[llvm] [AMDGPU] prevent shrinking udiv/urem if operands exceed smax_bitwidth (PR #116733)

[via llvm-commits]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-backend-amdgpu

Author: choikwa (choikwa)

<details>
<summary>Changes</summary>

Do this by using ComputeKnownBits and checking for !isNonNegative and isUnsigned. This rejects shrinking unsigned div/rem if operands exceed smax_bitwidth since we know NumSignBits will be always 0.

---
Full diff: https://github.com/llvm/llvm-project/pull/116733.diff


2 Files Affected:

- (modified) llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp (+12) 
- (modified) llvm/test/CodeGen/AMDGPU/amdgpu-codegenprepare-idiv.ll (+57) 


``````````diff
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp b/llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp
index c49aab823b44a4..bb9dbda200a524 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp
@@ -1193,6 +1193,18 @@ int AMDGPUCodeGenPrepareImpl::getDivNumBits(BinaryOperator &I, Value *Num,
                                             Value *Den, unsigned AtLeast,
                                             bool IsSigned) const {
   const DataLayout &DL = Mod->getDataLayout();
+  if (!IsSigned) {
+    KnownBits Known = computeKnownBits(Num, DL, 0, AC, &I);
+    // We know all bits are used for division for Operand > smax_bitwidth
+    if (!Known.isNonNegative())
+      return -1;
+
+    Known = computeKnownBits(Den, DL, 0, AC, &I);
+    if (!Known.isNonNegative())
+      return -1;
+  }
+
+
   unsigned LHSSignBits = ComputeNumSignBits(Num, DL, 0, AC, &I);
   if (LHSSignBits < AtLeast)
     return -1;
diff --git a/llvm/test/CodeGen/AMDGPU/amdgpu-codegenprepare-idiv.ll b/llvm/test/CodeGen/AMDGPU/amdgpu-codegenprepare-idiv.ll
index b7436aeb1d5302..283a3524c2ff93 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgpu-codegenprepare-idiv.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgpu-codegenprepare-idiv.ll
@@ -9999,3 +9999,60 @@ define <2 x i64> @v_udiv_i64_exact(<2 x i64> %num) {
    %result = udiv exact <2 x i64> %num, <i64 4096, i64 1024>
    ret <2 x i64> %result
 }
+
+define i64 @udiv_i64_gt_smax(i8 %size) {
+; GFX6-LABEL: udiv_i64_gt_smax:
+; GFX6:       ; %bb.0:
+; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    v_bfe_i32 v0, v0, 0, 8
+; GFX6-NEXT:    v_ashrrev_i32_e32 v1, 31, v0
+; GFX6-NEXT:    v_not_b32_e32 v1, v1
+; GFX6-NEXT:    v_not_b32_e32 v0, v0
+; GFX6-NEXT:    s_mov_b32 s4, 0xcccccccd
+; GFX6-NEXT:    v_mul_lo_u32 v3, v1, s4
+; GFX6-NEXT:    v_mul_hi_u32 v4, v0, s4
+; GFX6-NEXT:    s_mov_b32 s6, 0xcccccccc
+; GFX6-NEXT:    v_mul_hi_u32 v5, v1, s4
+; GFX6-NEXT:    v_mul_hi_u32 v2, v0, s6
+; GFX6-NEXT:    v_mul_lo_u32 v0, v0, s6
+; GFX6-NEXT:    v_add_i32_e32 v3, vcc, v3, v4
+; GFX6-NEXT:    v_addc_u32_e32 v4, vcc, 0, v5, vcc
+; GFX6-NEXT:    v_add_i32_e32 v0, vcc, v0, v3
+; GFX6-NEXT:    v_addc_u32_e32 v0, vcc, 0, v2, vcc
+; GFX6-NEXT:    v_mul_lo_u32 v2, v1, s6
+; GFX6-NEXT:    v_mul_hi_u32 v1, v1, s6
+; GFX6-NEXT:    v_add_i32_e32 v0, vcc, v4, v0
+; GFX6-NEXT:    v_addc_u32_e64 v3, s[4:5], 0, 0, vcc
+; GFX6-NEXT:    v_add_i32_e32 v0, vcc, v2, v0
+; GFX6-NEXT:    v_addc_u32_e32 v1, vcc, v1, v3, vcc
+; GFX6-NEXT:    v_alignbit_b32 v0, v1, v0, 3
+; GFX6-NEXT:    v_lshrrev_b32_e32 v1, 3, v1
+; GFX6-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: udiv_i64_gt_smax:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v1, 31
+; GFX9-NEXT:    v_not_b32_sdwa v4, sext(v0) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0
+; GFX9-NEXT:    s_mov_b32 s4, 0xcccccccd
+; GFX9-NEXT:    v_ashrrev_i32_sdwa v1, v1, sext(v0) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_mul_hi_u32 v0, v4, s4
+; GFX9-NEXT:    v_not_b32_e32 v5, v1
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    s_mov_b32 s6, 0xcccccccc
+; GFX9-NEXT:    v_mad_u64_u32 v[2:3], s[4:5], v5, s4, v[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v6, v3
+; GFX9-NEXT:    v_mov_b32_e32 v3, v1
+; GFX9-NEXT:    v_mad_u64_u32 v[0:1], s[4:5], v4, s6, v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v0, v1
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, v6, v0
+; GFX9-NEXT:    v_addc_co_u32_e64 v1, s[4:5], 0, 0, vcc
+; GFX9-NEXT:    v_mad_u64_u32 v[0:1], s[4:5], v5, s6, v[0:1]
+; GFX9-NEXT:    v_alignbit_b32 v0, v1, v0, 3
+; GFX9-NEXT:    v_lshrrev_b32_e32 v1, 3, v1
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  %esize = sext i8 %size to i64
+  %minus = sub nuw nsw i64 -1, %esize
+  %div = udiv i64 %minus, 10
+  ret i64 %div
+}

``````````

</details>


https://github.com/llvm/llvm-project/pull/116733