[llvm] [AMDGPU] Disallow negative offset when addressing constant memory (PR #76009)
via llvm-commits
llvm-commits at lists.llvm.org
Tue Dec 19 21:26:21 PST 2023
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-backend-amdgpu
Author: None (vangthao95)
<details>
<summary>Changes</summary>
When loading from scalar memory, the load instruction is illegal and undefined if the immediate offset is negative and is less than M0 or SGPR[offset] according to SPG. This patch is a workaround that makes negative offset illegal for GFX9+ when addressing constant memory. It is not a full fix as we must also check if there is use of M0 or soffset present. I believe there also needs to be a fix during instruction selection regarding this issue but would like to open this to discussion as how to address this.
---
Full diff: https://github.com/llvm/llvm-project/pull/76009.diff
2 Files Affected:
- (modified) llvm/lib/Target/AMDGPU/SIISelLowering.cpp (+7)
- (modified) llvm/test/CodeGen/AMDGPU/cgp-addressing-modes-smem.ll (+16-25)
``````````diff
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index 4f4bc45e49b43e..f4dbb852cf9487 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -1462,6 +1462,13 @@ bool SITargetLowering::isLegalAddressingMode(const DataLayout &DL,
// for S_BUFFER_* instructions).
if (!isInt<21>(AM.BaseOffs))
return false;
+
+ // FIXME: When addressing scalar memory, it is illegal and undefined for
+ // IOFFSET + (M0 or soffset) to be negative. This also need to
+ // verify that M0 or soffset is present. For now it is only rejecting
+ // all negative offsets.
+ if (AM.BaseOffs < 0)
+ return false;
} else {
// On GFX12, all offsets are signed 24-bit in bytes.
if (!isInt<24>(AM.BaseOffs))
diff --git a/llvm/test/CodeGen/AMDGPU/cgp-addressing-modes-smem.ll b/llvm/test/CodeGen/AMDGPU/cgp-addressing-modes-smem.ll
index f5846c3d6db737..add148c2f43064 100644
--- a/llvm/test/CodeGen/AMDGPU/cgp-addressing-modes-smem.ll
+++ b/llvm/test/CodeGen/AMDGPU/cgp-addressing-modes-smem.ll
@@ -279,31 +279,19 @@ end:
}
define amdgpu_cs void @test_sink_smem_offset_neg400(ptr addrspace(4) inreg %ptr, i32 inreg %val) {
-; GFX678-LABEL: test_sink_smem_offset_neg400:
-; GFX678: ; %bb.0: ; %entry
-; GFX678-NEXT: s_add_u32 s0, s0, 0xfffffe70
-; GFX678-NEXT: s_addc_u32 s1, s1, -1
-; GFX678-NEXT: .LBB5_1: ; %loop
-; GFX678-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX678-NEXT: s_waitcnt lgkmcnt(0)
-; GFX678-NEXT: s_load_dword s3, s[0:1], 0x0
-; GFX678-NEXT: s_add_i32 s2, s2, -1
-; GFX678-NEXT: s_cmp_lg_u32 s2, 0
-; GFX678-NEXT: s_cbranch_scc1 .LBB5_1
-; GFX678-NEXT: ; %bb.2: ; %end
-; GFX678-NEXT: s_endpgm
-;
-; GFX9-LABEL: test_sink_smem_offset_neg400:
-; GFX9: ; %bb.0: ; %entry
-; GFX9-NEXT: .LBB5_1: ; %loop
-; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: s_load_dword s3, s[0:1], -0x190
-; GFX9-NEXT: s_add_i32 s2, s2, -1
-; GFX9-NEXT: s_cmp_lg_u32 s2, 0
-; GFX9-NEXT: s_cbranch_scc1 .LBB5_1
-; GFX9-NEXT: ; %bb.2: ; %end
-; GFX9-NEXT: s_endpgm
+; GFX6789-LABEL: test_sink_smem_offset_neg400:
+; GFX6789: ; %bb.0: ; %entry
+; GFX6789-NEXT: s_add_u32 s0, s0, 0xfffffe70
+; GFX6789-NEXT: s_addc_u32 s1, s1, -1
+; GFX6789-NEXT: .LBB5_1: ; %loop
+; GFX6789-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6789-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6789-NEXT: s_load_dword s3, s[0:1], 0x0
+; GFX6789-NEXT: s_add_i32 s2, s2, -1
+; GFX6789-NEXT: s_cmp_lg_u32 s2, 0
+; GFX6789-NEXT: s_cbranch_scc1 .LBB5_1
+; GFX6789-NEXT: ; %bb.2: ; %end
+; GFX6789-NEXT: s_endpgm
;
; GFX12-LABEL: test_sink_smem_offset_neg400:
; GFX12: ; %bb.0: ; %entry
@@ -331,3 +319,6 @@ loop:
end:
ret void
}
+;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+; GFX678: {{.*}}
+; GFX9: {{.*}}
``````````
</details>
https://github.com/llvm/llvm-project/pull/76009
More information about the llvm-commits
mailing list