26, s18
-; CHECK-GI-NEXT:    sxtb w0, w0
-; CHECK-GI-NEXT:    mov b21, v1.b[7]
-; CHECK-GI-NEXT:    mov b18, v4.b[4]
-; CHECK-GI-NEXT:    fmov w7, s3
-; CHECK-GI-NEXT:    mov b3, v4.b[1]
-; CHECK-GI-NEXT:    fmov w12, s17
-; CHECK-GI-NEXT:    fmov w5, s5
-; CHECK-GI-NEXT:    mov b19, v4.b[3]
-; CHECK-GI-NEXT:    fmov w4, s20
-; CHECK-GI-NEXT:    fmov w3, s7
-; CHECK-GI-NEXT:    sxtb w29, w7
-; CHECK-GI-NEXT:    mov b17, v4.b[5]
-; CHECK-GI-NEXT:    fmov w1, s16
-; CHECK-GI-NEXT:    sxtb w5, w5
-; CHECK-GI-NEXT:    mov b16, v4.b[6]
-; CHECK-GI-NEXT:    fmov w18, s22
-; CHECK-GI-NEXT:    mov b7, v4.b[7]
-; CHECK-GI-NEXT:    fmov w17, s21
-; CHECK-GI-NEXT:    mov b5, v4.b[8]
-; CHECK-GI-NEXT:    mov b20, v4.b[9]
-; CHECK-GI-NEXT:    fmov w27, s6
-; CHECK-GI-NEXT:    mov b6, v4.b[10]
-; CHECK-GI-NEXT:    mov b21, v4.b[11]
-; CHECK-GI-NEXT:    fmov w21, s18
-; CHECK-GI-NEXT:    mov b18, v4.b[12]
-; CHECK-GI-NEXT:    mov b22, v4.b[13]
-; CHECK-GI-NEXT:    mov b23, v4.b[14]
-; CHECK-GI-NEXT:    fmov w10, s4
+; CHECK-GI-NEXT:    mov b7, v1.b[5]
+; CHECK-GI-NEXT:    mov b6, v3.b[3]
+; CHECK-GI-NEXT:    sxtb w11, w11
+; CHECK-GI-NEXT:    fmov w12, s23
+; CHECK-GI-NEXT:    mov b22, v1.b[6]
+; CHECK-GI-NEXT:    mov b23, v1.b[7]
+; CHECK-GI-NEXT:    mov b20, v3.b[4]
+; CHECK-GI-NEXT:    fmov w28, s4
+; CHECK-GI-NEXT:    fmov s4, w26
+; CHECK-GI-NEXT:    fmov w14, s16
+; CHECK-GI-NEXT:    fmov w27, s17
+; CHECK-GI-NEXT:    fmov w5, s18
+; CHECK-GI-NEXT:    sxtb w12, w12
+; CHECK-GI-NEXT:    fmov w4, s19
+; CHECK-GI-NEXT:    mov b19, v3.b[5]
+; CHECK-GI-NEXT:    sxtb w28, w28
+; CHECK-GI-NEXT:    fmov w3, s21
+; CHECK-GI-NEXT:    mov b18, v3.b[6]
 ; CHECK-GI-NEXT:    sxtb w27, w27
-; CHECK-GI-NEXT:    mov b24, v4.b[15]
-; CHECK-GI-NEXT:    fmov s4, w25
-; CHECK-GI-NEXT:    fmov w30, s3
-; CHECK-GI-NEXT:    fmov s3, w28
-; CHECK-GI-NEXT:    fmov w9, s5
-; CHECK-GI-NEXT:    sxtb w10, w10
-; CHECK-GI-NEXT:    fmov w7, s7
-; CHECK-GI-NEXT:    mov b7, v2.b[1]
+; CHECK-GI-NEXT:    sxtb w5, w5
+; CHECK-GI-NEXT:    fmov w1, s7
+; CHECK-GI-NEXT:    mov b16, v3.b[7]
+; CHECK-GI-NEXT:    fmov w0, s22
+; CHECK-GI-NEXT:    mov b17, v3.b[8]
+; CHECK-GI-NEXT:    fmov w17, s23
+; CHECK-GI-NEXT:    mov b7, v3.b[9]
+; CHECK-GI-NEXT:    fmov w30, s5
+; CHECK-GI-NEXT:    mov b5, v3.b[10]
+; CHECK-GI-NEXT:    mov b21, v3.b[11]
+; CHECK-GI-NEXT:    fmov w25, s6
+; CHECK-GI-NEXT:    mov b6, v3.b[12]
+; CHECK-GI-NEXT:    fmov w23, s20
+; CHECK-GI-NEXT:    mov b20, v3.b[13]
+; CHECK-GI-NEXT:    mov b22, v3.b[14]
+; CHECK-GI-NEXT:    fmov w6, s3
+; CHECK-GI-NEXT:    mov b23, v3.b[15]
+; CHECK-GI-NEXT:    fmov s3, w8
+; CHECK-GI-NEXT:    fmov w8, s1
 ; CHECK-GI-NEXT:    mov v4.h[1], w24
-; CHECK-GI-NEXT:    fmov w24, s1
-; CHECK-GI-NEXT:    fmov w8, s20
-; CHECK-GI-NEXT:    sxtb w9, w9
-; CHECK-GI-NEXT:    mov v3.h[1], w29
-; CHECK-GI-NEXT:    fmov w29, s6
-; CHECK-GI-NEXT:    fmov s6, w10
-; CHECK-GI-NEXT:    fmov w10, s2
-; CHECK-GI-NEXT:    fmov w19, s16
-; CHECK-GI-NEXT:    sxtb w24, w24
+; CHECK-GI-NEXT:    fmov w21, s19
+; CHECK-GI-NEXT:    mov b19, v2.b[1]
+; CHECK-GI-NEXT:    fmov w9, s17
+; CHECK-GI-NEXT:    fmov w24, s6
+; CHECK-GI-NEXT:    fmov w7, s16
+; CHECK-GI-NEXT:    mov b16, v2.b[2]
 ; CHECK-GI-NEXT:    sxtb w8, w8
-; CHECK-GI-NEXT:    mov b16, v2.b[3]
+; CHECK-GI-NEXT:    mov v3.h[1], w19
+; CHECK-GI-NEXT:    sxtb w19, w29
+; CHECK-GI-NEXT:    sxtb w9, w9
+; CHECK-GI-NEXT:    fmov w29, s5
+; CHECK-GI-NEXT:    mov v4.h[2], w22
+; CHECK-GI-NEXT:    sxtb w22, w6
+; CHECK-GI-NEXT:    fmov s5, w8
+; CHECK-GI-NEXT:    fmov w10, s7
+; CHECK-GI-NEXT:    fmov s7, w9
+; CHECK-GI-NEXT:    fmov w9, s16
+; CHECK-GI-NEXT:    fmov w20, s18
 ; CHECK-GI-NEXT:    sxtb w29, w29
-; CHECK-GI-NEXT:    fmov w23, s19
-; CHECK-GI-NEXT:    mov b19, v2.b[2]
+; CHECK-GI-NEXT:    fmov s6, w22
+; CHECK-GI-NEXT:    fmov w22, s2
 ; CHECK-GI-NEXT:    sxtb w10, w10
-; CHECK-GI-NEXT:    fmov s5, w24
-; CHECK-GI-NEXT:    sxtb w24, w30
-; CHECK-GI-NEXT:    mov v3.h[2], w6
-; CHECK-GI-NEXT:    sxtb w6, w26
-; CHECK-GI-NEXT:    fmov w28, s21
-; CHECK-GI-NEXT:    sxtb w23, w23
-; CHECK-GI-NEXT:    mov v6.h[1], w24
-; CHECK-GI-NEXT:    fmov w24, s7
-; CHECK-GI-NEXT:    fmov s7, w9
-; CHECK-GI-NEXT:    fmov w9, s19
-; CHECK-GI-NEXT:    mov v5.h[1], w6
-; CHECK-GI-NEXT:    mov v4.h[2], w22
-; CHECK-GI-NEXT:    fmov w20, s17
-; CHECK-GI-NEXT:    mov b17, v2.b[4]
-; CHECK-GI-NEXT:    sxtb w24, w24
-; CHECK-GI-NEXT:    mov v3.h[3], w14
-; CHECK-GI-NEXT:    sxtb w14, w2
+; CHECK-GI-NEXT:    mov v5.h[1], w27
+; CHECK-GI-NEXT:    sxtb w27, w30
 ; CHECK-GI-NEXT:    sxtb w9, w9
-; CHECK-GI-NEXT:    mov v7.h[1], w8
-; CHECK-GI-NEXT:    fmov w8, s16
-; CHECK-GI-NEXT:    fmov s16, w10
-; CHECK-GI-NEXT:    mov v6.h[2], w27
+; CHECK-GI-NEXT:    mov b18, v2.b[3]
+; CHECK-GI-NEXT:    mov v3.h[2], w19
+; CHECK-GI-NEXT:    sxtb w22, w22
+; CHECK-GI-NEXT:    mov v6.h[1], w27
+; CHECK-GI-NEXT:    fmov w27, s19
+; CHECK-GI-NEXT:    mov v7.h[1], w10
+; CHECK-GI-NEXT:    fmov w26, s21
+; CHECK-GI-NEXT:    mov b17, v2.b[4]
+; CHECK-GI-NEXT:    fmov s16, w22
 ; CHECK-GI-NEXT:    mov v5.h[2], w5
-; CHECK-GI-NEXT:    fmov w25, s18
-; CHECK-GI-NEXT:    mov v4.h[3], w0
-; CHECK-GI-NEXT:    sxtb w0, w4
-; CHECK-GI-NEXT:    sxtb w8, w8
-; CHECK-GI-NEXT:    mov b18, v2.b[5]
-; CHECK-GI-NEXT:    fmov w10, s17
-; CHECK-GI-NEXT:    mov v16.h[1], w24
+; CHECK-GI-NEXT:    sxtb w5, w25
+; CHECK-GI-NEXT:    sxtb w27, w27
+; CHECK-GI-NEXT:    fmov w10, s18
+; CHECK-GI-NEXT:    mov v3.h[3], w13
+; CHECK-GI-NEXT:    sxtb w13, w4
+; CHECK-GI-NEXT:    mov v6.h[2], w28
+; CHECK-GI-NEXT:    fmov w8, s20
+; CHECK-GI-NEXT:    mov v16.h[1], w27
 ; CHECK-GI-NEXT:    mov v7.h[2], w29
-; CHECK-GI-NEXT:    mov v3.h[4], w14
-; CHECK-GI-NEXT:    sxtb w14, w25
-; CHECK-GI-NEXT:    ldp x29, x30, [sp, #16] // 16-byte Folded Reload
+; CHECK-GI-NEXT:    mov b20, v2.b[5]
 ; CHECK-GI-NEXT:    sxtb w10, w10
-; CHECK-GI-NEXT:    mov v6.h[3], w23
-; CHECK-GI-NEXT:    mov v5.h[3], w0
-; CHECK-GI-NEXT:    fmov w26, s22
-; CHECK-GI-NEXT:    mov b19, v2.b[6]
-; CHECK-GI-NEXT:    fmov w27, s18
+; CHECK-GI-NEXT:    ldp x29, x30, [sp, #16] // 16-byte Folded Reload
+; CHECK-GI-NEXT:    sxtb w8, w8
+; CHECK-GI-NEXT:    fmov w22, s17
+; CHECK-GI-NEXT:    mov v5.h[3], w13
+; CHECK-GI-NEXT:    sxtb w13, w2
+; CHECK-GI-NEXT:    mov v6.h[3], w5
+; CHECK-GI-NEXT:    mov b21, v2.b[6]
 ; CHECK-GI-NEXT:    mov v16.h[2], w9
-; CHECK-GI-NEXT:    sxtb w9, w28
-; CHECK-GI-NEXT:    fmov w22, s23
+; CHECK-GI-NEXT:    sxtb w9, w18
+; CHECK-GI-NEXT:    sxtb w18, w23
+; CHECK-GI-NEXT:    mov v3.h[4], w13
+; CHECK-GI-NEXT:    sxtb w13, w24
+; CHECK-GI-NEXT:    fmov w27, s20
+; CHECK-GI-NEXT:    ldp x24, x23, [sp, #64] // 16-byte Folded Reload
+; CHECK-GI-NEXT:    mov v4.h[3], w9
+; CHECK-GI-NEXT:    sxtb w9, w26
+; CHECK-GI-NEXT:    ldp x26, x25, [sp, #48] // 16-byte Folded Reload
+; CHECK-GI-NEXT:    mov v16.h[3], w10
+; CHECK-GI-NEXT:    sxtb w10, w3
+; CHECK-GI-NEXT:    mov v6.h[4], w18
+; CHECK-GI-NEXT:    ldr w18, [sp, #4] // 4-byte Folded Reload
+; CHECK-GI-NEXT:    mov v7.h[3], w9
+; CHECK-GI-NEXT:    sxtb w9, w16
+; CHECK-GI-NEXT:    sxtb w16, w22
+; CHECK-GI-NEXT:    mov v5.h[4], w10
+; CHECK-GI-NEXT:    sxtb w10, w15
+; CHECK-GI-NEXT:    sxtb w18, w18
+; CHECK-GI-NEXT:    mov v4.h[4], w9
+; CHECK-GI-NEXT:    sxtb w9, w21
+; CHECK-GI-NEXT:    ldp x22, x21, [sp, #80] // 16-byte Folded Reload
+; CHECK-GI-NEXT:    mov v16.h[4], w16
+; CHECK-GI-NEXT:    mov v7.h[4], w13
+; CHECK-GI-NEXT:    ldr w13, [sp, #8] // 4-byte Folded Reload
+; CHECK-GI-NEXT:    mov v6.h[5], w9
+; CHECK-GI-NEXT:    sxtb w9, w1
+; CHECK-GI-NEXT:    mov v3.h[5], w11
+; CHECK-GI-NEXT:    sxtb w11, w27
+; CHECK-GI-NEXT:    fmov w19, s22
+; CHECK-GI-NEXT:    fmov w28, s21
+; CHECK-GI-NEXT:    sxtb w13, w13
 ; CHECK-GI-NEXT:    mov b17, v2.b[7]
-; CHECK-GI-NEXT:    fmov w6, s24
+; CHECK-GI-NEXT:    mov v5.h[5], w9
+; CHECK-GI-NEXT:    sxtb w9, w0
+; CHECK-GI-NEXT:    mov v4.h[5], w10
+; CHECK-GI-NEXT:    sxtb w10, w20
+; CHECK-GI-NEXT:    mov v7.h[5], w8
+; CHECK-GI-NEXT:    mov v16.h[5], w11
+; CHECK-GI-NEXT:    sxtb w8, w14
+; CHECK-GI-NEXT:    sxtb w11, w28
+; CHECK-GI-NEXT:    mov v6.h[6], w10
+; CHECK-GI-NEXT:    sxtb w10, w19
+; CHECK-GI-NEXT:    fmov w6, s23
+; CHECK-GI-NEXT:    mov v5.h[6], w9
+; CHECK-GI-NEXT:    fmov w9, s17
+; CHECK-GI-NEXT:    mov v3.h[6], w18
+; CHECK-GI-NEXT:    mov v4.h[6], w8
+; CHECK-GI-NEXT:    sxtb w8, w7
+; CHECK-GI-NEXT:    mov v7.h[6], w10
+; CHECK-GI-NEXT:    mov v16.h[6], w11
+; CHECK-GI-NEXT:    sxtb w10, w6
 ; CHECK-GI-NEXT:    mov v0.s[1], wzr
-; CHECK-GI-NEXT:    mov v7.h[3], w9
-; CHECK-GI-NEXT:    sxtb w9, w11
-; CHECK-GI-NEXT:    sxtb w11, w21
-; CHECK-GI-NEXT:    fmov w24, s19
-; CHECK-GI-NEXT:    mov v16.h[3], w8
-; CHECK-GI-NEXT:    sxtb w8, w16
-; CHECK-GI-NEXT:    sxtb w16, w3
-; CHECK-GI-NEXT:    mov v6.h[4], w11
-; CHECK-GI-NEXT:    ldr w11, [sp, #4] // 4-byte Folded Reload
-; CHECK-GI-NEXT:    mov v3.h[5], w9
-; CHECK-GI-NEXT:    sxtb w9, w15
-; CHECK-GI-NEXT:    sxtb w15, w27
-; CHECK-GI-NEXT:    mov v7.h[4], w14
-; CHECK-GI-NEXT:    sxtb w14, w1
-; CHECK-GI-NEXT:    sxtb w11, w11
-; CHECK-GI-NEXT:    mov v4.h[4], w8
-; CHECK-GI-NEXT:    sxtb w8, w20
+; CHECK-GI-NEXT:    mov v6.h[7], w8
+; CHECK-GI-NEXT:    sxtb w8, w17
+; CHECK-GI-NEXT:    sxtb w9, w9
+; CHECK-GI-NEXT:    mov v3.h[7], w13
 ; CHECK-GI-NEXT:    ldp x28, x27, [sp, #32] // 16-byte Folded Reload
-; CHECK-GI-NEXT:    mov v5.h[4], w16
-; CHECK-GI-NEXT:    mov v16.h[4], w10
-; CHECK-GI-NEXT:    sxtb w10, w26
-; CHECK-GI-NEXT:    ldp x26, x25, [sp, #48] // 16-byte Folded Reload
-; CHECK-GI-NEXT:    mov v6.h[5], w8
-; CHECK-GI-NEXT:    ldr w8, [sp, #8] // 4-byte Folded Reload
-; CHECK-GI-NEXT:    mov v7.h[5], w10
-; CHECK-GI-NEXT:    sxtb w10, w12
-; CHECK-GI-NEXT:    sxtb w12, w18
-; CHECK-GI-NEXT:    mov v4.h[5], w9
-; CHECK-GI-NEXT:    sxtb w9, w19
-; CHECK-GI-NEXT:    mov v5.h[5], w14
-; CHECK-GI-NEXT:    sxtb w8, w8
-; CHECK-GI-NEXT:    mov v16.h[5], w15
-; CHECK-GI-NEXT:    mov v3.h[6], w11
-; CHECK-GI-NEXT:    sxtb w11, w22
-; CHECK-GI-NEXT:    mov v6.h[6], w9
-; CHECK-GI-NEXT:    sxtb w9, w13
-; CHECK-GI-NEXT:    sxtb w13, w24
+; CHECK-GI-NEXT:    mov v4.h[7], w12
+; CHECK-GI-NEXT:    mov v5.h[7], w8
+; CHECK-GI-NEXT:    mov v7.h[7], w10
+; CHECK-GI-NEXT:    mov v16.h[7], w9
+; CHECK-GI-NEXT:    smov w8, v1.b[8]
+; CHECK-GI-NEXT:    smov w9, v2.b[8]
 ; CHECK-GI-NEXT:    mov v0.s[2], wzr
-; CHECK-GI-NEXT:    mov v7.h[6], w11
-; CHECK-GI-NEXT:    fmov w11, s17
-; CHECK-GI-NEXT:    mov v4.h[6], w10
-; CHECK-GI-NEXT:    sxtb w10, w7
-; CHECK-GI-NEXT:    mov v5.h[6], w12
-; CHECK-GI-NEXT:    mov v16.h[6], w13
-; CHECK-GI-NEXT:    mov v3.h[7], w8
-; CHECK-GI-NEXT:    sxtb w8, w6
-; CHECK-GI-NEXT:    smov w12, v1.b[8]
-; CHECK-GI-NEXT:    mov v6.h[7], w10
-; CHECK-GI-NEXT:    sxtb w10, w17
-; CHECK-GI-NEXT:    sxtb w11, w11
-; CHECK-GI-NEXT:    mov v4.h[7], w9
-; CHECK-GI-NEXT:    mov v7.h[7], w8
-; CHECK-GI-NEXT:    smov w8, v2.b[8]
-; CHECK-GI-NEXT:    mov v5.h[7], w10
-; CHECK-GI-NEXT:    ldp x22, x21, [sp, #80] // 16-byte Folded Reload
-; CHECK-GI-NEXT:    mov v16.h[7], w11
-; CHECK-GI-NEXT:    mov v...
[truncated]

arsenm
arsenm reviewed Nov 11, 2025
View reviewed changes
llvm/lib/CodeGen/GlobalISel/GISelValueTracking.cpp
Comment on lines +2092 to +2096
KnownBits Known1 = getKnownBits(Src1, DemandedElts, Depth + 1);
KnownBits Known2 = getKnownBits(Src2, DemandedElts, Depth + 1);

if (Known1.isZero() || Known2.isZero())
return TyBits;
Copy link
Contributor

@arsenm arsenm Nov 11, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I don't think you need to call getKnownBits at all here. The DAG version doesn't. 0 also shouldn't need a special case

llvm/lib/CodeGen/GlobalISel/GISelValueTracking.cpp
auto C1 = getIConstantVRegValWithLookThrough(Src1, MRI);
auto C2 = getIConstantVRegValWithLookThrough(Src2, MRI);

if (C1 && C2) {
Copy link
Contributor

@arsenm arsenm Nov 11, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Don't need to worry about trying to handle constants specially

llvm/lib/CodeGen/GlobalISel/GISelValueTracking.cpp
Comment on lines +2112 to +2116
unsigned Src2NumSignBits =
computeNumSignBits(Src2, DemandedElts, Depth + 1);
if (Src2NumSignBits == 1) {
return 1;
}
Copy link
Contributor

@arsenm arsenm Nov 11, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Compute src2 before src1, as it's canonically cheaper

@davemgreen davemgreen mentioned this pull request Nov 13, 2025
Open
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@arsenm arsenm arsenm left review comments

Assignees

No one assigned

Labels

backend:AArch64 backend:RISC-V llvm:globalisel

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

3 participants

@AnushaK6 @llvmbot @arsenm