[llvm] [AMDGPU] Reject misaligned SGPR constraints for inline asm (PR #123590)

[Fabian Ritter via llvm-commits]

https://github.com/ritter-x2a created https://github.com/llvm/llvm-project/pull/123590

The indices of SGPR register pairs need to be 2-aligned and SGPR quadruplets need to be 4-aligned. With this patch, we report an error when inline asm register constraints specify a misaligned register index, instead of silently dropping the specified index.

Fixes #123208

>From 9b8bee2dc6bd290354b02639daf16850763a309d Mon Sep 17 00:00:00 2001
From: Fabian Ritter <fabian.ritter at amd.com>
Date: Mon, 20 Jan 2025 06:10:18 -0500
Subject: [PATCH] [AMDGPU] Reject misaligned SGPR constraints for inline asm

The indices of SGPR register pairs need to be 2-aligned and SGPR
quadruplets need to be 4-aligned. With this patch, we report an error
when inline asm register constraints specify a misaligned register
index, instead of silently dropping the specified index.

Fixes #123208
---
 llvm/lib/Target/AMDGPU/SIISelLowering.cpp     |  6 +++
 .../AMDGPU/inlineasm-mismatched-size-error.ll | 51 +++++++++++++++++++
 2 files changed, 57 insertions(+)

diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index c4b1038b12d042..592a72e109543e 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -15877,6 +15877,12 @@ SITargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI_,
             RC = TRI->getAGPRClassForBitWidth(Width);
           if (RC) {
             Reg = TRI->getMatchingSuperReg(Reg, AMDGPU::sub0, RC);
+            if (Reg == 0U) {
+              // The register class does not contain the requested register,
+              // e.g., because it is an SGPR pair that would violate alignment
+              // requirements.
+              return std::pair(0U, nullptr);
+            }
             return std::pair(Reg, RC);
           }
         }
diff --git a/llvm/test/CodeGen/AMDGPU/inlineasm-mismatched-size-error.ll b/llvm/test/CodeGen/AMDGPU/inlineasm-mismatched-size-error.ll
index 723e0f2e7152d0..e8accc1c8a0f31 100644
--- a/llvm/test/CodeGen/AMDGPU/inlineasm-mismatched-size-error.ll
+++ b/llvm/test/CodeGen/AMDGPU/inlineasm-mismatched-size-error.ll
@@ -102,3 +102,54 @@ define <2 x i8> @inline_asm_2xi8_in_s_def() {
   %r = and <2 x i8> %phys, %virt
   ret <2 x i8> %r
 }
+
+
+; The register is wide enough, but it does not satisfy alignment constraints:
+
+; ERR: error: couldn't allocate input reg for constraint '{s[1:2]}'
+define void @misaligned_sgpr_2xi32_in(<2 x i32> inreg %arg0) {
+  call void asm sideeffect "; use $0", "{s[1:2]}"(<2 x i32> %arg0)
+  ret void
+}
+
+; ERR: error: couldn't allocate input reg for constraint '{s[23:24]}'
+define void @misaligned_sgpr_2xi32_in_23(<2 x i32> inreg %arg0) {
+  call void asm sideeffect "; use $0", "{s[23:24]}"(<2 x i32> %arg0)
+  ret void
+}
+
+; ERR: error: couldn't allocate input reg for constraint '{s[1:4]}'
+define void @misaligned_sgpr_4xi32_in(<4 x i32> inreg %arg0) {
+  call void asm sideeffect "; use $0", "{s[1:4]}"(<4 x i32> %arg0)
+  ret void
+}
+
+; ERR: error: couldn't allocate input reg for constraint '{s[2:5]}'
+define void @misaligned_sgpr_4xi32_in_2(<4 x i32> inreg %arg0) {
+  call void asm sideeffect "; use $0", "{s[2:5]}"(<4 x i32> %arg0)
+  ret void
+}
+
+; ERR: error: couldn't allocate output register for constraint '{s[1:2]}'
+define <2 x i32> @misaligned_sgpr_2xi32_out() {
+  %asm = call <2 x i32> asm sideeffect "; def $0", "={s[1:2]}"()
+  ret <2 x i32> %asm
+}
+
+; ERR: error: couldn't allocate output register for constraint '{s[23:24]}'
+define <2 x i32> @misaligned_sgpr_2xi32_out_23() {
+  %asm = call <2 x i32> asm sideeffect "; def $0", "={s[23:24]}"()
+  ret <2 x i32> %asm
+}
+
+; ERR: error: couldn't allocate output register for constraint '{s[1:4]}'
+define <4 x i32> @misaligned_sgpr_4xi32_out() {
+  %asm = call <4 x i32> asm sideeffect "; def $0", "={s[1:4]}"()
+  ret <4 x i32> %asm
+}
+
+; ERR: error: couldn't allocate output register for constraint '{s[2:5]}'
+define <4 x i32> @misaligned_sgpr_4xi32_out_2() {
+  %asm = call <4 x i32> asm sideeffect "; def $0", "={s[2:5]}"()
+  ret <4 x i32> %asm
+}