diff --git a/src/amd/common/ac_llvm_build.c b/src/amd/common/ac_llvm_build.c
index c325dbbc8cb..81f9e706887 100644
--- a/src/amd/common/ac_llvm_build.c
+++ b/src/amd/common/ac_llvm_build.c
@@ -1028,6 +1028,26 @@ LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
 				  ac_get_load_intr_attribs(can_speculate));
 }
 
+LLVMValueRef ac_build_buffer_load_format_gfx9_safe(struct ac_llvm_context *ctx,
+                                                  LLVMValueRef rsrc,
+                                                  LLVMValueRef vindex,
+                                                  LLVMValueRef voffset,
+                                                  bool can_speculate)
+{
+	LLVMValueRef elem_count = LLVMBuildExtractElement(ctx->builder, rsrc, LLVMConstInt(ctx->i32, 2, 0), "");
+	LLVMValueRef stride = LLVMBuildExtractElement(ctx->builder, rsrc, LLVMConstInt(ctx->i32, 1, 0), "");
+	stride = LLVMBuildLShr(ctx->builder, stride, LLVMConstInt(ctx->i32, 16, 0), "");
+
+	LLVMValueRef new_elem_count = LLVMBuildSelect(ctx->builder,
+	                                              LLVMBuildICmp(ctx->builder, LLVMIntUGT, elem_count, stride, ""),
+	                                              elem_count, stride, "");
+
+	LLVMValueRef new_rsrc = LLVMBuildInsertElement(ctx->builder, rsrc, new_elem_count,
+	                                               LLVMConstInt(ctx->i32, 2, 0), "");
+
+	return ac_build_buffer_load_format(ctx, new_rsrc, vindex, voffset, can_speculate);
+}
+
 /**
  * Set range metadata on an instruction.  This can only be used on load and
  * call instructions.  If you know an instruction can only produce the values
diff --git a/src/amd/common/ac_llvm_build.h b/src/amd/common/ac_llvm_build.h
index b2792099335..655f25df84f 100644
--- a/src/amd/common/ac_llvm_build.h
+++ b/src/amd/common/ac_llvm_build.h
@@ -225,6 +225,14 @@ LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
 					 LLVMValueRef voffset,
 					 bool can_speculate);
 
+/* load_format that handles the stride & element count better if idxen is
+ * disabled by LLVM. */
+LLVMValueRef ac_build_buffer_load_format_gfx9_safe(struct ac_llvm_context *ctx,
+                                                  LLVMValueRef rsrc,
+                                                  LLVMValueRef vindex,
+                                                  LLVMValueRef voffset,
+                                                  bool can_speculate);
+
 LLVMValueRef
 ac_get_thread_id(struct ac_llvm_context *ctx);
 
diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c
index 3e6bc5ec4d7..b9e0e94b8fc 100644
--- a/src/amd/common/ac_nir_to_llvm.c
+++ b/src/amd/common/ac_nir_to_llvm.c
@@ -2308,11 +2308,19 @@ static LLVMValueRef build_tex_intrinsic(struct ac_nir_context *ctx,
 					struct ac_image_args *args)
 {
 	if (instr->sampler_dim == GLSL_SAMPLER_DIM_BUF) {
-		return ac_build_buffer_load_format(&ctx->ac,
-						   args->resource,
-						   args->addr,
-						   ctx->ac.i32_0,
-						   true);
+		if (ctx->abi->gfx9_stride_size_workaround) {
+			return ac_build_buffer_load_format_gfx9_safe(&ctx->ac,
+			                                             args->resource,
+			                                             args->addr,
+			                                             ctx->ac.i32_0,
+			                                             true);
+		} else {
+			return ac_build_buffer_load_format(&ctx->ac,
+			                                   args->resource,
+			                                   args->addr,
+			                                   ctx->ac.i32_0,
+			                                   true);
+		}
 	}
 
 	args->opcode = ac_image_sample;
@@ -3685,8 +3693,23 @@ static void visit_image_store(struct ac_nir_context *ctx,
 		glc = ctx->ac.i1true;
 
 	if (glsl_get_sampler_dim(type) == GLSL_SAMPLER_DIM_BUF) {
+		LLVMValueRef rsrc = get_sampler_desc(ctx, instr->variables[0], AC_DESC_BUFFER, NULL, true, true);
+
+		if (ctx->abi->gfx9_stride_size_workaround) {
+			LLVMValueRef elem_count = LLVMBuildExtractElement(ctx->ac.builder, rsrc, LLVMConstInt(ctx->ac.i32, 2, 0), "");
+			LLVMValueRef stride = LLVMBuildExtractElement(ctx->ac.builder, rsrc, LLVMConstInt(ctx->ac.i32, 1, 0), "");
+			stride = LLVMBuildLShr(ctx->ac.builder, stride, LLVMConstInt(ctx->ac.i32, 16, 0), "");
+
+			LLVMValueRef new_elem_count = LLVMBuildSelect(ctx->ac.builder,
+			                                              LLVMBuildICmp(ctx->ac.builder, LLVMIntUGT, elem_count, stride, ""),
+			                                              elem_count, stride, "");
+
+			rsrc = LLVMBuildInsertElement(ctx->ac.builder, rsrc, new_elem_count,
+			                              LLVMConstInt(ctx->ac.i32, 2, 0), "");
+		}
+
 		params[0] = ac_to_float(&ctx->ac, get_src(ctx, instr->src[2])); /* data */
-		params[1] = get_sampler_desc(ctx, instr->variables[0], AC_DESC_BUFFER, NULL, true, true);
+		params[1] = rsrc;
 		params[2] = LLVMBuildExtractElement(ctx->ac.builder, get_src(ctx, instr->src[0]),
 						    ctx->ac.i32_0, ""); /* vindex */
 		params[3] = ctx->ac.i32_0; /* voffset */
@@ -6829,6 +6852,7 @@ LLVMModuleRef ac_translate_nir_to_llvm(LLVMTargetMachineRef tm,
 	ctx.abi.load_ssbo = radv_load_ssbo;
 	ctx.abi.load_sampler_desc = radv_get_sampler_desc;
 	ctx.abi.clamp_shadow_reference = false;
+	ctx.abi.gfx9_stride_size_workaround = ctx.ac.chip_class == GFX9;
 
 	if (shader_count >= 2)
 		ac_init_exec_full_mask(&ctx.ac);
diff --git a/src/amd/common/ac_shader_abi.h b/src/amd/common/ac_shader_abi.h
index 8d0738b6f94..68e45056acb 100644
--- a/src/amd/common/ac_shader_abi.h
+++ b/src/amd/common/ac_shader_abi.h
@@ -145,6 +145,10 @@ struct ac_shader_abi {
 	/* Whether to clamp the shadow reference value to [0,1]on VI. Radeonsi currently
 	 * uses it due to promoting D16 to D32, but radv needs it off. */
 	bool clamp_shadow_reference;
+
+	/* Whether to workaround GFX9 ignoring the stride for the buffer size if IDXEN=0
+	* and LLVM optimizes an indexed load with constant index to IDXEN=0. */
+	bool gfx9_stride_size_workaround;
 };
 
 #endif /* AC_SHADER_ABI_H */