diff --git a/bl32/tsp/aarch64/tsp_entrypoint.S b/bl32/tsp/aarch64/tsp_entrypoint.S
index 2c32578..489183c 100644
--- a/bl32/tsp/aarch64/tsp_entrypoint.S
+++ b/bl32/tsp/aarch64/tsp_entrypoint.S
@@ -43,10 +43,7 @@
 	msr	spsr_el1, \reg2
 	.endm
 
-	.section	.text, "ax"
-	.align 3
-
-func tsp_entrypoint
+func tsp_entrypoint _align=3
 
 	/* ---------------------------------------------
 	 * Set the exception vector to something sane.
diff --git a/include/common/asm_macros_common.S b/include/common/asm_macros_common.S
index b529246..dbc9e2d 100644
--- a/include/common/asm_macros_common.S
+++ b/include/common/asm_macros_common.S
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2013-2016, ARM Limited and Contributors. All rights reserved.
+ * Copyright (c) 2013-2017, ARM Limited and Contributors. All rights reserved.
  *
  * SPDX-License-Identifier: BSD-3-Clause
  */
@@ -11,9 +11,12 @@
 	 * code into a separate text section based on the function name
 	 * to enable elimination of unused code during linking. It also adds
 	 * basic debug information to enable call stack printing most of the
-	 * time.
+	 * time. The optional _align parameter can be used to force a
+	 * non-standard alignment (indicated in powers of 2). Do *not* try to
+	 * use a raw .align directive. Since func switches to a new section,
+	 * this would not have the desired effect.
 	 */
-	.macro func _name
+	.macro func _name, _align=-1
 	/*
 	 * Add Call Frame Information entry in the .debug_frame section for
 	 * debugger consumption. This enables callstack printing in debuggers.
@@ -33,6 +36,9 @@
 	 * .debug_frame
 	 */
 	.cfi_startproc
+	.if (\_align) != -1
+		.align \_align
+	.endif
 	\_name:
 	.endm
 
diff --git a/plat/hisilicon/hikey/hisi_pwrc_sram.S b/plat/hisilicon/hikey/hisi_pwrc_sram.S
index 1fb63ea..f9e1de4 100644
--- a/plat/hisilicon/hikey/hisi_pwrc_sram.S
+++ b/plat/hisilicon/hikey/hisi_pwrc_sram.S
@@ -15,8 +15,7 @@
 	.global v7_asm
 	.global v7_asm_end
 
-	.align	3
-func pm_asm_code
+func pm_asm_code _align=3
 	mov	x0, 0
 	msr	oslar_el1, x0
 
diff --git a/plat/nvidia/tegra/common/aarch64/tegra_helpers.S b/plat/nvidia/tegra/common/aarch64/tegra_helpers.S
index e5e5126..691b90a 100644
--- a/plat/nvidia/tegra/common/aarch64/tegra_helpers.S
+++ b/plat/nvidia/tegra/common/aarch64/tegra_helpers.S
@@ -307,8 +307,7 @@
 	 * Secure entrypoint function for CPU boot
 	 * ----------------------------------------
 	 */
-	.align 6
-func tegra_secure_entrypoint
+func tegra_secure_entrypoint _align=6
 
 #if ERRATA_TEGRA_INVALIDATE_BTB_AT_BOOT
 
diff --git a/plat/nvidia/tegra/soc/t186/plat_trampoline.S b/plat/nvidia/tegra/soc/t186/plat_trampoline.S
index 4841aa2..6a17c33 100644
--- a/plat/nvidia/tegra/soc/t186/plat_trampoline.S
+++ b/plat/nvidia/tegra/soc/t186/plat_trampoline.S
@@ -12,11 +12,10 @@
 
 #define TEGRA186_SMMU_CTX_SIZE		0x420
 
-	.align 4
 	.globl	tegra186_cpu_reset_handler
 
 /* CPU reset handler routine */
-func tegra186_cpu_reset_handler
+func tegra186_cpu_reset_handler _align=4
 	/*
 	 * The TZRAM loses state during System Suspend. We use this
 	 * information to decide if the reset handler is running after a
diff --git a/plat/rockchip/common/aarch64/plat_helpers.S b/plat/rockchip/common/aarch64/plat_helpers.S
index 1c8aefc..abfb5a7 100644
--- a/plat/rockchip/common/aarch64/plat_helpers.S
+++ b/plat/rockchip/common/aarch64/plat_helpers.S
@@ -112,8 +112,7 @@
 	 * cpus online or resume enterpoint
 	 * --------------------------------------------------------------------
 	 */
-	.align	16
-func platform_cpu_warmboot
+func platform_cpu_warmboot _align=16
 	mrs	x0, MPIDR_EL1
 	and	x19, x0, #MPIDR_CPU_MASK
 	and	x20, x0, #MPIDR_CLUSTER_MASK